Citation
An econometric model of interstate labor force migration

Material Information

Title:
An econometric model of interstate labor force migration
Creator:
Engler, Sheldon ( Sheldon Donald ), 1955-
Publication Date:
Language:
English

Subjects

Subjects / Keywords:
Climate models ( jstor )
Forecasting models ( jstor )
Forecasting techniques ( jstor )
Labor migration ( jstor )
Mathematical variables ( jstor )
Spouses ( jstor )
Statistical forecasts ( jstor )
Statistics ( jstor )
Variable coefficients ( jstor )
Workforce ( jstor )

Record Information

Source Institution:
University of Florida
Holding Location:
University of Florida
Rights Management:
Copyright Sheldon Donald Engler. Permission granted to the University of Florida to digitize, archive and distribute this item for non-profit research and educational purposes. Any reuse of this item in excess of fair use or other copyright exemptions requires permission of the copyright holder.
Resource Identifier:
6496799 ( OCLC )
0023327970 ( ALEPH )

Downloads

This item has the following downloads:


Full Text











AN ECONOMETRIC MODEL OF INTERSTATE
LABOR FORCE MIGRATION












BY

SHELDON DONALD ENGLER


















A DISSERTATION PRESENTED TO THE GRADUATE COUNCIL
OF THE UNIVERSITY OF FLORIDA IN
PARTIAL FULFILLMENT OF THE REQUIREMENTS
FOR THE DEGREE OF DOCTOR OF PHILOSOPHY


UNIVERSITY OF FLORIDA

1979

















TO MY PARENTS














ACKNOWLEDGEMENTS


The author wishes to thank Professor Jerome Milliman for his guid-

ance and support throughout the course of study. Special thanks are

given to Professor Henry Fishkind for his intellectual contributions

and his continued friendship. Acknowledgements are also extended to

Professors Stanley Smith, David Denslow, John Henretta, Angela O'Rand,

and William Tyler. Finally, the author would like to thank Alene

Williams and Doreen Willmeroth, whose typing and editing skills were

invaluable to the production of the finished paper.



































iii















TABLE OF CONTENTS


PAGE

ACKNOWLEDGEMENTS................................. ................iii

LIST OF TABLES........................... ..... ............. vi

LIST OF FIGURES.... ............................................... ix

ABSTRACT.......................................................... x

CHAPTER

1 INTRODUCTION.................................................. 1

Problem Statement.. ............................... ........... 1
Literature Review ......................... ................. 2
Overview of the Study....................................... 16

2 THEORY AND METHODOLOGY...... ................................. 18

Theoretical Model........................................... 18
Methodology................................................. 22
A Forecasting Methodology.................................. 34
Sources of Data ............................................. 38

3 EMPIRICAL RESULTS: DECISION TO MOVE.......................... 41

Interstate Migration: 1968-1977............................ 41
Comparison of Migrants and Nonmigrants...................... 45
Empirical Results........................................... 55
Application to Forecasting................................. 72

4 EMPIRICAL RESULTS: DESTINATION CHOICE........................ 86

Destination Choice: 1968-1977.............................. 86
Logit Results: Destination Choice.......................... 86
Application to Forecasting................................. 96
Income, Climate, and Migration ...............................107

5 SUMMARY AND CONCLUSIONS....................................... 110

Introduction. ............................................... 110
Review of Findings: Decision to Move....................... 111
Review of Findings: Destination Choice....................115
Strengths and Weaknesses of the Study.......................116
Implications of the Study.....................................118


iv








PAGE

APPENDIX

A ALTERNATIVE MODEL FORMULATIONS: DECISION TO MOVE............. 120

Combined State Results..................................... 120
Annual Migration Results.....................................122
Redefinition of Qualitative Explanatory Variables...........122

B ALTERNATIVE MODEL FORMULATIONS: DESTINATION CHOICE........... 128

Nominal Wage Model...........................................128
Real Wage Model.............................................. 128

BIBLIOGRAPHY .....................................................132

BIOGRAPHICAL SKETCH................................... ................136














































v














LIST OF TABLES


PAGE
Table 1.1 Greenwood Gross Migration Regression
Results.................... ......................... 4

Table 1.2 Lowry Net Migration Regression Results............... 6

Table 1.3 Graves Net Migration Regression Results.............. 10

Table 1.4 DaVanzo Destination Choice Results................... 15

Table 3.1 Summary of Panel Study Migration, 1968-1977.......... 42

Table 3.2 Florida In-migration by State of Origin,
1968-1977 ........................................... 44

Table 3.3 Variable Definitions for Decision to
Migrate Analysis..................................... 46

Table 3.4 Comparison of Migrant and Nonmigrant
Characteristics: New York, Ohio,
Pennsylvania, and South Carolina...................... 48

Table 3.5 Results of Logit Analysis: Decision to
Move, Ohio, 1968-1977................................ 60

Table 3.6 Results of Logit Analysis: Decision to
Move, New York, 1968-1977............................ 63

Table 3.7 Results of Logit Analysis: Decision to
Move, South Carolina, 1968-1977...................... 65

Table 3.8 Results of Logit Analysis: Decision to
Move, Pennsylvania, 1968-1977........................ 67

Table 3.9 Ordinary Least Squares Results: Decision
to Move, 1968-1977................................... 70

Table 3.10 Results of Logit Analysis: Decision to
Move, 1968-1972 ................................... .. 73

Table 3.11 Comparison of Forecast to Actual Migration,
1972-1976.. ............................... ... 75

Table 3.12 Results of Logit Analysis: Decision to
Move, 1972-1976 ....................................... .77



vi








PAGE
Table 3.13 Forecast: Decision to Move,
1976-1980............................................. 80

Table 3.14 Aggregate Labor Force Migration:
1968-1972, 1972-1976, 1976-1980...................... 82

Table 3.15 Comparison of Panel Study Derived
Out-Migration Estimates to Census
Estimates ............................................ 84

Table 4.1 Destination Choice Summary:
1968-1977............................................. 87

Table 4.2 Variable Definitions: Destination
Choice Equations....................................... 88

Table 4.3 Logit Results: Destination Choice,
1968-1977, Model 1................................... 89

Table 4.4 Logit Results: Destination Choice,
1968-1977, Model 2................................... 93

Table 4.5 Logit Results: Destination Choice,
1968-1977, Climate Excluded.......................... 95

Table 4.6 Logit Results: Destination Choice,
1968-1972............................................. 97

Table 4.7 Comparison of Forecast to Actual
Destination Choice: 1972-1976 ....................... 99

Table 4.8 Logit Results: Destination Choice,
1972-1976.............................................. 101

Table 4.9 Forecast: Destination Choice,
1976-1980.............................................102

Table 4.10 Aggregate Destination Choice,
1976-1980.............................................. 104

Table 4.11 Comparison of Panel Study Derived
Migration Flows to Census Estimates................... 105

Table 4.12 Climate and Income: 1968, 1972,
1976 .................................................108










vii








PAGE
Table 5.1 Summary of Findings: Decision
to Move .............................................112

Table A.1 Combined State Results:
Decision to Move, 1968-1977,
Ordinary Least Squares............................... 121

Table A.2 Annual Migration Results:
Decision to Move, 1976-1977,
Ordinary Least Squares...............................123

Table A.3 Results of Logit Analysis with
Alternative Dummies, 1968-1977...................... 124

Table B.1 Nominal Wage Model: Destination
Choice, 1968-1977......... ..........................129

Table B.2 Real Wage Model: Destination
Choice, 1968-1977 ...................................... 131






































viii















LIST OF FIGURES


PAGE

Figure 2.1 Illustration of Forecasting Methodology:
Decision to Move....................................... 36

Figure 2.2 Illustration of Forecasting Methodology:
Destination Choice................................... 37















































ix















Abstract of Dissertation Presented to the Graduate Council
of the University of Florida in Partial Fulfillment of the
Requirements for the Degree of Doctor of Philosophy

AN ECONOMETRIC MODEL OF INTERSTATE LABOR FORCE MIGRATION

By

Sheldon Donald Engler

December 1979

Chairman: Jerome W. Milliman
Major Department: Economics

The purpose of this study is to analyze and forecast migration

decisions of labor force members. Location choice is viewed within

a theoretical framework which assumes that the individual is a utility

maximizer. Location bundles which include some goods available at all

locations and other goods which are location-specific can be varied by

moving. Each person reaches his (or her) optimum by first considering

the opportunities which are available at all potential locations.

Differences in individual characteristics and circumstances are

hypothesized to cause people to evaluate these opportunities in

alternative ways.

A two-step methodology is adopted for the empirical section of

this study. First, the "whether to move" decision is modeled for all

sample members. Then, the choice of "where to move" is analyzed for

the proportion of the sample who migrated. The first decision is

considered for four separate states. States were chosen based upon

historical rates of out-migration, particularly to Florida. The





x








second model focuses upon the destination choice of those who actually

moved from these states. Eight alternative states are considered as

destinations.

Although differences in migration determinants between states are

discovered, there are some general findings. The typical migrants in

the sample tend to be older and earn higher income than do nonmigrants.

Family relationships also appear to be important influences on the

migration decisions. Migrants are also less likely to be married than

nonmigrants. Contrary to expectations, families with two income earners

appear to move just as often as families where only one person is employed.

Destination choice is shown to be determined by income opportunity and

climate. Evidence that improved climate can sometimes only be obtained

at the cost of reduced income is discovered.

Overall, the migrant appears to be older, richer, and more willing

to take a cut in income for better climate than the nonmigrant. One

possible interpretation of the results is that many interstate moves are

being made in anticipation of retirement. Those who have earned greater

lifetime income are most able to absorb the decline in earnings consequent

upon moving prior to retirement. Economic theory postulates that higher

income people will demand greater quantities of leisure activities.

This increase can be obtained through the migration process. Locations

with warm weather have historically offered greater leisure possibilities

than places with colder, more variant climate.

The results are utilized to obtain interstate migration forecasts.

Rising family incomes, the process of population aging, and an assumption

of the continued importance of climate-related amenities in the migration





xi








decision lead to a forecast of further migration concentrated toward

Sunbelt destinations. Policies aimed at restricting this growth or

attracting migrants to other locations could alter this pattern.


























































xii















CHAPTER 1

INTRODUCTION

Problem Statement


During recent years experts on migration have begun to focus their

efforts upon understanding the decision to move at the individual level.

A continuous stream of detailed survey data has helped to bring about

this endeavor. In addition, increased availability of sophisticated

computer software packages and advances in econometric techniques have

facilitated the handling and analysis of this data. Until now the

research which has come about with the aid of these tools has been

limited to answering age-old questions in the migration literature.

What are the characteristics which distinguish migrants from nonmi-

grants? Do individuals move in the direction of higher wages? How does

the unemployment rate affect the migration decision? What role does

climate play in the choice of location? Does migration lead to regional

income convergence?

Since many of these issues are as yet unsettled, they will be

considered here. The ultimate test of a theory, however, is its ability

to predict the future. With the proportion of population change re-

sulting from natural increase on the decline, migration forecasts become

more of a necessity. This is particularly true for states which are

experiencing rapid population growth such as Florida, California, Texas,

and Arizona. In addition, forecasts of other economic events at the





1




2


regional level depend heavily upon the ability to predict population.

The demand for housing, the unemployment rate, and state and local

government tax collections are examples of economic variables which are

sensitive to the rate of population change. Knowledge of the level of

future population and its consequences for the rest of the economy is

crucial to success in planning for regional economic growth.

In this study, an econometric model is developed for use in ex-

plaining and forecasting the movement of workers and their families

between states. A general theoretical model of interstate migration is

derived within a microeconomic framework. This model is then placed

into an estimable form and tested for the period 1968-77. Individual

data obtained from the University of Michigan Panel Study of Income

Dynamics are used. Maximum likelihood estimation techniques are em-

ployed. Finally, the results of the analysis are utilized in devel-

oping an aggregate forecast of interstate labor force migration.

Throughout the study the case of Florida is emphasized. Separate

models are estimated for four typical origin states of Florida migrants.

The determinants of Florida's attractiveness relative to other desti-

nations which draw migrants from these states are analyzed.

In the next section the existing migration literature will be

brought up to date. Then an overview of this study will be presented.


Literature Review

Introduction

Existing studies will be reviewed according to two major cate-

gories. The first section which follows will summarize research that is

centered upon explaining aggregate flows of population. The second





3

section will focus upon literature that seeks to explain migration

decisions at the level of the individual.


Aggregate Migration Models

Aggregate models of migration are subclassified into two types

suggested by Greenwood (1975). The first type is concerned with gross

migration, defined as a single flow or the sum of unidirectional flows

of population. The second type is concerned with net migration, defined

as the difference between gross flows occurring in opposite directions.

For any region total net migration is simply the difference between the

number of in-migrants and the number of out-migrants over a specified

time period.

A typical gross-migration model is provided by Greenwood (1969) in

his analysis of migration determinants. Greenwood examines a cross

section of flows between all 48 mainland states during the 1955-60

period. His regression results are presented in Table 1.1. Positive

and significant variables in Greenwood's final equation are the average

level of education at origin locations, the unemployment rate at the

origin, the percentage of urban population at the destination relative

to the percentage of urban population at the origin, the mean annual

temperature at the destination relative to the mean annual temperature

at the origin, and the number of persons born at the origin who already

reside at the destination--called the migrant stock. Also significant,

but showing a negative sign, are the distance between origin and desti-

nation and the average level of education at the destination. The best

explanatory variable in Greenwood's model is the-migrant stock variable.

This variable is believed to be a proxy for the amount of information





4


Table 1.1 Greenwood Gross Migration Regression Results


Variable Coefficient t-Statistic

D.. -.300 -11.21

Y.. .160 1.27

E. 3.401 16.60
1

E. -.622 -2.95

Ui .705 8.44

U. -.057 -.66

R.j .771 7.52

T.. .903 6.49

MS.. .521 42.06
1J


Variable Definitions:

D.. -distance from origin to destination

Y.. -income at destination divided by income at origin
Ji

E. -median education at origin
1

E. -median education at destination

U. -unemployment rate at origin

U. -unemployment rate at destination

Rji -percent of urban population at destination divided by
percent of urban population at origin

Tji -mean annual temperature in principal city of destination
divided by mean annual temperature in principal city of
origin

MS.. -number of people residing at destination who were born at
origin





5

concerning the destination that is available at the origin. That is,

previous migrants channel information back to potential future migrants.

And, assuming that it is of an encouraging nature, more information

should lead to increased migration.

The interpretation of the migrant stock is subject to some con-

troversy. Laber (1972) argued that since the migrant stock is itself a

function of those factors influencing previous migration, it may be

acting as a proxy for lagged explanatory variables. If this is the

case, then a more appropriate specification may be a partial adjustment

model where lagged migration is included as an explanatory variable.

Dunlevy and Gemery (1977) estimate alternative specifications of the

model and conclude that it would be correct to include both migrant

stock and lagged migration variables. They argue that including one

variable while excluding the other results in the included variable

capturing parts of both effects. Thus, it is concluded that both the

information-creating effects of the migrant stock and a lagged migration

adjustment process are operating simultaneously.

A typical model of net migration is developed by Lowry (1966).

Lowry examines a cross section of Standard Metropolitan Statistical

Areas over the period 1950-60. His results are summarized in Table 1.2.

Significant variables in this model are the growth of the resident

working-age population (inversely related to net migration), employment

growth (positively related to net migration), and growth in armed

services personnel (positively related to net migration). Employment

growth is the best variable, the idea simply being that higher growth

reflects greater demand for labor, which translates into greater em-

ployment opportunities for potential in-migrants (making them more





6

Table 1.2 Lowry Net Migration Regression Results


Variable Coefficient t-Statistic

dP. -.62515 -5.24

dQi 1.49021 52.56

dA. 1.44409 7.36

Variable Definitions:

dP. -net change in the number of residents 15-64 years of age
in the absence of migration

dQi -net change in civilian nonagricultural employment

dA. -net change in number of armed services personnel





7

likely to migrate) and for residents (making them more likely to re-

main). Growth in the resident working-age population is important

because higher growth implies an increased supply of labor occurring

from within the region, thus reducing the number of job opportunities

available to potential in-migrants. Growth in armed services personnel

is included since it is a component of population change that is not

normally explained by the other variables in the equation.

Neoclassical regional growth models theorize that migration is

induced solely by the existence of interregional wage differentials.

That is, migration occurs from low-wage regions (where labor is plen-

tiful relative to capital) to high-wage regions (where labor is rela-

tively scarce). Increased supply of labor in high-wage regions is

hypothesized to put downward pressure on wages, while decreased supply

of labor in low-wage regions exerts upward pressure on wages. Thus, the

model leads to an equilibrium where wage differentials between regions

are eliminated. Smith (1975) includes a labor sector in his neoclas-

sical growth model which posits net migration to be a function of the

difference between the local wage and the national wage. He estimates

his model for various historical periods and generates mild support for

the convergence theory. This is in contrast to other tests of neoclas-

sical growth theory, such as that undertaken by Borts and Stein (1964),

in which the theory is rejected. But Smith also discovered a decrease

in responsiveness of potential migrants to income differentials over

time. This he partly attributes to the existence of unemployment which

discourages migration. Neoclassical growth models, whether regional or

national in scope, assume full employment of labor.





8

Richardson (1973) criticizes neoclassical models for assuming that

individuals are simply maximizing their income in the migration decision.

He proposes a model (which is as yet untested) in which net migration is

a function of wage differentials, agglomeration economies, and locational

preferences. Agglomeration economies are divided into two types.

Household agglomeration economies refer to the benefits and costs of

life in large cities for households. They include benefits accruing

from larger labor markets, availability of leisure and cultural facil-

ities, the quality of public service, and environmental amenities.

Business agglomeration economies refer to advantages which urban areas

offer businesses. They attract firms which lead to jobs which in turn

attract migrants. The level of agglomeration economies is thus hypoth-

esized to influence the rate of in-migration to a region. Locational

preferences are meant to explain why individuals may remain in low-

income regions despite the existence of improved income opportunities

elsewhere. Thus, they measure the retentive power of a region and are

related to the rate of out-migration. Examples of factors which in-

fluence locational preferences are community ties, sociocultural traditions,

and length of settlement.

In a recent study Graves (1979) rejects the theory that migration

is a response to interregional income differences. Rather, he advances

the view that any income differences which may exist between regions are

compensated for by differences in the amounts of other amenity-oriented

goods which are available. For example, an area which offers an income

advantage over other places is believed to offer a more objectionable

climate, allowing for less involvement in leisure or recreational

activities. Moving to a location with a better climate, under this





9

viewpoint, involves a sacrifice in income. Contrary to regional growth

models, Graves views the regional system as fundamentally being in

equilibrium so that utility is constant over space. Migration then

takes place as a result of changes in demand for location-fixed amen-

ities. These changes come about as a result of changed relative prices

and income. Under this system migration will not cause regional income

convergence since income differentials must be maintained in order to

compensate for differences in climate. Graves tests his theory by

analyzing net migration for a cross section of Standard Metropolitan

Statistical Areas during the 1960-70 period. A sample of his results

for white migrants is presented in Table 1.3. All climate variables are

significant. Median income is not significant, but it is more significant

than it is in a model with climate variables excluded. In addition,

when Graves disaggregated the sample by age, he found income to be a

very significant determinant of migration for some age groups. The

unemployment rate, as expected, is negative and significant. As with

income, when climate is excluded from the equation, the unemployment

rate loses significance. The results demonstrate than when employment

and income possibilities are the same across all locations, people

choose to move to more temperate climates. Alternatively, if climate is

held constant across alternatives, people will choose to move to places

offering greater income and employment opportunities. When climate is

excluded, and thus allowed to vary, income and employment possibilities

by themselves are no longer such important migration determinants.

Thus, there is evidence that opportunities and climate interact in the

way suggested by Graves.





10

Table 1.3 Graves Net Migration Regression Results


Variable Coefficient t-Statistic

Medinc .00162 1.07

Unemp -2.906 -4.26

Warmth* .0103 4.44

Cold .00686 4.27

Antmvr -.9989 -4.86

Annwnd -2.967 -4.49

Annhum -.7164 -4.74

Variable Definitions:

Medinc -1960 Median Income

Unemp -1960 Unemployment Rate

Warmth* -Mean annual number of cooling degree days (base = 650 F)

Cold -Mean annual number of heating degree days (base = 650 F)

Antmvr -Annual temperature variance (average daily maximum
July temperature average daily minimum January
temperature)

Annwnd -January and July average wind speed

Annhum -January and July average humidity



*All climate variables defined as 1931 to 1960 averages








One of the advantages of models of net migration is that data on

the dependent variable are easily computed for most areas. Net migra-

tion can be directly estimated by subtracting the population change due

to natural increase from the actual population change for any given

period. Measures of gross migration, in contrast, depend upon more

direct measurement techniques, such as surveys. However, as Lowry

(1966) notes, models of net migration reveal less about migratory be-

havior and the decision to migrate than do models of gross migration.

This is partly because variables which are important in determining

unidirectional flows are reduced in significance when the measure of

migration used includes flows occurring in opposite directions. Models

of net migration, however, are useful in forecasting population change.

They are of particular value to regions for which migration has been the

dominant component of population change. The next section will describe

research which has focused upon explaining migration at the individual

or family level.


Individual Migration Models

Rothenberg (1977) suggests an approach to the study of migration

that focuses upon the individual. He notes that the migrant is self-

selected. That is, given the availability of similar sets of oppor-

tunities, some individuals will migrate and some will not. The problem

is to determine the individual characteristics and circumstances that

cause people to evaluate their migration choice in different ways. Put

somewhat differently, one individual's maximizing decision may cause him

(or her) to migrate, while another individual facing similar opportuni-

ties may choose not to move. Both persons may be acting rationally.





12

In order to test the validity of this hypothesis, a model must be esti-

mated using individual data on non-migrants as well as migrants.

Morrison (1973) suggests the use of a theoretical framework which

combines two models of migration. First, he suggests a microeconomic

model which analyzes the decision "whether to move." This model, he

claims, should reflect the idea that individuals have variable decision

thresholds. That is, those with lower thresholds are more likely to

seek out and respond to opportunities elsewhere, and hence are more

likely to move. The level of threshold variability is determined by

such factors as the individual's position in the life cycle, occupa-

tionally induced contraints on movements, and prior migration experi-

ence. The second type of model suggested by Morrison is one which

allocates those who do move (as a result of the first decision) among

alternative destinations. Thus, this model addresses the question

"where to move." This model, he claims, is more macroeconomic in con-

tent.

An empirical study by DaVanzo (1977) divides the migration choice

into the two-stage process suggested above. She looked at a sample of

married couples from the University of Michigan Study of Income Dynamics.

First, she estimates "whether to migrate" equations. The dependent

variable in these equations is a zero-one dummy indicating whether or

not a move was made for each couple. A one-year time period (1971-72)

was selected. The explanatory variables can be categorized as employ-

ment status, returns to migration, composition of family earnings,

location-specific assets (such as homeownership), previous migration,

age, and education. The basic conclusions can be summarized as follows:

Families whose heads are unemployed or are dissatisfied with their jobs





13

are more likely to move than those whose heads are not searching for

work. Local employment conditions are more important in the migration

decisions of the unemployed than the employed. Unemployed persons and

others looking for work are more responsive to family income, origin

wage rates, and expected earnings increases than persons satisfied with

their jobs. Families are more likely to move if they have moved in the

recent past. Wives have a significant influence on the family's de-

cision to migrate. Age and education are relatively unimportant in

explaining the migration of married couples.

DaVanzo estimates her "whether to migrate" equations using ordinary

least squares (OLS). However, as acknowledged by the author, use of OLS

when the dependent variable is dichotomous leads to inefficient (although

unbiased) estimates. In addition, the fitted equation will yield pre-

dictions outside the zero-one range. This criticism is particularly

crucial to a model which is to be used for the purpose of forecasting.

DaVanzo, however, was only trying to identify the determinants of the

migration decision. She estimates one equation using a probit model, a

maximum likelihood technique, and demonstrates that there is little

difference between the derived coefficients and the coefficients that

result from using OLS.

After analyzing the decision "whether to migrate," DaVanzo esti-

mates choice of destination equations. She analyzes the choice among

eight regions of the United States. The explanatory variables are the

present value of the difference between what each family could earn at

each destination and what the family could earn if it stayed at the

origin, the unemployment rate at alternative destinations, the distance






14

between each destination and the origin, and an interaction term between

the present value of the wage differences and a dummy indicating whether

a person had resided at that destination in the recent past. The

results appear in Table 1.4. None of the variables turn out to be

significant at better than the .10 level. Only the wage variable is

even close to being significant, indicating that families are likely to

move to destinations where the earnings gains are greatest. The sign of

the interaction term indicates that families are more likely to move to

an area where they have lived before than to one where they have never

lived, especially if the family earnings they could receive there are

higher than what they could earn by staying where they are. The unemploy-

ment rate is insignificant and shows an unexpected positive sign. The

coefficient of distance shows that migrants are more likely to choose

closer destinations, although it too is insignificant. The model is

estimated using conditional logit, a maximum likelihood technique

designed to analyze multinomial choices among discrete alternatives.

DaVanzo considered the migration decision from the viewpoint of the

married couple. Mincer (1978) notes that there are so far very few

migration studies that consider the effect of family relations on this

decision. He points out that at the individual level a person should

choose to move to that location at which the return is at a maximum.

For a family the optimal move is one which maximizes the combined return

to the family. Whether this last criterion is satisfied or not, frequently

a family will end up at a location which does not reward every family

member with the maximum possible return. The members of the family who

do not reach their private optimum are called tied movers. Mincer

points out that the conflict which could result from tied migration can





15

Table 1.4 DaVanzo Destination Choice Results


Variable Coefficient Asymptotic t-Ratio

fam
PV.. .00548 1.53
1J
fam
PV.. Here Before .0139 1.46

Unemployment Rate .00909 .05

Ln Distance .. -.322 -.81


Variable Definitions:

fam
PV.. -present value of the difference between what the
S family could earn at destination j and what it
could earn if it stayed in its 1971 location, i

Here Before -dummy that indicates whether the family
resided in area j recently (between 1968
and 1970)

Unemployment Rate j -unemployment rate in 1971 at destination j


Ln Distance .. -natural logarithm of the distance between
origin i and destination j





16

only act as a deterrent to a possible move. This is particularly the

case where more than one family member is working. Thus, the increasing

proportion of women in the labor force is expected to have an inhibiting

effect upon migration. Mincer uses a scattering of data to test his

theories. He discovers, as did DaVanzo, that marriage itself reduces

migration and that migration rates are lower in families with employed

wives. The magnitude of the effect which the working wife has on the

migration decision also depends upon her share of family earnings.

Finally, he shows that tied migration of working wives frequently

results in lower earnings, unemployment, or labor force withdrawal.

Thus, male household heads usually dominate the choice of destination.


Overview of the Study

In Chapter 2 the migration decision is analyzed within the context

of utility theory. A theoretical model of interstate migration is then

derived. A methodology for estimating this model is put forth which

considers the decision to move and the choice of destination together in

one analytical framework. Problems that arise in implementing this

procedure are then discussed and an alternative is proposed. The new

method views the migration decision as a two-stage process. Finally, a

technique for forecasting migration decisions and destination choice is

proposed.

Chapters 3 and 4 present the empirical results. The decision to

move is analyzed in Chapter 3. Regression results are presented for

four states and forecasts of aggregate migration from these states are

developed. Chapter 4 deals with the choice of destination. The de-

terminants of this decision are analyzed and a forecast of migration





17

between 31 origin and destination combinations is presented. In both

chapters forecasting accuracy will be tested.

The final chapter will review the findings of the study and place

them into the context of the migration literature. The strengths and

weaknesses of the research will be emphasized. The implications of the

study for theory, policy, and future work will be discussed.















CHAPTER 2

THEORY AND METHODOLOGY

Theoretical Model


The principle of utility maximization can be adapted to the mi-

gration decision. This adaptation, however, is not straightforward.

Differences between the standard model of consumer behavior and the

approach to be taken here arise because of the nature of location as a

good. These differences are discussed in the following paragraphs.

The location decision involves a choice among mutually exclusive

alternatives. For some given time period, the individual lives at only

one location. Thus, the quantity consumed of various locations cannot

be adjusted in the same manner as are the relative amounts of other

goods. Any change in location can be viewed as a decision to give up

the entire quantity of one place in favor of consuming all of another

residence.1 This change, however, is not quite as drastic as it first

appears. To see why, location must be defined as a bundle of goods.

Some goods can then be common to more than one bundle. For example,

most food products can be purchased at any location in the United States.




iThe case of one person having two or more residences can be raised
as a contradiction to these statements. In defense, it can be argued
that this individual can only live at one place at a time and that each
move between residences involves a new choice among mutually exclusive
alternatives. Additionally, the person with multiple residences is
atypical.




18





19

A change in residence will, by itself, cause little change in quantity

of food consumed. Changes in food buying behavior will occur if a move

is accompanied by an income increase or by a change in relative food

prices.

There are also goods contained in the location bundle which can

sometimes only be obtained in varying quantities by moving. An example

is climate. Although typically defined as a characteristic of a place

rather than a good within a location bundle, climate does display

certain characteristics of goods. The quantity of climate can be

represented by measuring factors such as temperature or nearness to the

coast. The price can be measured in terms of moving costs or foregone

earnings involved in moving to a more temperate climate. So although it

is not produced by man, from the consumer's standpoint climate displays

characteristics that other goods possess. The need to move in order to

obtain substantial changes in consumption makes climate an important

determinant in the migration decision. This is true for other fixed-

location amenities such as symphonies, sporting events, and certain

public services.

Another important factor distinguishing locations for the consumer

is the existence of differential employment opportunities across space.

A person may have a potential job available in one place that will earn

him (or her) greater income than jobs at other locations may offer.

Higher income, in turn, allows for greater consumption. These increased

consumption possibilities include leisure activities which are related

to climate. Thus, although many of the goods in the location bundle are

available at all locations, different quantities may be consumed at

various locations because of the uneven spatial distribution of em-

ployment and income possibilities.





20

In addition to items contained in the location bundles, it is

posited that individual characteristics and circumstances are deter-

minants of migration decisions. These factors determine an individual's

responsiveness to opportunities contained in the location bundle. As an

example, older people move less often because of a shorter expected

lifetime over which they may experience migration-related gains and

because they are likely to possess location-specific capital, such as a

home. Or a person who has moved many times in the past may be more

susceptible to migration influences. An individual who has resided at

the same location for a long period of time may feel greater ties to

relatives, friends, and community. A married person may be less likely

to move than an unmarried individual, especially if the spouse works.

The location choice which maximizes utility for one member of the family

may not be the optimal choice for other family members. The end result

will depend upon the nature of family relations. The point here is only

that the existence of the marriage can create a conflict which affects

the migration decision. From the standpoint of the family as a unit the

optimal choice may still result.

Other factors entering into the migration decision will be dis-

cussed when the final model is specified. The above discussion has

served mainly to illustrate differences between the study of location

choice and the analysis of other choices in which economists are typically

interested. With these considerations in mind, a model of individual

location choice will be derived.

Assume that the individual chooses to reside at that location from

which he (or she) derives maximum utility. Further, assume that the

utility function is linear in parameters with additive disturbances.

Define Zti = (X P ), where X is a vector of independent variables





21

describing individual t, and P. is a vector of explanatory variables

describing alternative i. Also define et = (etx, tp), where 9tx and

tp are vectors of parameters assigned to Xt and P. respectively.

Finally, define Eti as a random disturbance for alternative i and

decision maker t. The utility which individual t derives from place i

can now be expressed as follows:

Uti = Ztit + ti (1)

Alternatively, it can be written:

Uti = Xttx + P t + ti (2)
ti t tx i tp ti
If the individual maximizes utility, then he or she will choose to live

at alternative i if Uti > Utj for all j not equal to i.

Moving costs can be incorporated into the model. Assume first that

we find individual t living at place i. Then we conclude that if he (or

she) is rational: Uti > Utj C.i for all j not equal to i where Cij is

the cost of moving from i to j. If at a later time person t migrates to

place j, then we conclude that something occurred during the interim

which caused t's utility evaluation to change so that: Utj Cij >

Uti. Throughout the remainder of the study, actual moving expenses will

be ignored since it is believed that they are insignificant in relation

to the other costs and benefits of moving.

The probability that a person will choose to live at location i can

be written as follows:

P(i) = P(Uti > Utj for all j # i) (3)

A major goal of this study will be to impute values of P(i) for in-

dividuals inside and outside the sample. Inferences about aggregate

behavior will also be made using these probabilities.






22

The model developed above belongs to a general classification

called random utility models. (For a discussion of these models see

Albright, Lerman, and Manski [1977].) Alternative methods of estimating

this model have been attempted in this study. The technique which at

first appeared to be ideal will be put forth in the next section.

Reasons for the abandonment of this approach will then be given. These

reasons are both theoretical and practical. The method finally settled

upon will be discussed in detail. Its strengths and weaknesses in

studying location choice will be highlighted.


Methodology

An Ideal Methodology

Equation (2), as it stands, cannot be estimated. This is because

utility is unobservable. We can, however, observe actual location

choices. Define I = 1 if individual t chooses alternative i. Define I

= 0 if some other location is chosen. Then the utility function can be

rewritten:

I = Xt+tx + Pi(tp + ti (4)

Given the assumptions of the theory, if I = 1,then we know that Uti >

Utj for all j not equal to i. If I = 0, we then conclude that another

location allows the individual to obtain a higher level of utility.

Equation (4) is an estimable equation. It can be applied to both

single-choice and multiple-choice situations. For multiple-choice

situations it is assumed that the individual is making any number of

concurrent binomial decisions. For example, a person residing in New

York and deciding whether to move to Florida, migrate to California, or

remain in New York can be said to be faced with three binomial decisions:






23

A. Migrating to Florida or not

B. Migrating to California or not

C. Staying in New York or not

Values of I can now be assigned to each decision. If this person

chooses to migrate to Florida, then for decision A, I = 1; for decision

B, I = 0; and for decision C, I = 0. The values of Xt will vary across

alternatives. Individual-alternative interaction terms can also be

introduced into the model.

If each binomial decision by each individual is treated as a single

observation, then the model can be estimated by ordinary least squares

(OLS). It is well known, however, that use of OLS when the dependent

variable is dichotomous leads to inefficient estimates and to predictions

outside the zero-one interval. In addition, OLS is unable to distinguish

between individuals and observations in the multinomial case. More

appropriate methods use the maximum likelihood technique of estimation.

McFadden (1973) developed a technique he calls conditional logit

for use in analyzing consumer choice among lumpy alternatives. He

demonstrates the applicability of conditional logit in the study of

urban travel demand. In a recent study, Falaris (1978) applies the same

technique to the migration decision. An attempt was made in this study

to adopt the same methodology. It did not prove to be useful for this

study. The conditional logit model seems best adapted to problems where

the choice frequencies are fairly well balanced. Location choice

(studied over reasonable periods of time) is biased heavily toward a

single alternative: staying where one already is. Thus, when a model is

estimated which includes staying as one of many alternatives, this

choice swamps all others. Consequently, it is difficult to obtain






24

variables that distinguish between the remaining alternatives. The

magnitude of this problem enlarges as the number of alternatives in-

creases. Falaris avoided this problem to some extent by restricting his

research to a choice among four broad regions of the United States.

Since this study will focus upon the state as the level of analysis, it

will be desirable to include a larger number of alternatives. A further

problem encountered with the model is that as new variables are added

when there are a large number of alternatives, the cost of estimation

rises substantially and the chances of early convergence diminish.

Falaris, in fact, considered only a small number of variables in his

study. Finally, testing various model specifications is difficult using

such high-cost procedures.


Proposed Methodology

The above considerations have led to the decision to analyze the

migration choice as a two-stage process. This is in line with the

approach suggested by Morrison (1973) and taken by DaVanzo (1977). The

first stage of the migration process is the decision "whether to move."

All members of the population are faced with this choice. The second

stage involves the decision "where to move." Only migrants face this

choice. The decision equation (4) is now divided into two:

II = Xttx + ti (5)

1 = Ptp + Ui (6)

where Uti is a random disturbance term and every other term is defined

as before. Equation (5) represents a binomial decision, with I = 1 if

the person moves and I = 0 if they remain where they are. Equation (6)

can be a multinomial decision, with the number of alternatives





25

depending upon the number of locations available to the individual.

Utilizing this framework relieves to some extent the econometric problems

encountered when the entire decision was placed in one equation. The

"whether to move" equation is now easier to estimate and interpret since

all migrants are grouped into one category. The "where to move" equation

also contains more balance among alternatives. In addition, since the

sample size has been reduced substantially by considering only migrants

in this equation, estimation is now less costly.

Viewing the migration decision under a two-stage regime involves

making the assumption that these decisions are independent of each

other. This is the major weakness of this approach. It can readily be

argued that the decision to move and destination choice occur jointly in

many individual cases. But, given currently available techniques, the

costs of modeling them together appear to outweigh the benefits derived

from this approach. A detailed description of the procedures used in

estimation and application of equations (5) and (6) begins in the next

section.


The Decision to Move

As discussed earlier, if two people are choosing from a set of

similar location bundles, one may move and one may remain at his (or

her) current residence. Characteristics of the individual, contained in

the vector Xt, are believed to determine who is selected to migrate.

These characteristics may represent differences in opportunities. For

instance, the older person may have a lower return from migration in

terms of lifetime income. Differences in individual attributes may also

represent differences in the way people evaluate the same opportunities.






26

As an example, those who have moved more often in the past are more

likely to respond positively to migration opportunities.

Measures of migration opportunities could be included in the

"whether to migrate" function. DaVanzo (1977) constructs a return to

migration variable that represents the maximum return available from

moving. It is derived by estimating potential wages, via a human

capital wage model, at each possible destination and at the origin.

Then the differences between potential wages at the origin and each

destination are obtained. The maximum difference is used as the es-

timate of migration returns. Problems arise with this variable.

Since the migration decision is itself a function of individual character-

istics, many of these being the same ones which determine wages, intro-

ducing this variable into the equation is somewhat redundant. If mi-

gration opportunities, such as potential income, are a function of

individual attributes, then the direct inclusion of these character-

istics in the equation seems to be the best approach. Using the returns

variable also introduces the possible presence of multicollinearity in

the equation since the variable is, in effect, just a linear combination

of some of the remaining variables in the equation. For these reasons a

similar variable will not be included in this model.

A final argument can be put forth in favor of including opportunity-

related variables in the decision equation. Differences in potential

returns from migration can be derived because individuals originate at

different locations. The potential return of moving from New York to




2More details on construction of the potential wage variable will
appear later since this variable is considered in the choice of desti-
nation analysis.






27

Florida may exceed the gain resulting from a California to Florida move.

Therefore, more people may migrate to Florida from New York than from

California, despite the fact that Californians and New Yorkers may have

similar characteristics. There is a way, used in this study, to allow

for this occurrence without including a returns variable in the equation.

The model can be estimated for a sample that is restricted to one

origin at a time. In this way, everyone in the sample faces the same

alternative choices even if these choices are viewed relative to their

present location. Differences in migration propensities now occur as a

result of differences in the way individual characteristics are valued

at alternative locations and because of the effects that individual

attributes have upon the way in which people evaluate and respond to

migration opportunities. Estimating the model for various places also

allows for a comparison of the determinants of the migration decision

over space. If the factors important in explaining the New Yorker's

decision to move are significantly different from the factors affecting

the Californian's choice, then there is an additional rationale for

having separate models. Such an approach will be taken in the main body

of this study. Results obtained when origin states are combined are

presented in Appendix A.

Lack of time and resources prevents estimation for all possible

origins. The model will be estimated for four states. These are New

York, Pennsylvania, Ohio, and South Carolina. These states were chosen

because, in the sample to be studied, they are the leading origins of

migrants to Florida, the destination of primary concern to this study.

In addition, all four are states with relatively large numbers of out-

migrants during the period under study.





28

The choice of time period is important in any migration study.

Shorter time periods at first seem optimal, since the determinants of

migration can be measured at or close to the time at which the actual

move is made or not made. Longer time periods lead to problems, since

migration may be taking place at many different points during the

period. Determination of when to measure the explanatory variables is

then difficult. Estimation problems, however, arise for short time

periods because the proportion of the population that moves is smaller

than it is for long periods. Estimations were first attempted for one-

year periods. Poor fits were obtained and few variables were significant

enough to distinguish between migrants and nonmigrants. A sample of

these results can be found in Appendix A. For the main part of the

study, three time periods were considered. First, the model was es-

timated for the 1968-77 period, the maximum amount of time for which

data is available in the sample. Then, for the purpose of forecasting

the models were re-estimated for the periods 1968-72 and 1972-76. The

explanatory variables were always measured in the first year of the

period under consideration. The migration variable took on the value of

one if during the last year of the period, an individual was living in a

state which differed from his (or her) state of residence during the

first year of the period. Otherwise it took the value of zero.

The unit of analysis in the study is the head of household who is a

member of the labor force. Since the survey from which the sample is

derived contains some information about other household members, var-

iables such as family size and spouse's employment status can also be

entered into the equations. Other studies (e.g., DaVanzo [1977])





29

have considered only married couples in their samples. Restricting the

sample this way allows for consideration of variables such as spouse's

wage and helps to increase the explanatory power of the equations. This

approach is rejected here, since the ultimate concern of this research

is applying the results to a forecast of migration for all labor force

members, whether married or unmarried.

The sample over which the model is estimated includes members of

the labor force only. Persons not attached to the labor force, such as

retirees, are believed to be subject to a set of influences that is

distinct from the mix of factors affecting labor force members. Health

and unearned income are examples of variables expected to be of greater

importance in the retiree migration decision. The retiree choice of

destination is also affected by different factors. Employment- or

earnings-related variables are clearly not appropriate while cost-of-

living variables may take an added significance in a model of retiree

migration. For this reason, a separate theoretical and empirical model

should be developed for this segment of the population.

In deriving the final sample, return migrants are excluded. A

return migrant is defined here as a person who is found to be moving

back to his (or her) state of birth. Other studies (for example Kau and

Sirmans [1977]) find that including return and nonreturn migrants in the

same equations leads to a specification bias. Return migrants appear to

be moving back home for reasons such as poor health or because they miss

friends or relatives--factors which do not ordinarily enter into non-

return migrant decisions. A finding of a recent study of elderly return

migration is that while nonreturn migration is positively selective,

return migration is negatively selective. Longino (1979) finds that





30

elderly migrants returning to their state of birth have lower socio-

economic characteristics than other movers. Thus, only nonreturn

migrants are considered in this analysis.

The decision equation will be estimated using a logit program

developed by Nerlove and Press (1973). Using this technique, the

estimate of Il in equation (5) is equal to the log odds of one of

the alternatives. The estimated equation can be written:

log (stay Xttx(7)
_-Pstay/

where Pstay is the probability of staying, i.e., not moving. We can

now solve for Pstay:

Pstay = 1 (8)
stay 1 + e-Xtetx

Since there are only two alternatives, the probability of moving is

Pmove = 1 Pstay = 1 Ytet (9)
1 + eitotx

This value will be derived for every individual in the sample and com-

pared with each person's actual choice. In addition to comparing actual

and fitted values, the overall fit of the model can be determined by

comparing the log likelihood value at model convergence with the log

likelihood value that results if all coefficients are restricted to

zero. It has been shown (see McFadden [1974]) that the statistic

-2 [ In (Ir) In (T) ] (10)

is distributed approximately chi-square, where Ir is the likelihood

value when the coefficients are restricted and 1 is the likelihood

value for unconstrained model at its maximum. The degrees of freedom

are equal to the number of restricted coefficients.





31

In addition to comparing the likelihood value at the unconstrained

maximum to the likelihood value when all coefficients are restricted to

zero, a more stringent test is proposed. First the right-hand side of

the decision equation is set equal to the mean value of the dependent

variable. This would be the best guess of the probability that some

individual in the sample would move, given that we have no other in-

formation. Then the resulting likelihood value is compared to the

likelihood value when all variables are included. The statistic put

forth above can then be calculated using these two likelihood values as

input.

The specific variables considered in the decision-to-move analysis

will be defined and discussed in Chapter 3. The next section puts forth

the methodology employed in the destination-choice analysis.


The Choice of Destination

Having determined the characteristics that distinguish interstate

movers from nonmovers, the next task is to examine the factors which

enter into the destination choice. Previous studies have considered

location choice primarily as a function of potential wages at alter-

native locations. DaVanzo (1977) and Falaris (1978) first estimate a

wage model for each alternative location. Wages are viewed as a func-

tion of the characteristics of those living at each place. Some of the

variables considered are age, race, sex, education, occupation, and

experience. These variables are similar to those typically employed in

human capital models of wage determination (for example Dalton and Ford

[1977]). After estimation, predicted values of-wages at each desti-

nation are obtained for everyone in the sample by plugging their





32

characteristics into the fitted equations. This variable is usually

found to be a significant determinant of location choice. A similar

variable has been constructed for use in this study. Although the sign

was positive, it did not turn out to be significant. One possible

explanation derives from the way in which the sample was chosen. As

mentioned earlier, the origin states considered are the leading senders

of migrants to Florida. Florida is not known as a high-wage state, yet

many people move here. In fact, many Florida migrants in the sample

have experienced declines in nominal and real (price deflated) wages. A

sample of results obtained when the wage variable is included is contained

in Appendix B.

Generally, earlier empirical models of individual migration choice

have not found aggregate location characteristics, such as the unemploy-

ment rate, to be important variables. But these studies have considered

only very large geographic areas as potential destinations. Differences

in location-specific characteristics then tend to get averaged out

across the region. Better results should be obtained if destinations

are defined as smaller areas, such as states. A disadvantage of this

approach is that all possible destinations cannot be considered since

the estimation technique puts limits upon the number of alternatives

that may be considered at once.

Essentially, the approach suggested by Morrison (1973) is taken

here. That is, while the decision to migrate is considered from a

microanalytic perspective, the decision where to go is modeled in a

macroanalytic framework. By considering only aggregate characteristics

of each location as variables, the model will, for example, predict the

probability of a New York migrant choosing Florida as a destination. It

will not, however, explain why one New Yorker chooses Florida and another





33

chooses California. Micro type variables, such as individual potential

wages, which take into account differences among people were tested.

Their failure to yield meaningful results led to the adoption of a more

macro approach.

Equation (6) was estimated for the period 1968-77. For a forecasting

test, it was estimated for the periods 1968-72 and 1972-76. The sample

was defined as it was for the decision-to-move analysis, except that

only migrants were considered. Migrants from New York, Ohio, Pennsylvania,

and South Carolina were then grouped together. The top eight destination

states were selected as potential alternatives. These included Florida,

California, New York, New Jersey, Michigan, Illinois, Texas, and Virginia.

Migrants choosing other states were excluded from the analysis.

The model is estimated using a multinomial "conditional" logit

program. Estimating equation (6) using this technique allows us to

derive predicted probability values. These probabilities can be rep-

resented as follows:

P.B
i tp
p. =e
SJ itp (10)
Ee
i=l

where i indexes alternatives and J is the total number of choices

available to the individual. The variables in the vector P. will be
i
defined as differences between the value at location i and the value at

the individual's origin of residence. Probabilities will be calculated

for migration between each possible origin and destination combination.

Since starting from any origin these values will-not vary across indi-

viduals, there will be a total of 32 probabilities calculated (8 alter-

natives x 4 origins).






34

All variables will be premigration measures of location aggregates.

Specific variables to be included in the choice of destination equations

will be defined and discussed in Chapter 4. The next section will

propose a method of obtaining a forecast from the completely estimated

model.


A Forecasting Methodology

In an application of logit analysis to transportation decisions,

McFadden (1974) states that if the sample under consideration represents

a random selection of the environments faced by the population as a

whole, the average of the predicted values over the sample is a best

estimate of aggregate demand. With this in mind a methodology is proposed

for forecasting the number of migrants from a given origin and their

destination choices. The accuracy of this methodology will be tested

within the sample.

For any given origin state, the coefficients of the 1968-72 decision-

to-migrate equation are applied to values of the explanatory variables

for individuals living in that state in 1972. In this way predictions

for the next four-year period, 1972-76, are obtained. Predicted proba-

bilities for each individual are obtained and all individuals are averaged.

This average probability is then applied to the total sample in 1972 to

get a forecast of the number of migrants between 1972 and 1976. This

number can then be compared to the actual (known) number of migrants in

the sample during the period. If successful, then 1972-76 coefficients

can be applied to the 1976 sample to obtain a 1976-80 forecast. The

next step involves application of the sample-derived average proba-

bilities to aggregate (outside the sample) measures of the labor





35

force in the state under consideration. In this way a forecast of the

actual number of labor force migrants from that state is obtained.

Figure 2.1 contains a flow chart which illustrates the procedure just

described.

A similar procedure is followed for forecasting destination choice.

The model is estimated for the 1968-72 period and the coefficients are

then applied to locational characteristics in 1972. Forecasts of the

probabilities of migration between each origin and each destination are

obtained. These probabilities are then multiplied times the number of

migrants leaving each origin location. The values obtained can be

compared to the actual origin-destination flows during the 1972-76

period. If successful, the model is re-estimated for the 1972-76

period and these coefficients are applied to the 1976 locational char-

acteristics. A 1976-80 aggregate forecast can be obtained by applying

the probabilities derived from the 1972-76 estimation to the aggregate

decision-to-move forecast. Thus, forecasts of actual flows of labor

force migrants to all destinations are derived. The destination-choice

forecasting methodology is illustrated in Figure 2.2.

In addition to forecasting applications, estimating the model for

two equal length time periods allows for observation of coefficient

stability. In this way it can be determined whether events such as the

energy crisis and the deep recession of 1974 have had any impact on

migration decisions and destination choice.3 Since earlier period




In fact this goal will be difficult to attain. Four-year migration
periods contain many events which cannot easily be disentangled from one
another. Estimation for shorter periods is even harder because of
(migrant) sample size problems.





36



1972-1976
decision to move 1976 characteristics of
coefficients, individual t originating
origin i in state i






Predicted probability of moving out
of origin i, individual t, 1976-1980






Total labor Average of predicted probabilities
force, state of moving out of origin i, all
i, 1976 individuals, 1976-1980






Forecast of labor force migration
from state i, 1976-1980




Figure 2.1 Illustration of Forecasting Methodology: Decision to Move





37


1972-1976 1976 characteristics of
destination choice destination j relative
coefficients to origin i






Forecast of Predicted probability of moving from
labor force origin i to destination j, 1976-1980
migration from
state i




Forecast of labor force migration
between origin i and destination j





Figure 2.2 Illustration of Forecasting Methodology: Destination Choice






38

migration coefficients are used to forecast later period migrant flows,

coefficient instability implies forecasting problems. Employing energy

shortage coefficients, for example, to project migration for a period in

which we expect no energy problems would be incorrect.

Finally, historical estimates of aggregate labor force migration

can be obtained from the sample by utilizing a methodology similar to

that proposed for forecasting. For the decision-to-move analysis, the

average probability of migration for each state during any given period

can be applied to a measure of the aggregate labor force at the beginning

of that period. The resulting estimates of out-migration can then be

distributed out among destinations by multiplying them times the prob-

abilities obtained from the destination-choice equations for the same

period. Thus, estimates of aggregate migration between all possible

origins and destinations are derived from the results of the entire

analysis. In the next section of this chapter, the sources of data will

briefly be described.


Sources of Data

The primary source of data for this study is the University of

Michigan's Panel Study of Income Dynamics. In this survey, 5,862

families have been interviewed each year since 1968. At present, 10

years of data are available on tape. Each year approximately 450

variables are available for each family. Categories under which the

variables can be grouped are family composition information, education,

transportation, housing, employment of head, housework, work for money

by wife, food and clothing, income, intelligence, feelings, and time

use.






39

Typical sources of data for migration research are the Continuous

Work History Sample of the Social Security Administration and the Public

Use Sample derived from the 1970 Census of Population. The Panel Study's

advantage over these sources lies in its richness of variables and in

its continuity and consistency of sampling. Its disadvantage is that it

is a much smaller sample. But since its patterns of movement over the

sample period closely resemble the movement identified in these other

sources, great confidence is placed in its use as a tool in migration

studies.

Aggregate data is obtained from various sources. Chief among these

is the Statistical Abstract of the United States. This is one of the

few publications that contains easily accessible and reasonably consistent

time-series data for all states. Other sources are Climatological Data,

National Summary, Cost of Living Indicators, and The Employment and

Training Report of the President.


Sample Size Limitations

Dividing up the sample according to state of origin reduces the

number of observations for each estimation substantially. For the state

of Ohio, 220 observations are drawn from the Panel Study tapes. Thirty-

seven of these individuals (and their families) moved out of Ohio in the

sample period. There are 195 New York observations, 29 of which are

migrants. Pennsylvania has 202 observations and 28 of these moved out

of state during the nine-year period. The sample size for South Carolina

is 222, with 32 being migrants. The destination-choice equations grouped

together 85 migrants from all four states who chose among eight potential

destinations.






40

Logit analysis depends upon the assumption that the dependent

variable (the logarithm of the odds that a particular choice will be

made) approximates the normal distribution. A large number of obser-

vations and sufficient repetitions for each possible choice assure that

this criterion will be met. For the decision-to-move analyses it is

believed that sample size is sufficient for employing logit techniques.

The destination choice results should be interpreted more cautiously. A

particular problem for these equations is the small number of observations

occurring in each possible category of choice.4

All equations were also estimated using ordinary least squares

techniques (see Chapter 3). The signs and magnitudes of the coefficients

obtained were very similar to those obtained with logit analysis.5

These similarities increase the degree of confidence placed in the major

findings of this study. In Chapter 3 the results of the decision-to-

move analysis are put forth and discussed.




SAn additional source of worry arises because origin states are
chosen according to their rates of migration to Florida. This may
introduce an upward bias in the predictions of migration to Florida.
Consequently, the probabilities of migration to other states will be
understated.

5DaVanzo (1977) compares probit estimates of decision to move
equations with ordinary least squares estimates. She also finds similar
results.














CHAPTER 3

EMPIRICAL RESULTS: DECISION TO MOVE

Interstate Migration: 1968-1977


In Table 3.1, a summary of interstate movement among Panel Study

respondents is presented. It can be seen that the states receiving the

largest number of migrants between 1968 and 1977 were California, Florida,

and Texas. The leading out-migration states were California, Ohio,

South Carolina, New York, and Pennsylvania. Net migration, defined as

the difference between the number of in-migrants and out-migrants, is

greatest in Florida and Texas. Twenty-two states registered negative

levels of net migration, with the greatest declines occurring in Ohio

and South Carolina. The negative value recorded in California is

surprising and leads to some concern about whether the Panel Study

Sample is representative or not. The fact that this study focuses upon

gross rather than net migration is some consolation, but the small

number of migrants in the sample is a worrisome factor.

States were chosen for this analysis according to their rates of

migration to Florida. Of major interest in this study are the determinants

of migration decisions and destination choice among residents of origin

states for Florida migrants. This will aid in developing a forecast of

migration from these states to Florida and to competing destinations

such as Texas and California. A summary of Florida in-migration by

state of origin appears in Table 3.2. There it is shown that 58.2

percent of all Florida migrants in the sample came from either New York,



41






42

Table 3.1 Summary of Panel Study Migration, 1968-1977


Number of In- Number of Out-
State Migrants Migrants Net Migration

Alabama 7 3 4
Arizona 11 6 5
Arkansas 3 11 -8
California 44 48 -4
Colorado 16 6 10
Connecticut 8 7 1
Delaware 3 0 3
District of
Columbia 9 12 -3
Florida 43 11 32
Georgia 9 5 4
Idaho 2 0 2
Illinois 23 22 1
Indiana 11 14 -3
Iowa 4 17 -13
Kansas 6 0 6
Kentucky 5 14 -9
Louisiana 5 13 -8
Maine 7 1 6
Maryland 18 9 9
Massachusetts 14 10 4
Michigan 10 10 0
Minnesota 5 7 -2
Mississippi 4 16 -12
Missouri 4 19 -15
Montana 0 0 0
Nebraska 7 3 4
Nevada 7 0 7
New Hampshire 5 0 5
New Jersey 21 15 6
New Mexico 4 0 4
New York 25 30 -5
North Carolina 5 12 -7
North Dakota 0 0 0
Ohio 11 37 -26
Oklahoma 5 4 1
Oregon 17 8 9
Pennsylvania 15 28 -13
Rhode Island 0 0 0
South Carolina 5 33 -28
South Dakota 1 5 -4
Tennessee 10 5 5
Texas 39 14 2-5
Utah 1 7 -6
Vermont 1 0 1
Virginia 22 15 7





43

Table 3.1 (continued)


Number of In- Number of Out-
State Migrants Migrants Net Migration

Washington 6 9 -3
West Virginia 2 0 2
Wisconsin 6 2 4
Wyoming 2 0 2





44

Table 3.2 Florida In-migration by State
of Origin, 1968-1977


Percentage of
Origin State Number of Florida Migrants Florida Migrants

New York 7 16.3

Pennsylvania 7 16.3

Ohio 6 14.0

South Carolina 5 11.6

Indiana 3 7.0

Kentucky 2 4.7

Virginia 2 4.7

Missouri 2 4.7

Alabama 1 2.3

Arizona 1 2.3

Arkansas 1 2.3

California 1 2.3

Connecticut 1 2.3

Illinois 1 2.3

Louisiana 1 2.3

Maryland 1 2.3

North Carolina 1 2.3





45

Pennsylvania, Ohio, or South Carolina. After California, these are also

the leading out-migration states in the sample. These four states were

thus selected as the major sample for this analysis.

Table 3.3 defines the variables to be considered in the decision-

to-move analysis. In the next section, the rationale for including each

variable will be discussed and a first look at the data will be provided.


Comparison of Migrants and Nonmigrants

Family Size

Table 3.4 presents a comparison of migrant and nonmigrant char-

acteristics in the four-state sample. A priori it is expected that

persons in larger families are less likely to move since additional

people represent ties to present location. Children in school or a

spouse who is working are both examples of migration ties. Comparison

of the mean values of family size among migrants and nonmigrants, however,

reveals very little difference. Both groups average about five persons

per family. Thus, at first glance, family size does not appear to be a

constraining force.


Sex

Households with male heads are expected to move more often than

households with female heads. One reason is that men are more likely to

be employed in occupations where job transfers are common. In addition,

higher rates of unemployment for women at many locations and lower

income opportunities would deter female migration. The mean value of

the sex variable, however, is only slightly higher for movers than it is

for nonmovers. Eighty-seven percent of the households that moved were

headed by men. It may be that since women are also likely to be earning






46

Table 3.3 Variable Definitions for Decision to Migrate Analysis


Dependent Variable Definition

Move Equals 1 if the person's state of residence
in 1977 differs from his or her state of
residence; equals 0 otherwise. (Defined
in 1968 for different periods in later
analysis.)

Explanatory Variables

Famsz Family size, actual number in household.

Sex Equals 1 for male; 2 for female.

Homeown Equals 1 if home is owned; equals 0
otherwise.

Lres Length of residence in current house or
apartment, takes on following values:

0 if length of residence < 1 year
1 if length of residence = 1 year
2 if length of residence = 2 years
3 if length of residence = 3 years
4 if length of residence = 4 years
5 if length of residence > 5 years and
< 10 years
6 if length of residence > 10 years and
< 15 years
7 if length of residence > 15 years and
< 25 years
8 if length of residence > 25 years

Lempl Length of time employed by present employer,
takes on following values:

0 if self-employed or unemployed
1 if employed > 0 months and < 6 months
2 if employed > 6 months and < 18 months
3 if employed > 18 months and < 42 months
4 if employed > 42 months and < 9 years
5 if employed > 9 years and < 19 years
6 if employed > 19 years

Emself Equals 1 if self-employed; equals 0
otherwise.

Avage Average age of husband and wife; if unmarried
Avage = age of household head.






47

Table 3.3 (continued)


Explanatory Variables Definition

Emst Employment status, equals 1 if employed;
equals 0 if unemployed.

Marst Marital status, equals 1 if married;
equals 0 otherwise.

Race Equals 1 if white; equals 0 otherwise.

Aveduc Average education level of husband and wife;
if unmarried Aveduc = education of household
head, takes on following values:

1 if education > 0 years and < 6 years
2 if education > 6 years and < 9 years
3 if education > 9 years and < 12 years
4 if education = 12 years
5 if education = 12 years and person has
some non-academic training
6 if person attended college, but did not
receive a degree
7 if person received a bachelor's degree
8 if person attended graduate school

Prevmig Previous migration, equals 1 if current
state of residence differs from state of
birth; equal 0 if current state of
residence is the same as state of birth.

Emstw Equals 1 if spouse works; equals 0
otherwise; if no spouse then Emstw = 0.

Famy Total family income, in thousands of dollars.





48

Table 3.4 Comparison of Migrant and Nonmigrant Characteristics:
New York, Ohio, Pennsylvania, and South Carolina


Variable Mean Value Migrants Mean Value Nonmigrants

Famsz 4.98 4.95

Sex 0.87 0.84

Homeown 0.64 0.58

Lres 5.93 5.87

Lempl 3.14 3.31

Emself 0.10 0.08

Avage 37.89 35.31

Emst 0.96 0.98

Marst 0.78 0.81

Race 0.71 0.68

Aveduc 4.11 3.78

Prevmig 0.31 0.23

Emstw 0.38 0.35

Famy 11.24 8.79





49

lower incomes at their current locations and have a higher probability

of being unemployed before moving, they are more responsive to oppor-

tunities elsewhere. This would counteract the negative effects of the

sex variable and lead to a more ambiguous result. No conclusions can be

drawn without more rigorous analysis.


Homeownership

Owning a home is usually expected to act as a tie to present location.

Time and transactions costs involved in selling a home and perhaps

purchasing another one in a new location would tend to deter migration.

On the other hand, homeownership represents wealth. If a wealth effect

operates, property owners may be more likely to move than renters. In

this sample 64 percent of the migrants owned homes prior to migration

while only 58 percent of the nonmigrants were homeowners. Thus, the

data give some support to the second hypothesis.


Length of Residence

Persons who have lived in their current houses or apartments for

long periods of time are expected to display a lower probability of

moving. Duration of past residence should act as a measure of will-

ingness to pull up stakes, leave friends and relatives, and start all

over somewhere else. An individual who remains in the same dwelling

unit for a long time displays an aversion to risk which leads one to

believe that he (or she) would not be likely to do something as filled

with uncertainty as changing states of residence. The data in Table 3.4

do not demonstrate this to be the case. There appears to be very

little difference in length of residence for migrants and nonmigrants.

The average length of residence for both groups is close to 10 years.





50

Length of Employment

As with the last variable, the length of employment would indicate

attachment to present environment. In addition, it can be viewed as a

measure of ability to hold a job. A person who has worked for his (or

her) present employer for only a short period of time is more likely to

have been periodically unemployed. The mean value of this variable is

indeed somewhat lower for migrants than for nonmigrants in the states

under consideration. The average length of employment before moving is

about 21 months for migrants.


Self-Employment

Individuals who are self-employed are expected to migrate less. The

self-employed person is more likely to have an occupation that involves

heavy investment in capital equipment. Dentists and printers are examples.

Although this equipment may be transportable, the costs of moving and

re-establishing at some other location are significant enough to deter

migration. Self-employed workers are also likely to be part of smaller

organizations where job transfer is uncommon. Finally, those who are

employed in professions requiring state licenses, such as medical and

legal fields, are bound to their state to some extent by their licenses.

The mean values for the four sample states show very little difference

in propensity to migrate based upon self-employment, with migrants

showing a slightly higher degree of self-employment. Ten percent of the

migrants from these states were self-employed before moving.





51

Average Age

In most migration studies age is expected to be negatively related

to migration since older people have a shorter work life over which to

realize the gains from migration. In addition, it is felt that older

people have greater ties to present location. That is, they are likely

to have stronger attachments to friends and relatives and are likely to

possess location-specific capital, such as a home. Only upon retirement

are older people expected to show higher migration rates. But since

this study is restricted to members of the labor force, the inverse

relationship is still expected here. It appears, however, that in this

sample, the average age is greater for migrants than for nonmigrants.

Perhaps this occurs because the sample is restricted to Florida-sending

states. Many of the movers to Florida and competing destinations, such

as California, may be anticipating future retirement. For example, a

person may move to Florida and accept a low-paying job knowing that (he

or she) will withdraw from the labor force after a predetermined number

of years. Individuals who have accumulated substantial savings during

their lifetime can afford to do this. In addition, many older migrants

may choose semiretirement rather than retirement, perhaps maintaining

part-time jobs after moving. If the preretirement or semiretirement

thesis is correct, we would expect labor force migration from the sample

states to consist of older migrants than migration from all other states.

In the Panel Study, it is found that the average age of all U.S. labor

force migrants is 36 years old. This compares to a mean age of 38 in

the sample of Florida-sending states. Thus, there is mild evidence in

favor of this hypothesis.





52

Employment Status

People who are unemployed are expected to be more responsive to

opportunities elsewhere than those who are employed. Particularly if

local economic conditions are poor, the process of job search for the

unemployed is likely to include alternative locations. If, however,

employment conditions at other locations are also depressed, the response

of the unemployed is uncertain. The effect of employment status upon

the migration decision seems to depend on local economic conditions

relative to alternative destinations. In the analysis to follow, relative

conditions will be held constant by considering only one state at a

time. The mean value of employment status for all four states combined

is slightly lower for migrants than for nonmigrants. Ninety-six percent

of the migrants in the sample were employed prior to moving.


Marital Status

Marriage is expected to act as a deterrent to migration. As with

the family-size variable, presence of a spouse represents an additional

tie to present location. This is particularly true if the spouse is

working. The data in Table 3.4 show that migrants are less likely to be

married than nonmigrants. Seventy-eight percent of the migrants in the

four-state sample were married, while 81 percent of the nonmigrants were

married.


Race

The expected sign of the race variable is ambiguous. On the one

hand it can be argued that since some blacks are more likely to earn

lower incomes and are more likely to be unemployed than whites, they are





53

more responsive to opportunities elsewhere. On the other hand, there

may also be fewer opportunities elsewhere for blacks because they are

concentrated in low-skill occupations and because of discrimination in

the labor market. In addition, blacks are less likely to be employed in

jobs requiring geographic transfers. For the sample to be considered

here it looks as if migrants are more likely to be white, indicating

that the latter two effects outweigh the first factor.


Average Education

More educated people are expected to move more often. The reason

is twofold. For one, education (in most cases) leads to increased

employment opportunities. Particularly it is believed that increased

education opens up employment possibilities which are more national in

scope resulting in greater migration propensities. In addition, those

who have attained higher education levels are expected to have more and

better information about migration opportunities. The extreme of this

phenomenon occurs in professions that form organizations which aid their

members in locating jobs. The economics profession is an example. The

data in Table 3.4 give preliminary evidence that this thesis is correct.

Migrants in the sample are likely to have at least started college,

while the average nonmigrant posesses only a high school education.


Previous Migration

Migrants are believed to be characterized by frequent mobility

during their lifetimes. It is hypothesized here that a person who has

changed state of residence at least once during his (or her) lifetime is

more likely to move again than an individual who has always lived at the

same location. People who have moved frequently are also less likely





54

to have developed attachments to their present state of residence. In

the four-state sample, 31 percent of the migrants have moved before.

Only 23 percent of the nonmigrants have previously changed their states

of residence. Thus, there is support for the hypothesis set forth

above.


Employment Status of Spouse

The presence of a working spouse is expected to deter migration.

If both partners have substantial earnings, any migration decision must

presumably maximize their combined return. The probability of finding a

location that meets this criterion is lower than the chances of finding

one which maximizes the returns of one family member. It is possible

that only one member of the couple will be properly rewarded upon moving

and the spouse will be tied to this choice. Whether joint utility is

maximized or not, the spouse's employment and income possibilities must

be taken into account in any decision to move. Theoretically, this

should act as a deterrent to migration; however, preliminary evidence in

Table 3.4 suggests that migrants in the sample were more likely to have

working spouses. Thirty-eight percent of the migrants' spouses worked

prior to moving.


Family Income

Migrants are expected to earn higher annual income than nonmigrants.

A few reasons can be given for this belief. First, higher incomes

usually imply greater demand for leisure activity. Changing location is

one way of obtaining more leisure. This is particularly true for this

sample of states which have typically sent large numbers of migrants to

Florida and California, states which possess climates amenable to





55

recreational activities. Higher income people are also more likely to

have visited other states as tourists. Thus, they would have obtained

information about these states which could lead to increased likelihood

of moving. The data in Table 3.4 show significant differences between

family income levels of movers and nonmovers. Migrants earned an average

of $11,240 in 1968. This compares to an average income level of $8,790

for nonmigrants.


Empirical Results

Introduction

The comparison of migrant and nonmigrant characteristics presented

so far serves only to provide a general picture of relationships in the

data. Much can be hidden in such a rudimentary analysis. Since all

four sample states were grouped together, differences in migrant selection

between states are hidden. As an example, migrants from New York may

tend to have larger families than New Yorkers who do not move. There

may, however, be little difference in the family size of movers and non-

movers from Ohio, Pennsylvania, and South Carolina. When all four

states are viewed together,very little difference in family size will

show up, even though this is an important determinant in New York. This

problem becomes particularly severe if the variable takes on opposite

signs in each state.

Theoretical reasons exist for expecting differences in the deter-

minants of migration between states of origin. The crux of the argument

is that residents of different states face varying relative migration

opportunities. Residents of Ohio, for example, will experience a greater

improvement in climate from moving to Florida than will residents of






56

South Carolina. If older people are more responsive to climate differ-

entials than younger people,we might then expect age to be a more impor-

tant migration determinant in Ohio than it is in South Carolina. Thus,

variables are significant in one state but not in another because of

differences in initial (origin) conditions. From now on, each state

will be analyzed by itself.

When migrants and nonmigrants are compared only according to the

mean values of a set of variables, other problems arise. First, there

are the usual weaknesses involved in using the mean as a measuring

device (its sensitivity to extreme values is one such weakness). More

importantly, when any variable is considered, the values of remaining

factors are not held constant. Thus, the true effect of each char-

acteristic is not being measured. Use of multiple regression techniques

will relieve these problems.

Before the results of the logit analysis of the decision to move

are presented, a discussion of the scaling of some of the qualitative

(dummy) explanatory variables is in order. Particularly, length of

residence, self-employment, and average education are defined in ways

which are somewhat out of the ordinary. Looking back at Table 3.3 it is

observed that these variables are scaled using an ordinal ranking system,

with each value representing a given level or interval. This scaling is

the same as that which appears on the Panel Study tapes. The conventional

method of modeling variables which are coded in such a fashion is to

include separate dummy variables for each category of data. In this

case, however, 21 variables would need to be added to the model. This





57

would result in reduction of degrees of freedom and, more importantly,

would cause logit estimation to become prohibitively expensive.1

A second method of handling the problem would be to redefine the

dummy variables so as to have fewer categories. In the limiting case

only one zero-one dummy is used. Some cutoff point is chosen and a

value of one is assigned to those observations for which the cutoff is

exceeded and a value of zero is assigned otherwise. In this study, for

example, the education variable could be defined so that those individuals

with 12 years or more of schooling were given values of one with the

remaining observations equaling zero. Assumed in such an approach is

that increases in the level of education up until the 12th grade have no

effect on the probability of moving. In addition it is assumed that

increases in education beyond high school have no effect. To validly

use such an approach some prior expectation of the proper cutoff point

is required. Appendix A presents the results of logit estimation when

this approach is adopted. Not having prior knowledge of the correct

cutoff point, the choices used are somewhat arbitrary.

The actual scaling used for the main part of this study (see Table

3.3) also imposes restrictions upon the model. To illustrate these

restrictions, this approach is compared to the conventional method of

defining such data. Suppose an explanatory variable contains three

categories. Following the conventional method, two dummy variables are

introduced. (If three dummies are used,then the constant term must be

eliminated.) Call the dependent variable Y and the two dummies X1 and X2.




1In fact, given the computer program limitations, logit estimation
would have been impossible.






58

Assuming no other variables in the model, the equation can be written as:

Y = a + BIX1 + A2X2 + E (1)

where a, 1, and A2 are coefficients and E is a random disturbance term.

Since X1 and X2 are zero-one variables we can further say that:

E (Y X = 1 and X2 = 0) = a + 1 (2)

E (Y X = 0 and X2 = 1) = a + 2 (3)

E (Y X1 = 0 and X2 = 0) = a (4)

Now suppose instead of using the above approach we define a single

variable, X, which takes on the values of 0, 1, and 2 for each of the

three categories. Now the equation can be written as:

Y = a + BX + E (5)

where a and B are parameters and E is the disturbance. Now we can

say that:

E (Y X = 0) = a (6)

E (Y X = 1) = a + B (7)

E (Y X = 2) = + 2 (8)

This formulation turns out to be equivalent to estimating the model of

equation (1) with the following restriction attached:

= 2 (9)

While this restriction should be considered, it may be preferable to the

restrictions imposed when the variable is collapsed into two categories

(as discussed earlier). Using the latter approach, information which

is available about alternative categories goes unused. The chosen

technique uses all of the information, but it imposes rather tight

restrictions on the relationship between categories.





59

Summarizing, it would be optimal to include separate dummy variables

for each category for which there is information. Program and cost limi-

tations, however, prevent using this approach. Alternatives include

collapsing the variable into fewer categories or including a single

ordinally ranked categorical variable. While the latter approach is

somewhat unprecedented2, it has been chosen because it utilizes more

information which is available in the data. Both alternatives impose

restrictions which may be considered severe without prior knowledge

of data. In the next four sections, the results of the logit analysis

of the decision to move will be presented and discussed.


Logit Results: Ohio

Table 3.5 presents the regression results for the state of Ohio.

Variables which are significant with at least 95 percent confidence

are sex, marital status, and family income. The sex variable, as

hypothesized, has a positive coefficient. Thus, households with male

heads from Ohio are more likely to migrate than households with female

heads from that state. The most significant variable is marital status.

As expected, unmarried individuals move more often than those who are

married. Thus, marriage does act as a deterrent to migration in Ohio.

Family income is also positive and significant. Higher income families

in Ohio are more likely to change states of residence than lower income

families. This is also the expected result.




2To test whether wage and price controls of the Nixon administration
had any effect upon wage formation, a three-leveled ordinal dummy was used
by Eckstein and Girola (1978). The variable took on the value of .5 for
Phase I of controls and 1 for Phase II of controls. Periods without con-
trols took on the value of 0. The variable was found to be insignificant.
The same variable was used in a price equation and was significant.





60

Table 3.5 Results of Logit Analysis: Decision to Move,
Ohio, 1968-1977


Variable Coefficient Asymptotic t-Ratio

Famsz .147 1.27

Sex 2.802 2.60

Homeown -3.65 .64

Lres .013 .06

Lempl .052 .31

Emself -.845 .73

Avage .054 1.53

Emst -

Marst -4.142 -3.39

Race .867 1.09

Aveduc -.024 .15

Prevmig .298 .69

Emstw .471 .96

Famy .057 2.06



Constant -5.000 -2.72

Summary Statistics:

Log of Likelihood Function = -86.0

Likelihood Statistic (at 0) = 133.0

Significance Level (at 0) = .005

Likelihood Statistic (at mean) = 27.3

Level of Significance (at mean) = .025

Percentage Correctly Predicted = 85%

Percentage of Migrants Correctly Predicted = 16%





61

The remaining variables are not significant. Family size, average

age, and race are variables which show some importance. Migrant families

in Ohio are more likely to be larger than nonmigrant. This result is

opposite from that expected. A possible explanation is that many

migrant families from Ohio have preschool children. While having

children in school may act as a deterrent to migration, families may

decide to move prior to their children's first enrollment in school.

Indeed, people may move so that their children can go to different

schools. Age shows a positive sign, lending support to the prere-

tirement hypothesis set forth earlier. Race is also positive, indi-

cating that households headed by white persons are more likely to move

interstate than households headed by nonwhites.

Although insignificant, the homeownership variable has the ex-

pected sign. People owning homes in Ohio are less likely to move than

those who rent. They are also more likely to have lived in their homes

and worked for the same employer for longer time periods than non-

migrants. These results are contrary to expectations. The migrant is

less likely to be self-employed. Thus, self-employment appears to act

as a migration deterrent. Those who chose to leave Ohio are more

likely to have moved previously, supporting the hypothesis put forth

earlier. Education carries an unexpected negative coefficient. Migrant

household heads are also likely to have spouses who are employed. This

variable has the opposite sign from that which was expected. Employment

status was not included in the analysis because there was no variation

of this characteristic among migrants. Estimation was thus impossible.

When the likelihood value of the decision equation is compared to

the value that results when.all coefficients are zeroed out, the equation

is significant at better than the .005 level. When the likelihood





62

value of the equation is tested against that which results when the

right-hand side of the equation is set equal to the mean of the dependent

variable, the equation is significant at the .025 level. Under either

test, the equation appears to produce a good fit.

A correct prediction is defined as one in which the predicted

probability is within 50 percent of the actual value of the dependent

variable.3 For migrants this means that Pmove > .50. For nonmigrants

this means Pmove < .50. Under this criterion, the equation predicted 85

percent of all (migrant and nonmigrant) individuals in the sample correctly.

Among migrants, 16 percent were correctly predicted.


Logit Results: New York

Table 3.6 presents the decision-to-move results for the state of

New York. Variables which are significant at better than the .05 level

are length of employment, average age, employment status, and spouse's

employment status. Length of employment carries a negative sign, implying

that those who have worked for the same employer for long periods of

time are less likely to move. This result is expected. Migrants from

New York are also older than nonmigrants from that state, providing

further support for the preretirement thesis. Migrant household heads

are less likely to be employed than nonmigrants, but their spouses are

likely to be employed. This last result, which was also found to be

true for Ohio, is contrary to expectations. Perhaps what we are observing




3This definition is admittedly arbitrary. Some cutoff had to be
chosen in order to summarize the results without listing every predicted
value.





63

Table 3.6 Results of Logit Analysis: Decision to Move,
New York, 1968-1977


Variable Coefficient Asymptotic t-Ratio

Famsz -.165 -1.37

Sex .127 .08

Homeown .773 1.38

Lres .374 1.51

Lempl -.561 -2.80

Emself -1.277 -1.40

Avage .074 2.30

Emst -2.960 -1.99

Marst -1.600 -1.03

Race -.890 .89

Aveduc .227 1.34

Prevmig -.448 .71

Emstw 1.682 2.77

Famy .084 1.88



Constant -2.580 -1.39

Summary Statistics:

Log of Likelihood Function = -62.8

Likelihood Statistic (at 0) = 144.8

Level of Significance (at 0) = .005

Likelihood Statistic (at mean) = 38.4

Level of Significance (at mean) = .005

Percentage Correctly Predicted = 83%

Percentage of Migrants Correctly Predicted = 21%






64

are married couples with both spouses being employed in highly mobile,

perhaps professional, occupations. In any case, employed spouses are

not acting as deterrents to migration.

Family income is significant at the 94 percent confidence level.

It carries its expected positive sign. The remaining variables are

insignificant. Sex, self-employment, marital status, and average

education have the signs which were hypothesized earlier. Homeownership

has a positive sign, indicating that migrants are more likely to own a

home than nonmigrants. They are also more likely to have lived in their

homes for longer periods of time. Race, although insignificant, carries

a negative sign. This lends support to the thesis that blacks are more

responsive to opportunities elsewhere. Previous migration also has the

opposite sign from that expected. Migrants from New York are less

likely to have moved before in their lifetimes. This is consistent with

the sign on the length-of-residence variables. Perhaps many of the New

York migrants are middle-aged or older people who have lived in New York

most of their lives and are now moving in anticipation of retirement.

The overall equation is significant at better than the .005 level.

This is true whether the model is compared to one where all coefficients

are zero or compared to a model where the right-hand side is equal to

zero. Eighty-three percent of the individuals in the sample are pre-

dicted correctly by the model, while 21 percent of the migrants are

predicted correctly.


Logit Results: South Carolina

The decision-to-move results for South Carolina appear in Table

3.7. Significant variables are family size, average age, marital





65

Table 3.7 Results of Logit Analysis: Decision to
Move, South Carolina, 1968-1977


Variable Coefficient Asymptotic t-Ratio

Famsz .252 1.95

Sex 1.231 1.04

Homeown -.180 .33

Lres .286 1.17

Lempl -.328 -1.76

Emself -2.713 -1.65

Avage .093 2.51

Emst -.678 .66

Marst -3.494 -2.46

Race -.371 .58

Aveduc .073 .38

Prevmig .803 1.24

Emstw -.528 .89

Famy .217 3.22



Constant -6.333 -2.96

Summary Statistics:

Log of Likelihood Function = -71.7

Likelihood Statistic (at 0) = 164.4

Level of Significance (at 0) = .005

Likelihood Statistic (at mean) = 39.8

Level of Significance (at mean) = .005

Percentage Correctly Predicted = 90%

Percentage of Migrants Correctly Predicted = 31%





66

status, and family income. Marital status and family income show the

expected signs. Contrary to expectations, family size is positive and

significant. Again average age carries a positive sign.

Although insignificant at the 95 percent confidence level, length

of employment and self-employment are important and carry the correct

signs. Of the remaining variables, sex, homeownership, employment

status, average education, previous migration, and spouse's employment

status influence migration decisions in the expected directions. As in

Ohio and New York, length of residence is positively related to migration.

Race displays a negative relationship, indicating that households

headed by blacks are more likely to leave the state.

The South Carolina equation provides a very good fit of the data.

The equation is significant by both criteria (at zero and at the mean)

at better than the .005 level. Ninety percent of the individuals in the

sample were correctly predicted by the model, with 31 percent of the

migrants' choices being predicted accurately.


Logit Results: Pennsylvania

In Table 3.8 are the results of the regression analysis for Penn-

sylvania. Variables which are significant at the .05 level are average

education and family income. Average education shows the expected

positive sign. Family income, however, is negatively related to the

probability of moving. This is truly unexpected for a variable that has

been positive and significant for the other three states.

Close to being significant are family size and marital status. As

in Ohio and South Carolina, family size takes on-a positive sign.

Marital status has the hypothesized negative sign, indicating marriage





67

Table 3.8 Results of Logit Analysis: Decision to
Move, Pennsylvania, 1968-1977


Variable Coefficient Asymptotic t-Ratio

Famsz .251 1.85

Sex 1.859 1.51

Homeown .105 .19

Lres .022 .11

Lempl .253 1.34

Emself 1.257 .88

Avage .021 .59

Emst

Marst -2.625 -1.84

Race .109 .17

Aveduc .638 2.88

Prevmig .225 .36

Emstw .869 1.66

Famy -.219 -2.65



Constant -5.650 -2.78

Summary Statistics:

Log of Likelihood Function = -71.6

Likelihood Statistic (at 0) = 57.2

Level of Significance (at 0) = .005

Likelihood Statistic (at mean) = 19.4

Level of Significance (at mean) = .10

Percentage Correctly Predicted = 87%

Percentage of Migrants Corrected Predictly = 7%





68

deters migration. Sex and previous migration are both positively related

to migration, as expected. Homeownership, length of residence, length

of employment, self-employment, and spouse's employment status all show

the wrong sign. Average age has a positive sign, giving further support

to the preretirement thesis, and race is positively related to migration,

indicating that white families leave Pennsylvania more often than do

black families.

The Pennsylvania equation provides a poorer fit than the decision

equations for the other three states. Although significant compared to

the model where all coefficients are zero, the more stringent test

indicates a weaker model. Setting the right-hand side of the equation

equal to the mean value of the dependent variable and comparing the

likelihood value which results with the unconstrained likelihood value

results in a significance level of .10 for the unconstrained model.

Although 87 percent of the sample was predicted correctly by the model,

only 7 percent of the migrants were predicted correctly. Thus, this

model has not been very successful in distinguishing between migrants

and nonmigrants.


Test for Separation of States

In Chapter 2, a theoretical argument was made for estimating

separate equations for each state of origin. In the preceding sections

of this chapter evidence was presented which indicated that there may be

differences between states in the determinants of the migration decision.

In this section, a formal test for equality of coefficients between

states is carried out.





69

The test developed by Chow (1960) for the equality of regression

coefficients between equations is applied to the decision-to-move

equations. First, ordinary least squares (OLS) estimates for each state

are determined. Then the observations from all four states are combined

and an OLS equation is estimated for the combined data.4 The results of

these estimates appear in Table 3.9. The following statistic is calculated:

[SEETotal (SEENY + SEEOH + SEESC + SEEpA)] / 3k

[SEENY + SEEOH + SEESC + SEEpA] / n-4k

where SEE stands for the sum of squared residuals, k refers to the

number of coefficients in each equation, and n refers to the size of the

combined sample. The subscript NY refers to the New York equation with

OH standing for Ohio, SC referring to South Carolina, and PA meaning

Pennsylvania. The subscript TOTAL refers to the combined equation. The

statistic is distributed according to an F-distribution. Its value in

this case is 1.67 which compares to a critical F value (14 degrees of

freedom in the numerator, and 779 in the denominator) of 1.5 at the .05

level of significance. Thus, the hypothesis of equality of regression

coefficients between equations is rejected at the 95 percent confidence

level. The belief that there are differences in migration determinants

across states is confirmed.




4OLS is chosen for the test because execution of the logit program
is impossible for the combined sample and because no comparable test
exists for comparing coefficients across logit equations.






70

Table 3.9 Ordinary Least Squares Results: Decision to Move, 1968-1977


Ohio New York Pennsylvania
Variable Coefficient t-Ratio Coefficient t-Ratio Coefficient t-Ratio

Famsz .015 1.02 -.020 -1.51 .023 1.60

Sex .494 3.32 .073 .50 .323 1.74

Homeown -.047 .70 .055 .95 .005 .09

Lres .002 .05 .060 1.32 .003 .15

Lempl .005 .21 -.047 -2.63 .024 1.18

Emself -.105 .77 -.061 .59 .161 1.06

Avage .006 1.46 .009 2.75 .002 .43

Emst .037 .19 -.348 -1.80 .009 .04

Marst -.626 -4.05 -.198 -1.35 -.402 -2.06

Race .069 .86 -.136 -1.47 -.012 .17

Aveduc -.003 .14 .020 1.16 .064 2.78

Prevmig .035 .63 -.041 .64 .015 .20

Emstw .044 .74 .128 2.21 .106 1.84

Famy .009 2.48 .009 1.71 -.016 -2.87


Constant -.210 .94 .366 1.69 -.210 .81

R = .13 R2 = .17 R = .09





71

Table 3.9 (continued)


South Carolina Combined States
Variable Coefficient t-Ratio Coefficient t-Ratio

Famsz .018 1.58 .010 1.54

Sex .174 1.35 .278 3.96

Homeown -.020 .40 .005 .18

Lres .040 1.10 .011 1.00

Lempl -.042 -2.21 -.014 -1.52

Emself -.228 -1.76 -.047 .78

Avage .009 2.83 .005 3.38

Emst -.032 .28 -.120 -1.47

Marst -.399 -3.16 -.275 -4.01

Race -.082 -1.32 -.016 .52

Aveduc .006 .29 .018 1.85

Prevmig .069 .98 .024 .81

Emstw -.066 -1.27 .041 1.47

Famy .026 4.03 .005 2.23


Constant -.045 .31 -.148 -1.17

R =.18 R =.06






72

Application to Forecasting

The models for all four states have been re-estimated for the

period 1968 through 1972. The results are presented in Table 3.10. For

the shorter time period there are, of course, fewer migrants. This led

to some problems. First, some of the variables which were used in the

nine-year migration analysis could not be included here. A requirement

for logit estimation is that there be variation within each category of

the dependent variable. Since there are fewer migrants in the sample,

there is less likelihood that some variables will meet this requirement

in the migrant category. Those that do not must be deleted. Second,

with a smaller proportion of migrants in the sample it is more difficult

to distinguish between migrants and nonmigrants. In general, poorer

fitting equations were obtained for the shorter time period. This is

because there were fewer variables that could be included in the model

and because of the smaller proportion of migrants in the sample.

Using the methodology outlined in Chapter 2 a forecast is developed.

The coefficients in Table 3.10 are applied to values of the explanatory

variables for individuals living in the sample states in 1972. Individual

predicted probabilities are calculated and averaged over the number of

individuals in each state. The mean probability for each state is then

multiplied by the total sample size in 1972. The results of this procedure

appear in Table 3.11.

The forecast values are compared to the actual (known) number of

movers during the 1972-76 period. For Ohio, 25 migrants are predicted.

In fact, there were 20 Ohio migrants in the sample between 1972 and

1976. For New York, 13 migrants are predicted compared to 17 actual










Table 3.10 Results of Logit Analysis: Decision to Move, 1968-1972


Ohio New York South Carolina Pennsylvania
Variable Coefficient t-Ratio Coefficient t-Ratio Coefficient t-Ratio Coefficient t-Ratio

Famsz -.813 .38 -.316 -1.66 .207 1.35 .164 .87

Sex 1.614 1.32 1.087 .89

Homeown .845 .71 1.118 1.36 -.541 .69 -.107 .15

Lres -

Lempl .570 1.55 -.743 -2.58 -.540 -2.23 .228 .94

Emself 2.885 1.30 -1.940 -1.57 -2.313 -1.41 -

Avage -.062 -1.03 .042 1.07 .074 1.59 -.010 .23

Emst -

Marst .350 .22 .859 .58 -3.76 -2.40 -1.798 -1.17

Race .249 .19 -1.239 -1.33 .405 .48

Aveduc -.193 .59 .388 1.63 .390 1.62 .647 2.35

Prevmig .204 .26 1.094 1.34 .960 1.18 1.338 1.82

Emstw .624 .77 1.617 2.14 -.876 .97 .711 .96

Famy .114 2.29 -.045 .67 .160 2.10 -.212 -1.86



Constant -4.770 -2.26 -4.949 -2.36 -4.868 -2.88 -5.266 -2.71










Table 3.10 (continued)


Summary Statistics Ohio New York South Carolina Pennsylvania

Log of Likelihood Function -38.5 -36.7 -43.9 -45.5

Likelihood Statistic (at 0) 243 246 245 210

Level of Significance (at 0) .005 .005 .005 .005

Likelihood Statistic (at mean) 11.4 23.4 29.8 18.0

Level of Significance (at mean) .10 .025 .01 .10

Percentage Correctly Predicted 95% 95% 94% 93%

Percentage of Migrants Correctly
Predicted 0% 15% 19% 7%





75

Table 3.11 Comparison of Forecast to Actual Migration, 1972-1976


State of Number of Number of Difference
Origin Migrants-Forecast Migrants-Actual (Forecast-Actual)

Ohio 25 20 5

New York 13 17 -4

South Carolina 27 15 12

Pennsylvania 10 15 -5

Total 75 67 8






76

movers. Twenty-seven South Carolina migrants are forecast. In fact,

only 15 left that state during the period. The model predicts 10

migrants from Pennsylvania, while 15 actually left the state. In total

75 migrants are forecast to have moved. This compares to 67 people who

actually moved. This is a 12 percent error. It is interesting to note

that the best fitting models in the historical period, 1968-72, do not

provide the most accurate forecasts. The best fitting equation is for

the state of South Carolina, yet the greatest forecast error occurs in

this case.

The decision equations for each state are estimated next for the

period 1972-76. The results appear in Table 3.12. It should first be

observed that these results look very different from the 1968-72 mi-

gration equation appearing in Table 3.10. Many variables which are

significant for the earlier period lose their significance in the latter

period and other variables gain significance. For example, for the

state of Ohio, the only significant variable in the 1968-72 equation was

family income. In the 1972-76 period, this variable is no longer

significant and length of employment becomes significant. Significant

variables in the earlier period New York equations are length of em-

ployment and employment status of spouse. In the later period, no

variables are significant at the 95 percent confidence level. Length of

employment, marital status, and family income were significant variables

in the first four-year period for the state of South Carolina. Only

family income remained significant in the 1972-76 period. For Penn-

sylvania, average education was the most significant 1968-72 variable.

Later, family size, length of employment, average age, and marital

status showed significance.. In addition, sign changes frequently











Table 3.12 Results of Logit Analysis: Decision to Move, 1972-1976


Ohio New York South Carolina Pennsylvania
Variable Coefficient t-Ratio Coefficient t-Ratio Coefficient t-Ratio Coefficient t-Ratio

Famsz .050 .34 .005 .03 .167 1.09 .307 1.96

Sex .617 .43 1.534 1.69

Homeown -.015 .02 -1.212 -1.66 -.710 -1.03 -.119 .17

Lres -

Lempl .610 3.13 -.244 -1.49 -.398 -1.71 -.380 -2.20

Emself -.715 .65 .921 .84 -

Avage -.019 .62 .016 .60 .004 .10 .084 2.50

Emst -.252 .19 -

Marst -1.108 -1.27 .475 .59 -.681 .53 -4.011 3.15

Race -.576 .71 -.902 .89 .680 .88

Aveduc .109 .55 .148 .78 -.126 .44 .378 1.31

Prevmig .280 .51 -.515 .74 1.387 1.62 .011 .012

Emstw .371 .56 -.432 .57 -.107 .15 1.516 1.64

Famy -.029 .70 .308 .74 .967 2.25 .0004 .005



Constant -3.805 -2.23 -2.520 -1.56 -2.148 .83 -7.128 -3.12










Table 3.12 (continued)


Summary Statistics Ohio New York South Carolina Pennsylvania

Log of Likelihood Function -58.6 -53.7 -47.0 -44.2

Likelihood Statistic (at 0) 218 89 85 182

Level of Significance (at 0) .005 .005 .005 .005

Likelihood Statistic (at mean) 17.4 10.2 11.0 17.4

Level of Significance (at mean) .10 .10 .10 .10

Percentage Correctly Predicted 90% 92% 92% 93%

Percentage of Migrants Correctly
Predicted 5% 0% 0% 7% co






79

occurred, although this was usually restricted to the insignificant

variables. It should be emphasized that all four-year period equations

are misspecified because some variables could not be included in the

estimation. This makes comparisons of the two periods less meaningful.

There is evidence, however, of instability of the decision-to-move

coefficients between the two time periods. This has negative impli-

cations for forecasting based upon previous coefficients. But the

reasonableness of the forecasts which have been obtained provides support

for this methodology. Coefficient changes which occur may be balanced

in such a way as to affect the average probability, the key variable in

the forecast, only slightly.

The final step in the analysis is to apply the 1972-76 coefficients

to the characteristics of people living in the sample states in 1976.

Thus, a 1976-80 forecast is obtained. These predictions are presented

in Table 3.13. The results show that 24 of the Panel Study members who

lived in New York in 1976 will have left by 1980. For both South Carolina

and Pennsylvania 19 migrants are forecast. Ohio is expected to lose 18

migrants during the period.

An aggregate forecast is also presented in Table 3.13. It is

obtained in the following manner. First, the number of persons in the

labor force for each state in 1976 is taken from the Employment and

Training Report of the President. Then, in order to be comparable to

the Panel Study sample, working spouses are removed from the measure.

This is accomplished by multiplying the labor force number by an esti-

mate of the proportion of working spouses in the labor force. The

estimates used are the proportions that occur in the Panel Study sample

states in 1976. The resulting measure of labor force household heads is





80

Table 3.13 Forecast: Decision to Move, 1976-1980


State of Number of Migrants- Number of Migrants-
Origin Panel Study Aggregate Labor Force

Ohio 18 833,000

New York 24 1,867,000

South Carolina 19 309,000

Pennsylvania 19 1,003,000

Total 80 4,012,000






81

then multiplied by average family size to obtain an estimate of the

actual number of potential migrants from each state. The average family

size measure is also estimated from the Panel Study sample.5 Finally,

the potential population at risk is multiplied by the average probability

of migraiton estimated for the Panel Study. The numbers in Table 3.13

result.

Using the procedure put forth in Chapter 2, estimates of aggregate

labor force migration were developed from sample characteristics for the

periods 1968-72 and 1972-76. Average predicted probabilities were

calculated for each period based upon the regression results. The

population at risk for each state was derived using the method discussed

in the last paragraph. The estimates for these two periods and the

forecast are presented together in Table 3.14. It can be seen that for

the 1968-72 period, New York is estimated to have had the largest amount

of out-migration. Pennsylvania is next, followed by Ohio and South

Carolina. For the 1972-76 period, total migration out of the four

states increased. Only New York did not share in this increase. The

largest jump occurred in Ohio which is estimated to have 457,000 more

out-migrants during this period than the number who left in the 1968-72

period. This increase was large enough to put that state in second

place to New York. Pennsylvania and South Carolina showed small in-

creases from the previous periods. For the 1976-80 period, total out-

migration is expected to decline from the previous period. This decline

will be felt in all states except New York. The greatest decline-




5Since this study has found family size to be insignificant in the
determination of the decision to move, the family-size measure is obtained
from the total (migrant and nonmigrant) sample.





82

Table 3.14 Aggregate Labor Force Migration: 1968-1972,
1972-1976, 1976-1980


State of Number of Migrants Number of Migrants Number of Migrants
Origin 1968-1972 1972-1976 1976-1980

Ohio 714,000 1,171,000 833,000

New York 1,708,000 1,643,000 1,867,000

South Carolina 310,000 341,000 309,000

Pennsylvania 1,121,000 1,131,000 1,003,000

Total 3,853,000 4,286,000 4,012,000





83

occurs in Ohio, the same state that was estimated to have an unusually

high level of out-migration in the 1972-76 period. Overall, about 4

million persons are forecast to leave the four states between 1976 and

1980, indicating that there will be a continuing large pool of migrants

from which Florida and other states may draw.

The reasonability of the method used to estimate and forecast

aggregate labor force migration can partially be determined by comparing

the results with measures of out-migration derived from the 1970 Census

of Population. These estimates are not directly comparable, however.

For one, the Census estimates cover the five-year period between 1965

and 1970, while the estimates in this study cover four-year periods.

Since there is some overlap the 1968-72 period is chosen for comparison.

An additional problem in comparing estimates from the two sources is

that the Census figures are not tabulated according to labor force

status in published reports.

Table 3.15 compares the 1965-70 Census estimates to the 1968-72

estimates of out-migration obtained in this study. It can be seen that

with the exception of Ohio, the estimates of migration derived in this

study exceed those obtained from the 1970 Census. Because the Census

estimates include all members of the population (whether they are in the

labor force or not) and because the Census data cover a longer period

the opposite result would be expected. On the other hand, it is generally

believed that there was significant undercount in the 1970 Census. In

addition, because of prosperous economic conditions the early 1970s may

have been years of greater rates of migration. It is known that in

Florida, for example, in-migration accelerated significantly during

this period. Since the origin states chosen for this study are some





84

Table 3.15 Comparison of Panel Study Derived Out-Migration
Estimates to Census Estimates


State of Origin Panel Study Estimate, 1968-72 Census Estimate, 1965-70

Ohio 714,000 787,546

New York 1,708,000 1,329,432

South Carolina 310,000 248,609

Pennsylvania 1,121,000 781,684

Total 3,853,000 3,147,271





85

of the strongest contributors of Florida migrants, these states are

likely to have experienced higher rates of out-migration. These latter

arguments provide some support for the finding of higher levels of

migration in this study. More detailed evaluation of the estimates is

impossible without having a comparable set of labor force migration

figures for the same period. The next chapter presents the results of

the destination-choice analysis and develops a forecast for the 1976-80

period.















CHAPTER 4

EMPIRICAL RESULTS: DESTINATION CHOICE

Destination Choice: 1968-1977


A summary of the destinations chosen by migrants in the four sample

states appears in Table 4.1. Overall, Florida is the leading destination,

with 20 percent of all out-migrants choosing this state. Florida is the

top location choice for New York, Pennsylvania, and Ohio migrants.

South Carolina migrants showed a preference for moving north, with New

York being the leading destination. Florida, however, followed closely

behind. With the exception of those choosing Florida, California, and

Texas, many migrants chose to move to nearby states. As an example, 14

percent of the Ohio migrants moved to Michigan and 17 percent of New

York's migrants chose to live in New Jersey. Because of technical and

cost limitations, not all destinations which were chosen could be in-

cluded in the analysis. The top eight destination states were selected.

These states accounted for 70 percent of the migrants who left the four-

state area of New York, Pennsylvania, Ohio, and South Carolina.

The variables used in the destination-choice equations are presented

in Table 4.2. In the next section the rationale for inclusion of each

variable will be discussed and the results are presented.


Logit Results: Destination Choice

The results of the first destination-choice model estimated appear

in Table 4.3. The sample consists of 88 migrants originating from




86





87

Table 4.1 Destination Choice Summary: 1968-1977

(Number of Migrants)


Destination State of Origin
State New York Pennsylvania Ohio South Carolina Total

Florida 7 7 6 5 25

California 4 3 5 1 13

New Jersey 5 5 2 1 13

New York 0 2 2 7 11

Illinois 1 1 2 4 8

Virginia 2 2 1 3 8

Michigan 0 1 5 0 6

Texas 0 0 4 0 4

All other
states 11 7 10 12 37


Total 30 28 37 33 125





88

Table 4.2 Variable Definitions: Destination Choice Equations


Dependent Variable Definition

Choice Equals 1 for chosen destination,
equals 0 for all other alternatives.

Explanatory Variables

Distance Straight line mileage between each
origin and each destination.

Tempdiff Annual average temperature at each
destination minus annual average
temperature at each origin.

Incdiff Annual per-capita personal income
at each destination minus annual
per-capita personal income at each
origin.

Hdiff First quarter housing price index at
each destination minus first quarter
housing price index at each origin.

Taxdiff Annual per-capita state and local
government tax collections at each
destination minus annual per-capita
state and local government tax
collections at each origin.

Unemdiff Annual average unemployment rate of
each destination minus annual average
unemployment rate at each origin.




Full Text
118
taneously would thus seem preferable to the two-stage methodology em
ployed here. But, using existing techniques, trying to estimate such a
model over a large sample while still considering many alternatives is
extremely expensive. In addition, since most individuals do not move
during any reasonably defined period, the observations in such a model
are heavily weighted toward the staying alternative. The current state
of the art in choice theory does not seem to possess the proper tools
for the analysis of decisions where the alternatives are so unbalanced.
Implications of the Study
The influence of age and family income on the decision to move and
the interaction of income and climate in determining destination choice
can be tied together in a consistent fashion. This study presents a
picture of an older than average person with relatively high income who
chooses to move, with his (or her) family to a state with a more tem
perate climate. Possibly this person is moving in anticipation of
retirement. In addition, this move involves a significant probability
of a reduction in income. Another equivalent interpretation is that
older people with higher incomes have a greater demand for location-
specific amenities than do other members of the labor force.
The aging of the "baby boom" generation combined with the recent
decline in the birth rate of the United States implies that an increas
ing absolute number and proportion of the population will be in older
age groups. This occurrence should lead to an increase in the number of
persons who are at risk to the process described above. The implications
for amenity-rich states such as Florida, California, and Texas are
startling. Population growth will continue in these states at perhaps'
even higher rates than have so far occurred. And this increase will
continue to be concentrated in older age groups.


109
The per-capita income data generally paint a picture of persistence
of state income inequality over time. Despite considerable levels of
out-migration from the four origin states and large numbers of in
migrants arriving at the eight destination states, the magnitude of
state income differences has been maintained over time. Thus, the
thesis that migration leads to convergence of regional per-capita income
levels is not supported here.
Despite cold weather, positive income gains available in such
states as Illinois, Michigan, New Jersey, and New York lead to continued
in-migration to these destinations. With the exception of migrants
originating from South Carolina, those arriving in Florida, Texas, and
Virginia can generally expect to experience declines in income but at
the same time see improvements in climate. Thus, support is generated
for the thesis of Graves (1979) that markets adjust to leave utility
constant over space and that migration involves a tradeoff between
income and climate-related amenities. Migration to California, on the
other hand, leads to greater income and better climate, although both
increases are usually moderate. Those leaving South Carolina can expect
to receive substantial income increases, whether moving to a warmer or
colder climate.
Earlier regression results showed that although both income and
climate work together in determining migration flows between states, it
appears that each changes in relative importance at different points in
time. Whenever a factor changes in significance, the distribution of
migrants by destination can be altered. In the final chapter, the
findings of the study will be reviewed and conclusions will be drawn.


107
Since climate and income have turned out to be such important
variables in the destination-choice decision, they deserve some special
attention. In the final section of this chapter, a closer look at these
variables over time will be provided.
Income, Climate, and Migration
Earlier it was discovered that climate became a more important
determinant of destination choice during the 1972-76 period than it was
for the 1968-72 period. It was also demonstrated that migration to such
warm weather states as Florida, California, and Texas reached its
maximum during the period 1972-76. Table 4.12 presents the actual
average temperature difference and per-capita income difference data for
the three points in time used in the previous analysis.
It can be observed that temperature differences between sending and
receiving regions are generally at their maximum in 1972. These larger
differences are not the result of lower temperatures at origin states
during that year. Rather, the temperatures in warm weather destination
regions were somewhat higher than usual. Further, some of the northern
receiving regions such as Michigan and Illinois had unusually cold
winters in 1972. Thus, migration to Sunbelt states increased relative
to northern competing regions. This increase in relative attractiveness
combined with the substantially larger number of out-migrants from
sending states, resulted in larger gains for Florida, California, and
Texas. By 1976, temperature differences declined to near normal levels.
This contributed to a forecast of somewhat reduced afcfcrativeness of
Sunbelt states relative to northern receiving regions.


13
are more likely to move than those whose heads are not searching for
work. Local employment conditions are more important in the migration
decisions of the unemployed than the employed. Unemployed persons and
others looking for work are more responsive to family income, origin
wage rates, and expected earnings increases than persons satisfied with
their jobs. Families are more likely to move if they have moved in the
recent past. Wives have a significant influence on the family's de
cision to migrate. Age and education are relatively unimportant in
explaining the migration of married couples.
DaVanzo estimates her "whether to migrate" equations using ordinary
least squares (OLS). However, as acknowledged by the author, use of OLS
when the dependent variable is dichotomous leads to inefficient (although
unbiased) estimates. In addition, the fitted equation will yield pre
dictions outside the zero-one range. This criticism is particularly
crucial to a model which is to be used for the purpose of forecasting.
DaVanzo, however, was only trying to identify the determinants of the
migration decision. She estimates one equation using a probit model, a
maximum likelihood technique, and demonstrates that there is little
difference between the derived coefficients and the coefficients that
result from using OLS.
After analyzing the decision "whether to migrate," DaVanzo esti
mates choice of destination equations. She analyzes the choice among
eight regions of the United States. The explanatory variables are the
present value of the difference between what each family could earn at
each destination and what the family could earn if it stayed at the
origin, the unemployment rate at alternative destinations, the distance


106
the viability of the results. The forecasting methodology is based on
the assumption that sample proportions represent population proportions.
When sample size is small the probability of staisfying this assumption
is low. This is particularly true when the number of choices is high,
such as in the destination-choice equations. The number of observations
for each alternative ranged from 25 (Florida) to 4 (Texas).
A final hypothesis is set forth for explaining differences between
Census destination proportions and Panel Study based estimates. The
Census includes the entire population regardless of labor force status.
People who are not labor force members may have higher rates of mi
gration to.states such as Florida or California. Retired people, for
example, have a high propensity to migrate to Sunbelt states. This
group raises the rate of migration to these states. Since the Panel
Study estimates are restricted to labor force members, lower rates of
migration to retirement states are expected.
For the 1976-80 period in-migration to Florida from the four-state
sample is projected to be 652,000. This is in line with current ex
pectations that migration to Florida will remain healthy but will not be
able to maintain recent historical peaks. Floridas major competitor,
California, is forecast to receive 504,000 new migrants between 1976 and
1980. One of its newest competitors, the state of Texas, will gain
311,000 in-migrants during the period.
The poor fits obtained in the shorter time period destination-
choice equations lead to some skepticism about outright acceptance of
the forecast and estimates put forth above. Even with this limitation,
however, these figures appear reasonable and are in line with historical
standards and current expectations of the future.


72
Application to Forecasting
The models for all four states have been re-estimated for the
period 1968 through 1972. The results are presented in Table 3.10. For
the shorter time period there are, of course, fewer migrants. This led
to some problems. First, some of the variables which were used in the
nine-year migration analysis could not be included here. A requirement
for logit estimation is that there be variation within each category of
the dependent variable. Since there are fewer migrants in the sample,
there is less likelihood that some variables will meet this requirement
in the migrant category. Those that do not must be deleted. Second,
with a smaller proportion of migrants in the sample it is more difficult
to distinguish between migrants and nonmigrants. In general, poorer
fitting equations were obtained for the shorter time period. This is
because there were fewer variables that could be included in the model
and because of the smaller proportion of migrants in the sample.
Using the methodology outlined in Chapter 2 a forecast is developed.
The coefficients in Table 3.10 are applied to values of the explanatory
variables for individuals living in the sample states in 1972. Individual
predicted probabilities are calculated and averaged over the number of
individuals in each state. The mean probability for each state is then
multiplied by the total sample size in 1972. The results of this procedure
appear in Table 3.11.
The forecast values are compared to the actual (known) number of
movers during the 1972-76 period. For Ohio, 25 migrants are predicted.
In fact, there were 20 Ohio migrants in the sample between 1972 and
1976. For New York, 13 migrants are predicted compared to 17 actual


100
variable has taken on added significance in the later period. This
would suggest that migration to Florida, a state for which climate is
the main attraction, would be greater than a forecast based on 1968-72
coefficients would suggest. This is, in fact, the case. The forecast
in Table 4.7 underestimated migration to Florida for all states except
Ohio.
Overall, the equation in Table 4.8 is significant at the .005 level
when compared to the model with coefficients set equal to zero. This is
a better fit than for the previous four-year period. But when the
unconstrained model is compared to one with the right-hand side of the
equation set equal to the mean value of the dependent variable, it is
significant at only the .10 level.
The coefficients for the 1972-76 period were applied to the values
of the explanatory variables in 1976. Predicted probabilities were
obtained and multiplied by estimates of the total number of migrants
leaving each origin state for the eight destination states between 1976
and 1980. The estimates of the total pool of migrants were derived by
taking the number of migrants forecast to leave each state (from Chapter
3) and subtracting out the proportion expected to choose states other
than the eight alternatives. As before, the proportion choosing other
alternatives is assumed constant from the previous period. The destination-
choice forecast is presented in Table 4.9.
The same procedure outlined above was used to distribute the 1976-
80 aggregate forecast for each origin state among the destinations. For
each state, the aggregate forecast obtained from the decision-to-move
analysis is first adjusted by subtracting out the number expected to
choose other alternative states. The remaining number is then mul-


103
tiplied by the 1976-80 destination-choice forecast probabilities. The
predicted distribution of migrants appears in Table 4.9.
The historical aggregate out-migration estimates were also distributed
out among the eight destinations, utilizing predicted probabilities from
the estimated models. These flows, along with the 1976-80 aggregate
forecast, appear in Table 4.10. As was done with the aggregate out
migration forecast from each state, the 1968-72 destination-choice
estimates are compared with 1965-70 Census estimated flows between
states. In Chapter 3 the observation was made that, in total, out
migration based upon Panel Study estimates was higher than Census
estimates despite the facts that the Census is not restricted to labor
force members and that it measures five-year flows. Differences were
partially attributed to Census undercounts and higher rates of migration
in the early 1970s. Remaining differences are attributed to weaknesses
in the methodology employed in the study.
Table 4.11 presents a comparison of Census estimates to Panel Study
based estimates. Since we already know that the absolute number of out-
migrants from each state (with the exception of Ohio) are estimated to
be higher in this study than they are in the 1970 Census, it will be
better to compare the relative proportions of migrants who choose
alternative destinations. Notice that the Census estimates larger
proportions of migrants from all states choosing Florida and California
as destinations than this study estimates. Part of this discrepancy may
arise because of time-period differences. It is believed, however, that
greater reliance should be placed upon the Census estimates. In this
study the small sample size of the destination-choice analysis reduces


Table 4.11 Comparison of Panel Study Derived Migration Flows to Census Estimates
Origin-Destination
Panel Study Estimate, 1968-1972
Census Estimate
i, 1965'
New York-Florida
210,000
(.16)
182,531
(.29)
New York-California
131,000
(.10)
122,766
(.19)
New York-New Jersey
355,000
(.27)
209,698
(.33)
New York-Illinois
224,000
(.17)
31,041
(.05)
New York-Virginia
131,000
(.10)
36,908
(.06)
New York-Michigan
171,000
(.13)
23,761
(.04)
New York-Texas
79,000
(.06)
30,174
(.04)
Pennsylvania-Florida
81,000
(.12)
58,528
(.15)
Pennsylvania-California
47,000
(.07)
51,612
(.13)
Pennsylvania-New Jersey
135,000
(.20)
100,011
(.28)
Pennsylvania-New York
188,000
(.28)
72,001
(.19)
Pennsylvania-Illinois
81,000
(.12)
21,308
(.06)
Pennsylvania-Virginia
47,000
(.07)
36,450
(.09)
Pennsylvania-Michigan
67,000
(.10)
18,036
(.05)
Pennsylvania-Texas
34,000
(.05)
18,105
(.05)
Ohio-Florida
63,000
(.12)
80,679
(.23)
Ohio-California
52,000
(.10)
70,366
(.20)
Ohio-New Jersey
78,000
(.15)
16,288
(.05)
Ohio-Neib York
109,000
(.21)
33,548
(.10)
Ohio-Illinois
89,000
(.17)
36,910
(.11)
Ohio-Virginia
36,000
(.07)
23,388
(.07)
Ohio-Michigan
63,000
(.12)
58,350
(.17)
Ohio-Texas
31,000
(.06)
25,278
(.07)
South Carolina-Florida
41,000
(.19)
18,951
(.24)
South Carolina-California
19,000
(.09)
11,301
(.15)
South Carolina-New Jersey
32,000
(.15)
6,311
(.08)
South Carolina-New York
43,000
(.20)
13,274
(.17)
South Carolina-Illinois
30,000
(.14)
3,481
(.04)
South Carolina-Virginia
17,000
(.08)
13,316
(.17)
South Carolina-Michigan
19,000
(.09)
3,293
(.04)
South Carolina-Texas
13,000
(.06)
7,808
(.10)
105


91
income over their present state of residence. Since the interest here
is aggregate flows, average per-capita income differences are used to
represent the opportunities available to the average person. This
variable is positive and significant at only the 90 percent confidence
level. This lends only mild support to the hypothesis that regional
income inequality stimulates migration.
The cost of living at alternative locations is also believed to
influence the location choice. At first, an attempt was made to deflate
the average-income variable by local price indices provided by the
American Chamber of Commerce Association. This, however, caused very
little change in the coefficient of the income difference variable. It
was later thought that just comparing the overall cost-of-living index
may be improper. Rather it may be particular components of the market
basket which are important in location choice. Housing costs and tax
levels were two variables which came to mind. Since moving often involves
selling a home and buying a new home at a chosen location, the differ
ence in state housing price indices between origin and destination
locations was used as a variable. This variable was constructed as an
average of all reporting cities in each state during the first quarter
of the year under consideration (in this case 1968). It was expected to
have a negative sign, meaning that people are expected to move from
areas of high housing costs to locations where housing costs are lower.
At the individual level a person would like to maximize the gain realized
through selling his (or her) present home and purchasing a new one.-
Since the price index also reflects rental costs, renters too are affected
by this variable. When plugged into the model, however, the variable
displayed the wrong sign and showed no significance.


CHAPTER 3
EMPIRICAL RESULTS: DECISION TO MOVE
Interstate Migration: 1968-1977
In Table 3.1, a summary of interstate movement among Panel Study
respondents is presented. It can be seen that the states receiving the
largest number of migrants between 1968 and 1977 were California, Florida,
and Texas. The leading out-migration states were California, Ohio,
South Carolina, New York, and Pennsylvania. Net migration, defined as
the difference between the number of in-migrants and out-migrants, is
greatest in Florida and Texas. Twenty-two states registered negative
levels of net migration, with the greatest declines occurring in Ohio
and South Carolina. The negative value recorded in California is
surprising and leads to some concern about whether the Panel Study
Sample is representative or not. The fact that this study focuses upon
gross rather than net migration is some consolation, but the small
number of migrants in the sample is a worrisome factor.
States were chosen for this analysis according to their rates of
migration to Florida. Of major interest in this study are the determinants
of migration decisions and destination choice among residents of origin
states for Florida migrants. This will aid in developing a forecast of
migration from these states to Florida and to competing destinations
such as Texas and California. A summary of Florida in-migration by
state of origin appears in Table 3.2. There it is shown that 58.2
percent of all Florida migrants in the sample came from either New York,
41


PAGE
Table 3.13 Forecast: Decision to Move,
1976-1980 80
Table 3.14 Aggregate Labor Force Migration:
1968-1972, 1972-1976, 1976-1980 82
Table 3.15 Comparison of Panel Study Derived
Out-Migration Estimates to Census
Estimates 84
Table 4.1 Destination Choice Summary:
1968-1977 87
Table 4.2 Variable Definitions: Destination
Choice Equations 88
Table 4.3 Logit Results: Destination Choice,
1968-1977, Model 1 89
Table 4.4 Logit Results: Destination Choice,
1968-1977, Model 2 93
Table 4.5 Logit Results: Destination Choice,
1968-1977, Climate Excluded 95
Table 4.6 Logit Results: Destination Choice,
1968-1972 97
Table 4.7 Comparison of Forecast to Actual
Destination Choice: 1972-1976 99
Table 4.8 Logit Results: Destination Choice,
1972-1976 101
Table 4.9 Forecast: Destination Choice,
1976-1980 102
Table 4.10 Aggregate Destination Choice,
1976-1980 104
Table 4.11 Comparison of Panel Study Derived
Migration Flows to Census Estimates 105
Table 4.12 Climate and Income: 1968, 1972,
1976 108
vii


63
Variable
Famsz
Sex
Homeown
Lres
Lempl
Emself
Avage
Emst
Marst
Race
Aveduc
Prevmig
Emstw
Famy
Constant
Table 3.6 Results of Logit Analysis: Decision to Move,
New York, 1968-1977
Coefficient
-.165
.127
.773
.374
-.561
-1.277
.074
-2.960
-1.600
-.890
.227
-.448
1.682
.084
Asymptotic t-Ratio
-1.37
.08
1.38
1.51
-2.80
-1.40
2.30
-1.99
-1.03
- .89
1.34
- .71
2.77
1.88
-2.580 -1.39
Summary Statistics:
Log of Likelihood Function = -62.8
Likelihood Statistic (at 0) = 144.8
Level of Significance (at 0) = .005
Likelihood Statistic (at mean) = 38.4
Level of Significance (at mean) = .005
Percentage Correctly Predicted = 83%
Percentage of Migrants Correctly Predicted = 21%


44
Table 3.2 Florida In-migration by State
of Origin, 1968-1977
Origin State
Percentage of
Number of Florida Migrants Florida Migrants
New York
7 16.3
Pennsylvania
7 16.3
Ohio
6 14.0
South Carolina
5 11.6
Indiana
3 7.0
Kentucky
2 4.7
Virginia
2 4.7
Missouri
2 4.7
Alabama
1 2.3
Arizona
1 2.3
Arkansas
1 2.3
California
1 2.3
Connecticut
1 2.3
Illinois
1 2.3
Louisiana
1 2.3
Maryland
1 2.3
1 2.3
North Carolina


59
Summarizing, it would be optimal to include separate dummy variables
for each category for which there is information. Program and cost limi
tations, however, prevent using this approach. Alternatives include
collapsing the variable into fewer categories or including a single
ordinally ranked categorical variable. While the latter approach is
somewhat unprecedented2, it has been chosen because it utilizes more
information which is available in the data. Both alternatives impose
restrictions which may be considered severe without prior knowledge
of data. In the next four sections, the results of the logit analysis
of the decision to move will be presented and discussed.
Logit Results: Ohio
Table 3.5 presents the regression results for the state of Ohio.
Variables which are significant with at least 95 percent confidence
are sex, marital status, and family income. The sex variable, as
hypothesized, has a positive coefficient. Thus, households with male
heads from Ohio are more likely to migrate than households with female
heads from that state. The most significant variable is marital status.
As expected, unmarried individuals move more often than those who are
married. Thus, marriage does act as a deterrent to migration in Ohio.
Family income is also positive and significant. Higher income families
in Ohio are more likely to change states of residence than lower income
families. This is also the expected result.
2To test whether wage and price controls of the Nixon administration
had any effect upon wage formation, a three-leveled ordinal dummy was used
by Eckstein and Girla (1978). The variable took on the value of .5 for
Phase I of controls and 1 for Phase II of controls. Periods without con
trols took on the value of 0. The variable was found to be insignificant.
The same variable was used in a price equation and was significant.


Table 4.12 Climate and Income: 1968, 1972, and 1976
Origin- Average Temperature Difference (C) Per-Capita Income Difference (dollars)
Destination
1968
1972
1976
1968
1972
1976
New York-Florida
9.0
11.4
10.1
-967
-1131
-992
New York-California
5.0
5.3
6.0
-134
-317
64
New York-New Jersey
-1.3
-. 6
-1.0
-160
-193
169
New York-Illinois
-1.5
-2.9
-2.3
-134
-193
332
New York-Virginia
2.2
2.0
2.4
-1065
-1061
-824
New York-Michigan
-1.7
-3.0
-1.7
-424
-502
-106
New York-Texas
5.4
6.8
6.0
-1107
-1274
-857
Pennsylvania-Florida
9.0
11.2
9.7
-253
-259
-358
Pennsylvania-California
5.0
5.1
5.6
580
555
698
Pennsylvania-New Jersey
-1.3
-.8
-1.4
554
679
803
Pennsylvania-New York
0
-.2
-.4
714
872
634
Pennsylvania-Illinois
-1.5
-3.1
-2.7
580
679
966
Pennsylvania-Virginia
2.2
1.8
2.0
-351
-189
-190
Pennsylvania-Michigan
-1.7
-3.2
-2. 1
290
370
528
Pennsylvania-Texas
5.4
6.6
5.6
-393
-402
-223
Ohio-Florida
12.1
14.3
12.8
-341
-324
-324
Ohio-California
8.1
8.2
8.7
492
490
732
Ohio-New Jersey
1.8
2.3
1.7
466
614
837
Ohio-New York
3.1
2.9
2.7
626
807
668
Ohio-Illinois
1.6
0
.4
492
614
1,000
Ohio-Virginia
5.3
4.9
5.1
-439
-254
-156
Ohio-Michigan
1.4
-.10
1.0
202
305
562
Ohio-Texas
8.5
9.7
8.7
-481
-467
-189
South Carolina-Florida
4.6
6.2
5.1
784
740
982
South Carolina-California
. 6
. 1
1.0
1,617
1,554
2,038
South Carolina-New Jersey
-5.7
-5.8
-6.0
1,591
1,678
2,143
South Carolina-New York
-4.4
-5.2
-5.0
1,751
1,879
1,974
South Carolina-Illinois
-5.9
-8.1
-7.3
1,617
1,678
2,306
South Carolina-Virginia
-2.2
-3.2
-2.6
686
810
1,150
South Carolina-Michigan
-6.1
-8.2
-6.7
1,327
1,369
1,868
South Carolina-Texas
1.0
1.6
1.0
644
597
1,117
108


76
movers. Twenty-seven South Carolina migrants are forecast. In fact,
only 15 left that state during the period. The model predicts 10
migrants from Pennsylvania, while 15 actually left the state. In total
75 migrants are forecast to have moved. This compares to 67 people who
actually moved. This is a 12 percent error. It is interesting to note
that the best fitting models in the historical period, 1968-72, do not
provide the most accurate forecasts. The best fitting equation is for
the state of South Carolina, yet the greatest forecast error occurs in
this case.
The decision equations for each state are estimated next for the
period 1972-76. The results appear in Table 3.12. It should first be
observed that these results look very different from the 1968-72 mi
gration equation appearing in Table 3.10. Many variables which are
significant for the earlier period lose their significance in the latter
period and other variables gain significance. For example, for the
state of Ohio, the only significant variable in the 1968-72 equation was
family income. In the 1972-76 period, this variable is no longer
significant and length of employment becomes significant. Significant
variables in the earlier period New York equations are length of em
ployment and employment status of spouse. In the later period, no
variables are significant at the 95 percent confidence level. Length of
employment, marital status, and family income were significant variables
in the first four-year period for the state of South Carolina. Only
family income remained significant in the 1972-76 period. For Penn
sylvania, average education was the most significant 1968-72 variable.
Later, family size, length of employment, average age, and marital
status showed significance. In addition, sign changes frequently


87
Table 4.1 Destination Choice Summary: 1968-1977
(Number of Migrants)
Destination
State
of Origin
State
New York
Pennsylvania
Ohio South
Carolina
Total
Florida
7
7
6
5
25
California
4
3
5
1
13
New Jersey
5
5
2
1
13
New York
0
2
2
7
11
Illinois
1
1
2
4
8
Virginia
2
2
1
3
8
Michigan
0
1
5
0
6
Texas
0
0
4
0
4
All other
states
11
7
10
12
37
Total
30
28
37
33
125


CHAPTER 2
THEORY AND METHODOLOGY
Theoretical Model
The principle of utility maximization can be adapted to the mi
gration decision. This adaptation, however, is not straightforward.
Differences between the standard model of consumer behavior and the
approach to be taken here arise because of the nature of location as a
good. These differences are discussed in the following paragraphs.
The location decision involves a choice among mutually exclusive
alternatives. For some given time period, the individual lives at only
one location. Thus, the quantity consumed of various locations cannot
be adjusted in the same manner as are the relative amounts of other
goods. Any change in location can be viewed as a decision to give up
the entire quantity of one place in favor of consuming all of another
residence.1 This change, however, is not quite as drastic as it first
appears. To see why, location must be defined as a bundle of goods.
Some goods can then be common to more than one bundle. For example,
most food products can be purchased at any location in the United States.
LThe case of one person having two or more residences can be raised
as a contradiction to these statements. In defense, it can be argued
that this individual can only live at one place at a time and that each
move between residences involves a new choice among mutually exclusive
alternatives. Additionally, the person with multiple residences is
atypical.
18


51
Average Age
In most migration studies age is expected to be negatively related
to migration since older people have a shorter work life over which to
realize the gains from migration. In addition, it is felt that older
people have greater ties to present location. That is, they are likely
to have stronger attachments to friends and relatives and are likely to
possess location-specific capital, such as a home. Only upon retirement
are older people expected to show higher migration rates. But since
this study is restricted to members of the labor force, the inverse
relationship is still expected here. It appears, however, that in this
sample, the average age is greater for migrants than for nonmigrants.
Perhaps this occurs because the sample is restricted to Florida-sending
states. Many of the movers to Florida and competing destinations, such
as California, may be anticipating future retirement. For example, a
person may move to Florida and accept a low-paying job knowing that (he
or she) will withdraw from the labor force after a predetermined number
of years. Individuals who have accumulated substantial savings during
their lifetime can afford to do this. In addition, many older migrants
may choose semiretirement rather than retirement, perhaps maintaining
part-time jobs after moving. If the preretirement or semiretirement
thesis is correct, we would expect labor force migration from the sample
states to consist of older migrants than migration from all other states.
In the Panel Study, it is found that the average age of all U.S. labor
force migrants is 36 years old. This compares to a mean age of 38 in
the sample of Florida-sending states. Thus, there is mild evidence in
favor of this hypothesis.


133
Graves, P.E., "A Life-Cycle Empirical Analysis of Migration and
Climate, by Race," Journal of Urban Economics, Vol. 6 (April
1979), pp. 135-47.
Greenwood, M.J., "An Analysis of the Determinants of Geographic Labor
Mobility in the United States," The Review of Economics and
Statistics, Vol. 51 (May 1969), pp. 189-94.
Greenwood, M.J., "Research on Internal Migration in the United States:
A Survey," Journal of Economic Literature, Vol. 13 (June 1975),
pp. 397-433.
Johnston, J., Econometric Methods, McGraw-Hill, Inc., 1963.
Kau, J.B. and Sirmans, C.F., "The Influence of Information Cost and
Uncertainty on Migration: A Comparison of Migrant Types,"
Journal of Regional Science, Vol. 17 (April 1977), pp. 89-96.
Laber, G., "Lagged Response in the Decision to Migrate: A Comment,"
Journal of Regional Science, Vol. 12 (August 1972), pp. 307-10.
Longino, C.F., Jr., "Going Home: Aged Return Migration in the United
States 1965-70," Based on a paper presented at the 31st meeting
of the Gerontological Society, 1979.
Lowry, I.S., Migration and Metropolitan Growth: Two Analytical Models,
Chandler Publishing Company, 1966.
Maddala, G.S. and Roberts, R.B., "Estimation of Econometric Models
Involving Self-Selection," Paper for European Meetings of the
Econometric Society in Vienna, September 1977.
McFadden, D., "Conditional Logit Analysis of Qualitative Choice
Behavior," in Frontiers in Econometrics, Zarembka, P. (ed.),
Academic Press, 1973, pp. 105-42.
McFadden, D., "The Measurement of Urban Travel Demand," Journal of
Public Economics, Vol. 3 (June 1974), pp. 303-28.
Mincer, J., "Family Migration Decisions," Journal of Political Economy,
Vol. 86 (October 1978), pp. 749-73.
Morrison, P.A., "Theoretical Issues in the Design of Population Mobility
Models," Environment and Planning, Vol. 5 (1973), pp. 125-34.
Muth, R.F., "Differential Growth Among U.S. Cities," in Papers in
Quantitative Economics, Quirk, J.P. and Zarley, A.M. (eds.),
The University Press of Kansas, 1968, pp. 311-55.
Nerlove, M. and Press, S.J., "Univariate and Multivariate Log-Linear
and Logistic Models," Rand Corporation Technical Report R-1306-
EDA/NIH, Rand Corporation, December 1973.


4
Table 1.1 Greenwood Gross Migration Regression Results
Variable
Coefficient
t-Statistic
D. .
ij
-.300
-11.21
Y. .
Ji
.160
1.27
E.
l
3.401
16.60
E.
J
-.622
-2.95
U.
i
.705
8.44
u.
1
-.057
-.66
R. .
Ji
.771
7.52
T. .
li
.903
6.49
MS. .
.521
42.06
Variable
Definitions:
D. .
11
-distance from origin to destination
Y. .
li
-income at destination divided by income at
origin
E.
l
-median education at origin
E.
1
-median education at destination
U.
i
-unemployment rate at origin
U.
1
-unemployment rate at destination
R. .
li
-percent of urban population at destination
percent of urban population at origin
divided by
T. .
li
-mean annual temperature in principal city of destination
divided by mean annual temperature in principal city of
origin
MS. .
il
-number of people residing at destination who were born at
origin


29
have considered only married couples in their samples. Restricting the
sample this way allows for consideration of variables such as spouse's
wage and helps to increase the explanatory power of the equations. This
approach is rejected here, since the ultimate concern of this research
is applying the results to a forecast of migration for all labor force
members, whether married or unmarried.
The sample over which the model is estimated includes members of
the labor force only. Persons not attached to the labor force, such as
retirees, are believed to be subject to a set of influences that is
distinct from the mix of factors affecting labor force members. Health
and unearned income are examples of variables expected to be of greater
importance in the retiree migration decision. The retiree choice of
destination is also affected by different factors. Employment- or
earnings-related variables are clearly not appropriate while cost-of-
living variables may take an added significance in a model of retiree
migration. For this reason, a separate theoretical and empirical model
should be developed for this segment of the population.
In deriving the final sample, return migrants are excluded. A
return migrant is defined here as a person who is found to be moving
back to his (or her) state of birth. Other studies (for example Kau and
Sirmans [1977]) find that including return and nonreturn migrants in the
same equations leads to a specification bias. Return migrants appear to
be moving back home for reasons such as poor health or because they miss
friends or relativesfactors which do not ordinarily enter into non
return migrant decisions. A finding of a recent study of elderly return
migration is that while nonreturn migration is positively selective,
return migration is negatively selective. Longino (1979) finds that


26
As an example, those who have moved more often in the past are more
likely to respond positively to migration opportunities.
Measures of migration opportunities could be included in the
"whether to migrate" function. DaVanzo (1977) constructs a return to
migration variable that represents the maximum return available from
moving. It is derived by estimating potential wages, via a human
capital wage model, at each possible destination and at the origin.
Then the differences between potential wages at the origin and each
destination are obtained. The maximum difference is used as the es-
2
timate of migration returns. Problems arise with this variable.
Since the migration decision is itself a function of individual character
istics, many of these being the same ones which determine wages, intro
ducing this variable into the equation is somewhat redundant. If mi
gration opportunities, such as potential income, are a function of
individual attributes, then the direct inclusion of these character
istics in the equation seems to be the best approach. Using the returns
variable also introduces the possible presence of multicollinearity in
the equation since the variable is, in effect, just a linear combination
of some of the remaining variables in the equation. For these reasons a
similar variable will not be included in this model.
A final argument can be put forth in favor of including opportunity-
related variables in the decision equation. Differences in potential
returns from migration can be derived because individuals originate at
different locations. The potential return of moving from New York to
2
More details on construction of the potential wage variable will
appear later since this variable is considered in the choice of desti
nation analysis.


20
In addition to items contained in the location bundles, it is
posited that individual characteristics and circumstances are deter
minants of migration decisions. These factors determine an individual's
responsiveness to opportunities contained in the location bundle. As an
example, older people move less often because of a shorter expected
lifetime over which they may experience migration-related gains and
because they are likely to possess location-specific capital, such as a
home. Or a person who has moved many times in the past may be more
susceptible to migration influences. An individual who has resided at
the same location for a long period of time may feel greater ties to
relatives, friends, and community. A married person may be less likely
to move than an unmarried individual, especially if the spouse works.
The location choice which maximizes utility for one member of the family
may not be the optimal choice for other family members. The end result
will depend upon the nature of family relations. The point here is only
that the existence of the marriage can create a conflict which affects
the migration decision. From the standpoint of the family as a unit the
optimal choice may still result.
Other factors entering into the migration decision will be dis
cussed when the final model is specified. The above discussion has
served mainly to illustrate differences between the study of location
choice and the analysis of other choices in which economists are typically
interested. With these considerations in mind, a model of individual
location choice will be derived.
Assume that the individual chooses to reside at that location from
which he (or she) derives maximum utility. Further, assume that the
utility function is linear in parameters with additive disturbances.
Define Z^ = (X^_, P^), where X^_ is a vector of independent variables


37
Figure 2.2 Illustration of Forecasting Methodology: Destination Choice


47
Table 3.3 (continued)
Explanatory Variables
Definition
Ernst
Employment status, equals 1 if employed;
equals 0 if unemployed.
Marst
Marital status, equals 1 if married;
equals 0 otherwise.
Race
Equals 1 if white; equals 0 otherwise.
Aveduc
Average education level of husband and wife;
if unmarried Aveduc = education of household
head, takes on following values:
1 if education > 0 years and < 6 years
2 if education > 6 years and < 9 years
3 if education > 9 years and < 12 years
4 if education = 12 years
5 if education = 12 years and person has
some non-academic training
6 if person attended college, but did not
receive a degree
7 if person received a bachelor's degree
8 if person attended graduate school
Prevmig
Previous migration, equals 1 if current
state of residence differs from state of
birth; equal 0 if current state of
residence is the same as state of birth.
Ernst w
Equals 1 if spouse works; equals 0
otherwise; if no spouse then Emstw = 0.
Famy
Total family income, in thousands of dollars.


135
U.S. Department of Labor, Employment and Training Administration and
U.S. Department of Health, Education and Welfare, Employment and
Training Report of the President, U.S. Government Printing Office,
1978.
U.S. Department of Labor, Manpower Administration, Manpower Report of
the President, U.S. Government Printing Office, 1971.
Vanderkamp, J., "Return Migration: Its Significance and Behavior,"
Western Economic Journal, Vol. 10 (December 1972), pp. 460-65.



PAGE 1

AN ECONOMETRIC MODEL OF INTERSTATE LABOR FORCE MIGRATION BY SHELDON DONALD ENGLER A DISSERTATION PRESENTED TO THE GRADUATE COUNCIL OF THE UNIVERSITY OF FLORIDA IN PARTIAL FULFILLMENT OF THE REQUIREMENTS FOR THE DEGREE OF DOCTOR OF PHILOSOPHY UNIVERSITY OF FLORIDA 1979 .4.

PAGE 2

TO MY PARENTS

PAGE 3

ACKNOWLEDGEMENTS The author wishes to thank Professor Jerome Milliman for his guidance and support throughout the course of study. Special thanks are given to Professor Henry Flshkind for his intellectual contributions and his continued friendship. Acknowledgements are also extended to Professors Stanley Smith, David Denslow, John Henretta, Angela O'Rand, and William Tyler. Finally, the author would like to thank Alene Williams and Doreen Willmeroth, whose typing and editing skills were invaluable to the production of the finished paper. Xll

PAGE 4

TABLE OF CONTENTS PAGE ACKITOWLEDGEMENTS iii LIST OF TABLES vi LIST OF FIGURES Ix ABSTRACT x CHAPTER 1 INTRODUCTION 1 Problem Statement 1 Literature Review 2 Overview of the Study 16 2 THEORY AND METHODOLOGY 18 Theoretical Model 18 Methodology 22 A Forecasting Methodology 34 Sources of Data 38 3 EMPIRICAL RESULTS: DECISION TO MOVE 41 Interstate Migration: 1958-1977 41 Comparison of Migrants and Nonmigrants 45 Empirical Results 55 Application to Forecasting 72 4 EMPIRICAL RESULTS : DESTINATION CHOICE. 86 Destination Choice: 1968-1977 86 Logit Results: Destination Choice 86 Application to Forecasting 95 Income, Climate, and Migration 107 5 SUMMARY AND CONCLUSIONS 110 Introduction 110 Review of Findings: Decision to Move Ill Review of Findings: Destination Choice 115 Strengths and Weaknesses of the Study 116 Implications of the Study 118 IV

PAGE 5

PAGE APPENDIX A ALTERNATIVE MODEL FORJMULATIONS : DECISION TO MOVE 120 Combined State Results 120 Annual Migration Results 122 Redefinition of Qualitative Explanatory Variables 122 B ALTERNATIVE MODEL FORMULATIONS : DESTINATION CHOICE 128 Nominal Wage Model 128 Real Wage Model 128 BIBLIOGRAPHY 132 BIOGRAPHICAL SKETCH 136

PAGE 6

LIST OF TABLES PAGE Table 1.1 Greenwood Gross Migration Regression Results 4 Table 1.2 Lowry Net Migration Regression Results 6 Table 1 3 Graves Net Migration Regression Results 10 Table 1.4 DaVanzo Destination Choice Results 15 Table 3.1 Summary of Panel Study Migration, 1968-1977 42 Table 3.2 Florida In-migration by State of Origin, 1968-1977 44 Table 3.3 Variable Definitions for Decision to Migrate Analysis 46 Table 3.4 Comparison of Migrant and Nonmigrant Characteristics: New York, Ohio, Pennsylvania, and South Carolina 48 Table 3.5 Results of Logit Analysis: Decision to Move, Ohio, 1968-1977 60 Table 3.6 Results of Logit Analysis: Decision to Move, New York, 1968-1977 63 Table 3.7 Results of Logit Analysis: Decision to Move, South Carolina, 1968-1977 65 Table 3.8 Results of Logit Analysis: Decision to Move, Pennsylvania, 1968-1977 67 Table 3.9 Ordinary Least Squares Results: Decision to Move, 1968-1977 70 Table 3.10 Results of Logit Analysis: Decision to Move, 1968-1972 73 Table 3.11 Comparison of Forecast to Actual Migration, 1972-1976 75 Table 3.12 Results of Logit Analysis: Decision to Move, 1972-1976 77 VI

PAGE 7

PAGE Table 3.13 Forecast: Decision to Move, 1976-1980 80 Table 3.14 Aggregate Labor Force Migration: 1968-1972, 1972-1976, 1976-1980 82 Table 3.15 Comparison of Panel Study Derived Out -Migration Estimates to Census Estimates 84 Table 4.1 Destination Choice Summary: 1968-1977 87 Table 4.2 Variable Definitions: Destination Choice Equations 88 Table 4.3 Logit Results: Destination Choice, 1968-1977, Model 1 89 Table 4.4 Logit Results: Destination Choice, 1968-1977, Model 2 93 Table 4.5 Logit Results: Destination Choice, 1968-1977, Climate Excluded 95 Table 4.6 Logit Results: Destination Choice, 1968-1972 97 Table 4.7 Comparison of Forecast to Actual Destination Choice: 1972-1976 99 Table 4.8 Logit Results: Destination Choice, 1972-1976 101 Table 4.9 Forecast: Destination Choice, 1976-1980 102 Table 4.10 Aggregate Destination Choice, 1976-1980 104 Table 4.11 Comparison of Panel Study Derived Migration Flows to Census Estimates 105 Table 4.12 Climate and Income: 1968, 1972, 1976 108 vxi

PAGE 8

PAGE Table 5.1 Summary of Findings: Decision to Move 112 Table A. 1 Combined State Results: Decision to Move, 1968-1977, Ordinary Least Squares 121 Table A. 2 Annual Migration Results: Decision to Move, 1976-1977, Ordinary Least Squares 123 Table A. 3 Results of Logit Analysis with Alternative Dummies, 1968-1977 124 Table B.l Nominal Wage Model: Destination Choice, 1968-1977 129 Table B.2 Real Wage Model: Destination Choice, 1968-1977 131 vixi

PAGE 9

LIST OF FIGURES PAGE Figure 2,1 Illustration of Forecasting Methodology: Decision to Move 36 Figure 2.2 Illustration of Forecasting Methodology: Destination Choice 37 XX

PAGE 10

Abstract of Dissertation Presented to the Graduate Council of the University of Florida in Partial Fulfillment of the Requirements for the Degree of Doctor of Philosophy AN ECONOMETRIC MODEL OF INTERSTATE LABOR FORCE MIGRATION By Sheldon Donald Engler December 1979 Chairman: Jerome W. Milliman Major Department: Economics The purpose of this study is to analyze and forecast migration decisions of labor force members. Location choice is viewed within a theoretical framework which assumes that the individual is a utility maximizer. Location bundles which include some goods available at all locations and other goods which are location-specific can be varied by moving. Each person reaches his (or her) optimum by first considering the opportunities which are available at all potential locations. Differences in individual characteristics and circumstances are hypothesized to cause people to evaluate these opportunities in alternative ways. A two-step methodology is adopted for the empirical section of this study. First, the "whether to move" decision is modeled for all sample members. Then, the choice of "where to move" is analyzed for the proportion of the sample who migrated. The first decision is considered for four separate states. States were chosen based upon historical rates of out-migration, particularly to Florida. The

PAGE 11

second model focuses upon the destination choice of those who actually moved from these states. Eight alternative states are considered as destinations. Although differences in migration determinants between states are discovered, there are some general findings. The typical migrants in the sample tend to be older and earn higher income than do nonmigrants. Family relationships also appear to be important influences on the migration decisions. Migrants are also less likely to be married than J nonmigrants. Contrary to expectations, families with two income earners appear to move just as often as families where only one person is employed. Destination choice is shown to be determined by income opportunity and climate. Evidence that improved climate can sometimes only be obtained at the cost of reduced income is discovered. Overall, the migrant appears to be older, richer, and more willing to take a cut in income for better climate than the nonmigrant. One possible interpretation of the results is that many interstate moves are being made in anticipation of retirement. Those who have earned greater lifetime income are most able to absorb the decline in earnings consequent upon moving prior to retirement. Economic theory postulates that higher income people will demand greater quantities of leisure activities. This increase can be obtained through the migration process. Locations with warm weather have historically offered greater leisure possibilities than places with colder, more variant climate. The results are utilized to obtain interstate migration forecasts. Rising family incomes, the process of population aging, and an assumption of the continued importance of climate-related amenities in the migration XI

PAGE 12

decision lead to a forecast of further migration concentrated toward Sunbelt destinations. Policies aimed at restricting this growth or attracting migrants to other locations could alter this pattern. Xll

PAGE 13

CHAPTER 1 INTRODUCTION Problem Statement During recent years experts on migration have begun to focus their efforts upon understanding the decision to move at the Individual level. A continuous stream of detailed survey data has helped to bring about this endeavor. In addition, Increased availability of sophisticated computer software packages and advances In econometric techniques have facilitated the handling and analysis of this data. Until now the research which has come about with the aid of these tools has been limited to answering age-old questions In the migration literature. What are the characteristics which distinguish migrants from nonmlgrants? Do individuals move in the direction of higher wages? How does the unemployment rate affect the migration decision? What role does climate play in the choice of location? Does migration lead to regional income convergence? Since many of these issues are as yet unsettled, they will be considered here. The ultimate test of a theory, however, is its ability to predict the future. With the proportion of population change resulting from natural increase on the decline, migration forecasts become more of a necessity. This is particularly true for states which are experiencing rapid population growth such as FloTlda, California, Texas, and Arizona. In addition, forecasts of other economic events at the

PAGE 14

regional level depend heavily upon the ability to predict population. The demand for housing, the unemployment rate, and state and local government tax collections are examples of economic variables which are sensitive to the rate of population change. Knowledge of the level of future population and its consequences for the rest of the economy is crucial to success in planning for regional economic growth. In this study, an econometric model is developed for use in explaining and forecasting the movement of workers and their families between states. A general theoretical model of interstate migration is derived within a microeconomic framework. This model is then placed into an estimable form and tested for the period 1968-77. Individual data obtained from the University of Michigan Panel Study of Income Dynamics are used. Maximum likelihood estimation techniques are employed. Finally, the results of the analysis are utilized in developing an aggregate forecast of interstate labor force migration. Throughout the study the case of Florida is emphasized. Separate models are estimated for four typical origin states of Florida migrants. The determinants of Florida's attractiveness relative to other destinations which draw migrants from these states are analyzed. In the next section the existing migration literature will be brought up to date. Then an overview of this study will be presented. Literature Review Introduction Existing studies will be reviewed according to two major categories. The first section which follows will summarize research that is centered upon explaining aggregate flows of population. The second

PAGE 15

section will focus upon literature that seeks to explain migration decisions at the level of the individual. Aggregate Migration Models Aggregate models of migration are subclassif ied into two types suggested by Greenwood (1975). The first type is concerned with gross migration, defined as a single flow or the sum of unidirectional flows of population. The second type is concerned with net migration, defined as the difference between gross flows occurring in opposite directions. For any region total net migration is simply the difference between the number of in-migrants and the number of out-migrants over a specified time period. A typical gross-migration model is provided by Greenwood (1969) in his analysis of migration determinants. Greenwood examines a cross section of flows between all 48 mainland states during the 1955-60 period. His regression results are presented in Table 1.1. Positive and significant variables in Greenwood's final equation are the average level of education at origin locations, the unemployment rate at the origin, the percentage of urban population at the destination relative to the percentage of urban population at the origin, the mean annual temperature at the destination relative to the mean annual temperature at the origin, and the number of persons born at the origin who already reside at the destination — called the migrant stock. Also significant, \ but showing a negative sign, are the distance between origin and desti\A I nation and the average level of education at the destination. The best explanatory variable in Greenwood's model is the migrant stock variable. This variable is believed to be a proxy for the amount of information

PAGE 16

t-Statistic -11 .21 1 .27 16 .60 -2 .95 8 44 66 7 52 6 49 42. 06 Table 1.1 Greenwood Gross Migration Regression Results Variable Coefficient D^ -.300 Y., .160 E^ 3.401 E. -.622 U^ .705 U -.057 R. .771 T .903 MS .521 Variable Definitions: D -distance from origin to destination Y -income at destination divided by income at oricrin E. -median education at origin E, -median education at destination U. -unemplojrment rate at origin U. -unemployment rate at destination R. ^ -percent of urban population at destination divided by percent of urban population at origin T. ^ -mean annual temperature in principal city of destination divided by mean annual temperature in principal city of origin ^^ii ~ri"™bsr of people residing at destination who were born at origin

PAGE 17

concerning the destination that is available at the origin. That is, previous migrants channel information back to potential future migrants. And, assuming that it is of an encouraging nature, more information should lead to increased migration. The interpretation of the migrant stock is subject to some controversy. Laber (1972) argued that since the migrant stock is itself a function of those factors influencing previous migration, it may be acting as a proxy for lagged explanatory variables. If this is the case, then a more appropriate specification may be a partial adjustment model where lagged migration is included as an explanatory variable. Dunlevy and Gemery (1977) estimate alternative specifications of the model and conclude that it would be correct to include both migrant stock and lagged migration variables. They argue that including one variable while excluding the other results in the included variable capturing parts of both effects. Thus, it is concluded that both the information-creating effects of the migrant stock and a lagged migration adjustment process are operating simultaneously. A typical model of net migration is developed by Lowry (1966). Lowry examines a cross section of Standard Metropolitan Statistical Areas over the period 1950-60. His results are summarized in Table 1.2. Significant variables in this model are the growth of the resident working-age population (inversely related to net migration) employment growth (positively related to net migration) and growth in armed services personnel (positively related to net migration) Employment growth is the best variable, the idea simply being that higher growth reflects greater demand for labor, which translates into greater employment opportunities for potential in-migrants (making them more

PAGE 18

Table 1.2 Lowry Net Migration Regression Results Variable Coefficient t-Statistic dP^ • -.62515 -5.24 dQ^ 1.49021 52.56 dA^ 1.44409 7.36 Variable Definitions: dP. -net change in the number of residents 15-64 years of age in the absence of migration dQ. -net change in civilian nonagricultural employment dA. -net change in number of armed services personnel

PAGE 19

likely to migrate) and for residents (making them more likely to remain). Growth in the resident working-age population is important because higher growth implies an Increased supply of labor occurring from within the region, thus reducing the number of job opportunities available to potential in-migrants. Growth in armed services personnel is included since it is a component of population change that is not normally explained by the other variables in the equation. Neoclassical regional growth models theorize that migration is induced solely by the existence of interregional wage differentials. That is, migration occurs from low-wage regions (where labor is plentiful relative to capital) to high-wage regions (where labor is relatively scarce) Increased supply of labor in high-wage regions is hypothesized to put downward pressure on wages, while decreased supply of labor in low-wage regions exerts upward pressure on wages. Thus, the model leads to an equilibrium where wage differentials between regions are eliminated. Smith (1975) includes a labor sector in his neoclassical growth model which posits net migration to be a function of the difference between the local wage and the national wage. He estimates his model for various historical periods and generates mild support for the convergence theory. This is in contrast to other tests of neoclassical growth theory, such as that undertaken by Borts and Stein (1964), in which the theory is rejected. But Smith also discovered a decrease in responsiveness of potential migrants to income differentials over time. This he partly attributes to the existence of unemployment which discourages migration. Neoclassical growth models, whether regional or national in scope, assume full employment of labor.

PAGE 20

8 Richardson (1973) criticizes neoclassical models for assuming that individuals are simply maximizing their income in the migration decision. He proposes a model (which is as yet untested) in which net migration is a function of wage differentials, agglomeration economies, and locational preferences. Agglomeration economies are divided into two types. Household agglomeration economies refer to the benefits and costs of life in large cities for households. They include benefits accruing from larger labor markets, availability of leisure and cultural facilities, the quality of public service, and environmental amenities. Business agglomeration economies refer to advantages which urban areas offer businesses. They attract firms which lead to jobs which in turn attract migrants. The level of agglomeration economies is thus hypothesized to influence the rate of in-migration to a region. Locational preferences are meant to explain why individuals may remain in lowincome regions despite the existence of improved income opportunities elsewhere. Thus, they measure the retentive power of a region and are related to the rate of out -migration. Examples of factors which influence locational preferences are community ties, sociocultural traditions, and length of settlement. In a recent study Graves (1979) rejects the theory that migration is a response to interregional income differences. Rather, he advances the view that any income differences which may exist between regions are compensated for by differences in the amounts of other amenity-oriented goods which are available. For example, an area which offers an income advantage over other places is believed to offer a more objectionable climate, allowing for less involvement in leisure or recreational activities. Moving to a location with a better climate, under this

PAGE 21

viewpoint, involves a sacrifice in income. Contrary to regional growth models, Graves views the regional system as fundamentally being in equilibrium so that utility is constant over space. Migration then takes place as a result of changes in demand for location-fixed amenities. These changes come about as a result of changed relative prices and income. Under this system migration will not cause regional income convergence since income differentials must be maintained in order to compensate for differences in climate. Graves tests his theory by analyzing net migration for a cross section of Standard Metropolitan Statistical Areas during the 1960-70 period. A sample of his results for white migrants is presented in Table 1.3. All climate variables are significant. Median income is not significant, but it is more significant than it is in a model with climate variables excluded. In addition, when Graves disaggregated the sample by age, he found income to be a very significant determinant of migration for some age groups. The unemployment rate, as expected, is negative and significant. As with income, when climate is excluded from the equation, the unemployment rate loses significance. The results demonstrate than when employment and income possibilities are the same across all locations, people choose to move to more temperate climates. Alternatively, if climate is held constant across alternatives, people will choose to move to places offering greater income and employment opportunities. When climate is excluded, and thus allowed to vary, income and employment possibilities by themselves are no longer such important migration determinants. Thus, there is evidence that opportunities and climate interact in the way suggested by Graves.

PAGE 22

10 Table 1.3 Graves Net Migration Regression Results Variable Coefficient Me dine .00162 Unemp -2.906 Warmth* .0103 Cold .00686 Antmvr -.9989 Annwnd -2.967 Annhum -.7164 Variable Definitions: Med Lnc -1960 Median Income Unemp -1960 Unemp 1 Dyment Rate t-Statistic 1 07 -4 26 4 44 4 27 -4 86 -4 49 -4 74 Warmth* -Mean annual number of cooling degree days (base = 65 F) Cold -Mean annual number of heating degree days (base = 65 F) Antmvr -Annual temperature variance (average daily maximum July temperature average daily minimum January temperature) Annwnd -January and July average wind speed Annhum -January and July average humidity *A11 climate variables defined as 1931 to 1960 averages

PAGE 23

11 One of the advantages of models of net migration is that data on the dependent variable are easily computed for most areas. Net migration can be directly estimated by subtracting the population change due to natural increase from the actual population change for any given period. Measures of gross migration, in contrast, depend upon more direct measurement techniques, such as surveys. However, as Lowry (1966) notes, models of net migration reveal less about migratory behavior and the decision to migrate than do models of gross migration. This is partly because variables which are important in determining unidirectional flows are reduced in significance when the measure of migration used includes flows occurring in opposite directions. Models of net migration, however, are useful in forecasting population change. They are of particular value to regions for which migration has been the dominant component of population change. The next section will describe research which has focused upon explaining migration at the individual or family level. Individual Migration Models Rothenberg (1977) suggests an approach to the study of migration that focuses upon the individual. He notes that the migrant is selfselected. That is, given the availability of similar sets of opportunities, some individuals will migrate and some will not. The problem is to determine the individual characteristics and circumstances that cause people to evaluate their migration choice in different ways. Put somewhat differently, one individual's maximizing decision may cause him (or her) to migrate, while another individual fax^ing similar opportunities may choose not to move. Both persons may be acting rationally.

PAGE 24

12 In order to test the validity of this hypothesis, a model must be estimated using individual data on non-migrants as well as migrants. Morrison (1973) suggests the use of a theoretical framework which combines two models of migration. First, he suggests a microeconomic model which analyzes the decision "whether to move." This model, he claims, should reflect the idea that individuals have variable decision thresholds. That is, those with lower thresholds are more likely to seek out and respond to opportunities elsewhere, and hence are more likely to move. The level of threshold variability is determined by such factors as the individual's position in the life cycle, occupationally induced contraints on movements, and prior migration experience. The second type of model suggested by Morrison is one which allocates those who do move (as a result of the first decision) among alternative destinations. Thus, this model addresses the question "where to move." This model, he claims, is more macroeconomic in content. An empirical study by DaVanzo (1977) divides the migration choice into the two-stage process suggested above. She looked at a sample of married couples from the University of Michigan Study of Income Dynamics. First, she estimates "whether to migrate" equations. The dependent variable in these equations is a zero-one dummy indicating whether or not a move was made for each couple. A one-year time period (1971-72) was selected. The explanatory variables can be categorized as employment status, returns to migration, composition of family earnings,, location-specific assets (such as homeownership) previous migration, age, and education. The basic conclusions can be summarized as follows: Families whose heads are unemployed or are dissatisfied with their jobs

PAGE 25

13 are more likely to move than those whose heads are not searching for work. Local employment conditions are more important in the migration decisions of the unemployed than the employed. Unemployed persons and others looking for work are more responsive to family income, origin wage rates, and expected earnings increases than persons satisfied with their jobs. Families are more likely to move if they have moved in the recent past. Wives have a significant influence on the family's decision to migrate. Age and education are relatively unimportant in explaining the migration of married couples. DaVanzo estimates her "whether to migrate" equations using ordinary least squares (OLS) However, as acknowledged by the author, use of OLS when the dependent variable is dichotomous leads to inefficient (although unbiased) estimates. In addition, the fitted equation will yield predictions outside the zero-one range. This criticism is particularly crucial to a model which is to be used for the purpose of forecasting. DaVanzo, however, was only trying to identify the determinants of the migration decision. She estimates one equation using a probit model, a maximum likelihood technique, and demonstrates that there is little difference between the derived coefficients and the coefficients that result from using OLS. After analyzing the decision "whether to migrate," DaVanzo estimates choice of destination equations. She analyzes the choice among eight regions of the United States. The explanatory variables are the present value of the difference between what each family could earn at each destination and what the family could earn if it stayed at the origin, the unemployment rate at alternative destinations, the distance

PAGE 26

14 between each destination and the origin, and an interaction term between the present value of the wage differences and a dummy indicating whether a person had resided at that destination in the recent past. The results appear in Table 1.4. None of the variables turn out to be significant at better than the .10 level. Only the wage variable is even close to being significant, indicating that families are likely to move to destinations where the earnings gains are greatest. The sign of the interaction term indicates that families are more likely to move to an area where they have lived before than to one where they have never lived, especially if the family earnings they could receive there are higher than what they could earn by staying where they are. The unemployment rate is insignificant and shows an unexpected positive sign. The coefficient of distance shows that migrants are more likely to choose closer destinations, although it too is insignificant. The model is estimated using conditional logit, a maximum likelihood technique designed to analyze multinomial choices among discrete alternatives. DaVanzo considered the migration decision from the viewpoint of the married couple. Mincer (1978) notes that there are so far very few migration studies that consider the effect of family relations on this decision. He points out that at the individual level a person should choose to move to that location at which the return is at a maximum. For a family the optimal move is one which maximizes the combined return to the family. Whether this last criterion is satisfied or not, frequently a family will end up at a location which does not reward every family member with the maximum possible return. The members of the family who do not reach their private optimum are called tied movers. Mincer ., points out that the conflict which could result from tied migration can

PAGE 27

15 Table 1.4 DaVanzo Destination Choice Results Variable f am PV. fam PV. Here Before Unemployment Rate Ln Distance 13 Coefficient .00548 .0139 .00909 -.322 Asymptotic t-Ratio 1.53 1.46 .05 -.81 Variable Definitions: fam PV, -present value of the difference between xjhat the ii family could earn at destination j and what it could earn if it stayed in its 1971 location, i Here Before -dummy that indicates whether the family resided in area j recently (between 1968 and 1970) Unemployment Rate -unemployment rate in 1971 at destination j Ln Distance -natural logarithm of the distance between origin i and destination j

PAGE 28

16 only act as a deterrent to a possible move. This is particularly the case where more than one family member is working. Thus, the increasing proportion of women in the labor force Is expected to have an inhibiting effect upon migration. Mincer uses a scattering of data to test his theories. He discovers, as did DaVanzo, that marriage itself reduces migration and that migration rates are lower in families with employed wives. The magnitude of the effect which the working wife has on the migration decision also depends upon her share of family earnings. Finally, he shows that tied migration of working wives frequently results in lower earnings, unemployment, or labor force withdrawal. Thus, male household heads usually dominate the choice of destination. Overview of the Study In Chapter 2 the migration decision is analyzed within the context of utility theory. A theoretical model of interstate migration is then derived. A methodology for estimating this model is put forth which considers the decision to move and the choice of destination together in one analytical framevjork. Problems that arise in implementing this procedure are then discussed and an alternative is proposed. The new method views the migration decision as a two-stage process. Finally, a technique for forecasting migration decisions and destination choice is proposed. Chapters 3 and 4 present the empirical results. The decision to move is analyzed in Chapter 3. Regression results are presented for four states and forecasts of aggregate migration from these states are developed. Chapter 4 deals with the choice of destination. The determinants of this decision are analyzed and a forecast of migration

PAGE 29

17 between 31 origin and destination combinations is presented. In both chapters forecasting accuracy will be tested. The final chapter will review the findings of the study and place them into the context of the migration literature. The strengths and weaknesses of the research will be emphasized. The implications of the study for theory, policy, and future work will be discussed.

PAGE 30

CHAPTER 2 THEORY AND METHODOLOGY Theoretical Model The principle of utility maximization can be adapted to the migration decision. This adaptation, however, is not straightforward. Differences between the standard model of consumer behavior and the approach to be taken here arise because of the nature of location as a good. These differences are discussed in the following paragraphs. The location decision involves a choice among mutually exclusive alternatives. For some given time period, the individual lives at only one location. Thus, the quantity consumed of various locations cannot be adjusted in the same manner as are the relative amounts of other goods. Any change in location can be viewed as a decision to give up the entire quantity of one place in favor of consuming all of another residence.-' This change, however, is not quite as drastic as it first appears. To see why, location must be defined as a bundle of goods. Some goods can then be common to more than one bundle. For example, most food products can be purchased at any location in the United States. ^The case of one person having two or more residences can be raised as a contradiction to these statements. In defense, it can be argued that this individual can only live at one place at a time and that each move between residences involves a new choice among mutually exclusive alternatives. Additionally, the person with multiple residences is atypical.

PAGE 31

19 A change in residence will, by itself, cause little change in quantity of food consumed. Changes in food buying behavior will occur if a move is accompanied by an income increase or by a change in relative food prices. There are also goods contained in the location bundle which can sometimes only be obtained in varying quantities by moving. An example is climate. Although typically defined as a characteristic of a place rather than a good within a location bundle, climate does display certain characteristics of goods. The quantity of climate can be represented by measuring factors such as temperature or nearness to the coast. The price can be measured in terms of moving costs or foregone earnings involved in moving to a more temperate climate. So although it is not produced by man, from the consumer's standpoint climate displays characteristics that other goods possess. The need to move in order to obtain substantial changes in consumption makes climate an important determinant in the migration decision. This is true for other fixedlocation amenities such as symphonies, sporting events, and certain public services. Another important factor distinguishing locations for the consumer is the existence of differential employment opportunities across space. A person may have a potential job available in one place that will earn him (or her) greater income than jobs at other locations may offer. Higher income, in turn, allows for greater consumption. These increased consumption possibilities include leisure activities which are related to climate. Thus, although many of the goods in the location bundle are available at all locations, different quantities may be consumed at various locations because of the uneven spatial distribution of employment and income possibilities.

PAGE 32

20 In addition to items contained in the location bundles, it is posited that individual characteristics and circumstances are determinants of migration decisions. These factors determine an individual's responsiveness to opportunities contained in the location bundle. As an example, older people move less often because of a shorter expected lifetime over which they may experience migration-related gains and because they are likely to possess location-specific capital, such as a home. Or a person who has moved many times in the past may be more susceptible to migration influences. An individual who has resided at the same location for a long period of time may feel greater ties to relatives, friends, and community. A married person may be less likely to move than an unmarried individual, especially if the spouse works. The location choice which maximizes utility for one member of the family may not be the optimal choice for other family members. The end result will depend upon the nature of family relations. The point here is only that the existence of the marriage can create a conflict which affects the migration decision. From the standpoint of the family as a unit the optimal choice may still result. Other factors entering into the migration decision will be discussed when the final model is specified. The above discussion has served mainly to illustrate differences between the study of location choice and the analysis of other choices in which economists are typically interested. With these considerations in mind, a model of individual location choice will be derived. Assume that the individual chooses to reside at that location from which he (or she) derives maximum utility. Further, assume that the utility function is linear in parameters with additive disturbances. Define Z = (X P.), where X is a vector of independent variables

PAGE 33

21 describing individual t, and P is a vector of explanatory variables describing alternative i. Also define 9 = (6 6 ), where and t tx, tp' tx are vectors of parameters assigned to X and P. respectively. Finally, define e as a random disturbance for alternative i and decision maker t. The utility which individual t derives from place i can now be expressed as follows: U = Z .9 + £ (I) tx ti t ti ^^^ Alternatively, it can be written: U^. = X^e^ + P.9 + £ (2) ti t tx X tp ti ^^ If the individual maximizes utility, then he or she will choose to live at alternative i if U > U^ for all i not equal to i. tl tj J H Moving costs can be incorporated into the model. Assume first that we find individual t living at place i. Then we conclude that if he (or she) is rational: U > U C for all j not equal to i where C. is '--^ ^J IJ XJ the cost of moving from i to j If at a later time person t migrates to place j then we conclude that something occurred during the interim which caused t's utility evaluation to change so that: U C > tj ij ^ti' Throughout the remainder of the study, actual moving expenses will be ignored since it is believed that they are insignificant in relation to the other costs and benefits of moving. The probability that a person will choose to live at location i can be written as follows: P(i) = P(U^^ > U^^ for all j ^ i) (3) A major goal of this study will be to impute values of P(i) for individuals inside and outside the sample. Inferences about aggregate behavior will also be made using these probabilities.

PAGE 34

22 The model developed above belongs to a general classification called random utility models. (For a discussion of these models see Albright, Lerman, and Manski [1977].) Alternative methods of estimating this model have been attempted in this study. The technique which at first appeared to be ideal will be put forth in the next section. Reasons for the abandonment of this approach will then be given. These reasons are both theoretical and practical. The method finally settled upon will be discussed in detail. Its strengths and weaknesses in studying location choice will be highlighted. Methodology An Ideal Methodology Equation (2), as it stands, cannot be estimated. This is because utility is unobservable. We can, however, observe actual location choices. Define I = 1 if individual t chooses alternative i. Define I = if some other location is chosen. Then the utility function can be rewritten; I = XtGtx + PiOtp + Eti (^^ Given the assumptions of the theory, if I = 1 then we know that U^-j^ > Utj for all 2 not equal to i. If I = 0, we then conclude that another location allows the individual to obtain a higher level of utility. Equation (4) is an estimable equation. It can be applied to both single-choice and multiple-choice situations. For multiple-choice situations it is assumed that the individual is making any number of concurrent binomial decisions. For example, a person residing in New York and deciding whether to move to Florida, migrate to California, or remain in New York can be said to be faced with three binomial decisions;

PAGE 35

23 A. Migrating to Florida or not B. Migrating to California or not C. Staying in New York or not Values of I can now be assigned to each decision. If this person chooses to migrate to Florida, then for decision A, I = 1; for decision B, I = 0; and for decision C, 1=0. The values of X will vary across alternatives. Individual-alternative interaction terms can also be introduced into the model. If each binomial decision by each individual is treated as a single observation, then the model can be estimated by ordinary least squares (OLS). It is well known, however, that use of OLS when the dependent variable is dichotomous leads to inefficient estimates and to predictions outside the zero-one interval. In addition, OLS is unable to distinguish between individuals and observations in the multinomial case. More appropriate methods use the maximum likelihood technique of estimation. McFadden (1973) developed a technique he calls conditional logit for use in analyzing consumer choice among lumpy alternatives. He demonstrates the applicability of conditional logit in the study of urban travel demand. In a recent study, Falaris (1978) applies the same technique to the migration decision. An attempt was made in this study to adopt the same methodology. It did not prove to be useful for this study. The conditional logit model seems best adapted to problems where the choice frequencies are fairly well balanced. Location choice (studied over reasonable periods of time) is biased heavily toward a single alternative: staying where one already is. Thus, when a model is estimated which includes staying as one of many alternatives, this choice swamps all others. Consequently, it is difficult to obtain

PAGE 36

24 variables that distinguish between the remaining alternatives. The magnitude of this problem enlarges as the number of alternatives increases. Falaris avoided this problem to some extent by restricting his research to a choice among four broad regions of the United States. Since this study will focus upon the state as the level of analysis, it will be desirable to include a larger number of alternatives. A further problem encountered with the model is that as new variables are added when there are a large number of alternatives, the cost of estimation rises substantially and the chances of early convergence diminish. Falaris, in fact, considered only a small number of variables in his study. Finally, testing various model specifications is difficult using such high-cost procedures. Proposed Methodology The above considerations have led to the decision to analyze the migration choice as a two-stage process. This is in line with the approach suggested by Morrison (1973) and taken by DaVanzo (1977). The first stage of the migration process is the decision "whether to move." All members of the population are faced with this choice. The second stage involves the decision "where to move." Only migrants face this choice. The decision equation (4) is now divided into two: I, = X 6 + e (^\ 1 t tx ti *'-'-' I„ = P. 9^ + U (6) 2 1 tp ti ^"' where U^^ is a random disturbance term and every other term is defined as before. Equation (5) represents a binomial decision, with I = 1 if the person moves and I = if they remain where tliey are. Equation (6) can be a multinomial decision, with the number of alternatives

PAGE 37

25 depending upon the number of locations available to the individual. Utilizing this framework relieves to some extent the econometric problems encountered when the entire decision was placed in one equation. The "whether to move" equation is now easier to estimate and interpret since all migrants are grouped into one category. The "where to move" equation also contains more balance among alternatives. In addition, since the sample size has been reduced substantially by considering only migrants in this equation, estimation is now less costly. Viewing the migration decision under a two-stage regime involves making the assumption that these decisions are independent of each other. This is the major weakness of this approach. It can readily be argued that the decision to move and destination choice occur jointly in many individual cases. But, given currently available techniques, the costs of modeling them together appear to outweigh the benefits derived from this approach. A detailed description of the procedures used in estimation and application of equations (5) and (6) begins in the next section. The Decision to Move As discussed earlier, if two people are choosing from a set of similar location bundles, one may move and one may remain at his (or her) current residence. Characteristics of the individual, contained in the vector X are believed to determine who is selected to migrate. These characteristics may represent differences in opportunities. For instance, the older person may have a lower return from migration in terms of lifetime income. Differences in individual attributes may also represent differences in the way people evaluate the same opportunities.

PAGE 38

26 As an example, those who have moved more often in the past are more likely to respond positively to migration opportunities. Measures of migration opportunities could be included in the "whether to migrate" function, DaVanzo (1977) constructs a return to migration variable that represents the maximum return available from moving. It is derived by estimating potential wages, via a human capital wage model, at each possible destination and at the origin. Then the differences between potential wages at the origin and each destination are obtained. The maximum difference is used as the estimate of migration returns.^ Problems arise with this variable. Since the migration decision is itself a function of individual characteristics, many of these being the same ones which determine wages, introducing this variable into the equation is somewhat redundant. If migration opportunities, such as potential income, are a function of individual attributes, then the direct inclusion of these characteristics in the equation seems to be the best approach. Using the returns variable also introduces the possible presence of multicollinearity in the equation since the variable is, in effect, just a linear combination of some of the remaining variables in the equation. For these reasons a similar variable will not be included in this model. A final argument can be put forth in favor of including opportunityrelated variables in the decision equation. Differences in potential returns from migration can be derived because individuals originate at different locations. The potential return of moving from New York to 2 More details on construction of the potential wage variable will appear later since this variable is considered in the choice of destination analysis.

PAGE 39

27 Florida may exceed the gain resulting from a California to Florida move. Therefore, more people may migrate to Florida from New York than from California, despite the fact that Calif ornians and New Yorkers may have similar characteristics. There is a way, used in this study, to allow for this occurrence without including a returns variable in the equation. The model can be estimated for a sample that is restricted to one origin at a time. In this way, everyone in the sample faces the same alternative choices even if these choices are viewed relative to their present location. Differences in migration propensities now occur as a result of differences in the way Individual characteristics are valued at alternative locations and because of the effects that individual attributes have upon the way in which people evaluate and respond to migration opportunities. Estimating the model for various places also allows for a comparison of the determinants of the migration decision over space. If the factors important in explaining the New Yorker's decision to move are significantly different from the factors affecting the Californlan's choice, then there is an additional rationale for having separate models. Such an approach will be taken in the main body of this study. Results obtained when origin states are combined are presented in Appendix A. Lack of time and resources prevents estimation for all possible origins. The model will be estimated for four states. These are New York, Pennsylvania, Ohio, and South Carolina. These states were chosen because, in the sample to be studied, they are the leading origins of migrants to Florida, the destination of primary concern to this study. In addition, all four are states with relatively large numbers of outmigrants during the period under study.

PAGE 40

28 The choice of time period is important in any migration study. Shorter time periods at first seem optimal, since the determinants of migration can be measured at or close to the time at which the actual move is made or not made. Longer time periods lead to problems, since migration may be taking place at many different points during the period. Determination of when to measure the explanatory variables is then difficult. Estimation problems, however, arise for short time periods because the proportion of the population that moves is smaller than it is for long periods. Estimations were first attempted for oneyear periods. Poor fits were obtained and few variables were significant enough to distinguish between migrants and nonmigrants. A sample of these results can be found in Appendix A. For the main part of the study, three time periods were considered. First, the model was estimated for the 1968-77 period, the maximum amount of time for which data is available in the sample. Then, for the purpose of forecasting the models were re-estimated for the periods 1968-72 and 1972-76. The explanatory variables were always measured in the first year of the period under consideration. The migration variable took on the value of one if during the last year of the period, an individual was living in a state which differed from his (or her) state of residence during the first year of the period. Otherwise it took the value of zero. The unit of analysis in the study is the head of household who is a member of the labor force. Since the survey from which the sample is derived contains some information about other household members, variables such as family size and spouse's employment status can also be entered into the equations. Other studies (e.g., DaVanzo [1977])

PAGE 41

29 have considered only married couples in their samples. Restricting the sample this way allows for consideration of variables such as spouse's wage and helps to increase the explanatory power of the equations. This approach is rejected here, since the ultimate concern of this research is applying the results to a forecast of migration for all labor force members, whether married or unmarried. The sample over which the model is estimated includes members of the labor force only. Persons not attached to the labor force, such as retirees, are believed to be subject to a set of influences that is distinct from the mix of factors affecting labor force members. Health and unearned income are examples of variables expected to be of greater importance in the retiree migration decision. The retiree choice of destination is also affected by different factors. Employmentor earnings-related variables are clearly not appropriate while cost-ofliving variables may take an added significance in a model of retiree migration. For this reason, a separate theoretical and empirical model should be developed for this segment of the population. In deriving the final sample, return migrants are excluded. A return migrant is defined here as a person who is found to be moving back to his (or her) state of birth. Other studies (for example Kau and Sirmans [1977]) find that including return and nonreturn migrants in the same equations leads to a specification bias. Return migrants appear to be moving back home for reasons such as poor health or because they miss friends or relatives — factors which do not ordinarily enter into nonreturn migrant decisions. A finding of a recent study of elderly return migration is that while nonreturn migration is positively selective, return migration is negatively selective. Longino (1979) finds that

PAGE 42

30 elderly migrants returning to their state of birth have lower socioeconomic characteristics than other movers. Thus, only nonreturn migrants are considered in this analysis. The decision equation will be estimated using a logit program developed by Nerlove and Press (1973). Using this technique, the estimate of Ii in equation (5) is equal to the log odds of one of the alternatives. The estimated equation can be written: = XtOtx (7) stay/ where Pstay is the probability of staying, i.e., not moving. We can now solve for ^stay: Pstay = ^ ^ ^ „ (8) 1 + e-AtBtx Since there are only two alternatives, the probability of moving is Pmove = 1 Pstay = -— y^ ^ (9) 1 + e-^totx This value will be derived for every individual in the sample and compared with each person's actual choice. In addition to comparing actual and fitted values, the overall fit of the model can be determined by comparing the log likelihood value at model convergence with the log likelihood value that results if all coefficients are restricted to zero. It has been shown (see McFadden [1974]) that the statistic -2 [ In (ir) In (1) ] (10) is distributed approximately chi-square, where 1^ is the likelihood value when the coefficients are restricted and 1 is the likelihood value for unconstrained model at its maximum. The degrees of freedom are equal to the number of restricted coefficients.

PAGE 43

31 In addition to comparing the likelihood value at the unconstrained maximum to the likelihood value when all coefficients are restricted to zero, a more stringent test is proposed. First the right-hand side of the decision equation is set equal to the mean value of the dependent variable. This would be the best guess of the probability that some individual in the sample would move, given that we have no other information. Then the resulting likelihood value is compared to the likelihood value when all variables are included. The statistic put forth above can then be calculated using these two likelihood values as input The specific variables considered in the decision-to-move analysis will be defined and discussed in Chapter 3. The next section puts forth the methodology employed in the destination-choice analysis. The Choice of Destination Having determined the characteristics that distinguish interstate movers from nonmovers, the next task is to examine the factors which enter into the destination choice. Previous studies have considered location choice primarily as a function of potential wages at alternative locations. DaVanzo (1977) and Falaris (1978) first estimate a wage model for each alternative location. Wages are viewed as a function of the characteristics of those living at each place. Some of the variables considered are age, race, sex, education, occupation, and experience. These variables are similar to those typically employed in human capital models of wage determination (for example Dalton and Ford [1977]). After estimation, predicted values of. .wages at each destination are obtained for everyone in the sample by plugging their

PAGE 44

32 characteristics into the fitted equations. This variable is usually found to be a significant determinant of location choice. A similar variable has been constructed for use in this study. Although the sign was positive, it did not turn out to be significant. One possible explanation derives from the way in which the sample was chosen. As mentioned earlier, the origin states considered are the leading senders of migrants to Florida. Florida is not known as a high-wage state, yet many people move here. In fact, many Florida migrants in the sample have experienced declines in nominal and real (price deflated) wages. A sample of results obtained when the wage variable is included is contained in Appendix B. Generally, earlier empirical models of individual migration choice have not found aggregate location characteristics, such as the unemployment rate, to be important variables. But these studies have considered only very large geographic areas as potential destinations. Differences in location-specific characteristics then tend to get averaged out across the region. Better results should be obtained if destinations are defined as smaller areas, such as states. A disadvantage of this approach is that all possible destinations cannot be considered since the estimation technique puts limits upon the number of alternatives that may be considered at once. Essentially, the approach suggested by Morrison (1973) is taken here. That is, while the decision to migrate is considered from a microanalytic perspective, the decision where to go is modeled in a macroanalytic framework. By considering only aggregate characteristics of each location as variables, the model will, for example, predict the probability of a New York migrant choosing Florida as a destination. It will not, however, explain why one New Yorker chooses Florida and another

PAGE 45

33 chooses California. Micro type variables, such as individual potential wages, which take into account differences among people were tested. Their failure to yield meaningful results led to the adoption of a more macro approach. Equation (6) was estimated for the period 1968-77. For a forecasting test, it was estimated for the periods 1968-72 and 1972-76. The sample was defined as it was for the decision-to-move analysis, except that only migrants were considered. Migrants from New York, Ohio, Pennsylvania, and South Carolina were then grouped together. The top eight destination states were selected as potential alternatives. These included Florida, California, New York, New Jersey, Michigan, Illinois, Texas, and Virginia. Migrants choosing other states were excluded from the analysis. The model is estimated using a multinomial "conditional" logit program. Estimating equation (6) using this technique allows us to derive predicted probability values. These probabilities can be represented as follows: p.e 1 tp Z e i=l where i Indexes alternatives and J is the total number of choices available to the individual. The variables in the vector P. will be 1 defined as differences between the value at location i and the value at the individual's origin of residence. Probabilities will be calculated for migration between each possible origin and destination combination. Since starting from any origin these values willnot vary across individuals, there will be a total of 32 probabilities calculated (8 alternatives X 4 origins) .-

PAGE 46

34 All variables will be premigration measures of location aggregates. Specific variables to be included in the choice of destination equations will be defined and discussed in Chapter 4. The next section will propose a method of obtaining a forecast from the completely estimated model. A Forecasting Methodology In an application of logit analysis to transportation decisions, McFadden (1974) states that if the sample under consideration represents a random selection of the environments faced by the population as a whole, the average of the predicted values over the sample is a best estimate of aggregate demand. With this in mind a methodology is proposed for forecasting the number of migrants from a given origin and their destination choices. The accuracy of this methodology will be tested within the sample. For any given origin state, the coefficients of the 1968-72 decisionto-mlgrate equation are applied to values of the explanatory variables for individuals living in that state in 1972. In this way predictions for the next four-year period, 1972-76, are obtained. Predicted probabilities for each individual are obtained and all Individuals are averaged. This average probability is then applied to the total sample in 1972 to get a forecast of the number of migrants between 1972 and 1976. This number can then be compared to the actual (known) number of migrants in the sample during the period. If successful, then 1972-76 coefficients can be applied to the 1976 sample to obtain a 1976-80 forecast. The next step involves application of the sample-derived average probabilities to aggregate (outside the sample) measures of the labor

PAGE 47

35 force in the state under consideration. In this way a forecast of the actual number of labor force migrants from that state is obtained. Figure 2.1 contains a flow chart which illustrates the procedure just described. A similar procedure is followed for forecasting destination choice. The model is estimated for the 1968-72 period and the coefficients are then applied to locational characteristics in 1972. Forecasts of the probabilities of migration between each origin and each destination are obtained. These probabilities are then multiplied times the number of migrants leaving each origin location. The values obtained can be compared to the actual origin-destination flows during the 1972-76 period. If successful, the model is re-estimated for the 1972-76 period and these coefficients are applied to the 1976 locational characteristics. A 1976-80 aggregate forecast can be obtained by applying the probabilities derived from the 1972-76 estimation to the aggregate decision-to-move forecast. Thus, forecasts of actual flows of labor force migrants to all destinations are derived. The destination-choice forecasting methodology is illustrated in Figure 2.2. In addition to forecasting applications, estimating the model for two equal length time periods allows for observation of coefficient stability. In this way it can be determined whether events such as the energy crisis and the deep recession of 1974 have had any impact on migration decisions and destination choice.^ Since earlier period 3 In fact this goal will be difficult to attain. Four-year migration periods contain many events which cannot easily be disentangled from one another. Estimation for shorter periods is even harder because of (migrant) sample size problems.

PAGE 48

36 1972-1976 decision to move coefficients, origin i 1976 characteristics of individual t originating in state i Predicted probability of moving out of origin i, individual t, 1976-1980 Total labor force, state i. 1976 Average of predicted probabilities of moving out of origin i, all individuals, 1976-1980 Forecast of labor force migration from state i, 1976-1980 Figure 2,1 Illustration of Forecasting Methodology: Decision to Move

PAGE 49

37 1972-1976 destination choice coefficients 1976 characteristics of destination j relative to origin i Forecast of labor force migration from state i Predicted probability of moving from origin i to destination j, 1976-1980 Forecast of labor force migration between origin i and destination j Figure 2.2 Illustration of Forecasting Methodology: Destination Choice

PAGE 50

38 migration coefficients are used to forecast later period migrant flows, coefficient instability implies forecasting problems. Employing energy shortage coefficients, for example, to project migration for a period in which we expect no energy problems would be incorrect. Finally, historical estimates of aggregate labor force migration can be obtained from the sample by utilizing a methodology similar to that proposed for forecasting. For the decision-to-move analysis, the average probability of migration for each state during any given period can be applied to a measure of the aggregate labor force at the beginning of that period. The resulting estimates of out-migration can then be distributed out among destinations by multiplying them times the probabilities obtained from the destination-choice equations for the same period. Thus, estimates of aggregate migration between all possible origins and destinations are derived from the results of the entire analysis. In the next section of this chapter, the sources of data will briefly be described. Sources of Data The primary source of data for this study is the University of Michigan's Panel Study of Income Dynamics. In this survey, 5,852 families have been interviewed each year since 1968. At present, 10 years of data are available on tape. Each year approximately 450 variables are available for each family. Categories under which the variables can be grouped are family composition information, education, transportation, housing, employment of head, housework, work for money by wife, food and clothing, income, intelligence", feelings, and time use.

PAGE 51

39 Typical sources of data for migration research are the Continuous Work History Sample of the Social Security Administration and the Public Use Sample derived from the 1970 Census of Population. The Panel Study's advantage over these sources lies in its richness of variables and in its continuity and consistency of sampling. Its disadvantage is that it is a much smaller sample. But since its patterns of movement over the sample period closely resemble the movement identified in these other sources, great confidence is placed in its use as a tool in migration studies. Aggregate data is obtained from various sources. Chief among these is the Statistical Abstract of the United States This is one of the few publications that contains easily accessible and reasonably consistent time-series data for all states. Other sources are Climatological Data National Summary Cost of Living Indicators and The Employment and Training Report of the President Sample Size Limitations Dividing up the sample according to state of origin reduces the number of observations for each estimation substantially. For the state of Ohio, 220 observations are drawn from the Panel Study tapes. Thirtyseven of these individuals (and their families) moved out of Ohio in the sample period. There are 195 New York observations, 29 of which are migrants. Pennsylvania has 202 observations and 28 of these moved out of state during the nine-year period. The sample size for South Carolina is 222, with 32 being migrants. The destination-choice equations grouped together 85 migrants from all four states who chose among eight potential destinations.

PAGE 52

40 Logit analysis depends upon the assumption that the dependent variable (the logarithm of the odds that a particular choice will be made) approximates the normal distribution. A large number of observations and sufficient repetitions for each possible choice assure that this criterion will be met. For the decision-to-move analyses it is believed that sample size is sufficient for employing logit techniques. The destination choice results should be interpreted more cautiously. A particular problem for these equations is the small number of observations occurring in each possible category of choice.'* All equations were also estimated using ordinary least squares techniques (see Chapter 3) The signs and magnitudes of the coefficients obtained were very similar to those obtained with logit analysis. These similarities increase the degree of confidence placed in the major findings of this study. In Chapter 3 the results of the decision-tomove analysis are put forth and discussed. '*An additional source of worry arises because origin states are chosen according to their rates of migration to Florida. This may introduce an upward bias in the predictions of migration to Florida. Consequently, the probabilities of migration to other states will be understated. ^DaVanzo (1977) compares probit estimates of decision to move equations with ordinary least squares estimates. She also finds similar results.

PAGE 53

CHAPTER 3 EMPIRICAL RESULTS: DECISION TO MOVE Interstate Migration: 1968-1977 In Table 3.1, a suiranary of interstate movement among Panel Study respondents is presented. It can be seen that the states receiving the largest number of migrants between 1968 and 1977 were California, Florida, and Texas. The leading out-migration states were California, Ohio, South Carolina, New York, and Pennsylvania. Net migration, defined as the difference between the number of in-migrants and out -migrants, is greatest in Florida and Texas. Twenty-two states registered negative levels of net migration, with the greatest declines occurring in Ohio and South Carolina. The negative value recorded in California is surprising and leads to some concern about whether the Panel Study Sample is representative or not. The fact that this study focuses upon gross rather than net migration is some consolation, but the small number of migrants in the sample is a worrisome factor. States were chosen for this analysis according to their rates of migration to Florida. Of major interest in this study are the determinants of migration decisions and destination choice among residents of origin states for Florida migrants. This will aid in developing a forecast of migration from these states to Florida and to competing destinations such as Texas and California. A summary of Florida in-migration by state of origin appears in Table 3.2. There it is shown that 58.2 percent of all Florida migrants in the sample came from either New York, 41

PAGE 54

42 Table 3.1 Summary of Panel Study Migration, 1968-1977 Number of InNumber of OutState Alabama Migrant s Migrants Net Migration 7 3 4 Arizona 11 6 5 Arkansas 3 11 -8 California 44 48 -4 Colorado 16 6 10 Connecticut 8 7 1 Delaware 3 3 District of Columbia 9 12 -3 Florida 43 11 32 Georgia 9 5 4 Idaho 2 2 Illinois 23 22 1 Indiana 11 14 -3 Iowa 4 17 -13 Kansas 6 6 Kentucky 5 14 -9 Louisiana 5 13 -8 Maine 7 1 6 Maryland IS 9 9 Massachusetts 14 10 4 Michigan 10 10 Minnesota 5 7 -2 Mississippi 4 16 -12 Missouri 4 19 -15 Montana 1 Nebraska 7 3 4 Nevada 7 7 New Hampshire 5 5 1 New Jersey 21 15 6 I New Mexico 4 4 j New York 25 30 -5 North Carolina 5 12 -7 North Dakota Ohio 11 37 -26 Oklahoma 5 4 1 Oregon 17 8 9 Pennsylvania 15 28 -13 Rhode Island South Carolina 5 33 -28 1 South Dakota 1 5 -4 1 Tennessee 10 5 5 Texas 39 14 2-5 Utah 1 7 -6 Vermont 1 1 Virginia 22 15 7

PAGE 55

43 Table 3.1 (continued) State Washington West Virginia Wisconsin Wyoming Number of InMigrants 6 2 6 2 Number of Out Migrant s 9 2 Net Migration -3 2 4 2

PAGE 56

44 Table 3.2 Florida In-migration by State of Origin, 1968-1977 Percentage of Origin State Number o f Florida Migrants Florida Migrants New York 7 16.3 Pennsylvania 7 16.3 Ohio 6 14.0 South Carolina 5 11.6 Indiana 3 7.0 Kentucky 2 4.7 Virginia 2 4.7 Missouri 2 4.7 Alabama 2.3 Arizona 2.3 Arkansas 2.3 California 2.3 Connecticut 2.3 Illinois 2.3 Louisiana 2.3 Maryland 2.3 North Carolina 2.3

PAGE 57

45 Pennsylvania, Ohio, or South Carolina. After California, these are also the leading out-migration states in the sample. These four states were thus selected as the major sample for this analysis. Table 3.3 defines the variables to be considered in the decisionto-m.ove analysis. In the next section, the rationale for including each variable will be discussed and a first look at the data will be provided. Comparison of Migrants and Nonmigrants Family Size Table 3.4 presents a comparison of migrant and nonmigrant characteristics in the four-state sample. A priori it is expected that persons in larger families are less likely to move since additional people represent ties to present location. Children in school or a spouse who is working are both examples of migration ties. Comparison of the mean values of family size among migrants and nonmigrants, however, reveals very little difference. Both groups average about five persons per family. Thus, at first glance, family size does not appear to be a constraining force. Sex Households with male heads are expected to move more often than households with female heads. One reason is that men are more likely to be employed in occupations where job transfers are common. In addition, higher rates of unemployment for women at many locations and lower income opportunities would deter female migration. The mean value of the sex variable, however, is only slightly higher for movers than it is for nonmovers. Eighty-seven percent of the households that moved were headed by men. It may be that since women are also likely to be earning

PAGE 58

46 Table 3.3 Variable Definitions for Decision to Migrate Analysis Dependent Variable Move Explanatory Variables Famsz Sex Homeown Lres Lempl Definition Equals 1 if the person's state of residence in 1977 differs from his or her state of residence; equals otherwise. (Defined in 1968 for different periods in later analysis. ) Family size, actual number in household. Equals 1 for male; 2 for female. Equals 1 if home is owned; equals otherwise. Length of residence in current house or apartment, takes on following values: if length of 1 if length of 2 if length of 3 if length of 4 if length of 5 if length of < 10 years 6 if length of < 15 years 7 if length of < 25 years 8 if length of residence < 1 year residence = 1 year residence = 2 years residence = 3 years residence = 4 years residence >^ 5 years and residence > 10 years and residence >_ 15 years and residence > 25 years Emself Length of time employed by present employer, takes on following values : if self-employed or unemployed 1 if employed > months and <^ 6 months 2 if employed > 6 months and <^ 18 months 3 if employed > 18 months and _< 42 months 4 if employed > 42 months and < 9 years 5 if employed > 9 years and < 19 years 6 if employed > 19 years Equals 1 if self-employed; equals otherwise. Avage Average age of husband and wife; if unmarried Avage = age of household head.

PAGE 59

47 Table 3.3 (continued) Explanatory Variables Ernst Marst Race Aveduc Prevmig Emstw Famy Definition Employment status, equals 1 if employed; equals if unemployed. Marital status, equals 1 if married; equals otherwise. Equals 1 if white; equals otherwise. Average education level of husband and wife; if unmarried Aveduc = education of household head, takes on following values: 1 if education > years and < 6 years 2 if education >^ 6 years and < 9 years 3 if education ^ 9 years and < 12 years 4 if education = 12 years 5 if education = 12 years and person has some non-academic training 6 if person attended college, but did not receive a degree 7 if person received a bachelor's degree 8 if person attended graduate school Previous migration, equals 1 if current state of residence differs from state of birth; equal if current state of residence is the same as state of birth. Equals 1 if spouse works; equals otherwise; if no spouse then Emstw = 0. Total family income, in thousands of dollars.

PAGE 60

48 Table 3.4 Comparison of Migrant and Nonmigrant Characteristics: New York, Ohio, Pennsylvania, and South Carolina Variable Mean Value F igrants Famsz 4.98 Sex 0.87 Homeown 0.64 Lres 5.93 Lempl 3.14 Emself 0.10 Avage 37.89 Ernst 0.96 Marst 0.78 Race 0.71 Aveduc 4.11 Prevmig 0.31 Emstw 0.38 Famy 11.24 Mean Value Nonmigrants 4. 95 0. 84 0. 58 5. 87 3 31 08 35 31 98 81 68 3 .78 .23 .35 8 .79

PAGE 61

49 lower incomes at their current locations and have a higher probability of being unemployed before moving, they are more responsive to opportunities elsewhere. This would counteract the negative effects of the sex variable and lead to a more ambiguous result. No conclusions can be draxsm without more rigorous analysis. Homeowner ship Owning a home is usually expected to act as a tie to present location. Time and transactions costs involved in selling a home and perhaps purchasing another one in a new location would tend to deter migration. On the other hand, homeownership represents wealth. If a wealth effect operates, property owners may be more likely to move than renters. In this sample 64 percent of the migrants owned homes prior to migration while only 58 percent of the nonmigrants were homeowners. Thus, the data give some support to the second hypothesis. Length of Residence Persons who have lived in their current houses or apartments for long periods of time are expected to display a lower probability of moving. Duration of past residence should act as a measure of willingness to pull up stakes, leave friends and relatives, and start all over somewhere else. An individual who remains in the same dwelling unit for a long time displays an aversion to risk which leads one to believe that he (or she) would not be likely to do something as filled with uncertainty as changing states of residence. The data in Table 3.4 do not demonstrate this to be the case. There appears to be very little difference in length of residence for migrants and nonmigrants. The average length of residence for both groups is close to 10 years.

PAGE 62

50 Length of Employment As with the last variable, the length of employment would indicate attachment to present environment. In addition, it can be viewed as a measure of ability to hold a job. A person who has worked for his (or her) present employer for only a short period of time is more likely to have been periodically unemployed. The mean value of this variable is indeed somewhat lower for migrants than for nonmigrants in the states under consideration. The average length of employment before moving is about 21 months for migrants. Self -Employment Individuals who are self-employed are expected to migrate less. The self-employed person is more likely to have an occupation that involves heavy investment in capital equipment. Dentists and printers are examples, Although this equipment may be transportable, the costs of moving and re-establishing at some other location are significant enough to deter migration. Self-employed workers are also likely to be part of smaller organizations where job transfer is uncommon. Finally, those who are employed in professions requiring state licenses, such as medical and legal fields, are bound to their state to some extent by their licenses. The mean values for the four sample states show very little difference in propensity to migrate based upon self-employment, with migrants showing a slightly higher degree of self-employment. Ten percent of the migrants from these states were self-employed before moving.

PAGE 63

51 Average Age In most migration studies age is expected to be negatively related to migration since older people have a shorter work life over which to realize the gains from migration. In addition, it is felt that older people have greater ties to present location. That is, they are likely to have stronger attachments to friends and relatives and are likely to possess location-specific capital, such as a home. Only upon retirement are older people expected to show higher migration rates. But since this study is restricted to members of the labor force, the inverse relationship is still expected here. It appears, however, that in this sample, the average age is greater for migrants than for nonmigrants. Perhaps this occurs because the sample is restricted to Florida-sending states. Many of the movers to Florida and competing destinations, such as California, may be anticipating future retirement. For example, a person may move to Florida and accept a low-paying job knowing that (he or she) will withdraw from the labor force after a predetermined number of years. Individuals who have accumulated substantial savings during their lifetime can afford to do this. In addition, many older migrants may choose semiretirement rather than retirement, perhaps maintaining part-time jobs after moving. If the preretirement or semiretirement thesis is correct, we would expect labor force migration from the sample states to consist of older migrants than migration from all other states. In the Panel Study, it is found that the average age of all U.S. labor force migrants is 36 years old. This compares to a mean age of 38 in the sample of Florida-sending states. Thus, there is mild evidence in favor of this hypothesis.

PAGE 64

52 Employment Status People who are unemployed are expected to be more responsive to opportunities elsewhere than those who are employed. Particularly if local economic conditions are poor, the process of job search for the unemployed is likely to Include alternative locations. If, however, employment conditions at other locations are also depressed, the response of the unemployed is uncertain. The effect of employment status upon the migration decision seems to depend on local economic conditions relative to alternative destinations. In the analysis to follow, relative conditions will be held constant by considering only one state at a time. The mean value of employment status for all four states combined is slightly lower for migrants than for nonmigrants. Ninety-six percent of the migrants in the sample were employed prior to moving. Marital Status Marriage is expected to act as a deterrent to migration. As with the family-size variable, presence of a spouse represents an additional tie to present location. This is particularly true if the spouse is working. The data in Table 3.4 show that migrants are less likely to be married than nonmigrants. Seventy-eight percent of the migrants in the four-state sample were married, while 81 percent of the nonmigrants were married. Race The expected sign of the race variable is ambiguous. On the one hand it can be argued that since some blacks are more likely to earn lower incomes and are more likely to be unemployed than x^hites, they are

PAGE 65

53 more responsive to opportunities elsewhere. On the other hand, there may also be fewer opportunities elsewhere for blacks because they are concentrated in low-skill occupations and because of discrimination in the labor market. In addition, blacks are less likely to be employed in jobs requiring geographic transfers. For the sample to be considered here it looks as if migrants are more likely to be white, indicating that the latter two effects outweigh the first factor. Average Education More educated people are expected to move more often. The reason is twofold. For one, education (in most cases) leads to increased employment opportunities. Particularly it is believed that increased education opens up employment possibilities which are more national in scope resulting in greater migration propensities. In addition, those who have attained higher education levels are expected to have more and better information about migration opportunities. The extreme of this phenomenon occurs in professions that form organizations which aid their members in locating jobs. The economics profession is an example. The data in Table 3.4 give preliminary evidence that this thesis is correct. Migrants in the sample are likely to have at least started college, while the average nonmigrant posesses only a high school education. Previous Migration Migrants are believed to be characterized by frequent mobility during their lifetimes. It is hypothesized here that a person who has changed state of residence at least once during his (or her) lifetime is more likely to move again than an individual who has always lived at the same location. People who have moved frequently are also less likely

PAGE 66

54 to have developed attachments to their present state of residence. In the four-state sample, 31 percent of the migrants have moved before. Only 23 percent of the nonmigrants have previously changed their states of residence. Thus, there is support for the hypothesis set forth above. Employment Status of Spouse The presence of a working spouse is expected to deter migration. If both partners have substantial earnings, any migration decision must presumably maximize their combined return. The probability of finding a location that meets this criterion is lower than the chances of finding one which maximizes the returns of one family member. It is possible that only one member of the couple will be properly rewarded upon moving and the spouse will be tied to this choice. Whether joint utility is maximized or not, the spouse's employTiient and income possibilities must be taken into account in any decision to move. Theoretically, this should act as a deterrent to migration; however, preliminary evidence in Table 3.4 suggests that migrants in the sample were more likely to have working spouses. Thirty-eight percent of the migrants' spouses worked prior to moving. Family Income Migrants are expected to earn higher annual income than nonmigrants. A few reasons can be given for this belief. First, higher incomes usually imply greater demand for leisure activity. Changing location is one way of obtaining more leisure. This is particularly true for this sample of states which have typically sent large numbers of migrants to Florida and California, states which possess climates amenable to

PAGE 67

55 recreational activities. Higher income people are also more likely to have visited other states as tourists. Thus, they would have obtained information about these states which could lead to increased likelihood of moving. The data in Table 3.4 show significant differences between family income levels of movers and nonmovers Migrants earned an average of $11,240 in 1968. This compares to an average income level of $8,790 for nonmigrants. Empirical Results Introduction The comparison of migrant and nonmigrant characteristics presented so far serves only to provide a general picture of relationships in the data. Much can be hidden in such a rudimentary analysis. Since all four sample states were grouped together, differences in migrant selection between states are hidden. As an example, migrants from New York may tend to have larger families than New Yorkers who do not move. There may, however, be little difference in the family size of movers and nonmovers from Ohio, Pennsylvania, and South Carolina. VJhen all four states are viewed together, very little difference in family size will show up, even though this is an important determinant in New York. This problem becomes particularly severe if the variable takes on opposite signs in each state. Theoretical reasons exist for expecting differences in the determinants of migration between states of origin. The crux of the argument is that residents of different states face varying relative migration opportunities. Residents of Ohio, for example, wirll experience a greater improvement in climate from moving to Florida than will residents of

PAGE 68

56 South Carolina. If older people are more responsive to climate differentials than younger people, we might then expect age to be a more important migration determinant in Ohio than it is in South Carolina. Thus, variables are significant in one state but not In another because of differences in initial (origin) conditions. From now on, each state will be analyzed by itself. When migrants and nonmigrants are compared only according to the mean values of a set of variables, other problems arise. First, there are the usual weaknesses involved in using the mean as a measuring device (its sensitivity to extreme values is one such weakness). More importantly, when any variable is considered, the values of remaining factors are not held constant. Thus, the true effect of each characteristic is not being measured. Use of multiple regression techniques will relieve these problems. Before the results of the logit analysis of the decision to move are presented, a discussion of the scaling of some of the qualitative (dummy) explanatory variables is in order. Particularly, length of residence, s elfemployment and average education are defined in ways which are somewhat out of the ordinary. Looking back at Table 3.3 it is observed that these variables are scaled using an ordinal ranking system, with each value representing a given level or interval. This scaling is the same as that which appears on the Panel Study tapes. The conventional method of modeling variables which are coded in such a fashion is to include separate dummy variables for each category of data. In this case, however, 21 variables would need to be added to the model. This

PAGE 69

57 would result in reduction of degrees of freedom and, more importantly, would cause logit estimation to become prohibitively expensive."^ A second method of handling the problem would be to redefine the dummy variables so as to have fewer categories. In the limiting case only one zero-one dummy is used. Some cutoff point is chosen and a value of one is assigned to those observations for which the cutoff is exceeded and a value of zero is assigned otherwise. In this study, for example, the education variable could be defined so that those individuals with 12 years or more of schooling were given values of one with the remaining observations equaling zero. Assumed in such an approach is that increases in the level of education up until the 12th grade have no effect on the probability of moving. In addition it is assumed that increases in education beyond high school have no effect. To validly use such an approach some prior expectation of the proper cutoff point is required. Appendix A presents the results of logit estimation when this approach is adopted. Not having prior knowledge of the correct cutoff point, the choices used are somewhat arbitrary. The actual scaling used for the main part of this study (see Table 3.3) also imposes restrictions upon the model. To illustrate these restrictions, this approach is compared to the conventional method of defining such data. Suppose an explanatory variable contains three categories. Following the conventional method, two dummy variables are introduced. (If three dummies are used, then the constant term must be eliminated.) Call the dependent variable Y and the two dummies X and X In fact, given the computer program limitations, logit estimation would have been impossible.

PAGE 70

58 Assuming no other variables in the model, the equation can be written as: Y = a + 3^X^ + 32X2 + E (1) wh ere a, 3, and 3o are coefficients and E is a random disturbance term. Since X and X^ are zero-one variables we can further say that: E (Y E (Y E (Y X = 1 and X = 0) = a + 3j_ X^ = and X2 = 1) = a + 32 X, = and X„ = 0) = a (2) (3) (4) Now suppose instead of using the above approach we define a single variable, X, which takes on the values of 0, 1, and 2 for each of the three categories. Now the equation can be written as: Y = a + 3X + E (5) where a and 3 are parameters and E is the disturbance. Now we can say that: E (Y X = 0) = a (6) E (Y X = 1) = a + 3 (7) E (Y X = 2) = a + (8) This formulation turns out to be equivalent to estimating the model of equation (1) with the following restriction attached: 3, = \ While this restriction should be considered, it may be preferable to the (9) restrictions imposed when the variable is collapsed into t\
PAGE 71

59 Summarizing, it would be optimal to include separate dummy variables for each category for which there is information. Program and cost limitations, however, prevent using this approach. Alternatives include collapsing the variable into fewer categories or including a single ordinally ranked categorical variable. While the latter approach is somewhat unprecedented^, it has been chosen because it utilizes more information which is available in the data. Both alternatives impose restrictions which may be considered severe without prior knowledge of data. In the next four sections, the results of the logit analysis of the decision to move will be presented and discussed. Logit Results: Ohio Table 3.5 presents the regression results for the state of Ohio. Variables which are significant with at least 95 percent confidence are sex, marital status, and family income. The sex variable, as hypothesized, has a positive coefficient. Thus, households with male heads from Ohio are more likely to migrate than households with female heads from that state. The most significant variable is marital status. As expected, unmarried individuals move more often than those who are married. Thus, marriage does act as a deterrent to migration in Ohio. Family income is also positive and significant. Higher income families in Ohio are more likely to change states of residence than lower income families. This is also the expected result. ^To test whether wage and price controls of the Nixon administration had any effect upon wage formation, a three-leveled ordinal dummy was used by Eckstein and Girola (1978). The variable took on the value of .5 for Phase I of controls and 1 for Phase II of controls. Periods without controls took on the value of 0. The variable was found to be insignificant. The same variable was used in a price equation and was significant.

PAGE 72

60 Table 3.5 Results of Logit Analysis; Ohio, 1968-1977 Decision to Move, Variable Famsz Sex Homeown Lres Lempl Emself Avage Ernst Marst Race Aveduc Prevmig Emstw Famy Coefficient .147 2.802 -3.65 .013 .052 -.845 .054 -4.142 .867 -.024 .298 .471 .057 Asymp totic t-Ratio 1.27 2.60 .64 .06 .31 .73 1.53 -3.39 1.09 .15 .69 .96 2.06 Constant -5,000 Summary Statistics: Log of Likelihood Function = -86.0 Likelihood Statistic (at 0) = 133.0 Significance Level (at 0) = .005 Likelihood Statistic (at mean) =27.3 Level of Significance (at mean) = .025 Percentage Correctly Predicted = 85% Percentage of Migrants Correctly Predicted = 16% -2.72

PAGE 73

61 The remaining variables are not significant. Family size, average age, and race are variables which show some importance. Migrant families in Ohio are more likely to be larger than nonmigrant. This result is opposite from that expected. A possible explanation is that many migrant families from Ohio have preschool children. While having children in school may act as a deterrent to migration, families may decide to move prior to their children's first enrollment in school. Indeed, people may move so that their children can go to different schools. Age shows a positive sign, lending support to the preretirement hypothesis set forth earlier. Race is also positive, indicating that households headed by white persons are more likely to move interstate than households headed by nonwhites. Although insignificant, the homeownership variable has the expected sign. People owning homes in Ohio are less likely to move than those who rent. They are also more likely to have lived in their homes and worked for the same employer for longer time periods than nonmigrants. These results are contrary to expectations. The migrant is less likely to be self-employed. Thus, self -employment appears to act as a migration deterrent. Those who chose to leave Ohio are more likely to have moved previously, supporting the hypothesis put forth earlier. Education carries an unexpected negative coefficient. Migrant household heads are also likely to have spouses who are employed. This variable has the opposite sign from that which was expected. Employment status was not included in the analysis because there was no variation of this characteristic among migrants. Estimation was thus impossible. When the likelihood value of the decision equation is compared to the value that results when all coefficients are zeroed out, the equation is significant at better than the .005 level. When the likelihood

PAGE 74

62 value of the equation is tested against that which results when the right-hand side of the equation is set equal to the mean of the dependent variable, the equation is significant at the .025 level. Under either test, the equation appears to produce a good fit. A correct prediction is defined as one in which the predicted probability is within 50 percent of the actual value of the dependent variable. For migrants this means that move > .50. For nonmigrants this means -^move < .50. Under this criterion, the equation predicted 85 percent of all (migrant and nonmigrant) individuals in the sample correctly. Among migrants, 16 percent were correctly predicted. Logit Results: New York Table 3.6 presents the decision-to-move results for the state of New York. Variables which are significant at better than the .05 level are length of employment, average age, employment status, and spouse's employment status. Length of employment carries a negative sign, implying that those who have worked for the same employer for long periods of time are less likely to move. This result is expected. Migrants from New York are also older than nonmigrants from that state, providing further support for the preretirement thesis. Migrant household heads are less likely to be employed than nonmigrants, but their spouses are likely to be employed. This last result, which was also found to be true for Ohio, is contrary to expectations. Perhaps what we are observing This definition is admittedly arbitrary. Some cutoff had to be chosen in order to summarize the results without listing every predicted value.

PAGE 75

63 Table 3.6 Results of Logit Analysis: Decision to Move, New York, 1968-1977 Variable Famsz Sex Homeown Lres Lempl Emself Avage Ernst Marst Race Aveduc Prevmig Emstw Famy Coefficient -.165 .127 .773 .374 -.561 -1.277 .074 -2.960 -1.600 -.890 .227 -.448 1.682 .084 Asymp tot ic t-Ratio -1 .37 .08 1 .38 1 .51 -2 .80 -1 .40 2 .30 -1 99 -1 03 89 1. 34 71 2. 77 1. 88 Constant -2.580 Summary Statistics: Log of Likelihood Function = -62.8 Likelihood Statistic (at 0) = 144.8 Level of Significance (at 0) = .005 Likelihood Statistic (at mean) =38.4 Level of Significance (at mean) = .005 Percentage Correctly Predicted = 83% Percentage of Migrants Correctly Predicted = 21% -1.39

PAGE 76

64 are married couples with both spouses being employed in highly mobile, perhaps professional, occupations. In any case, employed spouses are not acting as deterrents to migration. Family income is significant at the 94 percent confidence level. It carries its expected positive sign. The remaining variables are insignificant. Sex, sel f -employment marital status, and average education have the signs which were hypothesized earlier. Homeownership has a positive sign, indicating that migrants are more likely to own a home than nonmigrants. They are also more likely to have lived in their homes for longer periods of time. Race, although insignificant, carries a negative sign. This lends support to the thesis that blacks are more responsive to opportunities elsewhere. Previous migration also has the opposite sign from that expected. Migrants from New York are less likely to have moved before in their lifetimes. This is consistent with the sign on the length-of-residence variables. Perhaps many of the New York migrants are middle-aged or older people who have lived in New York most of their lives and are now moving in anticipation of retirement. The overall equation is significant at better than the .005 level. This is true whether the model is compared to one where all coefficients are zero or compared to a model where the right-hand side is equal to zero. Eighty-three percent of the individuals in the sample are predicted correctly by the model, while 21 percent of the migrants are predicted correctly. Logit Results: South Carolina The decision-to-move results for South Carolina appear in Table 3.7. Significant variables are family size, average age, marital

PAGE 77

65 Table 3.7 Results of Logit Analysis: Decision to Move, South Carolina, 1968-1977 Variable Famsz Sex Homeown Lres Lempl Emself Avage Ernst Marst Race Aveduc Prevmig Emstw Famy Coefficient .252 1.231 -.180 .286 -.328 -2.713 .093 -.678 -3.494 -.371 .073 .803 -.528 .217 Asymptotic t-Ratio 1.95 1.04 .33 1.17 -1.76 -1.65 2.51 .66 -2.46 .58 .38 1.24 .89 3.22 Constant -6.333 -2.96 Sunimary Statistics: Log of Likelihood Function = -71.7 Likelihood Statistic (at 0) = 164.4 Level of Significance (at 0) = .005 Likelihood Statistic (at mean) =39.8 Level of Significance (at mean) = .005 Percentage Correctly Predicted = 90% Percentage of Migrants Correctly Predicted = 31%

PAGE 78

66 status, and family Income. Marital status and family Income show the expected signs. Contrary to expectations, family size is positive and significant. Again average age carries a positive sign. Although insignificant at the 95 percent confidence level, length of employment and self-employment are important and carry the correct signs. Of the remaining variables, sex, homeowner ship, employment status, average education, previous migration, and spouse's employment status Influence migration decisions in the expected directions. As in Ohio and New York, length of residence is positively related to migration. Race displays a negative relationship, indicating that households headed by blacks are more likely to leave the state. The South Carolina equation provides a very good fit of the data. The equation is significant by both criteria (at zero and at the mean) at better than the .005 level. Ninety percent of the individuals in the sample were correctly predicted by the model, with 31 percent of the migrants' choices being predicted accurately. Logit Results: Pennsylvania In Table 3.8 are the results of the regression analysis for Pennsylvania. Variables which are significant at the .05 level are average education and family income. Average education shows the expected positive sign. Family income, however, is negatively related to the probability of moving. This is truly unexpected for a variable that has been positive and significant for the other three states. Close to being significant are family size and marital status. As in Ohio and South Carolina, family size takes on -a positive sign.' Marital status has the hypothesized negative sign, indicating marriage"

PAGE 79

67 Table 3.8 Results of Loglt Analysis: Decision to Move, Pennsylvania, 1968-1977 Variable Famsz Sex Homeown Lres Lempl Emself Avage Ernst Marst Race Aveduc Prevmig Emstw Famy Coefficient .251 1.859 .105 .022 .253 1.257 .021 -2.625 .109 .638 .225 .869 -.219 Asymptotic t-Ratio 1.85 1.51 .19 .11 1.34 .88 .59 -1.84 .17 2.88 .36 1.66 -2.65 Constant -5.650 -2.78 Summary Statistics: Log of Likelihood Function = -71.6 Likelihood Statistic (at 0) = 57.2 Level of Significance (at 0) = .005 Likelihood Statistic (at mean) = 19.4 Level of Significance (at mean) = 10 Percentage Correctly Predicted = 871 Percentage of Migrants Corrected Predictly = 7%

PAGE 80

68 deters migration. Sex and previous migration are both positively related to migration, as expected. Homeownership length of residence, length of employment, self-employment, and spouse's employment status all show the wrong sign. Average age has a positive sign, giving further support to the preretirement thesis, and race is positively related to migration, indicating that white families leave Pennsylvania more often than do black families. The Pennsylvania equation provides a poorer fit than the decision equations for the other three states. Although significant compared to the model where all coefficients are zero, the more stringent test indicates a weaker model. Setting the right-hand side of the equation equal to the mean value of the dependent variable and comparing the likelihood value which results with the unconstrained likelihood value results in a significance level of .10 for the unconstrained model. Although 87 percent of the sample was predicted correctly by the model, only 7 percent of the migrants were predicted correctly. Thus, this model has not been very successful in distinguishing between migrants and nonmigrants. Test for Separation of States In Chapter 2, a theoretical argument was made for estimating separate equations for each state of origin. In the preceding sections of this chapter evidence was presented which Indicated that there may be differences between states in the determinants of the migration decision. In this section, a formal test for equality of coefficients between states is carried out. —

PAGE 81

69 The test developed by Chow (1960) for the equality of regression coefficients between equations is applied to the decision-to-move equations. First, ordinary least squares (OLS) estimates for each state are determined. Then the observations from all four states are combined and an OLS equation is estimated for the combined data."* The results of these estimates appear in Table 3.9. The following statistic is calculated; [SEExotal (SEEny + SEEqh + SEEsc + SEEp^) ] / 3k [SEEi^Y + SEEq^ + SEEsc + SEEp^] / n-4k where SEE stands for the sum of squared residuals, k refers to the number of coefficients in each equation, and n refers to the size of the combined sample. The subscript NY refers to the New York equation with OH standing for Ohio, SC referring to South Carolina, and PA meaning Pennsylvania. The subscript TOTAL refers to the combined equation. The statistic is distributed according to an F-distribution. Its value in this case is 1,67 which compares to a critical F value (14 degrees of freedom in the numerator, and 779 in the denominator) of 1.5 at the .05 level of significance. Thus, the hypothesis of equality of regression coefficients between equations is rejected at the 95 percent confidence level. The belief that there are differences in migration determinants across states is confirmed. OLS is chosen for the test because execution of the logit program is impossible for the combined sample and because no comparable test exists for comparing coefficients across logit equations.

PAGE 82

70 Table 3.9 Ordinary Least Squares Results: Decision to Move, 1968-1977 Variable Famsz Sex Homeown Lres Lempl Emself Avage Ernst Marat Race Aveduc Prevmig Emstw Famy Constant -.210 .94 .366 1.69 -.210 R^ = .13 R^ = .17 R^ = .09 Ohio New York Penns ylvania Coefficient t-Ratio Coefficient t-Ratlo Coefficient t-Ratio .015 1.02 -.020 -1.51 .023 1.60 .494 3.32 .073 .50 .323 1.74 -.047 .70 .055 .95 .005 .09 .002 .05 .060 1.32 .003 .15 .005 .21 -.047 -2.63 .024 1.18 -.105 .77 -.061 .59 .161 1.06 .006 1.46 .009 2.75 .002 .43 .037 .19 -.348 -1.80 .009 .04 -.626 -4.05 -.198 -1.35 -.402 -2.06 .069 .86 -.136 -1.47 -.012 .17 -.003 .14 .020 1.16 .064 2.78 .035 .63 -.041 .64 .015 .20 .044 .74 .128 2.21 .106 1.84 .009 2.48 .009 1.71 -.016 -2.87

PAGE 83

71 Table 3.9 (continued) South Carolina Combined States Variable Famsz Sex Homeown Lres Lempl Emself Avage Ernst Marst Race Aveduc Prevmlg Emstw Famy Coefficient t -Ratio Coefficient t-Ratio .018 1.58 .010 1.54 .174 1.35 .278 3.96 -.020 .40 .005 .18 .040 1.10 .011 1.00 -.042 -2.21 -.014 -1.52 -.228 -1.76 -.047 .78 .009 2.83 .005 3.38 -.032 .28 -.120 -1.47 -.399 -3.16 -.275 -4.01 -.082 -1.32 -.016 .52 .006 .29 .018 1.85 .069 .98 .024 .81 -.066 -1.27 .041 1.47 .026 4.03 .005 2.23 Constant -.045 .31 -.148 -1.17 18 R"" = .06

PAGE 84

72 Application to Forecasting The models for all four states have been re-estimated for the period 1968 through 1972. The results are presented in Table 3.10. For the shorter time period there are, of course, fewer migrants. This led to some problems. First, some of the variables which were used in the nine-year migration analysis could not be included here. A requirement for logit estimation is that there be variation within each category of the dependent variable. Since there are fewer migrants in the sample, there is less likelihood that some variables will meet this requirement in the migrant category. Those that do not must be deleted. Second, with a smaller proportion of migrants in the sample it is more difficult to distinguish between migrants and nonmigrants. In general, poorer fitting equations were obtained for the shorter time period. This is because there were fewer variables that could be included in the model and because of the smaller proportion of migrants in the sample. Using the methodology outlined in Chapter 2 a forecast is developed. The coefficients in Table 3.10 are applied to values of the explanatory variables for individuals living in the sample states in 1972. Individual predicted probabilities are calculated and averaged over the number of individuals in each state. The mean probability for each state is then multiplied by the total sample size in 1972. The results of this procedure appear in Table 3.11. The forecast values are compared to the actual (known) number of movers during the 1972-76 period. For Ohio, 25 migrants are predicted. In fact, there were 20 Ohio migrants in the sample between 1972 and 1976. For New York, 13 migrants are predicted compared to 17 actual

PAGE 85

73 CM r-1—1 I 00 VD 0^ 0) > C O •rl M •H o O o H 4-J r-^ CTi m CO C30 C .— 1 Pi • cti 1 •H +J 1 C ffl > +J .H c >^ (U orr^ CD H ^ 00 o C O T— 1 o f-H c •H • • • cu M-( •—1 1 PM M-l OJ O U 0^ CO CNl CO O r—i O I 00 o I— H 00 00 — 1 I— t I CM 1 rH 1 1 — i c^) 1 i-H 1 rH i-H 1 CM CM 1 r~-* 1 1 1 o cn ~* v^ C7\ o o ^ o 00 C_) r— 1 -* ->D 1 1 — 1 4-i 1 ^ S-i •U >-' d OJ VO rs •H I— t Q) CJ n 2 •H M-1 OJ O O i' cn 00 00 IT) CM I o 00 CO cn r-^13 CO CM 1 CO o CM o 859 CO 00 cn o r-H u-5 o CTN CJ^ 1 .—1 1 1 — i t-H 1 1 Cfl 0) Pi cn cu r-l nj H o •H •U CO cn CO Pi 4-1 1 o U •w c a:: QJ CO o •H I-H O 00 H cw 1 U-i 0) o u in o cn 1 rsi c^ vO r^ CJ^ VD m CO o CN LTl CN r-^ CM CM .-H r—i • I-H 1 1 CsJ CM 1 00 o un CsJ r~-. CO vD in 00 o o LP) CO CO ~d, cn H CD nj cn U O 0) QJ cn e C (U B > E iS cd > M e nd o ^-^ W <; M S Pi
PAGE 86

74 cfl •H C ca > LO o m ^ • I— t o >^ Ln CSl o M ^ • c 1 c u-l lO 00 r—i >? &^ • — I 0-) 1 • CNI 0-) 0) H o LO CO m •H • oo ^ 00 CM o O CO 1 6-S 0^ &-5 O >> ^'-s c T3 4J d (fl 0) O /-^ m OJ 4J CU ^-^ O 0) B a h C o e •H u o 4-1 4-1 -d o •H 4-1 cti 4J cd 0) u 4J cd nj u a v_^ v_x (14 en d 0) 0) u 3 a a a CJ >, C f^ •H c •H S rH cO M 4-) CO 4.) CO 4-1 M a -d OT o m O o &C H O •H •H •H H cu •H 4-1 4-) M-l 4-1 M-l 5-4 g m ^ Ctl •H nj •H !-l •H •H 4-1 d 4-1 n o M-l U iH W M CO bO o o nJ 0) •H •H X) 4-1 ^ -13 C/3 X) OO cu , Xi o ^ o 4-J U -H u m •H •H P C T3 CO o iH iH 1—1 iH CU tU QJ s 0) 0) 0) 0) o O M s M M > M > !-i !-i Pu 3 o •H 0) •H 0) OJ a) CO hJ i-J hJ h-I J Ph Cm

PAGE 87

75 Table 3.11 Comparison of Forecast to Actual Migration, 1972-1976 State of Origin Ohio New York South Carolina Pennsylvania Total Number of Number of Difference Migrants-Forecast Mi grants-Actual (Forecast-Actual ) 25 20 5 13 17 -4 27 15 12 10 15 -5 75 67 8

PAGE 88

76 movers. Twenty-seven South Carolina migrants are forecast. In fact, only 15 left that state during the period. The model predicts 10 migrants from Pennsylvania, while 15 actually left the state. In total 75 migrants are forecast to have moved. This compares to 67 people who actually moved. This is a 12 percent error. It is interesting to note that the best fitting models in the historical period, 1968-72, do not provide the most accurate forecasts. The best fitting equation is for the state of South Carolina, yet the greatest forecast error occurs in this case. The decision equations for each state are estimated next for the period 1972-76. The results appear in Table 3.12. It should first be observed that these results look very different from the 1968-72 migration equation appearing in Table 3.10. Many variables which are significant for the earlier period lose their significance in the latter period and other variables gain significance. For example, for the state of Ohio, the only significant variable in the 1958-72 equation was family income. In the 1972-76 period, this variable is no longer significant and length of employment becomes significant. Significant variables in the earlier period New York equations are length of employment and employment status of spouse. In the later period, no variables are significant at the 95 percent confidence level. Length of employment, marital status, and family income were significant variables in the first four-year period for the state of South Carolina. Only family income remained significant in the 1972-76 period. For Pennsylvania, average education was the most significant 1968-72 variable. Later, family size, length of employment, average age, and marital status showed significance. In addition, sign changes frequently

PAGE 89

77 o •H u VD a\ rt a^ VD Pi • CO 1 .— 1 1 — 1 •H +-) C cd > +J rH fl f>. 0) r-^ o § •H CO •H a CO CO ca H &0 O o •H U o^ n o 1 C^ 1 — 1 CO ^ r~svD VD • i-H CO IX! i-q J w ^ (H o OJ 0) en F Cd cd > ^J B en Sh pc: <: P-i W tlH c cd +j C o u

PAGE 90

78 •H C > en S C QJ CM CM in • CO o -0r-H o 1 • 5-S B-2 CO r-a^ O in in • 00 o r^ o -* 1 Csl 6^ O T3 0) C •H J-J O O o rON in >-l • CO o 00 o 3: in • 0) 1 a O-J CM ON B-2 O 0) O ^ 00 in iH •H f-H o ^ Xi CO CM o n) O m • H 1 . ^^^ fi Td 4-1 d cd QJ a /'-^ ca QJ 4J QJ /--N o 0) g a u C! o g •H u O 4-1 4-1 -XS o •H 4-1 cd 4-1 CO QJ o M cd ^— cd V— IH O ^.^ PM CO C 0) QJ 4J =1 a o o O >. p! U^ •H d •H s r-{ cd M 4-1 03 4-J cd 4-1 U a 'T3 CO o M a CJ M •H O •H •H •H •H QJ •H U O 4-1 IW 4-1 y-i U s cn ^ cd •H cd t-4 U •H •H 4-1 C 4-1 a o 14H 4-1 iH CO M CO bO CJ o nj 0) •H •H TJ u ^ Ta CO T) CO 0) QJ Q) CO •rH o O bO bO 4J l-I o IM O 14-1 cd Cd a >N ^ O (-< o 4J 4J -H 5-1 M-l •iH ^ c C -13 cfl O rH M .H rH QJ QJ QJ g 0) 0) 0) Q) CJ CJ U g 00 ^ > ^ > u U FL, 3 o •H QJ •H QJ QJ QJ CO hJ hJ i-J 1-1 1-1 ei4 fL,

PAGE 91

79 occurred, although this was usually restricted to the insignificant variables. It should be emphasized that all four-year period equations are misspecified because some variables could not be included in the estimation. This makes comparisons of the two periods less meaningful. There is evidence, however, of instability of the decision-to-move coefficients between the two time periods. This has negative implications for forecasting based upon previous coefficients. But the reasonableness of the forecasts which have been obtained provides support for this methodology. Coefficient changes which occur may be balanced in such a way as to affect the average probability, the key variable in the forecast, only slightly. The final step in the analysis is to apply the 1972-76 coefficients to the characteristics of people living in the sample states in 1976. Thus, a 1976-80 forecast is obtained. These predictions are presented in Table 3.13. The results show that 24 of the Panel Study members who lived in New York in 1976 will have left by 1980. For both South Carolina and Pennsylvania 19 migrants are forecast. Ohio is expected to lose 18 migrants during the period. An aggregate forecast is also presented in Table 3.13. It is obtained in the following manner. First, the number of persons in the labor force for each state in 1976 is taken from the Employment and Training Report of the President Then, in order to be comparable to the Panel Study sample, working spouses are removed from the measure. This is accomplished by multiplying the labor force number by an estimate of the proportion of working spouses in the labor force. The estimates used are the proportions that occur in the Panel Study sample states in 1976. The resulting measure of labor force household heads is

PAGE 92

80 Table 3.13 Forecast: Decision to Move, 1976-1980 State of Origin Ohio New York South Carolina Pennsylvania Total Number of Migrant sPanel Study 18 24 19 19 80 Number of MigrantsAggregate Labor Force 833,000 1,867,000 309,000 1,003,000 4,012,000

PAGE 93

81 then multiplied by average family size to obtain an estimate of the actual number of potential migrants from each state. The average family size measure is also estimated from the Panel Study sample. Finally, the potential population at risk is multiplied by the average probability of migraiton estimated for the Panel Study. The numbers in Table 3.13 result Using the procedure put forth in Chapter 2, estimates of aggregate labor force migration were developed from sample characteristics for the periods 1968-72 and 1972-76. Average predicted probabilities were calculated for each period based upon the regression results. The population at risk for each state was derived using the method discussed in the last paragraph. The estimates for these two periods and the forecast are presented together in Table 3.14. It can be seen that for the 1968-72 period, New York is estimated to have had the largest amount of out -migration. Pennsylvania is next, followed by Ohio and South Carolina. For the 1972-76 period, total migration out of the four states increased. Only New York did not share in this increase. The largest jump occurred in Ohio which is estimated to have 457,000 more out-migrants during this period than the number who left in the 1968-72 period. This increase was large enough to put that state in second place to New York. Pennsylvania and South Carolina showed small increases from the previous periods. For the 1976-80 period, total outmigration is expected to decline from the previous period. This decline will be felt in all states except New York. The greatest decline Since this study has found family size to be insignificant in the determination of the decision to move, the family-size measure is obtained from the total (migrant and nonmigrant) sample.

PAGE 94

82 Table 3.14 Aggregate Labor Force Migration: 1968-1972, 1972-1976, 1976-1980 State of Number of Migrants Number of Migrants Number of Migrants Origin 1968-1972 1972-1976 1976-1980 Ohio 714,000 1,171,000 833,000 New York 1,708,000 1,643,000 1,867,000 South Carolina 310,000 341,000 309,000 Pennsylvania 1,121,000 1,131,000 1,003,000 Total 3,853,000 4,286,000 4,012,000

PAGE 95

83 occurs in Ohio, the same state that was estimated to have an unusually high level of out-migration in the 1972-76 period. Overall, about 4 million persons are forecast to leave the four states between 1976 and 1980, indicating that there will be a continuing large pool of migrants from which Florida and other states may draw. The reasonability of the method used to estimate and forecast aggregate labor force migration can partially be determined by comparing the results with measures of out-migration derived from the 1970 Census of Population. These estimates are not directly comparable, however. For one, the Census estimates cover the five-year period between 1965 and 1970, while the estimates in this study cover four-year periods. Since there is some overlap the 1968-72 period is chosen for comparison. An additional problem in comparing estimates from the two sources is that the Census figures are not tabulated according to labor force status in published reports. Table 3.15 compares the 1965-70 Census estimates to the 1968-72 estimates of out -migration obtained in this study. It can be seen that with the exception of Ohio, the estimates of migration derived in this study exceed those obtained from the 1970 Census. Because the Census estimates include all members of the population (whether they are in the labor force or not) and because the Census data cover a longer period the opposite result would be expected. On the other hand, it is generally believed that there was significant undercount in the 1970 Census. In addition, because of prosperous economic conditions the early 1970s may have been years of greater rates of migration. It is kno\-m that in Florida, for example, in-migration accelerated significantly during this period. Since the origin states chosen for this study are some

PAGE 96

84 Table 3.15 Comparison of Panel Study Derived Out -Migration Estimates to Census Estimates State of Origin Panel Study Estimate, 1968-72 Census Estimate, 1965-70 Ohio 714,000 787,546 New York 1,708,000 1,329,432 South Carolina 310,000 248,609 Pennsylvania 1,121,000 781,684 Total 3,853,000 3,147,271

PAGE 97

85 of the strongest contributors of Florida migrants, these states are likely to have experienced higher rates of out-migration. These latter arguments provide some support for the finding of higher levels of migration in this study. More detailed evaluation of the estimates is impossible without having a comparable set of labor force migration figures for the same period. The next chapter presents the results of the destination-choice analysis and develops a forecast for the 1976-80 period.

PAGE 98

CHAPTER 4 EMPIRICAL RESULTS: DESTINATION CHOICE Destination Choice: 1963-1977 A summary of the destinations chosen by migrants in the four sample states appears in Table 4.1. Overall, Florida is the leading destination, with 20 percent of all out -migrants choosing this state. Florida is the top location choice for New York, Pennsylvania, and Ohio migrants. South Carolina migrants showed a preference for moving north, with New York being the leading destination. Florida, however, followed closely behind. With the exception of those choosing Florida, California, and Texas, many migrants chose to move to nearby states. As an example, 14 percent of the Ohio migrants moved to Michigan and 17 percent of New York's migrants chose to live in New Jersey. Because of technical and cost limitations, not all destinations which were chosen could be included in the analysis. The top eight destination states were selected. These states accounted for 70 percent of the migrants who left the fourstate area of New York, Pennsylvania, Ohio, and South Carolina. The variables used in the destination-choice equations are presented in Table 4.2. In the next section the rationale for inclusion of each variable will be discussed and the results are presented. Logit Results: Destination Choice The results of the first destination-choice model estimated appear in Table 4.3. The sample consists of 88 migrants originating from 86

PAGE 99

87 Table 4.1 Destination Choice Summary: 1968-1977 (Number of Migrants) Destination State New Y Florida 7 California 4 New Jersey 5 New York Illinois 1 Virginia 2 Mchigan Texas All other states 11 State of Origin 'Ivania Ohio So jth Carolina Total 7 6 5 25 3 5 1 13 5' 2 1 13 2 2 7 11 1 2 4 8 2 1 3 8 1 5 6 4 4 10 12 37 Total 30 28 37 33 125

PAGE 100

Table 4.2 Variable Definitions: Destination Choice Equations Dependent Variable Choice Explanatory Variables Distance Tempdiff Incdif f Hdiff Taxdiff Unemdiff Definition Equals 1 for chosen destination, equals for all other alternatives. Straight line mileage between each origin and each destination. Annual average temperature at each destination minus annual average temperature at each origin. Annual per-capita personal income at each destination minus annual per-capita personal income at each origin. First quarter housing price index at each destination minus first quarter housing price index at each origin. Annual per-capita state and local government tax collections at each destination minus annual per-capita state and local government tax collections at each origin. Annual average unemployment rate of each destination minus annual average unemployment rate at each origin.

PAGE 101

89 Table 4.3 Logit Results: Destination Choice, 1968-1977, Model 1 Variable Coefficient Distance -.703 Tempdiff -229 Incdiff .128 Hdiff .020 Taxdiff .010 Unemdiff .047 Summary Statistics: Log of Likelihood Function = -162.5 Likelihood Statistic (at 0) = 23.4 Level of Significance (at 0) = .005 Likelihood Statistic (at mean) = 14.35 Level of Significance (at mean) = .05 Asymp totic t-Ratio -2.42 3.86 1.35 .44 .22 .17

PAGE 102

90 either New York, Ohio, Pennsylvania, or South Carolina. Each member of the sample chose one of the eight alternatives put forth in the previous section. As in other studies, distance is expected to be negatively related to destination choice. That is, all other things constant, people are expected to move to locations which are closer to their initial residence. Distance is believed to act as a proxy for the cost of moving and the cost of obtaining information about alternative locations. As an example, a New Yorker is likely to know more about what it would be like to live in New Jersey than he would know about the situation in Texas. In fact, distance turns up negative and significant at the 99 percent confidence level. It is expected that individuals are attracted to more temperate climates. Warmer temperatures, in addition, may represent other amenities such as recreational activities and access to the coast. Colder temperatures at origin locations are expected to act as a further stimulus to migration. It is hypothesized here that climate is a more important variable in deciding on a destination to persons originating in colder climates. Thus, the greater the temperature difference between origin location and a given destination, the greater is the expected probability of choosing that destination. This is, in fact, what occurs. The difference in average temperature variable turns out positive and significant at the .005 level of significance. As hypothesized in regional growth models, migration is believed to be a response to differences in income opportunities. Thus, holding other variables constant, individuals are believed to choose to move to that location which promises the maximum possible lifetime gain in

PAGE 103

91 income over their present state of residence. Since the interest here is aggregate flows, average per-capita income differences are used to represent the opportunities available to the average person. This variable is positive and significant at only the 90 percent confidence level. This lends only mild support to the hypothesis that regional income inequality stimulates migration. The cost of living at alternative locations is also believed to influence the location choice. At first, an attempt was made to deflate the average-income variable by local price indices provided by the American Chamber of Commerce Association. This, however, caused very little change in the coefficient of the income difference variable. It was later thought that just comparing the overall cost-of-living index may be improper. Rather it may be particular components of the market basket which are important in location choice. Housing costs and tax levels were two variables which came to mind. Since moving often involves selling a home and buying a new home at a chosen location, the difference in state housing price indices between origin and destination locations was used as a variable. This variable was constructed as an average of all reporting cities in each state during the first quarter of the year under consideration (in this case 1968). It was expected to have a negative sign, meaning that people are expected to move from areas of high housing costs to locations where housing costs are lower. At the individual level a person would like to maximize the gain realized through selling his (or her) present home and purchasing a new one.Since the price index also reflects rental costs, renters too are affected by this variable. I#ien plugged into the model, however, the variable displayed the wrong sign and showed no significance.

PAGE 104

92 Per-capita state and local tax collections are also believed to make a difference in the location choice. Frequently, migrants are escaping areas with high state income and local property taxes. Again a negative sign was expected here, implying that individuals would like to reduce their tax burden upon moving. But, as before, the variable showed up to be insignificant. The final variable in the model of Table 4.3 is one which reflects employment conditions in sending and receiving areas. It is believed that individuals will choose to live at locations where there are a relatively large number of employment opportunities. At the same time, an individual from an area of high unemployTnent is expected to be even more responsive to job availability at other places. This is particularly the case if he (or she) is currently unemployed. As in other studies, though, the unemployment-differential variable is of the wrong sign and insignificant. Other variables representing employment opportunities, such as the rate of employment growth, were substituted. Still there was no success. I^^hen the estimated equation is compared to one in which all coefficients are restricted to zero, the model turns out significant at the .005 level. When compared to an equation in which the right-hand side is set equal to the mean value of the dependent variable, the equation is only significant at the .05 level. The cost of living and unemployment variables are removed from the model and the equation is re-estimated. The results are shown in Table 4.4. Distance is now significant at the .005 level. The temperature difference variable remains extremely significant. The difference in average annual per-capita income variable is now also significant at the .005 level. .......

PAGE 105

93 Table 4.4 Logit Results: Destination Choice, 1968-1977, Model 2 Variable Coefficient Asymptotic t-Ratio Distance -.741 -3.36 Tempdiff .231 4.78 Incdiff .152 3.55 Summary Statistics: Log of Likelihood Function = -162.8 Likelihood Statistic (at 0) = 22.9 Level of Significance (at 0) = .005 Likelihood Statistic (at mean) = 13.75 Level of Significance (at mean) = .005

PAGE 106

94 The significance of temperature and income differences lends support to the Graves (1978) thesis which suggests that migration is just a response to an increased demand for amenities which are available at certain locations. The economy is believed to start out in spatial equilibrium, meaning that amenity-adjusted real income is the same at all locations. Differences in income across locations are said to exist because of differences in the availability of location-specific goods, such as recreation. An exogenous shock, such as an income increase, causes the demand for location-specific goods to increase. This creates a temporary disequilibrium which is restored through migration. The positive and significant coefficient on the income variable says that if temperature differences are the same for all possible moves, people will choose places which promise positive income gains. To test whether individuals will still choose higher Income places, even if temperature differences are allowed to vary, the model is re-estimated with the climate variable removed. The results of this estimation are presented in Table 4.5. The income difference variable is no longer significant. Omitting the climate variable causes a downward specification bias on the income variable. Thus, the Graves theory receives confirmation. If the income variable had remained significant, then support would have been generated for regional growth models which claim that migration is solely a function of interregional wage differences. But, instead, income is only significant in conjunction with climate, so that the trade-off hypothesized by Graves appears to be in operation. The equation of Table 4.4 is significant at _the .005 level whether it is compared to the model where all coefficients are zero or to the

PAGE 107

95 Table 4.5 Logit Results: Destination Choice, 1968-1977, Climate Excluded Variable Coefficient Asymptotic t-Ratio -.10 .02 Distance -.016 Incdiff .005 Summary Statistics: Log of Likelihood Function = 174.19 Likelihood Statistic (at 0) = .02 Level of Significance (at 0) = (not significant at any acceptable level) Likelihood Statistic (at mean) = -4.5 Level of Significance (at mean) = (not significant at any acceptable level)

PAGE 108

96 model that is evaluated at the mean value of the dependent variable. In the next section this version of the model of destination choice will be utilized to obtain a forecast of destination choice. Application to Forecasting The procedure used to forecast destination choices follows the description set forth in Chapter 2. The three-variable version of the destination-choice equation is first estimated for the 1968-72 period. The results appear in Table 4.6. Distance and income differences are still significant variables, with the temperature-differentials varriable losing much of its explanatory power. The overall equation is significant at only the .10 level. The 1968-72 coefficients were applied to values of the explanatory variables in order to derive a 1972-76 forecast for each pair of origin and destination locations. For example, for the New York to Florida forecast the distance between the two states (constant over time) the 1972 per-capita income difference between them, and the 1972 average temperature difference are substituted into the 1968-72 equation. The predicted probability is then applied to the total number of migrants leaving New York for the eight alternative states between 1972 and 1976. The total number of available migrants is determined by assuming that the proportion of migrants who chose alternatives other than the eight states considered here is the same as it was during the most recent four-year period. The number choosing alternatives is then subtracted from the total number of migrants leaving New York for all states. The resulting number is multiplied by the probability_predicted by themodel. This process is repeated for each origin and destination pair.

PAGE 109

97 Table 4.6 Logit Results: Destination Choice, 1968-1972 Variable Coefficient Asymptotic t-Ratio Distance -.001 -2.15 Tempdiff .120 1.59 Incdiff .001 2.20 Summary Statistics: Log of Likelihood Function = -74.0 Likelihood Statistic (at 0) = 7.4 Level of Significance (at 0) = .10 Likelihood Statistic (at mean) = 7.4 Level of Significance (at mean) = 10

PAGE 110

98 Table 4.7 presents the Panel Study forecast for the period 1972-76. The forecast is also compared to actual destination choices during that period. For New York, Pennsylvania, Ohio, and South Carolina combined, 11 migrants are forecast to come to Florida. This compares to 15 who actually moved to Florida. This is about a 27 percent error. For California, 3 migrants are predicted. There were actually 8 migrants from the four states who chose California. Nine migrants are predicted to go to New Jersey in comparison to 5 who actually did. It is forecast that 6 movers would go to Illinois, while actually there were none. Five migrants are expected to go to Virginia. This compares to 8 who moved there between 1972 and 1976. Four migrants are predicted for Michigan, compared to 2 people who actually moved there. Finally, 3 are forecast to move to Texas, the same number that actually moved there. The overall ability to forecast thus appears to vary from state to state. Part of the problem, it is believed, results from the small sample size. A better fitting model was obtained for the 1968-77 time period. But when the time period is reduced there are smaller numbers of migrants from which to distinguish. This is believed to contribute to the reduced significance of the model and the less than adequate forecasts which were derived from it. In order to produce a forecast for the 1976-80 period the model was estimated for the four -year period just prior to it. The results are presented in Table 4.8. Only the temperature-difference variable remains significant at the .005 level, with the distance and incomedifference variables falling off to the ,10 level. While levels of significance have changed, the coefficients do not appear to have changed very much since the 1968-72 period. Only the temperature-difference

PAGE 111

99 Table 4.7 Comparison of Forecast to Actual Destination Choice: 1972-1976 OriginDestination New York-Florida New York-California New York-New Jersey New York-Illinois New York-Virginia New York-Michigan New York-Texas Pennsylvania-Florida PennsylvaniaCalifornia Pennsylvania-New Jersey Pennsylvania-New York PennsylvaniaIllinois PennsylvaniaVirginia PennsylvaniaMichigan Pennsylvania-Texas Ohio-Florida Ohio -California Ohio-New Jersey Ohio-New York Ohio-Illinois Ohio-Virginia Ohio-Michigan Ohio-Texas South CarolinaFlorida South CarolinaCalifornia 1 South CarolinaNew Jersey 1 South CarolinaNew York 2 South CarolinaIllinois 1 South CarolinaVirginia 1 South CarolinaMichigan 1 i South CarolinaTexas 1 Number of Number of Difference Migrants-Forecast MigrantsActual (F Drecast -Actual) 3 4 -1 I 3 -2 4 2 2 2 2 2 3 -1 1 1 1 1 2 5 1 -3 1 -1 2 1 1 2 1 1 1 Q 1 1 2 -1 1 1 1 1 3 2 I I 3 -2 2 1 1 3 1 2 2 2 1 1 1 2 -I 1 3 -2 3 • 4 -1 1 1 1 1 2 3 r-l 1 1 2 -1 1 Q 1

PAGE 112

100 variable has taken on added significance in the later period. This would suggest that migration to Florida, a state for which climate is the main attraction, would be greater than a forecast based on 1968-72 coefficients would suggest. This is, in fact, the case. The forecast in Table 4.7 underestimated migration to Florida for all states except Ohio. Overall, the equation in Table 4.8 is significant at the .005 level when compared to the model with coefficients set equal to zero. This is a better fit than for the previous four-year period. But when the unconstrained model is compared to one with the right-hand side of the equation set equal to the mean value of the dependent variable, it is significant at only the .10 level. The coefficients for the 1972-76 period were applied to the values of the explanatory variables in 1976. Predicted probabilities were obtained and multiplied by estimates of the total number of migrants leaving each origin state for the eight destination states between 1976 and 1980. The estimates of the total pool of migrants were derived by taking the number of migrants forecast to leave each state (from Chapter 3) and subtracting out the proportion expected to choose states other than the eight alternatives. As before, the proportion choosing other alternatives is assumed constant from the previous period. The destinationchoice forecast is presented in Table 4.9. The same procedure outlined above was used to distribute the 197680 aggregate forecast for each origin state among the destinations. For each state, the aggregate forecast obtained from the decision-to-move analysis is first adjusted by subtracting out the number expected to choose other alternative states. The remaining number is then mul-

PAGE 113

101 Table 4.8 Logit Results: Destination Choice, 1972-1976 Variable Coefficient Asymptotic t-Ratio Distance -.0004 -1.55 Tempdiff .181 3.67 Incdiff .001 1.59 Summary Statistics: Log of Likelihood Function = -85.1 Likelihood Statistic (at 0) = 17.9 Level of Significance (at 0) = .005 Likelihood Statistic (at mean) =4.7 Level of Significance (at mean) = .10

PAGE 114

102 Table 4.9 Forecast: Destination Choice, 1976-1980 OriginNumber of MigrantsNumber of MigrantsDestination Panel Study Aggregate Labor Force New York-Florida 4 331,000 New York-California 3 252,000 New York-New Jersey3 199,000 New York-Illinois 2 146,000 New York-Virginia 2 132,000 New York-Michigan 2 119,000 New York-Texas 2 159,000 Pennsylvania-Florida 3 148,000 Pennsylvania-California 2 108,000 Pennsylvania-New Jersey 2 87,000 Pennsylvania-New York 2 87,000 Pennsylvania-Illinois 1 67,000 Pennsylvania-Virginia 1 54,000 Pennsylvania -Michigan 1 54,000 Pennsylvania-Texas 1 67,000 Ohio-Florida 3 114,000 Ohio-California 2 103,000 Ohio-New Jersey 1 60,000 Ohio -New York 1 60,000 Ohio-Illinois 1 60,000 Ohio-Virginia I 43,000 Ohio-Michigan 1 43,000 Ohio-Texas 1 60,000 South Carolina-Florida 4 59,000 South Carolina-California 3 41,000 ; South Carolina-New Jersey 2 25,000 South Carolina-New York 1 23,000 South Carolina-Illinois 1 20,000 South Carolina-Virginia 1 18,000 1 South Carolina-iyiLchigan 1 16,000 1 South Carolina-Texas 2 25,000

PAGE 115

103 tlplied by the 1976-80 destination-choice forecast probabilities. The predicted distribution of migrants appears in Table 4.9. The historical aggregate out-migration estimates were also distributed out among the eight destinations, utilizing predicted probabilities from the estimated models. These flows, along with the 1976-80 aggregate forecast, appear in Table 4.10. As was done with the aggregate outmigration forecast from each state, the 1968-72 destination-choice estimates are compared with 1965-70 Census estimated flows between states. In Chapter 3 the observation was made that, in total, outmigration based upon Panel Study estimates was higher than Census estimates despite the facts that the Census is not restricted to labor force members and that it measures five-year flows. Differences were partially attributed to Census undercounts and higher rates of migration in the early 1970s. Remaining differences are attributed to weaknesses in the methodology employed in the study. Table 4.11 presents a comparison of Census estimates to Panel Study based estimates. Since we already know that the absolute number of outmigrants from each state (with the exception of Ohio) are estimated to be higher in this study than they are in the 1970 Census, it will be better to compare the relative proportions of migrants who choose alternative destinations. Notice that the Census estimates larger proportions of migrants from all states choosing Florida and California as destinations than this study estimates. Part of this discrepancy may arise because of time-period differences. It is believed, however, that greater reliance should be placed upon the Census estimates. In this study the small sample size of the destination-choice analysis reduces

PAGE 116

104 o 00 0^ I CTN to u C crt J-i o Mcoi •H CT^ S i-H MH O O r^ a\ M T-^ OJ x> B D 13 o o o o o o o o o o o o o o o o o o o o o oooooooo oooooooo oooooooo oooooooo oooooooo oooooooo oooooooo oooooooo oooooooo — I CN CTs VO CN CT\ a^ ro Lo CTi -^ LTi --i I 00 w •u C ni h VO bOl^l •H C7^ S I— 1 M-l CM r^ (^ H t— 1 QJ Xi IZ) S o o o o o o o o o o o o o o o o o o o o o oooooooo oooooooo oooooooo oooooooo oooooooo oooooooo oooooooo oooooooo oooooooo CO CO CO CM r^ O O O CO t-H VOCr\COr-i00r-l ~d-o^^^O^LO^DCoa^ LOOojt^inOOO a^cOCMCM'-^c^.— ICO o m •H u O C ,a CO o U CM Mr^ c •H CJ\ o S •-< H 1 •U 14-1 00 ctf ^ a r— I CO LO CM CO r^ 1^ tM •— I CO Csl i-H ,— I rLTi 00 — I r^ r^ -i d tn ^ tn cd d •H QJ CO !-i yi M •H •H cd d tn ,i4 en CO d 13 QJ d 60 cd n M i-i •H •H cd •H M-i d t>-i d H H tn cd >-. X) QJ d 60 U •H •H 60 x; cd •H 0) •H U-l i-J >-< d H -H tn iH :3 13 H M X C tn CO Cd C H •H •H 60 4:: cd rH cd QJ OJ .H •H •H QJ Cd j-i M •H •H CO ^ IS S tH !m a >< fe s 3 H > S H Xt OJ d 60 iH cd 01 QJ .H H -H Q) Cd >^ 1 1 1 1 1 1 1 1 •rl 14-1 1-5 c •H •H tfl IXH u ;z; S M > s H H QJ cd cd cd cd cd cd cd cd U •H •H W) ^ CO 1 1 I 1 1 1 1 i d tn ,^ tn cd d d d d d d d d d M S r-i ^ X CO cd CO CO cd Cd cd cd cd M U H •H •H cd •H •H H •H •H •H •H •H C H cd QJ H H •H QJ •H •H •H •H •H •H -H •H Td 0) d 60 rH rH tH iH tH .H ^ M P4 u S M > S H C d d d d d d d •H m o l>^ d •H •H tn 1 •H 1 1 1 1 1 1 1 CO CO CO CO cd cd cd cd V4 •H •H 60 rd cd U !-l -}-l ^4 u M M >-i c 4.) M ^ v^ r^ > > > > > > > i—i 3s rH u X cd Cd Cd cd cd cd Cd CT3 H Cd !-i !-i )-l >-l u S-i f-i tH iH ^ <-t iH iH rH M ^ cd QJ QJ M •H •H QJ CJ u u u u M C >^ >> >, >. >^ >^ >, >^ flJ zj 3 :z; M > S H H •T-l >H >< >-l !5-t >l >-• >-i cn CO tn tn W to m tn i 1 1 1 1 1 1 4:: r-! x: 4:: 42 x x: X S-l 4-1 d d d d d d d d 4J 4-J 4-) 4-1 4-J 4-1 4-1 4-) 0} S & 5 is S :s [3 d d d d d d d d •H •H H •H •H •H •H •H 3 d d d d d d 3 OJ 0) OJ QJ QJ 0) CU QJ 0) QJ QJ QJ 0) 0) QJ QJ rXi ^ j:: 42 4:: f— r; ia E3 S S s 53 J3 Pm Ph Pm fl4 PM P^ PL, fi. CO CO CO CO CO CO CO CO

PAGE 117

105 0) w CO a 0) o IT\ a^ rO lO ^ -cf -— 1.— lO-— lO CNjr— !Oi— lOrlO' — -<'X)OOr-ICO.— i-cjCOCNi— I — I CO O ^o m O O uo 0-) O o en ^cocsiLocyiromcNi L^lO>—^r^00'-^a^o cr\ H !-i 01 O t>. 3 cu Oi o d o m •H M cd 6 o u 0^ I— 1 1 00 QJ 4-1 e •H U CD W Td VO o CM "—I CO vD .-I O r^ O 00 CN r~~ O CNl CM 1— I O O CMOLOi— ir-^r-^cM^ t— It-HrHCN'— fO' — ^O f-H CM <|00 C3^ '^ rH O O O O O O O O O O O O O O O O O o o o o o o o oooooooo oooooooo oooooooo oooooooo oooooooo oooooooo oooooooo oooooooo oooooooo r-H CO in CNj CO rr^ CN — I CO CM T-H "-H OO-CJ-cOCOCO-^vDCO cocNlooa^a^^Dco•— I vOLor-~ococo^oco c^ CM CO o r^ CT\ i-H CO — I "-H H cd t>^ cd •H d cu CO Ai CD cd d -H OJ cd n u M •H •H cd C CD ^ CD CO d 13 o cu o O d M CO ^-1 U ^-1 •H •H CO •H I+-I *-) PH d H •H CD cfl >> t3 o 0) o d M M •rH •H M) 4:: Cd d H 111 •H M-l i-n I>-| d •H -H CD n rH [3 [i rH M c) X u C M CO cfl d S-l •H •H M x; cd iH Cd a> fl) rH 'H H 0) H ttj !-i !-i H •H CO o .H 13 S: iH 5-1 cj X h C3 y, P^ H > a H J-) 'O U (U o d w; rH cO <1J tu iH •H -H (1) CO >. 1 1 1 1 1 1 1 1 03 •rH 4-1 >-) d H •H Cfi l^ U iz; s M > S H •H CU cd cd CO Cd CO CO cd CO a J-J •H H bJJ r* Cd 1 1 1 1 1 1 1 1 d CD AJ CD td d d d d d d d d d •H O tH is iH !-i O CO cd cO cO Cd CO cd CO CO 5^ U )-l •H •H cd •H •H •H •H •H •H •H •H 4-) .H d (i) M •^ •H OJ •H H H •H •H •H -H •H Tl o dj n n d tin rH iH rH rH <-\ rH rH rH C/J f^ u l=S M > ^ H d d d d d d d d •H U-l i-l !>| d •H •H CD n n n n n n n n 0) 1 1 1 1 1 1 1 CO (d cd CO CO CO CO CO ^1 •H •H tC JId Cd n !-l "H !-i !h u M M H -i"! ,^ ^ ^ ^ ^ ^ > > > > > > > > o .— 1 3: i'3 iH !-i C) 1^ cd CO cd cd cd cd cd cd 1 5-1 u u ;-! H U !-l .-H M M t-H <-i H .-1 iH rH CO CU CD tH •H •rH (11 C ) c; C) c } c > c ) c 3 C ) C O o o o O >^ >. >. >. >-, >, >, >. l-M u 2: a M > •?; H •H >i ;>^ l!-l '^ >^ !>-i >-i en tn CD CD CD CD CD CD 1 1 1 1 1 1 1 ,d jd ^ ji: x: ,^ X ,d bL d d d d d d d d n o o n n n n O j-i 4J j-j 4-J 4-1 4-1 4-1 4J H U is '^ •s 15 S d d d d d d d d •H H H •H •H •rH H •H d d d 3 d d d d )-i OJ CU OJ OJ 0) CU (U OJ CIJ (U OJ cu cu cu OJ ,d ^ r1 ^G ^ ,d ^ ,d o n n O n n <_) S '^ k; k; i4 '^ s Ph CM p-i PM p-i PM fM fM o O o O o o O O CO C/D 00 CO CO CO CO CO

PAGE 118

106 the viability of the results. The forecasting methodology is based on the assumption that sample proportions represent population proportions. When sample size is small the probability of staisfying this assumption is low. This is particularly true when the number of choices is high, such as in the destination-choice equations. The number of observations for each alternative ranged from 25 (Florida) to 4 (Texas). A final hypothesis is set forth for explaining differences between Census destination proportions and Panel Study based estimates. The Census includes the entire population regardless of labor force status. People who are not labor force members may have higher rates of migration to states such as Florida or California. Retired people, for example, have a high propensity to migrate to Sunbelt states. This group raises the rate of migration to these states. Since the Panel Study estimates are restricted to labor force members, lower rates of migration to retirement states are expected. For the 1976-80 period in-migration to Florida from the four-state sample is projected to be 652,000. This is in line with current expectations that migration to Florida will remain healthy but will not be able to maintain recent historical peaks. Florida's major competitor, California, is forecast to receive 504,000 new migrants between 1976 and 1980. One of its newest competitors, the state of Texas, will gain 311,000 in-migrants during the period. The poor fits obtained in the shorter time period destinationchoice equations lead to some skepticism about outright acceptance of the forecast and estimates put forth above. Even with this limitation, however, these figures appear reasonable and are in line with historical standards and current expectations of the future.

PAGE 119

107 Since climate and income have turned out to be such important variables in the destination-choice decision, they deserve some special attention. In the final section of this chapter, a closer look at these variables over time will be provided. Income, Climate, and Migration Earlier it was discovered that climate became a more important determinant of destination choice during the 1972-76 period than it was for the 1968-72 period. It was also demonstrated that migration to such warm weather states as Florida, California, and Texas reached its maximum during the period 1972-76. Table 4.12 presents the actual average temperature difference and per-capita income difference data for the three points in time used in the previous analysis. It can be observed that temperature differences between sending and receiving regions are generally at their maximum in 1972. These larger differences are not the result of lower temperatures at origin states during that year. Rather, the temperatures in warm weather destination regions were somewhat higher than usual. Further, some of the northern receiving regions such as Michigan and Illinois had unusually cold winters in 1972. Thus, migration to Sunbelt states increased relative to northern competing regions. This increase in relative attractiveness combined with the substantially larger number of out -migrants from sending states, resulted in larger gains for Florida, California, and Texas. By 1976, temperature differences declined to near normal levels. This contribiited to n forecast of somowhat roiflijeed atttativcneacs of Sunbelt states relative to northern receiving regions.

PAGE 120

108 M u nS iH H O T3 N.-> ^o r^ 0) o a .— 1 cu M (U IW 14-1 •H (>j P C3^ OJ .— f 6 o a ts H cti CO +J o •H C3^ ^ a I— t r~ct) u VO >— 1 c r^ • cN X) r~~. o> vo ^ rn csj o IT) CT\ --H C>0 CO — I 00 I III .— I r^ en CO i-H CO rH CT^ o^ i>0 .— I CO — I t-H o '-4 I I I —I I I CM rMCNj COvO00^C3>'— iLTiCM I I I c?^l^CJ^CMa^CJ^OCM Lnu-ir-.r~-r-.cor^o CMlOvDCXJvOt— icocyi CMuOuOr^LOCOCMCO I I I •< I I CMO>i— lOr— ILOOvO co~^^oo^r\iro>0 — I tN CM 1— I O ^11 I cr\ un I CM CM CM UO I I — I LO I CM CM CO >^ -H I I IT) I I CO — I en vo I I -* CO CM CM LO LO 00 CO 00 >— I I I t I I OcDcOLOojr-.^ OOcooinojr^cM-HO CJ^L^.— 1.— ICMr-HLO CJ^Lni-H III I CO LO -^r u-1 CM ^ I I I I I CU en en ca I I U !-i O O >-i ;>^ cd d •M cd H -H en &c ^ ed ^^ o X •H M-l cd u I tu I I o ^-1 o cd Cd •H -H C C cd cd > > >1 cu en ^ cu o CU CU I I cd cd H -H c d cd cd > > en cd H C H H 00 H !-i rH -H M > a I I I c cd 60 •H en ^ Cd a Jk; cd •H a u o H cu en ^ S-l ^-1 cu o cd cd H -H a d cd cd > > eu H I cd cd H -H d d cd cd > > cd cu en ^ •H m S 3 & S 3 & S (U (U cu cu cu cu QJ t3 E5 S S 2 S S cncneneocnenenen dddddddd dddddddd cUcUcUCUQJCUCUCU I o •H O cd 0) (U en cd d •H -1-1 cd d tio d -H -H H M ji; ca iH !-i O X iH -t-l -H M > S 1 I I o o o •H -H -H ^ ^ ji: ^ o o o o .H cd cu eu Pm o s s I I I I cd cd cd ca en cd d •H -H cd o d 00 •H -H tJO ^ r-( ^j a X d d •H -H d d •H H -H H > S H I I I Cd cd cd Cd d d d d H -H -H en cu H I o oooooooo cdcdtdcdcdcdcdcd 4-I4-I4-I+J-U4-1.U4-1 ddddpsdd oooooooo C/De/2tylCOC/)WCOCO

PAGE 121

109 The per-capita income data generally paint a picture of persistence of state income inequality over time. Despite considerable levels of out-migration from the four origin states and large numbers of inmigrants arriving at the eight destination states, the magnitude of state income differences has been maintained over time. Thus, the thesis that migration leads to convergence of regional per-capita income levels is not supported here. Despite cold weather, positive income gains available in such states as Illinois, Michigan, New Jersey, and New York lead to continued in-migration to these destinations. With the exception of migrants originating from South Carolina, those arriving in Florida, Texas, and Virginia can generally expect to experience declines in income but at the same time see improvements in climate. Thus, support is generated for the thesis of Graves (1979) that markets adjust to leave utility constant over space and that migration involves a tradeoff between income and climate-related amenities. Migration to California, on the other hand, leads to greater income and better climate, although both increases are usually moderate. Those leaving South Carolina can expect to receive substantial income increases, whether moving to a warmer or colder climate. Earlier regression results showed that although both income and climate work together in determining migration flows between states, it appears that each changes in relative importance at different points in time. I-Jhenever a factor changes in significance, the distribution of migrants by destination can be altered. In the final chapter, the findings of the study will be reviewed and conclusions will be drawn.

PAGE 122

CHAPTER 5 SUMt'IARY AND CONCLUSIONS Introduction In this study a model of interstate labor force migration has been developed and tested. First, decision-to-move models were estimated for four states which have recently experienced out-migration. Migrants from each state were found to have been selected by a different mix of characteristics and circumstances. The out-migrants from all four states were grouped together and a destination-choice model was estimated. Seventy percent of the movers were found to have chosen among eight alternative states. The factors which influenced this decision were determined. Employment and cost-of-living variables were found to be of little importance. Distance, income, and climate interacted to determine location choice. Thus, support was generated for the Graves (1979) hypothesis of constant spatial utility. Recent individual models of migration, such as those carried out by DaVanzo (1977) and Falaris (1978), have analyzed the choice among broad United States regions. I^^hen areas are defined In this way, measures of regional characteristics reflect averages of all states across the region. As such, these measures hide many locatlonal differences which may enter into the migration decision. As an example, Florida and Georgia are both located in the southeast region of the country, yet they are certainly distinct enough to be considered as Independent 110

PAGE 123

Ill alternatives to a potential migrant. For this reason, the state has been chosen as the level of locatlonal analysis in this study. An argument could be made for defining regions at an even smaller level, such as the urban area, but this would increase the number of alternatives that would need to be included in a thorough analysis. Technical and cost limitations prohibit this procedure. Utilizing the results of both the decision-to-move analysis and the destination-choice model, an aggregate migration forecast was obtained for each origin and destination combination. Tests of forecasting accuracy revealed mixed outcomes. The total number of out -migrants from the four-state sample was predicted reliably. But because of a small sample and poorly fitting, short time period equation the choice of destination was not accurately predicted. More specific findings of the study are reviewed in the following sections. Review of Findings: Decision to Move The results of the decision-to-move analysis are summarized by state in Table 5.1. For the state of Ohio, marriage deters migration. Mincer's (1978) finding that the presence of a spouse represents a tie to present location is thus supported. Migrants from this state are also likely to have higher family incomes. This result supports the viewpoint that higher income families are more likely to demand goods which can only be obtained by changing locations. Finally, Ohio migrants are likely to be members of households headed by males. As with Ohio migrants. New Yorkers with greater Incomes are more likely to change states of residence. Unemployed persons in NewYork are more likely to move than those who are employed, supporting the

PAGE 124

112 Table 5.1 Summary of Findings: Decision to Move Variable Ohio New York South Carolina Pennsylvania Famsz + +* + Sex +* + + + Homeown + + Lres + + + + Lempl + _A + Emself + Avage + +* +* + Emst • _* • Marst -k -A Race + + Aveduc + + +* Prevmig + + + Emstw + +ft + Famy +* + +* _* Significant at 95 percent confidence level or better. + Positive coefficient. Negative coefficient. • Not in equation.

PAGE 125

113 thesis that the unemployed are more responsive to migration opportunities. This, it is recalled, was one of the major conclusions of the DaVanzo (1977) study. Migrants from New York are also less likely to have held their most recent job for a long period of time. This variable acts as an indicator of attachment to present environment and job stability. Migrants from New York are generally older than nonmigrants. In most studies, migration rates are found to be lower in older age groups, except at retirement. The continual presence of positive coefficients in Florida-sending states equations, however, has led to the development of an alternative hypothesis. It is believed that many older migrants are moving in anticipation of retirement at a later date. The positive and significant sign on family income complements this theory well. Those with higher incomes are certainly likely to settle at their planned retirement location at an early age. This is especially true if moving involves a cut in pay or acceptance of a part-time job. Finally, household heads from New York with employed spouses are more likely to leave the state than those whose spouses do not work. Earlier it was theorized that employed wives deter migration of families. This is because any new location is less likely to present optimal employment and income opportunities to both spouses than it is for just one of them. The positive and significant coefficient on this variable, however, indicates that increasing female labor force participation could lead to more migration from New York. Many spouses who are employed prior to migration may no longer be employed after the move. Mincer (1978) found evidence that one family member frequently dominates the migration decision. Thus, spouses become tied to this decision regardless of the personal consequences.

PAGE 126

114 South Carolina migrants also tend to be older and earn higher incomes than nonmigrants from that state. Married persons are less likely to move than those who are not married. Finally, migrant families tend to be of larger size than families that remain in South Carolina. Household size, a priori, was expected to act as a constraint upon migration. Education proved to be an important determinant of migration in Pennsylvania, with more educated people being more likely to leave. But migrants from the state were also likely to earn lower family income than those who remained. For New York, Ohio, and South Carolina, the evidence confirmed the view that increases in income increased the demand for location-specific goods found in such states as Florida or California. For Pennsylvania, it appears that greater income increases the probability of remaining in that state. The positive Income effects found for three of the four states conflicts with DaVanzo's (1977) finding that the migration process, or the returns to it, are on balance an inferior good. The difference in results may arise because this study is only considering states which have typically sent migrants to Sunbelt states. Sunbelt migrants typically move in search of increased leisure activities, the demand for which is expected to increase with income. The additional discovery that migrants from these states are usually older than nonmigrants adds further credence to this theory. It is the older migrant planning for future retirement who is likely to be seeking out more temperate climates. Those who have in the past earned higher incomes have the greatest ability to make such long distance, and sometimes income sacrificing, moves.

PAGE 127

115 Over time the migration determinant coefficients did not appear to be very stable. But, over shorter time periods, there are fewer migrants available for study. In addition, fewer variables could be considered in logit estimation, leading to the possibility of misspecified equations. A decline in the total number of out-migrants is forecast for the 197680 period. This is in line with previous expectations that the migration surge of the early 1970s cannot be maintained. Review of Findings: Destination Choice All other things constant, migrants from Ohio, Pennsylvania, South Carolina, and New York chose to move to states offering greater percapita income. This result, at first, seems to offer support to regional growth theories, such as that of Smith (1975), which portray migration solely as a response to income or wage differentials between regions. But it was discovered that income is only significant when a variable reflecting differences in climate is included in the model. Climate turned out to be the most significant determinant of destination choice. The downward bias on the income coefficient when climate was excluded from the model indicates a strong negative correlation between income and climate. Support is thus lent to the Graves (1979) thesis that long-distance migration involves a trade-off between income and location-specific, climate-related amenities. When destination-choice models were estimated for shorter time periods, coefficient instability was discovered. For the 1972-76 period, climate was found to be a more important determinant of location choice than during the 1968-72 period. Income,^ on the other hand.

PAGE 128

116 lost significance. Migration to Sunbelt states also showed a large increase during the later period. A possible conclusion is that the demand for climate related amenities increased during the whole eightyear period, creating a disequilibrium situation, which led to greater levels of migration to warm weather states such as Florida, California, and Texas. A forecast of the distribution of migrants from the sample sending states was developed for the 1976-80 period. This forecast was based upon the distance, income, and climate relationship during the previous four-year period. Thus, climate is the primary determinant in this forecast, promising strong future levels of migration to Sunbelt states. To the extent, however, that earlier period migration levels were a response to a disequilibrium situation, the forecast may be overstated. If migration during the 1972-76 period, in fact, moved the economy toward a new spatial equilibrium, then that rate of sunbelt migration would be expected to slow during the adjustment process. New exogenous shocks could, however, lead to a continuation of current migration patterns. Given the assumption of the continued importance of climate, the 1976-80 forecast appears to be reasonable. l-Jhile maintaining strong rates of in-migration, Florida is predicted to lose some ground to its chief competitor, the state of California. Strengths and Weaknesses of the Study This has been the first study of individual migration that analyzes the decision to move at the state level. Since a different set of determinants was discovered for each state considered, this approach has proven fruitful. Studies which have grouped all states together may

PAGE 129

117 be hiding these differences. Ambiguous coefficients which appear in these studies may be the result of aggregation. Destination choice was also analyzed at the state level. Studies that have combined states into larger alternative regions are believed to cover up locational differences which the migrant would be likely to consider in deciding where to move. This may explain why economic variables rarely turn out to be significant in these studies. This has also been the first individual migration model which has been used for the purpose of forecasting. While tests of accuracy showed the decision-to-move prediction to be reliable, less success was obtained in forecasting destination choice. This was partly because of a small sample size which caused the number of individuals who chose any given alternative to be low. Poor fits for equations estimated for short time periods added to this problem. In addition, evidence of coefficient instability was discovered. The use of a nine-year time frame for studying the migration decision is the source of potential problems. Migration, in fact, took place during each year of the period, yet the independent variables were measured at the initial year. I'Jhile the choice of a shorter time period would have relieved this problem, other larger problems would result. Fewer people migrate during shorter periods of time. This makes the determination of variables which distinguish migrants from nonmigrants more difficult. In addition, with fewer numbers of migrants, some variables must be excluded for successful logit estimation. As stated earlier in the paper, it is not believed that the de. cision "whether to move" is always made independently of the decision "where to move." A methodology which analyzes both decisions simul-

PAGE 130

118 taneously would thus seem preferable to the two-stage methodology employed here. But, using existing techniques, trying to estimate such a model over a large sample while still considering many alternatives is extremely expensive. In addition, since most individuals do not move during any reasonably defined period, the observations in such a model are heavily weighted toward the staying alternative. The current state of the art in choice theory does not seem to possess the proper tools for the analysis of decisions where the alternatives are so unbalanced. Implications of the Study The influence of age and family income on the decision to move and the interaction of income and climate in determining destination choice can be tied together in a consistent fashion. This study presents a picture of an older than average person with relatively high income who chooses to move, with his (or her) family to a state with a more temperate climate. Possibly this person is moving in anticipation of retirement. In addition, this move involves a significant probability of a reduction in income. Another equivalent interpretation is that older people with higher incomes have a greater demand for locationspecific amenities than do other members of the labor force. The aging of the "baby boom" generation combined with the recent decline in the birth rate of the United States implies that an increasing absolute number and proportion of the population will be in older age groups. This occurrence should lead to an increase in the number of persons who are at risk to the process described above. The implications for amenity-rich states such as Florida, California, and Texas are startling. Population growth will continue in these states at perhaps even higher rates than have so far occurred. And this increase will .continue to be concentrated in older age groups.

PAGE 131

119 The theoretical model and methodology employed in this study can easily be applied to other states. The results obtained could then be compared to the findings for the four states considered here. Including more states in the decision-to-move analysis would also provide a larger sample for the analysis of destination choice. Better fitting models might thus be obtained and forecasting accuracy may be improved. Ultimately, detailed forecasts of labor force migration between all states could be developed. Although the focus of this study has been upon labor force migration, the model developed is general enough to be applied to other demographic groups. The study of, retiree migration would, for instance, be an important application. Factors that influence the decision to move and destination choice of the elderly population could be determined. Ultimately, a forecast of the number of older people who will move to retirement states such as Florida, California, Arizona, and Texas could be developed. This forecast could then be combined with labor force migration predictions for these states. Forecasts of these two components of population growth should aid in projecting state population figures and in planning for economic growth. The forecasts can also be used as input in projecting other economic activities at the state level. State econometric models, for example, have proven to be very sensitive to the assumptions made about population growth. This is especially true for states which are experiencing rapid rates of migration. It is believed that the findings of this study have added to the understanding of the migration process. It is hoped that this research has provided theoretical, methodological, and empirical contributions that will be of value to others who are interested in the study of migration.

PAGE 132

APPENDIX A ALTERNATIVE MODEL FORl-IULATIONS : DECISION TO MOVE Combined State Results Table A.l contains the results of the decision-to-move analysis when all four origin states are combined into a single sample. Ordinary least squares estimates appear in the table. A problem with available storage space prevented successful execution of the logit program for a sample this large. The variables are defined as they were in Chapter 3. Variables that are significant at better than the .05 level are sex, average age, family income, and marital status. These variables are the same ones that often turn up significant at the state level. But other factors, which only occasionally appeared to be important for separate states, no longer are significant. Examples are average j education (significant in Pennsylvania) and spouse's employment status (significant in New York). In addition, variables that have opposite signs for different states become averaged into the combined state coefficients. For example, in Chapter 3 it was found that Pennsylvania migrants are negatively selected according to income. The positive and significant coefficient on combined state family income hides I this discovery. 120

PAGE 133

121 Table A. 1 Combined State Results: Decision to Move, 1968-1977, Ordinary Least Squares Variable Coefficient Famsz .010 Sex ,278 Race -.016 Homeown .005 Lempl -.014 Emself -.047 Avage .005 Aveduc .018 Ernst -.120 Famy ,005 Prevmig .024 Emstw .041 Marst -.275 Lres .011 Constant -.148 R^ = .06 t-Ratio 1 .54 3 .96 .52 .18 -1 .52 .78 3 .38 1 .85 1 .47 2 23 81 -1 47 -4. 01 1. 00 -1.17

PAGE 134

122. Finally, the low R indicates that only 6 percent of the variation of the migration decision is being explained by the regression equation. Thus, the model does not fit well for all four states. Annual Migration Results In Table A. 2 ordinary least squares results appear for the decision-to-move model when the sample period is defined as the 1976-77 period. The logit program could not be used here either. A requirement for logit estimation is that each explanatory factor display variation within every category. Since there were so few movers during a one-year period, most of the variables did not meet this criterion within the migrant category. The results indicate that very little of the variation in annual migration is explained by the factors in the model. Only in the Ohio equation does more than one variable appear to be significant. The New York and Pennsylvania models contain no significant variables. One possible conclusion is that the wrong set of explanatory variables is used in the model and that it should be respecified. Instead, it is concluded that the proportion of migrants in the sample during a one-year time period is too small to permit any set of variables to properly distinguish between migrants and nonmigrants. Longer time periods should thus be chosen. • Redifinition of Qualitative Explanatory Variables Table A. 3 contains the results of the decision-to-move analysis when the length of residence, length of employment, and education variables are redefined to be two-category dummy variables.

PAGE 135

123 cd l' •H P c crt > 4-J tH c >1 OJ en •H C o C •H OJ M-l P-i m CU o u O o VX) ^ cr^ o^l cn cr\ \o en t-^ r— t CTv -3cn en t— t r-^ CvJ Pi CM CO en m C7^ en u~) f-H in 00 en yD CN r^ CO CNl I — 1 a^ 1 — 1 T— ) OJ VD 1 — 1 in i^ 00 r^ CNl r^ o o CNl •-H CO LO 00 CM o •H M-l o o o o O 1-H o o o o o o o o 1 1 1 1 1 1 I • • < M-l 0) tS ^ n) H oi N W B X CO a o 0) o ft e CU en B W CO o CU > < en B W B cd •H 0) PM en 4J Ci Crt 4-1 4J cn en JH C cd n s u

PAGE 136

124 I 00 crt fii nj 1 H 4J S crt > j-i iH C P^ OJ W •H c o C •H 0) M-l P-i m CU o u o o O (>0 U-1 CO o 1 o CT^ r-H CN r-H 1 i-H (—1 CM 1 CM 1 CO o^ r-H I— 1 CJ^ <^ CM CO C3^ ! K 4_1 CO CO b X o e 0) 6 OT CO QJ CO B (11 cn u o o TO 0) ^ o u 0) B > > B CO )H g fe C/l Ph w hJ hJ w < <1 fS fe P^ [S s

PAGE 137

125 > W I O O 00 rH VO m o LO 6~^ B-S • • o • o O^ U~l T-H lO o i-H o CO CNl 1 1^ • >ds c •H u a o o < 0) o -0^ o IT) B^ B-2 • • o • O -dr-ro -ao 00 o 00 1 .—1 • ro H o 00 oITl •H • o ^ ro l^ o O CO CO 00 LH &^ o LO o CO crs !>^ /-"v M z**^ C T) 4-1 c 03 0) a / — s 03 OJ 4-1 0) ''^ o Q) V u C O e H ^1 O 4-1 4-1 o •H 4-1 nJ 4-1 03 OJ CJ 4-1 03 ^— -* 03 S-i O ^— ^ PM m d Q) Q) 4J 3 O O O a >^ C h •H C •H C rH 03 tfl 4-1 cti 4J 03 U M a T3 m a m o o 60 •H O •H •H •H H 01 •H 4-1 O 4J M-l 4-1 4-1 u Cfl ^ CO •H 03 •H ^-1 ^^ H •H 4-1 C 4-1 C o M-l 4-J rH w 60 CO bO u o rt OJ •H H t3 u AJ Td CO T3 CO 0) 0) 0) on -H o M 60 4-1 1-1 o m O m 03 03 CJ >^ j3 o ^ o 4-1 4-1 -r-l u m H •H c C t3 nj o M .-H .H i-i 0) QJ 01 ^ Q) 0) 0) OJ CJ O >-i p M ^ > ky] > !-l f-l p-l 3 O H 0) •H 0) 0) QJ CO i-l hJ 1-1 1-1 1-1 FL, P^

PAGE 138

126 The length of residence variable is assigned a value of 1 for those living in their current house or apartment for three or more years. It is assigned a value of for those living in their homes for less than three years. The variable is significant for the states of New York and Pennsylvania. T-/hen defined as a multi-category variable, length of residence was insignificant for all states. For the Pennsylvania observations, the variable now takes on the expected (negative) sign, while for New York the sign is positive. The length of employment variable is assigned the value of 1 for those employed at the same place for 18 months or more. Those at the job for less than 18 months are given the value of 0. The variable appears to be negative and significant for New York and South Carolina. For New York this result is the same as that which occurred when the variable was alternatively defined. For South Carolina it takes on greater significance than under the previous formulation. For the states of Ohio and Pennsylvania, the variable is insignificant under both regions. The education variable is assigned the value of 1 for those with 12 or more years of education, while those with less than 12 years of education are given the value of 0. The variable is insignificant for all states. When education was defined under the old formulation it was positive and significant for the state of Pennsylvania and insignificant for all other states. For the length of residence and length of employment variables, redefinition has brought the results slightly closer to expectation. The education variable was brought further from its expected effect as a result of the revision. Overall, the equations are changed a little

PAGE 139

127 by the redefinitions. The Ohio and South Carolina equations provide a somewhat better fit as a result of the change while the New York and Pennsylvania equations fit slightly poorer than with the formulation used in Chapter 3.

PAGE 140

APPENDIX B ALTERNATIVE MODEL FORMULATIONS: DESTINATION CHOICE Nominal Wage Model Table B.l presents results obtained when a measure of individual returns to migration is included as an explanatory variable in the model of destination choice. Wages at each location are estimated as a function of age, race, sex, marital status, education, union membership, experience, and occupation. Using the procedures described in Chapter 2, values of potential wages at each origin and destination are estimated for every person in the sample. For each individual, the variable Wdiff is defined as the difference between his (or her) potential 1977 hourly wage at each possible destination and the potential wage at his (or her) origin location. The remaining variables are defined as they were in Chapter 4. In this model, only climate (Tempdiff) turns out to be a significant determinant of destination choice. Both distance and potential wage gains turn out insignificant. Overall, the model exhibits a poorer fit than that obtained when the individual returns variable is replaced with the average aggregate per-caplta income gain variable used in Chapter 4. Real Wage Model The wage-difference variable defined in the previous section was deflated by state cost-of-living indices to obtain a new variable 128

PAGE 141

129 Table B.l Nominal Wage Model: Destination Choice, 1968-1977 Variable Coefficient Asymptotic t-Ratio Distance -.264 -1.33 Tempdlff .111 3.47 Wdiff .157 1.67 Summary Statistics: Log of Likelihood Function = 168.4 Likelihood Statistic (at 0) = 11.6 Level of Significance (at 0) = .01 Likelihood Statistic (at mean) =2.6 Level of Significance (at mean) = (not significant at any acceptable level)

PAGE 142

130 called Rwdlff. Even when measured in real terms, the individual returns variable remained insignificant. The overall fit of the model remained poor. These results appear in Table B.2.

PAGE 143

131 Table B.2 Real Wage Model: Destination Choice, 1968-1977 Variable Coefficient Asymptotic t-Ratio Distance -.265 I.33 Tempdiff .110 3.47 RWdiff ,163 1.64 Summary Statistics: Log of Likelihood Function = 168.4 Likelihood Statistic (at 0) = 11.6 Level of Significance (at 0) = .01 Likelihood Statistic (at mean) =2.6 Level of Significance (at mean) = (not significant at any acceptable level)

PAGE 144

BIBLIOGRAPHY Albright, L. Lerman, S.R., and Manski, C.F., "An Introduction to the Multinominal Problt Model," Cambridge Systematics Inc. Technical Report, Cambridge Systematics Inc., August 1977. American Chamber of Commerce Researchers Association, Cost of Living Indicators: Inter-City Index Report American Chamber of Commerce Researchers Association, March 1968. Borts, G.H. and Stein, J.L., Economic Growth in a Free Market Columbia University Press, 1964. Burns, J.F. and James, M.K. "Migration Into Florida: 1940-1973," Work Paper No. 4, Urban and Regional Development Center, University of Florida, October 1973. Chow, G.C., "Tests of Equality Between Sets of Coefficients in Two Linear Regressions," Econometrica Vol. 28 (1960), pp. 591-605. Dalton, J. A. and Ford, E.J., Jr., "Concentration and Labor Earnings in Manufacturing and Utilities," Industrial and Labor Relations Review Vol. 31 (October 1977), pp. 45-60. DaVanzo, J., I-Jhy Families Move R&D Monograph 48, U.S. Department of Labor, Empl03nnent and Training Administration, U.S. Government Printing Office, 1977. Dunlevy, J. A. and Gemery, H.A. "The Role of Migrant Stock and Lagged Migration in the Settlement Patterns of Nineteenth Century Inmigrants," The Review of Economics and Statistics Vol. 59 (May 1977), pp. 137-44. Eckstein, 0. and Girola, I.A. "Long-Term Properties of the Price-Wage Mechanisms in the United States, 1891 to 1977," The Review of Economics and Statistics Vol. 60 (1978), pp. 323-33. Eldridge, H.T., "Primary, Secondary, and Return Migration in the United States, 1955-60," Demography Vol. 2 (1965), pp. 444-55. Engler, S.D., "Labor Force Migration to Florida: 1970 to 1976," The Florida Outlook Vol. 2 (September 1978), pp. 25-30. Falaris, E.M., "Migration: A Study of Choice Among Discrete Alternatives," Seminar paper presented at the University of Florida, October 1978. 132

PAGE 145

133 Graves, P.E., "A Life-Cycle Empirical Analysis of Migration and Climate, by Race," Journal of Urban Economics Vol. 6 (April 1979), pp. 135-47. ~"~ Greenwood, M.J., "An Analysis of the Determinants of Geographic Labor Mobility in the United States," The Review of Economics and Statistics Vol. 51 (May 1969), pp. 189-94. Greenwood, M.J^. "Research on Internal Migration in the United States: A Survey," Journal of Economic Literature Vol. 13 (June 1975) pp. 397-433^ ~~ Johnston, J., Econometric Methods McGraw-Hill, Inc., 1963. Kau, J.B. and Sirmans, C.F., "The Influence of Information Cost and Uncertainty on Migration: A Comparison of Migrant Types," Journal of Regional Science Vol. 17 (April 1977), pp. 89-96. Laber, G. "Lagged Response in the Decision to Migrate: A Comment," Journal of Regional Science Vol. 12 (August 1972), pp. 307-10. Longino, C.F., Jr., ||Going Home: Aged Return Migration in the United States 1965-70," Based on a paper presented at the 31st meeting of the Gerontological Society, 1979. Lowry, I.S., Migration and Metropolitan Growth: Two Analytical Models Chandler Publishing Company, 1966. ~' Maddala, G.S. and Roberts, R.B., "Estimation of Econometric Models Involving Self-Selection," Paper for European Meetings of the Econometric Society in Vienna, September 1977. McFadden, D. "Conditional Logit Analysis of Qualitative Choice Behavior," in Frontiers in Econometrics Zarembka, P. (ed.). Academic Press, 1973, pp. 105-42. McFadden, D. "The Measurement of Urban Travel Demand," Journal of Public Economics Vol. 3 (June 1974), pp. 303-28. \ Mincer, J., "Family Migration Decisions," Journal of P olitical Economy Vol. 86 (October 1978), pp. 749-73. Morrison, P.A. "Theoretical Issues in the Design of Population Mobility Models," Environment and Planning Vol. 5 (1973), pp. 125-34. Muth, R.F. "Differential Growth Among U.S. Cities," in Papers in Quantitative Economics Quirk, J. P. and Zarley, A.M. (eds.). The University Press of Kansas, 1968, pp. 311-55. Nerlove, M. and Press, S.J., "Univariate and Multivariate Log-Linear and Logistic Models," Rand Corporation Technical Report R-1306EDA/NIH, Rand Corporation, December 1973.

PAGE 146

134 Orcutt, G.H. Greenberger, M. Korbel, J., N. Rivlin, A.M., Microanalysis of Socioeconomic Systems: A Simulation Study Harper & Brothers, 1961. Pindyck, R.S. and Rubinfeld, D.L., Econometric Models and Economic Forecasts McGraw-Hill, Inc., 1976, Richardson, H. Regional Growth Theory John Wiley, 1973. Rothenberg, J., "On the Microeconomics of Internal Migration," in Internal Migration: A Comparative Perspective Brown, A. A. and Neuberger, E. (eds.), 1977, pp. 183-205. Smith, D., "Neoclassical Growth Models and Regional Growth in the U.S.," Journal of Regional Science Vol. 15 (August 1975), pp. 165-82. Schwartz, A., "Interpreting the Effect of Distance on Migration," Journal Political Economy Vol. 81 (September/October 1973), pp. 1153-69. U.S. Department of Commerce, Bureau of the Census, Census of Population: 1970, Subject Reports Final Report PC(2)-2E, "Migration Between State Economic Areas," U.S. Government Printing Office, 1972. U.S. Department of Commerce, Bureau of the Census, Current Population Reports Series P-25, No. 701, "Gross Migration by County: 1965-1970," U.S. Government Printing Office, 1977. U.S. Department of Commerce, Bureau of the Census, Statistical Abstract of the United States: 1978 U.S. Government Printing Office, 1978. U.S. Department of Commerce, Bureau of the Census, Statistical Abstract of the United States: 1974 U.S. Government Printing Office, 1974. U.S. Department of Commerce, Bureau of the Census, Statistical Abstract of the United States: 1970 U.S. Government Printing Office, 1970. U.S. Department of Commerce, National Oceanic and Atmospheric Administration, Climatological Data, National Summary, Annual 1976 U.S. Government Printing Office, 1976. U.S. Department of Commerce, National Oceanic and Atmospheric Administration, Climatological Data, National Summary, Annual 1972 U.S. Government Printing Office, 1972. U.S. Department of Commerce, National Oceanic and Atmospheric Administration, Climatological Data, National Summary, Annual 1968 U.S. Government Printing Office, 1968.

PAGE 147

135 U.S. Department of Labor, EmployTnent and Training Administration and U.S. Department of Health, Education and Welfare, Employment and Training Report of the President U.S. Government Printing Office, 1978. U.S. Department of Labor, Manpower Administration, Manpower Report of the President U.S. Government Printing Office, 1971. Vanderkamp, J., "Return Migration: Its Significance and Behavior," Western Economic Journal Vol. 10 (December 1972), pp. 450-65.

PAGE 148

BIOGRAPHICAL SKETCH Sheldon Donald Engler was born on March 30, 1955, in Philadelphia, Pennsylvania. In 1972, he graduated from North Miami Senior High School in North Miami, Florida. In 1975, he received a Bachelor of Arts degree in economics and sociology from the University of South Florida, Tampa, Florida. In 1978, he received a Master of Arts degree in Economics from the University of Florida. In August 1979, Mr. Engler accepted employment as a Research Economist at Louisiana State University, Baton Rouge, Louisiana. 136

PAGE 149

I certify that I have read this study and that in my opinion it conforms to acceptable standards of scholarly presentation and is fully adequate, in scope of quality, as a dissertation for the degree of Doctor of Philosophy. Jerome W. Milliman, Chairman Ptofessor of Economics I certify that I have read this study and that in my opinion it conforms to acceptable standards of scholarly presentation and is fully adequate, in scope of quality, as a dissertation for the degree of Doctor of Philosophy. / M/^^/9 of Economics I certify that I have read this study and that in my opinion it conforms to acceptable standards of scholarly presentation and is fully adequate, in scope of quality, as a dissertation for the degree of Doctor of Philosophy. David A. Denslow Associate Professor of Economics

PAGE 150

I certify that I have read this study and that in my opinion it conforms to acceptable standards of scholarly presentation and is fully adequate, in scope of quality, as a dissertation for the degree of Doctor of Philosophy. v"---Stanley K. Smith Assistant Professor of Economics I certify that I have read this study and that in my opinion it conforms to acceptable standards of scholarly presentation and is fully adequate, in scope of quality, as a dissertation for the degree of Doctor of Philosophy. C, VL:335SiJohnxf. Henretta Assistant Professor of Sociology This dissertation was submitted to the Graduate Faculty of the Department of Economics in the College of Business Administration and the Graduate Council, and was accepted as partial fulfillment of the requirements for the degree of Doctor of Philosophy, December 1979 Dean, Graduate School


31
In addition to comparing the likelihood value at the unconstrained
maximum to the likelihood value when all coefficients are restricted to
zero, a more stringent test is proposed. First the right-hand side of
the decision equation is set equal to the mean value of the dependent
variable. This would be the best guess of the probability that some
individual in the sample would move, given that we have no other in
formation. Then the resulting likelihood value is compared to the
likelihood value when all variables are included. The statistic put
forth above can then be calculated using these two likelihood values as
input.
The specific variables considered in the decision-to-move analysis
will be defined and discussed in Chapter 3. The next section puts forth
the methodology employed in the destination-choice analysis.
The Choice of Destination
Having determined the characteristics that distinguish interstate
movers from nonmovers, the next task is to examine the factors which
enter into the destination choice. Previous studies have considered
location choice primarily as a function of potential wages at alter
native locations. DaVanzo (1977) and Falaris (1978) first estimate a
wage model for each alternative location. Wages are viewed as a func
tion of the characteristics of those living at each place. Some of the
variables considered are age, race, sex, education, occupation, and
experience. These variables are similar to those typically employed in
human capital models of wage determination (for example Dalton and Ford
[1977]). After estimation, predicted values ofuwages at each desti
nation are obtained for everyone in the sample by plugging their


119
The theoretical model and methodology employed in this study can
easily be applied to other states. The results obtained could then be
compared to the findings for the four states considered here. Including
more states in the decision-to-move analysis would also provide a larger
sample for the analysis of destination choice. Better fitting models
might thus be obtained and forecasting accuracy may be improved. Ul
timately, detailed forecasts of labor force migration between all
states could be developed.
Although the focus of this study has been upon labor force mi
gration, the model developed is general enough to be applied to other
demographic groups. The study of, retiree migration would, for instance,
be an important application. Factors that influence the decision to
move and destination choice of the elderly population could be determined.
Ultimately, a forecast of the number of older people who will move to
retirement states such as Florida, California, Arizona, and Texas could
be developed. This forecast could then be combined with labor force
migration predictions for these states. Forecasts of these two components
of population growth should aid in projecting state population figures
and in planning for economic growth. The forecasts can also be used as
input in projecting other economic activities at the state level. State
econometric models, for example, have proven to be very sensitive to the
assumptions made about population growth. This is especially true for
states which are experiencing rapid rates of migration.
It is believed that the findings of this study have added to the
understanding of the migration process. It is hoped that this research
has provided theoretical, methodological, and empirical contributions
that will be of value to others who are interested in the study of
migration. -


second model focuses upon the destination choice of those who actually
moved from these states. Eight alternative states are considered as
destinations.
Although differences in migration determinants between states are
discovered, there are some general findings. The typical migrants in
the sample tend to be older and earn higher income than do nonmigrants.
Family relationships also appear to be important influences on the
migration decisions. Migrants are also less likely to be married than
nonmigrants. Contrary to expectations, families with two income earners
appear to move just as often as families where only one person is employed.
Destination choice is shown to be determined by income opportunity and
climate. Evidence that improved climate can sometimes only be obtained
at the cost of reduced income is discovered.
Overall, the migrant appears to be older, richer, and more willing
to take a cut in income for better climate than the nonmigrant. One
possible interpretation of the results is that many interstate moves are
being made in anticipation of retirement. Those who have earned greater
lifetime income are most able to absorb the decline in earnings consequent
upon moving prior to retirement. Economic theory postulates that higher
income people will demand greater quantities of leisure activities.
This increase can be obtained through the migration process. Locations
with warm weather have historically offered greater leisure possibilities
than places with colder, more variant climate.
The results are utilized to obtain interstate migration forecasts.
Rising family incomes, the process of population aging, and an assumption
of the continued importance of climate-related amenities in the migration
xi


69
The test developed by Chow (1960) for the equality of regression
coefficients between equations is applied to the decision-to-move
equations. First, ordinary least squares (OLS) estimates for each state
are determined. Then the observations from all four states are combined
and an OLS equation is estimated for the combined data.4 The results of
these estimates appear in Table 3.9. The following statistic is calculated:
[SEEq;ota]_ (SEE^y + SEEoH + SEEgc + SEEpA) ] / 3k
[SEENY + SEEQh + SEESC + SEEPA] / n-4k
where SEE stands for the sum of squared residuals, k refers to the
number of coefficients in each equation, and n refers to the size of the
combined sample. The subscript NY refers to the New York equation with
OH standing for Ohio, SC referring to South Carolina, and PA meaning
Pennsylvania. The subscript TOTAL refers to the combined equation. The
statistic is distributed according to an F-distribution. Its value in
this case is 1.67 which compares to a critical F value (14 degrees of
freedom in the numerator, and 779 in the denominator) of 1.5 at the .05
level of significance. Thus, the hypothesis of equality of regression
coefficients between equations is rejected at the 95 percent confidence
level. The belief that there are differences in migration determinants
across states is confirmed.
40LS is chosen for the test because execution of the logit program
is impossible for the combined sample and because no comparable test
exists for comparing coefficients across logit equations.


21
describing individual t, and P^ is a vector of explanatory variables
describing alternative i. Also define 9 = (0 9 ), where 9 and
t tx, tp tx
9 are vectors of parameters assigned to Xt and P^ respectively.
Finally, define e as a random disturbance for alternative i and
decision maker t. The utility which individual t derives from place i
can now be expressed as follows:
(1)
Alternatively, it can be written:
(2)
If the individual maximizes utility, then he or she will choose to live
U = Z^.6^ + £ .
ti tx t ti
U = X 0 + P.0 + e .
ti t tx i tp ti
at alternative i if U > U for all j not equal to i.
tx tj j u
Moving costs can be incorporated into the model. Assume first that
we find individual t living at place i. Then we conclude that if he (or
she) is rational: U > U C.. for all j not equal to i where C.. is
the cost of moving from i to j. If at a later time person t migrates to
place j, then we conclude that something occurred during the interim
which caused t's utility evaluation to change so that: U C.. >
13 ij
U Throughout the remainder of the study, actual moving expenses will
be ignored since it is believed that they are insignificant in relation
to the other costs and benefits of moving.
The probability that a person will choose to live at location i can
be written as follows:
(3)
A major goal of this study will be to impute values of P(i) for in
dividuals inside and outside the sample. Inferences about aggregate
behavior will also be made using these probabilities.
P(i) = P(U > U for all j t i)
tx tj J


80
Table 3.13 Forecast: Decision to Move, 1976-1980
State of
Origin
Number of Migrants-
Panel Study
Number of Migrants-
Aggregate Labor Force
Ohio
18
833,000
New York
24
1,867,000
South Carolina
19
309,000
Pennsylvania
19
1,003,000
Total
80
4,012,000


52
Employment Status
People who are unemployed are expected to be more responsive to
opportunities elsewhere than those who are employed. Particularly if
local economic conditions are poor, the process of job search for the
unemployed is likely to include alternative locations. If, however,
employment conditions at other locations are also depressed, the response
of the unemployed is uncertain. The effect of employment status upon
the migration decision seems to depend on local economic conditions
relative to alternative destinations. In the analysis to follow, relative
conditions will be held constant by considering only one state at a
time. The mean value of employment status for all four states combined
is slightly lower for migrants than for nonmigrants. Ninety-six percent
of the migrants in the sample were employed prior to moving.
Marital Status
Marriage is expected to act as a deterrent to migration. As with
the family-size variable, presence of a spouse represents an additional
tie to present location. This is particularly true if the spouse is
working. The data in Table 3.4 show that migrants are less likely to be
married than nonmigrants. Seventy-eight percent of the migrants in the
four-state sample were married, while 81 percent of the nonmigrants were
married.
Race
The expected sign of the race variable is ambiguous. On the one
hand it can be argued that since some blacks are more likely to earn
lower incomes and are more likely to be unemployed than whites, they are


112
Table 5.1 Summary of Findings: Decision to Move
Variable
Ohio
New York
South Carolina
Pennsylv
Famsz
+
-
+*
+
Sex
+*
+
+
+
Homeown
-
+
-
+
Lres
+
+
+
+
Lempl
+
-
+
Emself
-
-
-
+
Avage
+
+*
+
Ernst

JU
-

Marst
-*
-
_*
-
Race
+
-
-
+
Aveduc
-
+
+
+*
Prevmig
+
-
+
+
Ernst w
+
+*
-
+
Famy
+*
+
+*
* Significant at 95 percent confidence level or better.
+ Positive coefficient.
- Negative coefficient.
. Not in equation.


101
Table 4.8 Logit Results: Destination Choice, 1972-1976
Variable
Coefficient Asymptotic t-Ratio
Distance
-.0004 -1.55
Tempdiff
.181 3.67
Incdiff
.001 1.59
Summary Statistics:
Log of Likelihood Function = -85.1
Likelihood Statistic (at 0) = 17.9
Level of Significance (at 0) = .005
Likelihood Statistic (at mean) =4.7
Level of Significance (at mean) = .10


49
lower incomes at their current locations and have a higher probability
of being unemployed before moving, they are more responsive to oppor
tunities elsewhere. This would counteract the negative effects of the
sex variable and lead to a more ambiguous result. No conclusions can be
drawn without more rigorous analysis.
Homeownership
Owning a home is usually expected to act as a tie to present location.
Time and transactions costs involved in selling a home and perhaps
purchasing another one in a new location would tend to deter migration.
On the other hand, homeownership represents wealth. If a wealth effect
operates, property owners may be more likely to move than renters. In
this sample 64 percent of the migrants owned homes prior to migration
while only 58 percent of the nonmigrants were homeowners. Thus, the
data give some support to the second hypothesis.
Length of Residence
Persons who have lived in their current houses or apartments for
long periods of time are expected to display a lower probability of
moving. Duration of past residence should act as a measure of will
ingness to pull up stakes, leave friends and relatives, and start all
over somewhere else. An individual who remains in the same dwelling
unit for a long time displays an aversion to risk which leads one to
believe that he (or she) would not be likely to do something as filled
with uncertainty as changing states of residence. The data in Table 3.4
do not demonstrate this to be the case. There appears to be very
little difference in length of residence for migrants and nonmigrants.
The average length of residence for both groups is close to 10 years.


130
called Rwdiff. Even when measured in real terms, the individual
returns variable remained insignificant. The overall fit of the
model remained poor. These results appear in Table B.2


61
The remaining variables are not significant. Family size, average
age, and race are variables which show some importance. Migrant families
in Ohio are more likely to be larger than nonmigrant. This result is
opposite from that expected. A possible explanation is that many
migrant families from Ohio have preschool children. While having
children in school may act as a deterrent to migration, families may
decide to move prior to their children's first enrollment in school.
Indeed, people may move so that their children can go to different
schools. Age shows a positive sign, lending support to the prere
tirement hypothesis set forth earlier. Race is also positive, indi
cating that households headed by white persons are more likely to move
interstate than households headed by nonwhites.
Although insignificant, the homeownership variable has the ex
pected sign. People owning homes in Ohio are less likely to move than
those who rent. They are also more likely to have lived in their homes
and worked for the same employer for longer time periods than non
migrants. These results are contrary to expectations. The migrant is
less likely to be self-employed. Thus, self-employment appears to act
as a migration deterrent. Those who chose to leave Ohio are more
likely to have moved previously, supporting the hypothesis put forth
earlier. Education carries an unexpected negative coefficient. Migrant
household heads are also likely to have spouses who are employed. This
variable has the opposite sign from that which was expected. Employment
status was not included in the analysis because there was no variation
of this characteristic among migrants. Estimation was thus impossible.
When the likelihood value of the decision equation is compared to
the value that results when all coefficients are zeroed out, the equation
is significant at better than the .005 level. When the likelihood


60
Table 3.5 Results of Logit Analysis: Decision to Move,
Ohio, 1968-1977
Variable Coefficient Asymptotic t-Ratio
Famsz
.147
1.27
Sex
2.802
2.60
Homeown
-3.65
- .64
Lres
.013
.06
Lempl
.052
.31
Emself
-.845
- .73
Avage
.054
1.53
Ernst
-
-
Marst
-4.142
-3.39
Race
.867
1.09
Aveduc
-.024
- .15
Prevmig
.298
.69
Ernst w
.471
.96
Famy
.057
2.06
Constant
-5.000
-2.72
Summary Statistics:
Log of Likelihood Function = -86.0
Likelihood Statistic (at 0) = 133.0
Significance Level (at 0) = .005
Likelihood Statistic (at mean) =27.3
Level of Significance (at mean) = .025
Percentage Correctly Predicted = 85%
Percentage of Migrants Correctly Predicted = 16%


Table 4.10 Aggregate Destination Choice
Origin-
Destination
Number of Migrants
1968-1972
New York-Florida
New York-California
New York-New Jersey
New York-Illinois
New York-Virginia
New York-Michigan
New York-Texas
210,000
131,000
355,000
224,000
131,000
171,000
79,000
Pennsylvania-Florida
81,000
Pennsylvania-California
47,000
Pennsylvania-New Jersey
135,000
Pennsylvania-New York
188,000
Pennsylvania-Illinois
81,000
Pennsylvania-Virginia
47,000
Pennsylvania-Michigan
67,000
Pennsylvania-Texas
34,000
Ohio-Florida
63,000
Ohio-California
52,000
Ohio-Ney Jersey
78,000
Ohio-New York
109,000
Ohio-Illinois
89,000
Ohio-Virginia
36,000
Ohio-Michigan
63,000
Ohio-Texas
31,000
South
Carolina-Florida
41,000
South
Carolina-California
19,000
South
Carolina-New Jersey
32,000
South
Carolina-Nex^ York
43,000
South
Carolina-Illinois
30,000
South
Carolina-Virginia
17,000
South
Carolina-Michigan
19,000
South
Carolina-Texas
13,000
1968-1972, 1972-1976, 1976-1980
Number of Migrants
1972-1976
Number of Migrants
1976-1980
443,000
163,000
163,000
82,000
117,000
700,000
140,000
250,000
83,000
91,000
114,000
45,000
61,000
38,000
83,000
244,000
99,000
76,000
99,000
53,000
61,000
38,000
91,000
95,000
30,000
22,000
27,000
15,000
20,000
10,000
30,000
331,000
252,000
199,000
146,000
132,000
119,000
159,000
148,000
108,000
87,000
87,000
67,000
54,000
54,000
67,000
114,000
103,000
60,000
60,000
60,000
43,000
43,000
60,000
59,000
41,000
25,000
23,000
20,000
18,000
16,000
25,000
104


Ill
alternatives to a potential migrant. For this reason, the state has
been chosen as the level of locational analysis in this study. An
argument could be made for defining regions at an even smaller level,
such as the urban area, but this would increase the number of alter
natives that would need to be included in a thorough analysis. Tech
nical and cost limitations prohibit this procedure.
Utilizing the results of both the decision-to-move analysis and the
destination-choice model, an aggregate migration forecast was obtained
for each origin and destination combination. Tests of forecasting
accuracy revealed mixed outcomes. The total number of out-migrants from
the four-state sample was predicted reliably. But because of a small
sample and poorly fitting, short time period equation the choice of
destination was not accurately predicted. More specific findings of the
study are reviewed in the following sections.
Review of Findings: Decision to Move
The results of the decision-to-move analysis are summarized by
state in Table 5.1. For the state of Ohio, marriage deters migration.
Mincer's (1978) finding that the presence of a spouse represents a tie
to present location is thus supported. Migrants from this state are
also likely to have higher family incomes. This result supports the
viewpoint that higher income families are more likely to demand goods
which can only be obtained by changing locations. Finally, Ohio migrants
are likely to be members of households headed by males.
As with Ohio migrants, New Yorkers with greater incomes are more
likely to change states of residence. Unemployed persons in New-York
are more likely to move than those who are employed, supporting the


decision lead to a forecast of further migration concentrated toward
Sunbelt destinations. Policies aimed at restricting this growth or
attracting migrants to other locations could alter this pattern.
xii


56
South Carolina. If older people are more responsive to climate differ
entials than younger people, we might then expect age to be a more impor
tant migration determinant in Ohio than it is in South Carolina. Thus,
variables are significant in one state but not in another because of
differences in initial (origin) conditions. From now on, each state
will be analyzed by itself.
When migrants and nonmigrants are compared only according to the
mean values of a set of variables, other problems arise. First, there
are the usual weaknesses involved in using the mean as a measuring
device (its sensitivity to extreme values is one such weakness). More
importantly, when any variable is considered, the values of remaining
factors are not held constant. Thus, the true effect of each char
acteristic is not being measured. Use of multiple regression techniques
will relieve these problems.
Before the results of the logit analysis of the decision to move
are presented, a discussion of the scaling of some of the qualitative
(dummy) explanatory variables is in order. Particularly, length of
residence, self-employment, and average education are defined in ways
which are somewhat out of the ordinary. Looking back at Table 3.3 it is
observed that these variables are scaled using an ordinal ranking system,
with each value representing a given level or interval. This scaling is
the same as that which appears on the Panel Study tapes. The conventional
method of modeling variables which are coded in such a fashion is to
include separate dummy variables for each category of data. In this
case, however, 21 variables would need to be added to the model. This


35
of the strongest contributors of Florida migrants, these states are
likely to have experienced higher rates of out-migration. These latter
arguments provide some support for the finding of higher levels of
migration in this study. More detailed evaluation of the estimates is
impossible without having a comparable set of labor force migration
figures for the same period. The next chapter presents the results of
the destination-choice analysis and develops a forecast for the 1976-80
period.


40
Logit analysis depends upon the assumption that the dependent
variable (the logarithm of the odds that a particular choice will be
made) approximates the normal distribution. A large number of obser
vations and sufficient repetitions for each possible choice assure that
this criterion will be met. For the decision-to-move analyses it is
believed that sample size is sufficient for employing logit techniques.
The destination choice results should be interpreted more cautiously. A
particular problem for these equations is the small number of observations
occurring in each possible category of choice.4
All equations were also estimated using ordinary least squares
techniques (see Chapter 3). The signs and magnitudes of the coefficients
obtained were very similar to those obtained with logit analysis.5
These similarities increase the degree of confidence placed in the major
findings of this study. In Chapter 3 the results of the decision-to-
move analysis are put forth and discussed.
4An additional source of worry arises because origin states are
chosen according to their rates of migration to Florida. This may
introduce an upward bias in the predictions of migration to Florida.
Consequently, the probabilities of migration to other states will be
understated.
sDaVanzo (1977) compares probit estimates of decision to move
equations with ordinary least squares estimates. She also finds similar
results.


Table A.3 Results of Logit Analysis With Alternative Dummies, 1968-1977
Variable
Ohio
Coefficient
t-Ratio
New York
Coefficient t-Ratio
South Carolina
Coefficient t-Ratio
Pennsylvania
Coefficient t-Ratio
Famsz
.231
1.93
-.251
-2.03
.205
1.70
.138
1.06
Sex
3.440
2.97
.712
.43
1.352
1.19
2.559
2.06
Race
1.116
1.28
-1.877
-1.70
-.480
- .75
.371
.56
Homeown
-.081
- .13
.903
1.52
-.212
- .37
.671
1.24
Lres
-.538
-1.03
1.096
2.00
.235
.44
-1.289
-2.41
Lempl
-1.036
-1.61
-1.750
-2.67
-1.382
-2.52
.014
.02
Emself
-1.726
-1.79
-.690
- .84
-2.513
-1.57
.692
.52
Avage
.085
2.37
.065
1.91
.084
2.22
.033
1.08
Aveduc
.823
1.51
.621
1.04
.438
.74
.349
.64
Ernst i
-
-
-3.450
-2.38
-.627
- .76
-
-
Famy
.035
1.46
. 134
2.75
.210
3.32
-.113
-1.60
Prevmig
.322
.73
-.510
- .82
.525
.86
.627
.97
Ernst w
.827
1.64
1.253
2.36
-.591
- .99
.538
.95
Marst
l
-5.213
-4.06
-1.452
- .85
-3.401
-2.51
-3.538
-2.54
Constant
-5.862
-4.58
-.094
- .06
-4.362
-3.30
-2.703
-2.53
124


2
regional level depend heavily upon the ability to predict population.
The demand for housing, the unemployment rate, and state and local
government tax collections are examples of economic variables which are
sensitive to the rate of population change. Knowledge of the level of
future population and its consequences for the rest of the economy is
crucial to success in planning for regional economic growth.
In this study, an econometric model is developed for use in ex
plaining and forecasting the movement of workers and their families
between states. A general theoretical model of interstate migration is
derived within a microeconomic framework. This model is then placed
into an estimable form and tested for the period 1968-77. Individual
data obtained from the University of Michigan Panel Study of Income
Dynamics are used. Maximum likelihood estimation techniques are em
ployed. Finally, the results of the analysis are utilized in devel
oping an aggregate forecast of interstate labor force migration.
Throughout the study the case of Florida is emphasized. Separate
models are estimated for four typical origin states of Florida migrants.
The determinants of Florida's attractiveness relative to other desti
nations which draw migrants from these states are analyzed.
In the next section the existing migration literature will be
brought up to date. Then an overview of this study will be presented.
Literature Review
Introduction
Existing studies will be reviewed according to two major cate
gories. The first section which follows will summarize research that is
centered upon explaining aggregate flows of population. The second


28
The choice of time period is important in any migration study.
Shorter time periods at first seem optimal, since the determinants of
migration can be measured at or close to the time at which the actual
move is made or not made. Longer time periods lead to problems, since
migration may be taking place at many different points during the
period. Determination of when to measure the explanatory variables is
then difficult. Estimation problems, however, arise for short time
periods because the proportion of the population that moves is smaller
than it is for long periods. Estimations were first attempted for one-
year periods. Poor fits were obtained and few variables were significant
enough to distinguish between migrants and nonmigrants. A sample of
these results can be found in Appendix A. For the main part of the
study, three time periods were considered. First, the model was es
timated for the 1968-77 period, the maximum amount of time for which
data is available in the sample. Then, for the purpose of forecasting
the models were re-estimated for the periods 1968-72 and 1972-76. The
explanatory variables were always measured in the first year of the
period under consideration. The migration variable took on the value of
one if during the last year of the period, an individual was living in a
state which differed from his (or her) state of residence during the
first year of the period. Otherwise it took the value of zero.
The unit of analysis in the study is the head of household who is a
member of the labor force. Since the survey from which the sample is
derived contains some information about other household members, var
iables such as family size and spouses employment status can also be
entered into the equations. Other studies (e.g., DaVanzo [1977])


TABLE OF CONTENTS
PAGE
ACKNOWLEDGEMENTS iii
LIST OF TABLES vi
LIST OF FIGURES ix
ABSTRACT x
CHAPTER
1 INTRODUCTION 1
Problem Statement 1
Literature Review 2
Overview of the Study 16
2 THEORY AND METHODOLOGY 18
Theoretical Model 18
Methodology 22
A Forecasting Methodology 34
Sources of Data 38
3 EMPIRICAL RESULTS: DECISION TO MOVE 41
Interstate Migration: 1968-1977 41
Comparison of Migrants and Nonmigrants 45
Empirical Results 55
Application to Forecasting 72
4 EMPIRICAL RESULTS: DESTINATION CHOICE.. 86
Destination Choice: 1968-1977 86
Logit Results: Destination Choice 86
Application to Forecasting 96
Income, Climate, and Migration 107
5 SUMMARY AND CONCLUSIONS 110
Introduction 110
Review of Findings: Decision to Move Ill
Review of Findings: Destination Choice 115
Strengths and Weaknesses of the Study 116
Implications of the Study 118
iv


PAGE
APPENDIX
A ALTERNATIVE MODEL FORMULATIONS: DECISION TO MOVE 120
Combined State Results 120
Annual Migration Results 122
Redefinition of Qualitative Explanatory Variables 122.
B ALTERNATIVE MODEL FORMULATIONS: DESTINATION CHOICE 128
Nominal Wage Model 128
Real Wage Model 128
BIBLIOGRAPHY 132
BIOGRAPHICAL SKETCH 136
v


23
A. Migrating to Florida or not
B. Migrating to California or not
C. Staying in New York or not
Values of I can now be assigned to each decision. If this person
chooses to migrate to Florida, then for decision A, I = 1; for decision
B, I = 0; and for decision C, 1=0. The values of X will vary across
alternatives. Individual-alternative interaction terms can also be
introduced into the model.
If each binomial decision by each individual is treated as a single
observation, then the model can be estimated by ordinary least squares
(OLS). It is well known, however, that use of OLS when the dependent
variable is dichotomous leads to inefficient estimates and to predictions
outside the zero-one interval. In addition, OLS is unable to distinguish
between individuals and observations in the multinomial case. More
appropriate methods use the maximum likelihood technique of estimation.
McFadden (1973) developed a technique he calls conditional logit
for use in analyzing consumer choice among lumpy alternatives. He
demonstrates the applicability of conditional logit in the study of
urban travel demand. In a recent study, Falaris (1978) applies the same
technique to the migration decision. An attempt was made in this study
to adopt the same methodology. It did not prove to be useful for this
study. The conditional logit model seems best adapted to problems where
the choice frequencies are fairly well balanced. Location choice
(studied over reasonable periods of time) is biased heavily toward a
single alternative: staying where one already is. Thus, when a model is
estimated which includes staying as one of many alternatives, this
choice swamps all others. Consequently, it is difficult to obtain


16
only act as a deterrent to a possible move. This is particularly the
case where more than one family member is working. Thus, the increasing
proportion of women in the labor force is expected to have an inhibiting
effect upon migration. Mincer uses a scattering of data to test his
theories. He discovers, as did DaVanzo, that marriage itself reduces
migration and that migration rates are lower in families with employed
wives. The magnitude of the effect which the working wife has on the
migration decision also depends upon her share of family earnings.
Finally, he shows that tied migration of working wives frequently
results in lower earnings, unemployment, or labor force withdrawal.
Thus, male household heads usually dominate the choice of destination.
Overview of the Study
In Chapter 2 the migration decision is analyzed within the context
of utility theory. A theoretical model of interstate migration is then
derived. A methodology for estimating this model is put forth which
considers the decision to move and the choice of destination together in
one analytical framework. Problems that arise in implementing this
procedure are then discussed and an alternative is proposed. The new
method views the migration decision as a two-stage process. Finally, a
technique for forecasting migration decisions and destination choice is
proposed.
Chapters 3 and 4 present the empirical results. The decision to
move is analyzed in Chapter 3. Regression results are presented for
four states and forecasts of aggregate migration from these states are
developed. Chapter 4 deals with the choice of destination. The de
terminants of this decision are analyzed and a forecast of migration


35
force in the state under consideration. In this way a forecast of the
actual number of labor force migrants from that state is obtained.
Figure 2.1 contains a flow chart which illustrates the procedure just
described.
A similar procedure is followed for forecasting destination choice.
The model is estimated for the 1968-72 period and the coefficients are
then applied to locational characteristics in 1972. Forecasts of the
probabilities of migration between each origin and each destination are
obtained. These probabilities are then multiplied times the number of
migrants leaving each origin location. The values obtained can be
compared to the actual origin-destination flows during the 1972-76
period. If successful, the model is re-estimated for the 1972-76
period and these coefficients are applied to the 1976 locational char
acteristics. A 1976-80 aggregate forecast can be obtained by applying
the probabilities derived from the 1972-76 estimation to the aggregate
decision-to-move forecast. Thus, forecasts of actual flows of labor
force migrants to all destinations are derived. The destination-choice
forecasting methodology is illustrated in Figure 2.2.
In addition to forecasting applications, estimating the model for
two equal length time periods allows for observation of coefficient
stability. In this way it can be determined whether events such as the
energy crisis and the deep recession of 1974 have had any impact on
migration decisions and destination choice.3 Since earlier period
3In fact this goal will be difficult to attain. Four-year migration
periods contain many events which cannot easily be disentangled from one
another. Estimation for shorter periods is even harder because of
(migrant) sample size problems.


36
Figure 2.1 Illustration of Forecasting Methodology: Decision to Move


34
All variables will be premigration measures of location aggregates.
Specific variables to be included in the choice of destination equations
will be defined and discussed in Chapter 4. The next section will
propose a method of obtaining a forecast from the completely estimated
model.
A Forecasting Methodology
In an application of logit analysis to transportation decisions,
McFadden (1974) states that if the sample under consideration represents
a random selection of the environments faced by the population as a
whole, the average of the predicted values over the sample is a best
estimate of aggregate demand. With this in mind a methodology is proposed
for forecasting the number of migrants from a given origin and their
destination choices. The accuracy of this methodology will be tested
within the sample.
For any given origin state, the coefficients of the 1968-72 decision-
to-migrate equation are applied to values of the explanatory variables
for individuals living in that state in 1972. In this way predictions
for the next four-year period, 1972-76, are obtained. Predicted proba
bilities for each individual are obtained and all individuals are averaged.
This average probability is then applied to the total sample in 1972 to
get a forecast of the number of migrants between 1972 and 1976. This
number can then be compared to the actual (known) number of migrants in
the sample during the period. If successful, then 1972-76 coefficients
can be applied to the 1976 sample to obtain a 1976-80 forecast. The
next step involves application of the sample-derived average proba
bilities to aggregate (outside the sample) measures of the labor


32
characteristics into the fitted equations. This variable is usually
found to be a significant determinant of location choice. A similar
variable has been constructed for use in this study. Although the sign
was positive, it did not turn out to be significant. One possible
explanation derives from the way in which the sample was chosen. As
mentioned earlier, the origin states considered are the leading senders
of migrants to Florida. Florida is not known as a high-wage state, yet
many people move here. In fact, many Florida migrants in the sample
have experienced declines in nominal and real (price deflated) wages. A
sample of results obtained when the wage variable is included is contained
in Appendix B.
Generally, earlier empirical models of individual migration choice
have not found aggregate location characteristics, such as the unemploy
ment rate, to be important variables. But these studies have considered
only very large geographic areas as potential destinations. Differences
in location-specific characteristics then tend to get averaged out
across the region. Better results should be obtained if destinations
are defined as smaller areas, such as states. A disadvantage of this
approach is that all possible destinations cannot be considered since
the estimation technique puts limits upon the number of alternatives
that may be considered at once.
Essentially, the approach suggested by Morrison (1973) is taken
here. That is, while the decision to migrate is considered from a
microanalytic perspective, the decision where to go is modeled in a
macroanalytic framework. By considering only aggregate characteristics
of each location as variables, the model will, for example, predict the
probability of a New York migrant choosing Florida as a destination. It
will not, however, explain why one New Yorker chooses Florida and another


I certify that I have read this study and that in my opinion it
conforms to acceptable standards of scholarly presentation and is
fully adequate, in scope of quality, as a dissertation for the degree
of Doctor of Philosophy.
Stanley K. Smith
Assistant Professor of Economics
I certify that I have read this study and that in my opinion it
conforms to acceptable standards of scholarly presentation and is
fully adequate, in scope of quality, as a dissertation for the degree
of Doctor of Philosophy.
e.
John\f. Henretta
Assistant Professor of
Sociology
This dissertation was submitted to the Graduate Faculty of the Department
of Economics in the College of Business Administration and the Graduate
Council, and was accepted as partial fulfillment of the requirements for
the degree of Doctor of Philosophy.
December 1979
Dean, Graduate School


46
Table 3.3 Variable Definitions for Decision to Migrate Analysis
Dependent Variable
Definition
Move Equals 1 if the person's state of residence
in 1977 differs from his or her state of
residence; equals 0 otherwise. (Defined
in 1968 for different periods in later
analysis.)
Explanatory Variables
Famsz
Family size, actual number in household.
Sex Equals 1 for male; 2 for female.
Homeown Equals 1 if home is owned; equals 0
otherwise.
Lres
Lempl
Emself
Avage
Length of residence in current house or
apartment, takes on following values:
0 if length of residence < 1 year
1 if length of residence = 1 year
2 if length of residence = 2 years
3 if length of residence = 3 years
4 if length of residence = 4 years
5 if length of residence 5 years and
< 10 years
6 if length of residence > 10 years and
< 15 years
7 if length of residence > 15 years and
< 25 years
8 if length of residence > 25 years
Length of time employed by present employer,
takes on following values:
0 if self-employed or unemployed
1 if employed > 0 months and < 6 months
2 if employed > 6 months and < 18 months
3 if employed > 18 months and <_ 42 months
4 if employed > 42 months and < 9 years
5 if employed > 9 years and < 19 years
6 if employed > 19 years
Equals 1 if self-employed; equals 0
otherwise.
Average age of husband and wife; if unmarried
Avage = age of household head.


11
One of the advantages of models of net migration is that data on
the dependent variable are easily computed for most areas. Net migra
tion can be directly estimated by subtracting the population change due
to natural increase from the actual population change for any given
period. Measures of gross migration, in contrast, depend upon more
direct measurement techniques, such as surveys. However, as Lowry
(1966) notes, models of net migration reveal less about migratory be
havior and the decision to migrate than do models of gross migration.
This is partly because variables which are important in determining
unidirectional flows are reduced in significance when the measure of
migration used includes flows occurring in opposite directions. Models
of net migration, however, are useful in forecasting population change.
They are of particular value to regions for which migration has been the
dominant component of population change. The next section will describe
research which has focused upon explaining migration at the individual
or family level.
Individual Migration Models
Rothenberg (1977) suggests an approach to the study of migration
that focuses upon the individual. He notes that the migrant is self-
selected. That is, given the availability of similar sets of oppor
tunities, some individuals will migrate and some will not. The problem
is to determine the individual characteristics and circumstances that
cause people to evaluate their migration choice in different ways. Put
somewhat differently, one individual's maximizing decision may cause him
(or her) to migrate, while another individual facing similar opportuni
ties may choose not to move. Both persons may be acting rationally.


65
Variable
Famsz
Sex
Homeown
Lres
Lempl
Emself
Ava ge
Emst
Marst
Race
Aveduc
Prevmig
Emstw
Famy
Constant
Table 3.7 Results of Logit Analysis: Decision to
Move, South Carolina, 1968-1977
Coefficient
.252
1.231
-.180
.286
-.328
-2.713
.093
-.678
-3.494
-.371
.073
.803
-.528
.217
Asymptotic t-Ratio
1.95
1.04
- .33
1.17
-1.76
-1.65
2.51
- 66
-2.46
- .58
.38
1.24
- .89
3.22
-6.333 -2.96
Summary Statistics:
Log of Likelihood Function = -71.7
Likelihood Statistic (at 0) = 164.4
Level of Significance (at 0) = .005
Likelihood Statistic (at mean) = 39.8
Level of Significance (at mean) = .005
Percentage Correctly Predicted = 90%
Percentage of Migrants Correctly Predicted = 31%


xml version 1.0 encoding UTF-8
REPORT xmlns http:www.fcla.edudlsmddaitss xmlns:xsi http:www.w3.org2001XMLSchema-instance xsi:schemaLocation http:www.fcla.edudlsmddaitssdaitssReport.xsd
INGEST IEID EXP7OWFA6_55FQ2A INGEST_TIME 2014-12-03T19:13:19Z PACKAGE AA00026470_00001
AGREEMENT_INFO ACCOUNT UF PROJECT UFDC
FILES


94
The significance of temperature and income differences lends support
to the Graves (1978) thesis which suggests that migration is just a
response to an increased demand for amenities which are available at
certain locations. The economy is believed to start out in spatial
equilibrium, meaning that amenity-adjusted real income is the same at
all locations. Differences in income across locations are said to exist
because of differences in the availability of location-specific goods,
such as recreation. An exogenous shock, such as an income increase,
causes the demand for location-specific goods to increase. This creates
a temporary disequilibrium which is restored through migration. The
positive and significant coefficient on the income variable says that if
temperature differences are the same for all possible moves, people will
choose places which promise positive income gains. To test whether
individuals will still choose higher income places, even if temperature
differences are allowed to vary, the model is re-estimated with the
climate variable removed. The results of this estimation are presented
in Table 4.5. The income difference variable is no longer significant.
Omitting the climate variable causes a downward specification bias on
the income variable. Thus, the Graves theory receives confirmation. If
the income variable had remained significant, then support would have
been generated for regional growth models which claim that migration is
solely a function of interregional wage differences. But, instead,
income is only significant in conjunction with climate, so that the
trade-off hypothesized by Graves appears to be in operation.
The equation of Table 4.4 is. significant at _the .005 level whether
it is compared to the model where all coefficients are zero or to the


APPENDIX A
ALTERNATIVE MODEL FORMULATIONS: DECISION TO MOVE
Combined State Results
Table A.l contains the results of the decision-to-move analysis
when all four origin states are combined into a single sample.
Ordinary least squares estimates appear in the table. A problem
with available storage space prevented successful execution of the
logit program for a sample this large. The variables are defined
as they were in Chapter 3.
Variables that are significant at better than the .05 level are
sex, average age, family income, and marital status. These variables
are the same ones that often turn up significant at the state level.
But other factors, which only occasionally appeared to be important
for separate states, no longer are significant. Examples are average
education (significant in Pennsylvania) and spouse's employment status
(significant in New York). In addition, variables that have opposite
signs for different states become averaged into the combined state
coefficients. For example, in Chapter 3 it was found that Pennsylvania
migrants are negatively selected according to income. The positive
and significant coefficient on combined state family income hides
this discovery.
120


93
Table 4.4 Logit Results: Destination Choice,
1968-1977, Model 2
Variable
Coefficient
Distance -.741
Tempdiff .231
Incdiff .152
Summary Statistics:
Log of Likelihood Function = -162.8
Likelihood Statistic (at 0) = 22.9
Level of Significance (at 0) = .005
Likelihood Statistic (at mean) = 13.75
Asymptotic t-Ratio
-3.36
4.78
3.55
Level of Significance (at mean) = .005


10
Table 1.3 Graves Net Migration Regression Results
Variable
Coefficient
t-Statistic
Me dine
.00162
1.07
Unemp
-2.906
-4.26
Warmth*
.0103
4.44
Cold
.00686
4.27
Antmvr
-.9989
-4.86
Annwnd
-2.967
-4.49
Annhum
-.7164
-4.74
Variable Definitions:
Medinc
-1960 Median Income
Unemp
-1960 Unemployment Rate
Warmth*
-Mean annual number of cooling degree
days (base = 65
F)
Cold
-Mean annual number of heating degree
days (base = 65
F)
Antmvr
-Annual temperature variance (average daily maximum
July temperature average daily minimum January
temperature)
Annwnd
-January and July average wind speed
Annhum
-January and July average humidity
*A11 climate variables defined as 1931 to 1960 averages


121
Table A.1
- Combined State Results:
Ordinary Least
Decision to Move,
Squares
1968-1977
Variable
Coefficient
t-Ratio
Famsz
.010
1.54
Sex
.278
3.96
Race
-.016
- .52
Homeown
.005
.18
Lempl
-.014
-1.52
Emself
-.047
-.78
Avage
.005
3.38
Aveduc
.018
1.85
Ernst
-.120
1.47
Famy
.005
2.23
Prevmig
.024
.81
Ernst w
.041
-1.47
Marst
-.275
-4.01
Lres
.011
1.00
Constant
-.148
-1.17
R2 =
.06


53
more responsive to opportunities elsewhere. On the other hand, there
may also be fewer opportunities elsewhere for blacks because they are
concentrated in low-skill occupations and because of discrimination in
the labor market. In addition, blacks are less likely to be employed in
jobs requiring geographic transfers. For the sample to be considered
here it looks as if migrants are more likely to be white, indicating
that the latter two effects outweigh the first factor.
Average Education
More educated people are expected to move more often. The reason
is twofold. For one, education (in most cases) leads to increased
employment opportunities. Particularly it is believed that increased
education opens up employment possibilities which are more national in
scope resulting in greater migration propensities. In addition, those
who have attained higher education levels are expected to have more and
better information about migration opportunities. The extreme of this
phenomenon occurs in professions that form organizations which aid their
members in locating jobs. The economics profession is an example. The
data in Table 3.4 give preliminary evidence that this thesis is correct.
Migrants in the sample are likely to have at least started college,
while the average nonmigrant posesses only a high school education.
Previous Migration
Migrants are believed to be characterized by frequent mobility
during their lifetimes. It is hypothesized here that a person who has
changed state of residence at least once during his (or her) lifetime is
more likely to move again than an individual who has always lived at the
same location. People who have moved frequently are also less likely


BIOGRAPHICAL SKETCH
Sheldon Donald Engler was born on March 30, 1955, in Philadelphia,
Pennsylvania. In 1972, he graduated from North Miami Senior High School
in North Miami, Florida. In 1975, he received a Bachelor of Arts degree
in economics and sociology from the University of South Florida, Tampa,
Florida. In 1978, he received a Master of Arts degree in Economics from
the University of Florida. In August 1979, Mr. Engler accepted employment
as a Research Economist at Louisiana State University, Baton Rouge,
Louisiana.
136


88
Table 4.2 Variable Definitions: Destination Choice Equations
Dependent Variable
Definition
Choice
Equals 1 for chosen destination,
equals 0 for all other alternatives.
Explanatory Variables
Distance
Straight line mileage between each
origin and each destination.
Tempdiff
Annual average temperature at each
destination minus annual average
temperature at each origin.
Incdiff
Annual per-capita personal income
at each destination minus annual
per-capita personal income at each
origin.
Hdiff
First quarter housing price index at
each destination minus first quarter
housing price index at each origin.
Taxdiff
Annual per-capita state and local
government tax collections at each
destination minus annual per-capita
state and local government tax
collections at each origin.
Unemdiff
Annual average unemployment rate of
each destination minus annual average
unemployment rate at each origin.


24
variables that distinguish between the remaining alternatives. The
magnitude of this problem enlarges as the number of alternatives in
creases. Falaris avoided this problem to some extent by restricting his
research to a choice among four broad regions of the United States.
Since this study will focus upon the state as the level of analysis, it
will be desirable to include a larger number of alternatives. A further
problem encountered with the model is that as new variables are added
when there are a large number of alternatives, the cost of estimation
rises substantially and the chances of early convergence diminish.
Falaris, in fact, considered only a small number of variables in his
study. Finally, testing various model specifications is difficult using
such high-cost procedures.
Proposed Methodology
The above considerations have led to the decision to analyze the
migration choice as a two-stage process. This is in line with the
approach suggested by Morrison (1973) and taken by DaVanzo (1977). The
first stage of the migration process is the decision "whether to move."
All members of the population are faced with this choice. The second
stage involves the decision "where to move." Only migrants face this
choice. The decision equation (4) is now divided into two:
(5)
(6)
I. = x 0 + e. .
l t tx tx
l = p.e, + u .
2 i tp ti
where U ^ is a random disturbance term and every other term is defined
as before. Equation (5) represents a binomial decision, with I = 1 if
the person moves and I = 0 if they remain where they are. Equation (6)
can be a multinomial decision, with the number of alternatives


17
between 31 origin and destination combinations is presented. In both
chapters forecasting accuracy will be tested.
The final chapter will review the findings of the study and place
them into the context of the migration literature. The strengths and
weaknesses of the research will be emphasized. The implications of the
study for theory, policy, and future work will be discussed.


92
Per-capita state and local tax collections are also believed to
make a difference in the location choice. Frequently, migrants are
escaping areas with high state income and local property taxes. Again a
negative sign was expected here, implying that individuals would like to
reduce their tax burden upon moving. But, as before, the variable
showed up to be insignificant.
The final variable in the model of Table 4.3 is one which reflects
employment conditions in sending and receiving areas. It is believed
that individuals will choose to live at locations where there are a
relatively large number of employment opportunities. At the same time,
an individual from an area of high unemployment is expected to be even
more responsive to job availability at other places. This is particularly
the case if he (or she) is currently unemployed. As in other studies,
though, the unemployment-differential variable is of the wrong sign and
insignificant. Other variables representing employment opportunities,
such as the rate of employment growth, were substituted. Still there
was no success.
When the estimated equation is compared to one in which all co
efficients are restricted to zero, the model turns out significant at
the .005 level. When compared to an equation in which the right-hand
side is set equal to the mean value of the dependent variable, the
equation is only significant at the .05 level.
The cost of living and unemployment variables are removed from the
model and the equation is re-estimated. The results are shown in Table
4.4. Distance is now significant at the .005 level. The temperature
difference variable remains extremely significant. The difference in
average annual per-capita income variable is now also significant at the
.005 level.


42
Table 3.1 Summary of Panel Study Migration, 1968-1977
Number of In- Number of Out-
State
Migrants
Migrants
Net Migration
Alabama
7
3
4
Arizona
11
6
5
Arkansas
3
11
-8
California
44
43
-4
Colorado
16
6
10
Connecticut
8
7
1
Delaware
District of
3
0
3
Columbia
9
12
-3
Florida
43
11
32
Georgia
9
5
4
Idaho
2
0
2
Illinois
23
22
1
Indiana
11
14
-3
Iowa
4
17
-13
Kansas
6
0
6
Kentucky
5
14
-9
Louisiana
5
13
-8
Maine
7
1
6
Maryland
18
9
9
Massachusetts
14
10
4
Michigan
10
10
0
Minnesota
5
7
-2
Mississippi
4
16
-12
Missouri
4
19
-15
Montana
0
0
0
Nebraska
7
3
4
Nevada
7
0
7
New Hampshire
5
0
5
New Jersey
21
15
6
New Mexico
4
0
4
New York
25
30
-5
North Carolina
5
12
-7
North Dakota
0
0
0
Ohio
11
37
-26
Oklahoma
5
4
1
Oregon
17
8
9
Pennsylvania
15
28
-13
Rhode Island
0
0
0
South Carolina
5
33
-28
South Dakota
1
5
-4
Tennessee
10
5
5
Texas
39
14
2-5
Utah
1
7
-6
Vermont
1
0
1
Virginia
22
15
7


55
recreational activities. Higher income people are also more likely to
have visited other states as tourists. Thus, they would have obtained
information about these states which could lead to increased likelihood
of moving. The data in Table 3.4 show significant differences between
family income levels of movers and nonmovers. Migrants earned an average
of $11,240 in 1968. This compares to an average income level of $8,790
for nonmigrants.
Empirical Results
Introduction
The comparison of migrant and nonmigrant characteristics presented
so far serves only to provide a general picture of relationships in the
data. Much can be hidden in such a rudimentary analysis. Since all
four sample states were grouped together, differences in migrant selection
between states are hidden. As an example, migrants from New York may
tend to have larger families than New Yorkers who do not move. There
may, however, be little difference in the family size of movers and non
movers from Ohio, Pennsylvania, and South Carolina. When all four
states are viewed together, very little difference in family size will
show up, even though this is an important determinant in New York. This
problem becomes particularly severe if the variable takes on opposite
signs in each state.
Theoretical reasons exist for expecting differences in the deter
minants of migration between states of origin. The crux of the argument
is that residents of different states face varying relative migration
opportunities. Residents of Ohio, for example, will experience a greater
improvement in climate from moving to Florida than will residents of


83
occurs in Ohio, the same state that was estimated to have an unusually
high level of out-migration in the 1972-76 period. Overall, about 4
million persons are forecast to leave the four states between 1976 and
1980, indicating that there will be a continuing large pool of migrants
from which Florida and other states may draw.
The reasonability of the method used to estimate and forecast
aggregate labor force migration can partially be determined by comparing
the results with measures of out-migration derived from the 1970 Census
of Population. These estimates are not directly comparable, however.
For one, the Census estimates cover the five-year period between 1965
and 1970, while the estimates in this study cover four-year periods.
Since there is some overlap the 1968-72 period is chosen for comparison.
An additional problem in comparing estimates from the two sources is
that the Census figures are not tabulated according to labor force
status in published reports.
Table 3.15 compares the 1965-70 Census estimates to the 1968-72
estimates of out-migration obtained in this study. It can be seen that
with the exception of Ohio, the estimates of migration derived in this
study exceed those obtained from the 1970 Census. Because the Census
estimates include all members of the population (whether they are in the
labor force or not) and because the Census data cover a longer period
the opposite result would be expected. On the other hand, it is generally
believed that there was significant undercount in the 1970 Census. In
addition, because of prosperous economic conditions the early 1970s may
have been years of greater rates of migration. It is known that in
Florida, for example, in-migration accelerated significantly during
this period. Since the origin states chosen for this study are some


9
viewpoint, involves a sacrifice in income. Contrary to regional growth
models, Graves views the regional system as fundamentally being in
equilibrium so that utility is constant over space. Migration then
takes place as a result of changes in demand for location-fixed amen
ities. These changes come about as a result of changed relative prices
and income. Under this system migration will not cause regional income
convergence since income differentials must be maintained in order to
compensate for differences in climate. Graves tests his theory by
analyzing net migration for a cross section of Standard Metropolitan
Statistical Areas during the 1960-70 period. A sample of his results
for white migrants is presented in Table 1.3. All climate variables are
significant. Median income is not significant, but it is more significant
than it is in a model with climate variables excluded. In addition,
when Graves disaggregated the sample by age, he found income to be a
very significant determinant of migration for some age groups. The
unemployment rate, as expected, is negative and significant. As with
income, when climate is excluded from the equation, the unemployment
rate loses significance. The results demonstrate than when employment
and income possibilities are the same across all locations, people
choose to move to more temperate climates. Alternatively, if climate is
held constant across alternatives, people will choose to move to places
offering greater income and employment opportunities. When climate is
excluded, and thus allowed to vary, income and employment possibilities
by themselves are no longer such important migration determinants.
Thus, there is evidence that opportunities and climate interact in the
way suggested by Graves.


22
The model developed above belongs to a general classification
called random utility models. (For a discussion of these models see
Albright, Lerman, and Manski [1977].) Alternative methods of estimating
this model have been attempted in this study. The technique which at
first appeared to be ideal will be put forth in the next section.
Reasons for the abandonment of this approach will then be given. These
reasons are both theoretical and practical. The method finally settled
upon will be discussed in detail. Its strengths and weaknesses in
studying location choice will be highlighted.
Methodology
An Ideal Methodology
Equation (2), as it stands, cannot be estimated. This is because
utility is unobservable. We can, however, observe actual location
choices. Define I = 1 if individual t chooses alternative i. Define I
= 0 if some other location is chosen. Then the utility function can be
rewritten:
I = Xtetx + Pi0tp + eti (4)
Given the assumptions of the theory, if I = 1,then we know that >
Utj for all j not equal to i. If I = 0, we then conclude that another
location allows the individual to obtain a higher level of utility.
Equation (4) is an estimable equation. It can be applied to both
single-choice and multiple-choice situations. For multiple-choice
situations it is assumed that the individual is making any number of
concurrent binomial decisions. For example, a person residing in New
York and deciding whether to move to Florida, migrate to California, or
remain in New York can be said to be faced with three binomial decisions:


97
Table 4.6 Logit Results: Destination Choice, 1968-1972
Variable
Coefficient Asymptotic t-Ratio
Distance
-.001 -2.15
Tempdiff
.120 1.59
Incdiff
.001 2.20
Summary Statistics
Log of Likelihood
Function = -74.0
Likelihood Statistic (at 0) = 7.4
Level of Significance (at 0) = .10
Likelihood Statistic (at mean) = 7.4
Level of Significance (at mean) = .10


102
Table 4.9 Forecast: Destination Choice, 1976-1980
Origin-
Number of Migrants-
Number of Migrants-
Destination
Panel Study
Aggregate Labor Force
New York-Florida
4
331,000
New York-California
3
252,000
New York-New Jersey
3
199,000
New York-Illinois
2
146,000
New York-Virginia
2
132,000
New York-Michigan
2
119,000
New York-Texas
2
159,000
Pennsylvania-Florida
3
148,000
Pennsylvania-California
2
108,000
Pennsylvania-New Jersey
2
87,000
Pennsylvania-New York
2
87,000
Pennsylvania-Illinois
1
67,000
Pennsylvania-Virginia
1
54,000
Pennsylvania-Michigan
1
54,000
Pennsylvania-Texas
1
67,000
Ohio-Florida
3
114,000
Ohio-California
2
103,000
Ohio-New Jersey
1
60,000
Ohio-New York
1
60,000
Ohio-Illinois
1
60,000
Ohio-Virginia
1
43,000
Ohio-Michigan
1
43,000
Ohio-Texas
1
60,000
South Carolina-Florida
4
59,000
South Carolina-California
3
41,000
South Carolina-New Jersey
2
25,000
South Carolina-New York
1
23,000
South Carolina-Illinois
1
20,000
South Carolina-Virginia
1
18,000
South Carolina-Michigan
1
16,000
South Carolina-Texas
2
25,000


114
South Carolina migrants also tend to be older and earn higher
incomes than nonmigrants from that state. Married persons are less
likely to move than those who are not married. Finally, migrant fam
ilies tend to be of larger size than families that remain in South
Carolina. Household size, a priori, was expected to act as a constraint
upon migration.
Education proved to be an important determinant of migration in
Pennsylvania, with more educated people being more likely to leave. But
migrants from the state were also likely to earn lower family income
than those who remained. For New York, Ohio, and South Carolina, the
evidence confirmed the view that increases in income increased the
demand for location-specific goods found in such states as Florida or
California. For Pennsylvania, it appears that greater income increases
the probability of remaining in that state.
The positive income effects found for three of the four states
conflicts with DaVanzo's (1977) finding that the migration process, or
the returns to it, are on balance an inferior good. The difference in
results may arise because this study is only considering states which
have typically sent migrants to Sunbelt states. Sunbelt migrants
typically move in search of increased leisure activities, the demand for
which is expected to increase with income. The additional discovery
that migrants from these states are usually older than nonmigrants adds
further credence to this theory. It is the older migrant planning for
future retirement who is likely to be seeking out more temperate climates.
Those who have in the past earned higher incomes have the greatest
ability to make such long distance, and sometimes income sacrificing.
moves.


3
section will focus upon literature that seeks to explain migration
decisions at the level of the individual.
Aggregate Migration Models
Aggregate models of migration are subclassified into two types
suggested by Greenwood (1975). The first type is concerned with gross
migration, defined as a single flow or the sum of unidirectional flows
of population. The second type is concerned with net migration, defined
as the difference between gross flows occurring in opposite directions.
For any region total net migration is simply the difference between the
number of in-migrants and the number of out-migrants over a specified
time period.
A typical gross-migration model is provided by Greenwood (1969) in
his analysis of migration determinants. Greenwood examines a cross
section of flows between all 48 mainland states during the 1955-60
period. His regression results are presented in Table 1.1. Positive
and significant variables in Greenwood's final equation are the average
level of education at origin locations, the unemployment rate at the
origin, the percentage of urban population at the destination relative
to the percentage of urban population at the origin, the mean annual
temperature at the destination relative to the mean annual temperature
at the origin, and the number of persons born at the origin who already
reside at the destinationcalled the migrant stock. Also significant,
but showing a negative sign, are the distance between origin and desti
nation and the average level of education at the destination. The best
explanatory variable in Greenwood's model is the migrant stock variable.
This variable is believed to be a proxy for the amount of information


82
Table 3.14 Aggregate Labor Force Migration: 1968-1972,
1972-1976, 1976-1980
State of Number of Migrants Number of Migrants Number of Migrants
Origin 1968-1972 1972-1976 1976-1980
Ohio
714,000
1,171,000
833,000
New York
1,708,000
1,643,000
1,867,000
South Carolina
310,000
341,000
309,000
Pennsylvania
1,121,000
1,131,000
1,003,000
Total
3,853,000
4,286,000
4,012,000


115
Over time the migration determinant coefficients did not appear to
be very stable. But, over shorter time periods, there are fewer migrants
available for study. In addition, fewer variables could be considered
in logit estimation, leading to the possibility of misspecified equations.
A decline in the total number of out-migrants is forecast for the 1976
80 period. This is in line with previous expectations that the migration
surge of the early 1970s cannot be maintained.
Review of Findings: Destination Choice
All other things constant, migrants from Ohio, Pennsylvania, South
Carolina, and New York chose to move to states offering greater per-
capita income. This result, at first, seems to offer support to re
gional growth theories, such as that of Smith (1975), which portray
migration solely as a response to income or wage differentials between
regions. But it was discovered that income is only significant when a
variable reflecting differences in climate is included in the model.
Climate turned out to be the most significant determinant of destination
choice. The downward bias on the income coefficient when climate was
excluded from the model indicates a strong negative correlation between
income and climate. Support is thus lent to the Graves (1979) thesis
that long-distance migration involves a trade-off between income and
location-specific, climate-related amenities.
When destination-choice models were estimated for shorter time
periods, coefficient instability was discovered. For the 1972-76
period, climate was found to be a more important determinant of location
choice than during the 1968-72 period. Income,on the other hand,


APPENDIX B
ALTERNATIVE MODEL FORMULATIONS: DESTINATION CHOICE
Nominal Wage Model
Table B.l presents results obtained when a measure of individual
returns to migration is included as an explanatory variable in the
model of destination choice. Wages at each location are estimated
as a function of age, race, sex, marital status, education, union
membership, experience, and occupation. Using the procedures described
in Chapter 2, values of potential wages at each origin and destination
are estimated for every person in the sample. For each individual,
the variable Wdiff is defined as the difference between his (or her)
potential 1977 hourly wage at each possible destination and the potential
wage at his (or her) origin location. The remaining variables are
defined as they were in Chapter 4.
In this model, only climate (Tempdiff) turns out to be a signi
ficant determinant of destination choice. Both distance and potential
wage gains turn out insignificant. Overall, the model exhibits a
poorer fit than that obtained when the individual returns variable is
replaced with the average aggregate per-capita income gain variable used
in Chapter 4.
Real Wage Model
The wage-difference variable defined in the previous section was
deflated by state cost-of-living indices to obtain a new variable
128


48
Table 3.4 Comparison of Migrant and Nonmigrant Characteristics:
New York, Ohio, Pennsylvania, and South Carolina
Variable
Mean Value Migrants
Mean Value Nonmigrants
Famsz
4.98
4.95
Sex
0.87
0.84
Homeown
0.64
0.58
Lres
5.93
5.87
Lempl
3.14
3.31
Emself
0.10
0.08
Avage
37.89
35.31
Ernst
0.96
0.98
Marst
0.78
0.81
Race
0.71
0.68
Aveduc
4.11
3.78
Prevmig
0.31
0.23
Emstw
0.38
0.35
Famy
11.24
8.79


62
value of the equation is tested against that which results when the
right-hand side of the equation is set equal to the mean of the dependent
variable, the equation is significant at the .025 level. Under either
test, the equation appears to produce a good fit.
A correct prediction is defined as one in which the predicted
probability is within 50 percent of the actual value of the dependent
3 p
variable. For migrants this means that move > .50. For nonmigrants
this means ^move < .50. Under this criterion, the equation predicted 85
percent of all (migrant and nonmigrant) individuals in the sample correctly.
Among migrants, 16 percent were correctly predicted.
Logit Results: New York
Table 3.6 presents the decision-to-move results for the state of
New York. Variables which are significant at better than the .05 level
are length of employment, average age, employment status, and spouse's
employment status. Length of employment carries a negative sign, implying
that those who have worked for the same employer for long periods of
time are less likely to move. This result is expected. Migrants from
New York are also older than nonmigrants from that state, providing
further support for the preretirement thesis. Migrant household heads
are less likely to be employed than nonmigrants, but their spouses are
likely to be employed. This last result, which was also found to be
true for Ohio, is contrary to expectations. Perhaps what we are observing
3This definition is admittedly arbitrary. Some cutoff had to be
chosen in order to summarize the results without listing every predicted
value.


30
elderly migrants returning to their state of birth have lower socio
economic characteristics than other movers. Thus, only nonreturn
migrants are considered in this analysis.
The decision equation will be estimated using a logit program
developed by Nerlove and Press (1973). Using this technique, the
estimate of I\ in equation (5) is equal to the log odds of one of
the alternatives. The estimated equation can be written:
log I ^stay
1-P
= X
stay/
ttx
(7)
where Pstay is the probability of staying, i.e., not moving. We can
now solve for Pstay:
Pstay = ---- y a
1 + e~Xt0tx
Since there are only two alternatives, the probability of moving is
Pmove = 1 Pstay = --T a (9)
1 + e^-ttx
This value will be derived for every individual in the sample and com
pared with each person's actual choice. In addition to comparing actual
and fitted values, the overall fit of the model can be determined by
comparing the log likelihood value at model convergence with the log
likelihood value that results if all coefficients are restricted to
zero. It has been shown (see McFadden [1974]) that the statistic
-2 [ In (1^) In (1) ] (10)
is distributed approximately chi-square, where lr is the likelihood
/s
value when the coefficients are restricted and 1 is the likelihood
value for unconstrained model at its maximum. The degrees of freedom
are equal to the number of restricted coefficients.


54
to have developed attachments to their present state of residence. In
the four-state sample, 31 percent of the migrants have moved before.
Only 23 percent of the nonmigrants have previously changed their states
of residence. Thus, there is support for the hypothesis set forth
above.
Employment Status of Spouse
The presence of a working spouse is expected to deter migration.
If both partners have substantial earnings, any migration decision must
presumably maximize their combined return. The probability of finding a
location that meets this criterion is lower than the chances of finding
one which maximizes the returns of one family member. It is possible
that only one member of the couple will be properly rewarded upon moving
and the spouse will be tied to this choice. Whether joint utility is
maximized or not, the spouse's employment and income possibilities must
be taken into account in any decision to move. Theoretically, this
should act as a deterrent to migration; however, preliminary evidence in
Table 3.4 suggests that migrants in the sample were more likely to have
working spouses. Thirty-eight percent of the migrants' spouses worked
prior to moving.
Family Income
Migrants are expected to earn higher annual income than nonmigrants.
A few reasons can be given for this belief. First, higher incomes
usually imply greater demand for leisure activity. Changing location is
one way of obtaining more leisure. This is particularly true for this
sample of states which have typically sent large numbers of migrants to
Florida and California, states which possess climates amenable to


6
Table 1.2 Lowry Net Migration Regression Results
Variable Coefficient t-Statistic
dP. -.62515 -5.24
i
dQ. 1.49021 52.56
i
dA. 1.44409 7.36
i
Variable Definitions:
dP. -net change in the number of residents 15-64 years of age
in the absence of migration
dQ_^ -net change in civilian nonagricultural employment
dA_^ -net change in number of armed services personnel


25
depending upon the number of locations available to the individual.
Utilizing this framework relieves to some extent the econometric problems
encountered when the entire decision was placed in one equation. The
"whether to move" equation is now easier to estimate and interpret since
all migrants are grouped into one category. The "where to move" equation
also contains more balance among alternatives. In addition, since the
sample size has been reduced substantially by considering only migrants
in this equation, estimation is now less costly.
Viewing the migration decision under a two-stage regime involves
making the assumption that these decisions are independent of each
other. This is the major weakness of this approach. It can readily be
argued that the decision to move and destination choice occur jointly in
many individual cases. But, given currently available techniques, the
costs of modeling them together appear to outweigh the benefits derived
from this approach. A detailed description of the procedures used in
estimation and application of equations (5) and (6) begins in the next
section.
The Decision to Move
As discussed earlier, if two people are choosing from a set of
similar location bundles, one may move and one may remain at his (or
her) current residence. Characteristics of the individual, contained in
the vector X^, are believed to determine who is selected to migrate.
These characteristics may represent differences in opportunities. For
instance, the older person may have a lower return from migration in
terms of lifetime income. Differences in individual attributes may also
represent differences in the way people evaluate the same opportunities.


39
Typical sources of data for migration research are the Continuous
Work History Sample of the Social Security Administration and the Public
Use Sample derived from the 1970 Census of Population. The Panel Study's
advantage over these sources lies in its richness of variables and in
its continuity and consistency of sampling. Its disadvantage is that it
is a much smaller sample. But since its patterns of movement over the
sample period closely resemble the movement identified in these other
sources, great confidence is placed in its use as a tool in migration
studies.
Aggregate data is obtained from various sources. Chief among these
is the Statistical Abstract of the United States. This is one of the
few publications that contains easily accessible and reasonably consistent
time-series data for all states. Other sources are Climatological Data,
National Summary, Cost of Living Indicators, and The Employment and
Training Report of the President.
Sample Size Limitations
Dividing up the sample according to state of origin reduces the
number of observations for each estimation substantially. For the state
of Ohio, 220 observations are drawn from the Panel Study tapes. Thirty-
seven of these individuals (and their families) moved out of Ohio in the
sample period. There are 195 New York observations, 29 of which are
migrants. Pennsylvania has 202 observations and 28 of these moved out
of state during the nine-year period. The sample size for South Carolina
is 222, with 32 being migrants. The destination-choice equations grouped
together 85 migrants from all four states who chose among eight potential
destinations.


AN ECONOMETRIC MODEL OF INTERSTATE
LABOR FORCE MIGRATION
BY
SHELDON DONALD ENGLER
A DISSERTATION PRESENTED TO THE GRADUATE COUNCIL
OF THE UNIVERSITY OF FLORIDA IN
PARTIAL FULFILLMENT OF THE REQUIREMENTS
FOR THE DEGREE OF DOCTOR OF PHILOSOPHY
UNIVERSITY OF FLORIDA
1979

TO MY PARENTS

ACKNOWLEDGEMENTS
The author wishes to thank Professor Jerome Milliman for his guid
ance and support throughout the course of study. Special thanks are
given to Professor Henry Fishkind for his intellectual contributions
and his continued friendship. Acknowledgements are also extended to
Professors Stanley Smith, David Denslow, John Henretta, Angela O'Rand,
and William Tyler. Finally, the author would like to thank Alee
Williams and Doreen Willmeroth, whose typing and editing skills were
invaluable to the production of the finished paper.
iii

TABLE OF CONTENTS
PAGE
ACKNOWLEDGEMENTS iii
LIST OF TABLES vi
LIST OF FIGURES ix
ABSTRACT x
CHAPTER
1 INTRODUCTION 1
Problem Statement 1
Literature Review 2
Overview of the Study 16
2 THEORY AND METHODOLOGY 18
Theoretical Model 18
Methodology 22
A Forecasting Methodology 34
Sources of Data 38
3 EMPIRICAL RESULTS: DECISION TO MOVE 41
Interstate Migration: 1968-1977 41
Comparison of Migrants and Nonmigrants 45
Empirical Results 55
Application to Forecasting 72
4 EMPIRICAL RESULTS: DESTINATION CHOICE.. 86
Destination Choice: 1968-1977 86
Logit Results: Destination Choice 86
Application to Forecasting 96
Income, Climate, and Migration 107
5 SUMMARY AND CONCLUSIONS 110
Introduction 110
Review of Findings: Decision to Move Ill
Review of Findings: Destination Choice 115
Strengths and Weaknesses of the Study 116
Implications of the Study 118
iv

PAGE
APPENDIX
A ALTERNATIVE MODEL FORMULATIONS: DECISION TO MOVE 120
Combined State Results 120
Annual Migration Results 122
Redefinition of Qualitative Explanatory Variables 122.
B ALTERNATIVE MODEL FORMULATIONS: DESTINATION CHOICE 128
Nominal Wage Model 128
Real Wage Model 128
BIBLIOGRAPHY 132
BIOGRAPHICAL SKETCH 136
v

LIST OF TABLES
PAGE
Table 1.1 Greenwood Gross Migration Regression
Results 4
Table 1.2 Lowry Net Migration Regression Results 6
Table 1.3 Graves Net Migration Regression Results 10
Table 1.4 DaVanzo Destination Choice Results 15
Table 3.1 Summary of Panel Study Migration, 1968-1977 42
Table 3.2 Florida In-migration by State of Origin,
1968-1977 44
Table 3.3 Variable Definitions for Decision to
Migrate Analysis 46
Table 3.4 Comparison of Migrant and Nonmigrant
Characteristics: New York, Ohio,
Pennsylvania, and South Carolina 48
Table 3.5 Results of Logit Analysis: Decision to
Move, Ohio, 1968-1977 60
Table 3.6 Results of Logit Analysis: Decision to
Move, New York, 1968-1977 63
Table 3.7 Results of Logit Analysis: Decision to
Move, South Carolina, 1968-1977 65
Table 3.8 Results of Logit Analysis: Decision to
Move, Pennsylvania, 1968-1977 67
Table 3.9 Ordinary Least Squares Results: Decision
to Move, 1968-1977 70
Table 3.10 Results of Logit Analysis: Decision to
Move, 1968-1972 73
Table 3.11 Comparison of Forecast to Actual Migration,
1972-1976 75
Table 3.12 Results of Logit Analysis: Decision to
Move, 1972-1976 77
vi

PAGE
Table 3.13 Forecast: Decision to Move,
1976-1980 80
Table 3.14 Aggregate Labor Force Migration:
1968-1972, 1972-1976, 1976-1980 82
Table 3.15 Comparison of Panel Study Derived
Out-Migration Estimates to Census
Estimates 84
Table 4.1 Destination Choice Summary:
1968-1977 87
Table 4.2 Variable Definitions: Destination
Choice Equations 88
Table 4.3 Logit Results: Destination Choice,
1968-1977, Model 1 89
Table 4.4 Logit Results: Destination Choice,
1968-1977, Model 2 93
Table 4.5 Logit Results: Destination Choice,
1968-1977, Climate Excluded 95
Table 4.6 Logit Results: Destination Choice,
1968-1972 97
Table 4.7 Comparison of Forecast to Actual
Destination Choice: 1972-1976 99
Table 4.8 Logit Results: Destination Choice,
1972-1976 101
Table 4.9 Forecast: Destination Choice,
1976-1980 102
Table 4.10 Aggregate Destination Choice,
1976-1980 104
Table 4.11 Comparison of Panel Study Derived
Migration Flows to Census Estimates 105
Table 4.12 Climate and Income: 1968, 1972,
1976 108
vii

PAGE
Table 5.1 Summary of Findings: Decision
to Move 112
Table A.1 Combined State Results:
Decision to Move, 1968-1977,
Ordinary Least Squares 121
Table A.2 Annual Migration Results:
Decision to Move, 1976-1977,
Ordinary Least Squares 123
Table A.3 Results of Logit Analysis with
Alternative Dummies, 1968-1977 .... 124
Table B.l Nominal Wage Model: Destination
Choice, 1968-1977 129
Table B.2 Real Wage Model: Destination
Choice, 1968-1977 131
viii

LIST OF FIGURES
PAGE
Figure 2.1 Illustration of Forecasting Methodology:
Decision to Move 36
Figure 2.2 Illustration of Forecasting Methodology:
Destination Choice 37
xx

Abstract of Dissertation Presented to the Graduate Council
of the University of Florida in Partial Fulfillment of the
Requirements for the Degree of Doctor of Philosophy
AN ECONOMETRIC MODEL OF INTERSTATE LABOR FORCE MIGRATION
By
Sheldon Donald Engler
December 1979
Chairman: Jerome W. Milliman
Major Department: Economics
The purpose of this study is to analyze and forecast migration
decisions of labor force members. Location choice is viewed within
a theoretical framework which assumes that the individual is a utility
maximizer. Location bundles which include some goods available at all
locations and other goods which are location-specific can be varied by
moving. Each person reaches his (or her) optimum by first considering
the opportunities which are available at all potential locations.
Differences in individual characteristics and circumstances are
hypothesized to cause people to evaluate these opportunities in
alternative ways.
A two-step methodology is adopted for the empirical section of
this study. First, the "whether to move" decision is modeled for all
sample members. Then, the choice of "where to move" is analyzed for
the proportion of the sample who migrated. The first decision is
considered for four separate states. States were chosen based upon
historical rates of out-migration, particularly to Florida. The
x

second model focuses upon the destination choice of those who actually
moved from these states. Eight alternative states are considered as
destinations.
Although differences in migration determinants between states are
discovered, there are some general findings. The typical migrants in
the sample tend to be older and earn higher income than do nonmigrants.
Family relationships also appear to be important influences on the
migration decisions. Migrants are also less likely to be married than
nonmigrants. Contrary to expectations, families with two income earners
appear to move just as often as families where only one person is employed.
Destination choice is shown to be determined by income opportunity and
climate. Evidence that improved climate can sometimes only be obtained
at the cost of reduced income is discovered.
Overall, the migrant appears to be older, richer, and more willing
to take a cut in income for better climate than the nonmigrant. One
possible interpretation of the results is that many interstate moves are
being made in anticipation of retirement. Those who have earned greater
lifetime income are most able to absorb the decline in earnings consequent
upon moving prior to retirement. Economic theory postulates that higher
income people will demand greater quantities of leisure activities.
This increase can be obtained through the migration process. Locations
with warm weather have historically offered greater leisure possibilities
than places with colder, more variant climate.
The results are utilized to obtain interstate migration forecasts.
Rising family incomes, the process of population aging, and an assumption
of the continued importance of climate-related amenities in the migration
xi

decision lead to a forecast of further migration concentrated toward
Sunbelt destinations. Policies aimed at restricting this growth or
attracting migrants to other locations could alter this pattern.
xii

CHAPTER 1
INTRODUCTION
Problem Statement
During recent years experts on migration have begun to focus their
efforts upon understanding the decision to move at the individual level.
A continuous stream of detailed survey data has helped to bring about
this endeavor. In addition, increased availability of sophisticated
computer software packages and advances in econometric techniques have
facilitated the handling and analysis of this data. Until now the
research which has come about with the aid of these tools has been
limited to answering age-old questions in the migration literature.
What are the characteristics which distinguish migrants from nonmi
grants? Do individuals move in the direction of higher wages? How does
the unemployment rate affect the migration decision? What role does
climate play in the choice of location? Does migration lead to regional
income convergence?
Since many of these issues are as yet unsettled, they will be
considered here. The ultimate test of a theory, however, is its ability
to predict the future. With the proportion of population change re
sulting from natural increase on the decline, migration forecasts become
more of a necessity. This is particularly true for states which are
experiencing rapid population growth such as Florida, California, Texas,
and Arizona. In addition, forecasts of other economic events at the
1

2
regional level depend heavily upon the ability to predict population.
The demand for housing, the unemployment rate, and state and local
government tax collections are examples of economic variables which are
sensitive to the rate of population change. Knowledge of the level of
future population and its consequences for the rest of the economy is
crucial to success in planning for regional economic growth.
In this study, an econometric model is developed for use in ex
plaining and forecasting the movement of workers and their families
between states. A general theoretical model of interstate migration is
derived within a microeconomic framework. This model is then placed
into an estimable form and tested for the period 1968-77. Individual
data obtained from the University of Michigan Panel Study of Income
Dynamics are used. Maximum likelihood estimation techniques are em
ployed. Finally, the results of the analysis are utilized in devel
oping an aggregate forecast of interstate labor force migration.
Throughout the study the case of Florida is emphasized. Separate
models are estimated for four typical origin states of Florida migrants.
The determinants of Florida's attractiveness relative to other desti
nations which draw migrants from these states are analyzed.
In the next section the existing migration literature will be
brought up to date. Then an overview of this study will be presented.
Literature Review
Introduction
Existing studies will be reviewed according to two major cate
gories. The first section which follows will summarize research that is
centered upon explaining aggregate flows of population. The second

3
section will focus upon literature that seeks to explain migration
decisions at the level of the individual.
Aggregate Migration Models
Aggregate models of migration are subclassified into two types
suggested by Greenwood (1975). The first type is concerned with gross
migration, defined as a single flow or the sum of unidirectional flows
of population. The second type is concerned with net migration, defined
as the difference between gross flows occurring in opposite directions.
For any region total net migration is simply the difference between the
number of in-migrants and the number of out-migrants over a specified
time period.
A typical gross-migration model is provided by Greenwood (1969) in
his analysis of migration determinants. Greenwood examines a cross
section of flows between all 48 mainland states during the 1955-60
period. His regression results are presented in Table 1.1. Positive
and significant variables in Greenwood's final equation are the average
level of education at origin locations, the unemployment rate at the
origin, the percentage of urban population at the destination relative
to the percentage of urban population at the origin, the mean annual
temperature at the destination relative to the mean annual temperature
at the origin, and the number of persons born at the origin who already
reside at the destinationcalled the migrant stock. Also significant,
but showing a negative sign, are the distance between origin and desti
nation and the average level of education at the destination. The best
explanatory variable in Greenwood's model is the migrant stock variable.
This variable is believed to be a proxy for the amount of information

4
Table 1.1 Greenwood Gross Migration Regression Results
Variable
Coefficient
t-Statistic
D. .
ij
-.300
-11.21
Y. .
Ji
.160
1.27
E.
l
3.401
16.60
E.
J
-.622
-2.95
U.
i
.705
8.44
u.
1
-.057
-.66
R. .
Ji
.771
7.52
T. .
li
.903
6.49
MS. .
.521
42.06
Variable
Definitions:
D. .
11
-distance from origin to destination
Y. .
li
-income at destination divided by income at
origin
E.
l
-median education at origin
E.
1
-median education at destination
U.
i
-unemployment rate at origin
U.
1
-unemployment rate at destination
R. .
li
-percent of urban population at destination
percent of urban population at origin
divided by
T. .
li
-mean annual temperature in principal city of destination
divided by mean annual temperature in principal city of
origin
MS. .
il
-number of people residing at destination who were born at
origin

5
concerning the destination that is available at the origin. That is,
previous migrants channel information back to potential future migrants.
And, assuming that it is of an encouraging nature, more information
should lead to increased migration.
The interpretation of the migrant stock is subject to some con
troversy. Laber (1972) argued that since the migrant stock is itself a
function of those factors influencing previous migration, it may be
acting as a proxy for lagged explanatory variables. If this is the
case, then a more appropriate specification may be a partial adjustment
model where lagged migration is included as an explanatory variable.
Dunlevy and Gemery (1977) estimate alternative specifications of the
model and conclude that it would be correct to include both migrant
stock and lagged migration variables. They argue that including one
variable while excluding the other results in the included variable
capturing parts of both effects. Thus, it is concluded that both the
information-creating effects of the migrant stock and a lagged migration
adjustment process are operating simultaneously.
A typical model of net migration is developed by Lowry (1966).
Lowry examines a cross section of Standard Metropolitan Statistical
Areas over the period 1950-60. His results are summarized in Table 1.2.
Significant variables in this model are the growth of the resident
working-age population (inversely related to net migration), employment
growth (positively related to net migration), and growth in armed
services personnel (positively related to net migration). Employment
growth is the best variable, the idea simply being that higher growth
reflects greater demand for labor, which translates into greater em
ployment opportunities for potential in-migrants (making them more

6
Table 1.2 Lowry Net Migration Regression Results
Variable Coefficient t-Statistic
dP. -.62515 -5.24
i
dQ. 1.49021 52.56
i
dA. 1.44409 7.36
i
Variable Definitions:
dP. -net change in the number of residents 15-64 years of age
in the absence of migration
dQ_^ -net change in civilian nonagricultural employment
dA_^ -net change in number of armed services personnel

7
likely to migrate) and for residents (making them more likely to re
main) Growth in the resident working-age population is important
because higher growth implies an increased supply of labor occurring
from within the region, thus reducing the number of job opportunities
available to potential in-migrants. Growth in armed services personnel
is included since it is a component of population change that is not
normally explained by the other variables in the equation.
Neoclassical regional growth models theorize that migration is
induced solely by the existence of interregional wage differentials.
That is, migration occurs from low-wage regions (where labor is plen
tiful relative to capital) to high-wage regions (where labor is rela
tively scarce). Increased supply of labor in high-wage regions is
hypothesized to put downward pressure on wages, while decreased supply
of labor in low-wage regions exerts upward pressure on wages. Thus, the
model leads to an equilibrium where wage differentials between regions
are eliminated. Smith (1975) includes a labor sector in his neoclas
sical growth model which posits net migration to be a function of the
difference between the local wage and the national wage. He estimates
his model for various historical periods and generates mild support for
the convergence theory. This is in contrast to other tests of neoclas
sical growth theory, such as that undertaken by Borts and Stein (1964),
in which the theory is rejected. But Smith also discovered a decrease
in responsiveness of potential migrants to income differentials over
time. This he partly attributes to the existence of unemployment which
discourages migration. Neoclassical growth models, whether regional or
national in scope, assume full employment of labor.

8
Richardson (1973) criticizes neoclassical models for assuming that
individuals are simply maximizing their income in the migration decision.
He proposes a model (which is as yet untested) in which net migration is
a function of wage differentials, agglomeration economies, and locational
preferences. Agglomeration economies are divided into two types.
Household agglomeration economies refer to the benefits and costs of
life in large cities for households. They include benefits accruing
from larger labor markets, availability of leisure and cultural facil
ities, the quality of public service, and environmental amenities.
Business agglomeration economies refer to advantages which urban areas
offer businesses. They attract firms which lead to jobs which in turn
attract migrants. The level of agglomeration economies is thus hypoth
esized to influence the rate of in-migration to a region. Locational
preferences are meant to explain why individuals may remain in low-
income regions despite the existence of improved income opportunities
elsewhere. Thus, they measure the retentive power of a region and are
related to the rate of out-migration. Examples of factors which in
fluence locational preferences are community ties, sociocultural traditions,
and length of settlement.
In a recent study Graves (1979) rejects the theory that migration
is a response to interregional income differences. Rather, he advances
the view that any income differences which may exist between regions are
compensated for by differences in the amounts of other amenity-oriented
goods which are available. For example, an area which offers an income
advantage over other places is believed to offer a more objectionable
climate, allowing for less involvement in leisure or recreational
activities. Moving to a location with a better climate, under this

9
viewpoint, involves a sacrifice in income. Contrary to regional growth
models, Graves views the regional system as fundamentally being in
equilibrium so that utility is constant over space. Migration then
takes place as a result of changes in demand for location-fixed amen
ities. These changes come about as a result of changed relative prices
and income. Under this system migration will not cause regional income
convergence since income differentials must be maintained in order to
compensate for differences in climate. Graves tests his theory by
analyzing net migration for a cross section of Standard Metropolitan
Statistical Areas during the 1960-70 period. A sample of his results
for white migrants is presented in Table 1.3. All climate variables are
significant. Median income is not significant, but it is more significant
than it is in a model with climate variables excluded. In addition,
when Graves disaggregated the sample by age, he found income to be a
very significant determinant of migration for some age groups. The
unemployment rate, as expected, is negative and significant. As with
income, when climate is excluded from the equation, the unemployment
rate loses significance. The results demonstrate than when employment
and income possibilities are the same across all locations, people
choose to move to more temperate climates. Alternatively, if climate is
held constant across alternatives, people will choose to move to places
offering greater income and employment opportunities. When climate is
excluded, and thus allowed to vary, income and employment possibilities
by themselves are no longer such important migration determinants.
Thus, there is evidence that opportunities and climate interact in the
way suggested by Graves.

10
Table 1.3 Graves Net Migration Regression Results
Variable
Coefficient
t-Statistic
Me dine
.00162
1.07
Unemp
-2.906
-4.26
Warmth*
.0103
4.44
Cold
.00686
4.27
Antmvr
-.9989
-4.86
Annwnd
-2.967
-4.49
Annhum
-.7164
-4.74
Variable Definitions:
Medinc
-1960 Median Income
Unemp
-1960 Unemployment Rate
Warmth*
-Mean annual number of cooling degree
days (base = 65
F)
Cold
-Mean annual number of heating degree
days (base = 65
F)
Antmvr
-Annual temperature variance (average daily maximum
July temperature average daily minimum January
temperature)
Annwnd
-January and July average wind speed
Annhum
-January and July average humidity
*A11 climate variables defined as 1931 to 1960 averages

11
One of the advantages of models of net migration is that data on
the dependent variable are easily computed for most areas. Net migra
tion can be directly estimated by subtracting the population change due
to natural increase from the actual population change for any given
period. Measures of gross migration, in contrast, depend upon more
direct measurement techniques, such as surveys. However, as Lowry
(1966) notes, models of net migration reveal less about migratory be
havior and the decision to migrate than do models of gross migration.
This is partly because variables which are important in determining
unidirectional flows are reduced in significance when the measure of
migration used includes flows occurring in opposite directions. Models
of net migration, however, are useful in forecasting population change.
They are of particular value to regions for which migration has been the
dominant component of population change. The next section will describe
research which has focused upon explaining migration at the individual
or family level.
Individual Migration Models
Rothenberg (1977) suggests an approach to the study of migration
that focuses upon the individual. He notes that the migrant is self-
selected. That is, given the availability of similar sets of oppor
tunities, some individuals will migrate and some will not. The problem
is to determine the individual characteristics and circumstances that
cause people to evaluate their migration choice in different ways. Put
somewhat differently, one individual's maximizing decision may cause him
(or her) to migrate, while another individual facing similar opportuni
ties may choose not to move. Both persons may be acting rationally.

12
In order to test the validity of this hypothesis, a model must be esti
mated using individual data on non-migrants as well as migrants.
Morrison (1973) suggests the use of a theoretical framework which
combines two models of migration. First, he suggests a microeconomic
model which analyzes the decision "whether to move." This model, he
claims, should reflect the idea that individuals have variable decision
thresholds. That is, those with lower thresholds are more likely to
seek out and respond to opportunities elsewhere, and hence are more
likely to move. The level of threshold variability is determined by
such factors as the individual's position in the life cycle, occupa
tionally induced contraints on movements, and prior migration experi
ence. The second type of model suggested by Morrison is one which
allocates those who do move (as a result of the first decision) among
alternative destinations. Thus, this model addresses the question
"where to move." This model, he claims, is more macroeconomic in con
tent .
An empirical study by DaVanzo (1977) divides the migration choice
into the two-stage process suggested above. She looked at a sample of
married couples from the University of Michigan Study of Income Dynamic
First, she estimates "whether to migrate" equations. The dependent
variable in these equations is a zero-one dummy indicating whether or
not a move was made for each couple. A one-year time period (1971-72)
was selected. The explanatory variables can be categorized as employ
ment status, returns to migration, composition of family earnings,
location-specific assets (such as homeownership), previous migration,
age, and education. The basic conclusions can be summarized as follows
Families whose heads are unemployed or are dissatisfied with their jobs

13
are more likely to move than those whose heads are not searching for
work. Local employment conditions are more important in the migration
decisions of the unemployed than the employed. Unemployed persons and
others looking for work are more responsive to family income, origin
wage rates, and expected earnings increases than persons satisfied with
their jobs. Families are more likely to move if they have moved in the
recent past. Wives have a significant influence on the family's de
cision to migrate. Age and education are relatively unimportant in
explaining the migration of married couples.
DaVanzo estimates her "whether to migrate" equations using ordinary
least squares (OLS). However, as acknowledged by the author, use of OLS
when the dependent variable is dichotomous leads to inefficient (although
unbiased) estimates. In addition, the fitted equation will yield pre
dictions outside the zero-one range. This criticism is particularly
crucial to a model which is to be used for the purpose of forecasting.
DaVanzo, however, was only trying to identify the determinants of the
migration decision. She estimates one equation using a probit model, a
maximum likelihood technique, and demonstrates that there is little
difference between the derived coefficients and the coefficients that
result from using OLS.
After analyzing the decision "whether to migrate," DaVanzo esti
mates choice of destination equations. She analyzes the choice among
eight regions of the United States. The explanatory variables are the
present value of the difference between what each family could earn at
each destination and what the family could earn if it stayed at the
origin, the unemployment rate at alternative destinations, the distance

14
between each destination and the origin, and an interaction term between
the present value of the wage differences and a dummy indicating whether
a person had resided at that destination in the recent past. The
results appear in Table 1.4. None of the variables turn out to be
significant at better than the .10 level. Only the wage variable is
even close to being significant, indicating that families are likely to
move to destinations where the earnings gains are greatest. The sign of
the interaction term indicates that families are more likely to move to
an area where they have lived before than to one where they have never
lived, especially if the family earnings they could receive there are
higher than what they could earn by staying where they are. The unemploy
ment rate is insignificant and shows an unexpected positive sign. The
coefficient of distance shows that migrants are more likely to choose
closer destinations, although it too is insignificant. The model is
estimated using conditional logit, a maximum likelihood technique
designed to analyze multinomial choices among discrete alternatives.
DaVanzo considered the migration decision from the viewpoint of the
married couple. Mincer (1978) notes that there are so far very few
migration studies that consider the effect of family relations on this
decision. He points out that at the individual level a person should
choose to move to that location at which the return is at a maximum.
For a family the optimal move is one which maximizes the combined return
to the family. Whether this last criterion is satisfied or not, frequently
a family will end up at a location which does not reward every family
member with the maximum possible return. The members of the family who
do not reach their private optimum are called tied movers. Mincer
points out that the conflict which could result from tied migration can

15
Table 1.4 DaVanzo Destination Choice Results
Variable
Coefficient
Asymptotic t-Ratio
fam
PV. .
11
.00548
1.53
fam
PV.. Here Before .
il 1
.0139
1.46
Unemployment Rate
.00909
.05
Ln Distance ..
il
-.322
-.81
Variable Definitions:
f am
PV,, -present value of the difference between what the
family could earn at destination j and what it
could earn if it stayed in its 1971 location, i
Here Before -dummy that indicates whether the family
^ resided in area j recently (between 1968
and 1970)
Unemployment Rate -unemployment rate in 1971 at destination j
Ln Distance .. -natural logarithm of the distance between
origin i and destination j

16
only act as a deterrent to a possible move. This is particularly the
case where more than one family member is working. Thus, the increasing
proportion of women in the labor force is expected to have an inhibiting
effect upon migration. Mincer uses a scattering of data to test his
theories. He discovers, as did DaVanzo, that marriage itself reduces
migration and that migration rates are lower in families with employed
wives. The magnitude of the effect which the working wife has on the
migration decision also depends upon her share of family earnings.
Finally, he shows that tied migration of working wives frequently
results in lower earnings, unemployment, or labor force withdrawal.
Thus, male household heads usually dominate the choice of destination.
Overview of the Study
In Chapter 2 the migration decision is analyzed within the context
of utility theory. A theoretical model of interstate migration is then
derived. A methodology for estimating this model is put forth which
considers the decision to move and the choice of destination together in
one analytical framework. Problems that arise in implementing this
procedure are then discussed and an alternative is proposed. The new
method views the migration decision as a two-stage process. Finally, a
technique for forecasting migration decisions and destination choice is
proposed.
Chapters 3 and 4 present the empirical results. The decision to
move is analyzed in Chapter 3. Regression results are presented for
four states and forecasts of aggregate migration from these states are
developed. Chapter 4 deals with the choice of destination. The de
terminants of this decision are analyzed and a forecast of migration

17
between 31 origin and destination combinations is presented. In both
chapters forecasting accuracy will be tested.
The final chapter will review the findings of the study and place
them into the context of the migration literature. The strengths and
weaknesses of the research will be emphasized. The implications of the
study for theory, policy, and future work will be discussed.

CHAPTER 2
THEORY AND METHODOLOGY
Theoretical Model
The principle of utility maximization can be adapted to the mi
gration decision. This adaptation, however, is not straightforward.
Differences between the standard model of consumer behavior and the
approach to be taken here arise because of the nature of location as a
good. These differences are discussed in the following paragraphs.
The location decision involves a choice among mutually exclusive
alternatives. For some given time period, the individual lives at only
one location. Thus, the quantity consumed of various locations cannot
be adjusted in the same manner as are the relative amounts of other
goods. Any change in location can be viewed as a decision to give up
the entire quantity of one place in favor of consuming all of another
residence.1 This change, however, is not quite as drastic as it first
appears. To see why, location must be defined as a bundle of goods.
Some goods can then be common to more than one bundle. For example,
most food products can be purchased at any location in the United States.
LThe case of one person having two or more residences can be raised
as a contradiction to these statements. In defense, it can be argued
that this individual can only live at one place at a time and that each
move between residences involves a new choice among mutually exclusive
alternatives. Additionally, the person with multiple residences is
atypical.
18

19
A change in residence will, by itself, cause little change in quantity
of food consumed. Changes in food buying behavior will occur if a move
is accompanied by an income increase or by a change in relative food
prices.
There are also goods contained in the location bundle which can
sometimes only be obtained in varying quantities by moving. An example
is climate. Although typically defined as a characteristic of a place
rather than a good within a location bundle, climate does display
certain characteristics of goods. The quantity of climate can be
represented by measuring factors such as temperature or nearness to the
coast. The price can be measured in terms of moving costs or foregone
earnings involved in moving to a more temperate climate. So although it
is not produced by man, from the consumer's standpoint climate displays
characteristics that other goods possess. The need to move in order to
obtain substantial changes in consumption makes climate an important
determinant in the migration decision. This is true for other fixed-
location amenities such as symphonies, sporting events, and certain
public services.
Another important factor distinguishing locations for the consumer
is the existence of differential employment opportunities across space.
A person may have a potential job available in one place that will earn
him (or her) greater income than jobs at other locations may offer.
Higher income, in turn, allows for greater consumption. These increased
consumption possibilities include leisure activities which are related
to climate. Thus, although many of the goods in the location bundle are
available at all locations, different quantities may be consumed at
various locations because of the uneven spatial distribution of em
ployment and income possibilities.

20
In addition to items contained in the location bundles, it is
posited that individual characteristics and circumstances are deter
minants of migration decisions. These factors determine an individual's
responsiveness to opportunities contained in the location bundle. As an
example, older people move less often because of a shorter expected
lifetime over which they may experience migration-related gains and
because they are likely to possess location-specific capital, such as a
home. Or a person who has moved many times in the past may be more
susceptible to migration influences. An individual who has resided at
the same location for a long period of time may feel greater ties to
relatives, friends, and community. A married person may be less likely
to move than an unmarried individual, especially if the spouse works.
The location choice which maximizes utility for one member of the family
may not be the optimal choice for other family members. The end result
will depend upon the nature of family relations. The point here is only
that the existence of the marriage can create a conflict which affects
the migration decision. From the standpoint of the family as a unit the
optimal choice may still result.
Other factors entering into the migration decision will be dis
cussed when the final model is specified. The above discussion has
served mainly to illustrate differences between the study of location
choice and the analysis of other choices in which economists are typically
interested. With these considerations in mind, a model of individual
location choice will be derived.
Assume that the individual chooses to reside at that location from
which he (or she) derives maximum utility. Further, assume that the
utility function is linear in parameters with additive disturbances.
Define Z^ = (X^_, P^), where X^_ is a vector of independent variables

21
describing individual t, and P^ is a vector of explanatory variables
describing alternative i. Also define 9 = (0 9 ), where 9 and
t tx, tp tx
9 are vectors of parameters assigned to Xt and P^ respectively.
Finally, define e as a random disturbance for alternative i and
decision maker t. The utility which individual t derives from place i
can now be expressed as follows:
(1)
Alternatively, it can be written:
(2)
If the individual maximizes utility, then he or she will choose to live
U = Z^.6^ + £ .
ti tx t ti
U = X 0 + P.0 + e .
ti t tx i tp ti
at alternative i if U > U for all j not equal to i.
tx tj j u
Moving costs can be incorporated into the model. Assume first that
we find individual t living at place i. Then we conclude that if he (or
she) is rational: U > U C.. for all j not equal to i where C.. is
the cost of moving from i to j. If at a later time person t migrates to
place j, then we conclude that something occurred during the interim
which caused t's utility evaluation to change so that: U C.. >
13 ij
U Throughout the remainder of the study, actual moving expenses will
be ignored since it is believed that they are insignificant in relation
to the other costs and benefits of moving.
The probability that a person will choose to live at location i can
be written as follows:
(3)
A major goal of this study will be to impute values of P(i) for in
dividuals inside and outside the sample. Inferences about aggregate
behavior will also be made using these probabilities.
P(i) = P(U > U for all j t i)
tx tj J

22
The model developed above belongs to a general classification
called random utility models. (For a discussion of these models see
Albright, Lerman, and Manski [1977].) Alternative methods of estimating
this model have been attempted in this study. The technique which at
first appeared to be ideal will be put forth in the next section.
Reasons for the abandonment of this approach will then be given. These
reasons are both theoretical and practical. The method finally settled
upon will be discussed in detail. Its strengths and weaknesses in
studying location choice will be highlighted.
Methodology
An Ideal Methodology
Equation (2), as it stands, cannot be estimated. This is because
utility is unobservable. We can, however, observe actual location
choices. Define I = 1 if individual t chooses alternative i. Define I
= 0 if some other location is chosen. Then the utility function can be
rewritten:
I = Xtetx + Pi0tp + eti (4)
Given the assumptions of the theory, if I = 1,then we know that >
Utj for all j not equal to i. If I = 0, we then conclude that another
location allows the individual to obtain a higher level of utility.
Equation (4) is an estimable equation. It can be applied to both
single-choice and multiple-choice situations. For multiple-choice
situations it is assumed that the individual is making any number of
concurrent binomial decisions. For example, a person residing in New
York and deciding whether to move to Florida, migrate to California, or
remain in New York can be said to be faced with three binomial decisions:

23
A. Migrating to Florida or not
B. Migrating to California or not
C. Staying in New York or not
Values of I can now be assigned to each decision. If this person
chooses to migrate to Florida, then for decision A, I = 1; for decision
B, I = 0; and for decision C, 1=0. The values of X will vary across
alternatives. Individual-alternative interaction terms can also be
introduced into the model.
If each binomial decision by each individual is treated as a single
observation, then the model can be estimated by ordinary least squares
(OLS). It is well known, however, that use of OLS when the dependent
variable is dichotomous leads to inefficient estimates and to predictions
outside the zero-one interval. In addition, OLS is unable to distinguish
between individuals and observations in the multinomial case. More
appropriate methods use the maximum likelihood technique of estimation.
McFadden (1973) developed a technique he calls conditional logit
for use in analyzing consumer choice among lumpy alternatives. He
demonstrates the applicability of conditional logit in the study of
urban travel demand. In a recent study, Falaris (1978) applies the same
technique to the migration decision. An attempt was made in this study
to adopt the same methodology. It did not prove to be useful for this
study. The conditional logit model seems best adapted to problems where
the choice frequencies are fairly well balanced. Location choice
(studied over reasonable periods of time) is biased heavily toward a
single alternative: staying where one already is. Thus, when a model is
estimated which includes staying as one of many alternatives, this
choice swamps all others. Consequently, it is difficult to obtain

24
variables that distinguish between the remaining alternatives. The
magnitude of this problem enlarges as the number of alternatives in
creases. Falaris avoided this problem to some extent by restricting his
research to a choice among four broad regions of the United States.
Since this study will focus upon the state as the level of analysis, it
will be desirable to include a larger number of alternatives. A further
problem encountered with the model is that as new variables are added
when there are a large number of alternatives, the cost of estimation
rises substantially and the chances of early convergence diminish.
Falaris, in fact, considered only a small number of variables in his
study. Finally, testing various model specifications is difficult using
such high-cost procedures.
Proposed Methodology
The above considerations have led to the decision to analyze the
migration choice as a two-stage process. This is in line with the
approach suggested by Morrison (1973) and taken by DaVanzo (1977). The
first stage of the migration process is the decision "whether to move."
All members of the population are faced with this choice. The second
stage involves the decision "where to move." Only migrants face this
choice. The decision equation (4) is now divided into two:
(5)
(6)
I. = x 0 + e. .
l t tx tx
l = p.e, + u .
2 i tp ti
where U ^ is a random disturbance term and every other term is defined
as before. Equation (5) represents a binomial decision, with I = 1 if
the person moves and I = 0 if they remain where they are. Equation (6)
can be a multinomial decision, with the number of alternatives

25
depending upon the number of locations available to the individual.
Utilizing this framework relieves to some extent the econometric problems
encountered when the entire decision was placed in one equation. The
"whether to move" equation is now easier to estimate and interpret since
all migrants are grouped into one category. The "where to move" equation
also contains more balance among alternatives. In addition, since the
sample size has been reduced substantially by considering only migrants
in this equation, estimation is now less costly.
Viewing the migration decision under a two-stage regime involves
making the assumption that these decisions are independent of each
other. This is the major weakness of this approach. It can readily be
argued that the decision to move and destination choice occur jointly in
many individual cases. But, given currently available techniques, the
costs of modeling them together appear to outweigh the benefits derived
from this approach. A detailed description of the procedures used in
estimation and application of equations (5) and (6) begins in the next
section.
The Decision to Move
As discussed earlier, if two people are choosing from a set of
similar location bundles, one may move and one may remain at his (or
her) current residence. Characteristics of the individual, contained in
the vector X^, are believed to determine who is selected to migrate.
These characteristics may represent differences in opportunities. For
instance, the older person may have a lower return from migration in
terms of lifetime income. Differences in individual attributes may also
represent differences in the way people evaluate the same opportunities.

26
As an example, those who have moved more often in the past are more
likely to respond positively to migration opportunities.
Measures of migration opportunities could be included in the
"whether to migrate" function. DaVanzo (1977) constructs a return to
migration variable that represents the maximum return available from
moving. It is derived by estimating potential wages, via a human
capital wage model, at each possible destination and at the origin.
Then the differences between potential wages at the origin and each
destination are obtained. The maximum difference is used as the es-
2
timate of migration returns. Problems arise with this variable.
Since the migration decision is itself a function of individual character
istics, many of these being the same ones which determine wages, intro
ducing this variable into the equation is somewhat redundant. If mi
gration opportunities, such as potential income, are a function of
individual attributes, then the direct inclusion of these character
istics in the equation seems to be the best approach. Using the returns
variable also introduces the possible presence of multicollinearity in
the equation since the variable is, in effect, just a linear combination
of some of the remaining variables in the equation. For these reasons a
similar variable will not be included in this model.
A final argument can be put forth in favor of including opportunity-
related variables in the decision equation. Differences in potential
returns from migration can be derived because individuals originate at
different locations. The potential return of moving from New York to
2
More details on construction of the potential wage variable will
appear later since this variable is considered in the choice of desti
nation analysis.

27
Florida may exceed the gain resulting from a California to Florida move.
Therefore, more people may migrate to Florida from New York than from
California, despite the fact that Californians and New Yorkers may have
similar characteristics. There is a way, used in this study, to allow
for this occurrence without including a returns variable in the equation.
The model can be estimated for a sample that is restricted to one
origin at a time. In this way, everyone in the sample faces the same
alternative choices even if these choices are viewed relative to their
present location. Differences in migration propensities now occur as a
result of differences in the way individual characteristics are valued
at alternative locations and because of the effects that individual
attributes have upon the way in which people evaluate and respond to
migration opportunities. Estimating the model for various places also
allows for a comparison of the determinants of the migration decision
over space. If the factors important in explaining the New Yorker's
decision to move are significantly different from the factors affecting
the Californian's choice, then there is an additional rationale for
having separate models. Such an approach will be taken in the main body
of this study. Results obtained when origin states are combined are
presented in Appendix A.
Lack of time and resources prevents estimation for all possible
origins. The model will be estimated for four states. These are New
York, Pennsylvania, Ohio, and South Carolina. These states were chosen
because, in the sample to be studied, they are the leading origins- of
migrants to Florida, the destination of primary concern to this study.
In addition, all four are states with relatively large numbers of out-
migrants during the period under study.

28
The choice of time period is important in any migration study.
Shorter time periods at first seem optimal, since the determinants of
migration can be measured at or close to the time at which the actual
move is made or not made. Longer time periods lead to problems, since
migration may be taking place at many different points during the
period. Determination of when to measure the explanatory variables is
then difficult. Estimation problems, however, arise for short time
periods because the proportion of the population that moves is smaller
than it is for long periods. Estimations were first attempted for one-
year periods. Poor fits were obtained and few variables were significant
enough to distinguish between migrants and nonmigrants. A sample of
these results can be found in Appendix A. For the main part of the
study, three time periods were considered. First, the model was es
timated for the 1968-77 period, the maximum amount of time for which
data is available in the sample. Then, for the purpose of forecasting
the models were re-estimated for the periods 1968-72 and 1972-76. The
explanatory variables were always measured in the first year of the
period under consideration. The migration variable took on the value of
one if during the last year of the period, an individual was living in a
state which differed from his (or her) state of residence during the
first year of the period. Otherwise it took the value of zero.
The unit of analysis in the study is the head of household who is a
member of the labor force. Since the survey from which the sample is
derived contains some information about other household members, var
iables such as family size and spouses employment status can also be
entered into the equations. Other studies (e.g., DaVanzo [1977])

29
have considered only married couples in their samples. Restricting the
sample this way allows for consideration of variables such as spouse's
wage and helps to increase the explanatory power of the equations. This
approach is rejected here, since the ultimate concern of this research
is applying the results to a forecast of migration for all labor force
members, whether married or unmarried.
The sample over which the model is estimated includes members of
the labor force only. Persons not attached to the labor force, such as
retirees, are believed to be subject to a set of influences that is
distinct from the mix of factors affecting labor force members. Health
and unearned income are examples of variables expected to be of greater
importance in the retiree migration decision. The retiree choice of
destination is also affected by different factors. Employment- or
earnings-related variables are clearly not appropriate while cost-of-
living variables may take an added significance in a model of retiree
migration. For this reason, a separate theoretical and empirical model
should be developed for this segment of the population.
In deriving the final sample, return migrants are excluded. A
return migrant is defined here as a person who is found to be moving
back to his (or her) state of birth. Other studies (for example Kau and
Sirmans [1977]) find that including return and nonreturn migrants in the
same equations leads to a specification bias. Return migrants appear to
be moving back home for reasons such as poor health or because they miss
friends or relativesfactors which do not ordinarily enter into non
return migrant decisions. A finding of a recent study of elderly return
migration is that while nonreturn migration is positively selective,
return migration is negatively selective. Longino (1979) finds that

30
elderly migrants returning to their state of birth have lower socio
economic characteristics than other movers. Thus, only nonreturn
migrants are considered in this analysis.
The decision equation will be estimated using a logit program
developed by Nerlove and Press (1973). Using this technique, the
estimate of I\ in equation (5) is equal to the log odds of one of
the alternatives. The estimated equation can be written:
log I ^stay
1-P
= X
stay/
ttx
(7)
where Pstay is the probability of staying, i.e., not moving. We can
now solve for Pstay:
Pstay = ---- y a
1 + e~Xt0tx
Since there are only two alternatives, the probability of moving is
Pmove = 1 Pstay = --T a (9)
1 + e^-ttx
This value will be derived for every individual in the sample and com
pared with each person's actual choice. In addition to comparing actual
and fitted values, the overall fit of the model can be determined by
comparing the log likelihood value at model convergence with the log
likelihood value that results if all coefficients are restricted to
zero. It has been shown (see McFadden [1974]) that the statistic
-2 [ In (1^) In (1) ] (10)
is distributed approximately chi-square, where lr is the likelihood
/s
value when the coefficients are restricted and 1 is the likelihood
value for unconstrained model at its maximum. The degrees of freedom
are equal to the number of restricted coefficients.

31
In addition to comparing the likelihood value at the unconstrained
maximum to the likelihood value when all coefficients are restricted to
zero, a more stringent test is proposed. First the right-hand side of
the decision equation is set equal to the mean value of the dependent
variable. This would be the best guess of the probability that some
individual in the sample would move, given that we have no other in
formation. Then the resulting likelihood value is compared to the
likelihood value when all variables are included. The statistic put
forth above can then be calculated using these two likelihood values as
input.
The specific variables considered in the decision-to-move analysis
will be defined and discussed in Chapter 3. The next section puts forth
the methodology employed in the destination-choice analysis.
The Choice of Destination
Having determined the characteristics that distinguish interstate
movers from nonmovers, the next task is to examine the factors which
enter into the destination choice. Previous studies have considered
location choice primarily as a function of potential wages at alter
native locations. DaVanzo (1977) and Falaris (1978) first estimate a
wage model for each alternative location. Wages are viewed as a func
tion of the characteristics of those living at each place. Some of the
variables considered are age, race, sex, education, occupation, and
experience. These variables are similar to those typically employed in
human capital models of wage determination (for example Dalton and Ford
[1977]). After estimation, predicted values ofuwages at each desti
nation are obtained for everyone in the sample by plugging their

32
characteristics into the fitted equations. This variable is usually
found to be a significant determinant of location choice. A similar
variable has been constructed for use in this study. Although the sign
was positive, it did not turn out to be significant. One possible
explanation derives from the way in which the sample was chosen. As
mentioned earlier, the origin states considered are the leading senders
of migrants to Florida. Florida is not known as a high-wage state, yet
many people move here. In fact, many Florida migrants in the sample
have experienced declines in nominal and real (price deflated) wages. A
sample of results obtained when the wage variable is included is contained
in Appendix B.
Generally, earlier empirical models of individual migration choice
have not found aggregate location characteristics, such as the unemploy
ment rate, to be important variables. But these studies have considered
only very large geographic areas as potential destinations. Differences
in location-specific characteristics then tend to get averaged out
across the region. Better results should be obtained if destinations
are defined as smaller areas, such as states. A disadvantage of this
approach is that all possible destinations cannot be considered since
the estimation technique puts limits upon the number of alternatives
that may be considered at once.
Essentially, the approach suggested by Morrison (1973) is taken
here. That is, while the decision to migrate is considered from a
microanalytic perspective, the decision where to go is modeled in a
macroanalytic framework. By considering only aggregate characteristics
of each location as variables, the model will, for example, predict the
probability of a New York migrant choosing Florida as a destination. It
will not, however, explain why one New Yorker chooses Florida and another

33
chooses California. Micro type variables, such as individual potential
wages, which take into account differences among people were tested.
Their failure to yield meaningful results led to the adoption of a more
macro approach.
Equation (6) was estimated for the period 1968-77. For a forecasting
test, it was estimated for the periods 1968-72 and 1972-76. The sample
was defined as it was for the decision-to-move analysis, except that
only migrants were considered. Migrants from New York, Ohio, Pennsylvania,
and South Carolina were then grouped together. The top eight destination
states were selected as potential alternatives. These included Florida,
California, New York, New Jersey, Michigan, Illinois, Texas, and Virginia.
Migrants choosing other states were excluded from the analysis.
The model is estimated using a multinomial "conditional" logit
program. Estimating equation (6) using this technique allows us to
derive predicted probability values. These probabilities can be rep
resented as follows:
P.
i
P.0
l tp
J P.0
i tp
£ e
i=l
(10)
where i indexes alternatives and J is the total number of choices
available to the individual. The variables in the vector P. will be
x
defined as differences between the value at location i and the value at
the individual's origin of residence. Probabilities will be calculated
for migration between each possible origin and destination combination.
Since starting from any origin these values will-not vary across "indi
viduals, there will be a total of 32 probabilities calculated (8 alter
natives x 4 origins).

34
All variables will be premigration measures of location aggregates.
Specific variables to be included in the choice of destination equations
will be defined and discussed in Chapter 4. The next section will
propose a method of obtaining a forecast from the completely estimated
model.
A Forecasting Methodology
In an application of logit analysis to transportation decisions,
McFadden (1974) states that if the sample under consideration represents
a random selection of the environments faced by the population as a
whole, the average of the predicted values over the sample is a best
estimate of aggregate demand. With this in mind a methodology is proposed
for forecasting the number of migrants from a given origin and their
destination choices. The accuracy of this methodology will be tested
within the sample.
For any given origin state, the coefficients of the 1968-72 decision-
to-migrate equation are applied to values of the explanatory variables
for individuals living in that state in 1972. In this way predictions
for the next four-year period, 1972-76, are obtained. Predicted proba
bilities for each individual are obtained and all individuals are averaged.
This average probability is then applied to the total sample in 1972 to
get a forecast of the number of migrants between 1972 and 1976. This
number can then be compared to the actual (known) number of migrants in
the sample during the period. If successful, then 1972-76 coefficients
can be applied to the 1976 sample to obtain a 1976-80 forecast. The
next step involves application of the sample-derived average proba
bilities to aggregate (outside the sample) measures of the labor

35
force in the state under consideration. In this way a forecast of the
actual number of labor force migrants from that state is obtained.
Figure 2.1 contains a flow chart which illustrates the procedure just
described.
A similar procedure is followed for forecasting destination choice.
The model is estimated for the 1968-72 period and the coefficients are
then applied to locational characteristics in 1972. Forecasts of the
probabilities of migration between each origin and each destination are
obtained. These probabilities are then multiplied times the number of
migrants leaving each origin location. The values obtained can be
compared to the actual origin-destination flows during the 1972-76
period. If successful, the model is re-estimated for the 1972-76
period and these coefficients are applied to the 1976 locational char
acteristics. A 1976-80 aggregate forecast can be obtained by applying
the probabilities derived from the 1972-76 estimation to the aggregate
decision-to-move forecast. Thus, forecasts of actual flows of labor
force migrants to all destinations are derived. The destination-choice
forecasting methodology is illustrated in Figure 2.2.
In addition to forecasting applications, estimating the model for
two equal length time periods allows for observation of coefficient
stability. In this way it can be determined whether events such as the
energy crisis and the deep recession of 1974 have had any impact on
migration decisions and destination choice.3 Since earlier period
3In fact this goal will be difficult to attain. Four-year migration
periods contain many events which cannot easily be disentangled from one
another. Estimation for shorter periods is even harder because of
(migrant) sample size problems.

36
Figure 2.1 Illustration of Forecasting Methodology: Decision to Move

37
Figure 2.2 Illustration of Forecasting Methodology: Destination Choice

38
migration coefficients are used to forecast later period migrant flows,
coefficient instability implies forecasting problems. Employing energy
shortage coefficients, for example, to project migration for a period in
which we expect no energy problems would be incorrect.
Finally, historical estimates of aggregate labor force migration
can be obtained from the sample by utilizing a methodology similar to
that proposed for forecasting. For the decision-to-move analysis, the
average probability of migration for each state during any given period
can be applied to a measure of the aggregate labor force at the beginning
of that period. The resulting estimates of out-migration can then be
distributed out among destinations by multiplying them times the prob
abilities obtained from the destination-choice equations for the same
period. Thus, estimates of aggregate migration between all possible
origins and destinations are derived from the results of the entire
analysis. In the next section of this chapter, the sources of data will
briefly be described.
Sources of Data
The primary source of data for this study is the University of
Michigan's Panel Study of Income Dynamics. In this survey, 5,862
families have been interviewed each year since 1968. At present, 10
years of data are available on tape. Each year approximately 450
variables are available for each family. Categories under which the
variables can be grouped are family composition information, education,
transportation, housing, employment of head, housework, work for money
by wife, food and clothing, income, intelligence", feelings, and time
use.

39
Typical sources of data for migration research are the Continuous
Work History Sample of the Social Security Administration and the Public
Use Sample derived from the 1970 Census of Population. The Panel Study's
advantage over these sources lies in its richness of variables and in
its continuity and consistency of sampling. Its disadvantage is that it
is a much smaller sample. But since its patterns of movement over the
sample period closely resemble the movement identified in these other
sources, great confidence is placed in its use as a tool in migration
studies.
Aggregate data is obtained from various sources. Chief among these
is the Statistical Abstract of the United States. This is one of the
few publications that contains easily accessible and reasonably consistent
time-series data for all states. Other sources are Climatological Data,
National Summary, Cost of Living Indicators, and The Employment and
Training Report of the President.
Sample Size Limitations
Dividing up the sample according to state of origin reduces the
number of observations for each estimation substantially. For the state
of Ohio, 220 observations are drawn from the Panel Study tapes. Thirty-
seven of these individuals (and their families) moved out of Ohio in the
sample period. There are 195 New York observations, 29 of which are
migrants. Pennsylvania has 202 observations and 28 of these moved out
of state during the nine-year period. The sample size for South Carolina
is 222, with 32 being migrants. The destination-choice equations grouped
together 85 migrants from all four states who chose among eight potential
destinations.

40
Logit analysis depends upon the assumption that the dependent
variable (the logarithm of the odds that a particular choice will be
made) approximates the normal distribution. A large number of obser
vations and sufficient repetitions for each possible choice assure that
this criterion will be met. For the decision-to-move analyses it is
believed that sample size is sufficient for employing logit techniques.
The destination choice results should be interpreted more cautiously. A
particular problem for these equations is the small number of observations
occurring in each possible category of choice.4
All equations were also estimated using ordinary least squares
techniques (see Chapter 3). The signs and magnitudes of the coefficients
obtained were very similar to those obtained with logit analysis.5
These similarities increase the degree of confidence placed in the major
findings of this study. In Chapter 3 the results of the decision-to-
move analysis are put forth and discussed.
4An additional source of worry arises because origin states are
chosen according to their rates of migration to Florida. This may
introduce an upward bias in the predictions of migration to Florida.
Consequently, the probabilities of migration to other states will be
understated.
sDaVanzo (1977) compares probit estimates of decision to move
equations with ordinary least squares estimates. She also finds similar
results.

CHAPTER 3
EMPIRICAL RESULTS: DECISION TO MOVE
Interstate Migration: 1968-1977
In Table 3.1, a summary of interstate movement among Panel Study
respondents is presented. It can be seen that the states receiving the
largest number of migrants between 1968 and 1977 were California, Florida,
and Texas. The leading out-migration states were California, Ohio,
South Carolina, New York, and Pennsylvania. Net migration, defined as
the difference between the number of in-migrants and out-migrants, is
greatest in Florida and Texas. Twenty-two states registered negative
levels of net migration, with the greatest declines occurring in Ohio
and South Carolina. The negative value recorded in California is
surprising and leads to some concern about whether the Panel Study
Sample is representative or not. The fact that this study focuses upon
gross rather than net migration is some consolation, but the small
number of migrants in the sample is a worrisome factor.
States were chosen for this analysis according to their rates of
migration to Florida. Of major interest in this study are the determinants
of migration decisions and destination choice among residents of origin
states for Florida migrants. This will aid in developing a forecast of
migration from these states to Florida and to competing destinations
such as Texas and California. A summary of Florida in-migration by
state of origin appears in Table 3.2. There it is shown that 58.2
percent of all Florida migrants in the sample came from either New York,
41

42
Table 3.1 Summary of Panel Study Migration, 1968-1977
Number of In- Number of Out-
State
Migrants
Migrants
Net Migration
Alabama
7
3
4
Arizona
11
6
5
Arkansas
3
11
-8
California
44
43
-4
Colorado
16
6
10
Connecticut
8
7
1
Delaware
District of
3
0
3
Columbia
9
12
-3
Florida
43
11
32
Georgia
9
5
4
Idaho
2
0
2
Illinois
23
22
1
Indiana
11
14
-3
Iowa
4
17
-13
Kansas
6
0
6
Kentucky
5
14
-9
Louisiana
5
13
-8
Maine
7
1
6
Maryland
18
9
9
Massachusetts
14
10
4
Michigan
10
10
0
Minnesota
5
7
-2
Mississippi
4
16
-12
Missouri
4
19
-15
Montana
0
0
0
Nebraska
7
3
4
Nevada
7
0
7
New Hampshire
5
0
5
New Jersey
21
15
6
New Mexico
4
0
4
New York
25
30
-5
North Carolina
5
12
-7
North Dakota
0
0
0
Ohio
11
37
-26
Oklahoma
5
4
1
Oregon
17
8
9
Pennsylvania
15
28
-13
Rhode Island
0
0
0
South Carolina
5
33
-28
South Dakota
1
5
-4
Tennessee
10
5
5
Texas
39
14
2-5
Utah
1
7
-6
Vermont
1
0
1
Virginia
22
15
7

43
Table 3.1 (continued)
State
Washington
West Virginia
Wisconsin
Wyoming
Number of In-
Migrants
6
2
6
2
Number of Out-
Migrants
9
0
2
0
Net Migration
-3
2
4
2

44
Table 3.2 Florida In-migration by State
of Origin, 1968-1977
Origin State
Percentage of
Number of Florida Migrants Florida Migrants
New York
7 16.3
Pennsylvania
7 16.3
Ohio
6 14.0
South Carolina
5 11.6
Indiana
3 7.0
Kentucky
2 4.7
Virginia
2 4.7
Missouri
2 4.7
Alabama
1 2.3
Arizona
1 2.3
Arkansas
1 2.3
California
1 2.3
Connecticut
1 2.3
Illinois
1 2.3
Louisiana
1 2.3
Maryland
1 2.3
1 2.3
North Carolina

45
Pennsylvania, Ohio, or South Carolina. After California, these are also
the leading out-migration states in the sample. These four states were
thus selected as the major sample for this analysis.
Table 3.3 defines the variables to be considered in the decision-
to-move analysis. In the next section, the rationale for including each
variable will be discussed and a first look at the data will be provided.
Comparison of Migrants and Nonmigrants
Family Size
Table 3.4 presents a comparison of migrant and nonmigrant char
acteristics in the four-state sample. A priori it is expected that
persons in larger families are less likely to move since additional
people represent ties to present location. Children in school or a
spouse who is working are both examples of migration ties. Comparison
of the mean values of family size among migrants and nonmigrants, however,
reveals very little difference. Both groups average about five persons
per family. Thus, at first glance, family size does not appear to be a
constraining force.
Sex
Households with male heads are expected to move more often than
households with female heads. One reason is that men are more likely to
be employed in occupations where job transfers are common. In addition,
higher rates of unemployment for women at many locations and lower
income opportunities would deter female migration. The mean value of
the sex variable, however, is only slightly higher for movers than it is
for nonmovers. Eighty-seven percent of the households that moved were
headed by men. It may be that since women are also likely to be earning

46
Table 3.3 Variable Definitions for Decision to Migrate Analysis
Dependent Variable
Definition
Move Equals 1 if the person's state of residence
in 1977 differs from his or her state of
residence; equals 0 otherwise. (Defined
in 1968 for different periods in later
analysis.)
Explanatory Variables
Famsz
Family size, actual number in household.
Sex Equals 1 for male; 2 for female.
Homeown Equals 1 if home is owned; equals 0
otherwise.
Lres
Lempl
Emself
Avage
Length of residence in current house or
apartment, takes on following values:
0 if length of residence < 1 year
1 if length of residence = 1 year
2 if length of residence = 2 years
3 if length of residence = 3 years
4 if length of residence = 4 years
5 if length of residence 5 years and
< 10 years
6 if length of residence > 10 years and
< 15 years
7 if length of residence > 15 years and
< 25 years
8 if length of residence > 25 years
Length of time employed by present employer,
takes on following values:
0 if self-employed or unemployed
1 if employed > 0 months and < 6 months
2 if employed > 6 months and < 18 months
3 if employed > 18 months and <_ 42 months
4 if employed > 42 months and < 9 years
5 if employed > 9 years and < 19 years
6 if employed > 19 years
Equals 1 if self-employed; equals 0
otherwise.
Average age of husband and wife; if unmarried
Avage = age of household head.

47
Table 3.3 (continued)
Explanatory Variables
Definition
Ernst
Employment status, equals 1 if employed;
equals 0 if unemployed.
Marst
Marital status, equals 1 if married;
equals 0 otherwise.
Race
Equals 1 if white; equals 0 otherwise.
Aveduc
Average education level of husband and wife;
if unmarried Aveduc = education of household
head, takes on following values:
1 if education > 0 years and < 6 years
2 if education > 6 years and < 9 years
3 if education > 9 years and < 12 years
4 if education = 12 years
5 if education = 12 years and person has
some non-academic training
6 if person attended college, but did not
receive a degree
7 if person received a bachelor's degree
8 if person attended graduate school
Prevmig
Previous migration, equals 1 if current
state of residence differs from state of
birth; equal 0 if current state of
residence is the same as state of birth.
Ernst w
Equals 1 if spouse works; equals 0
otherwise; if no spouse then Emstw = 0.
Famy
Total family income, in thousands of dollars.

48
Table 3.4 Comparison of Migrant and Nonmigrant Characteristics:
New York, Ohio, Pennsylvania, and South Carolina
Variable
Mean Value Migrants
Mean Value Nonmigrants
Famsz
4.98
4.95
Sex
0.87
0.84
Homeown
0.64
0.58
Lres
5.93
5.87
Lempl
3.14
3.31
Emself
0.10
0.08
Avage
37.89
35.31
Ernst
0.96
0.98
Marst
0.78
0.81
Race
0.71
0.68
Aveduc
4.11
3.78
Prevmig
0.31
0.23
Emstw
0.38
0.35
Famy
11.24
8.79

49
lower incomes at their current locations and have a higher probability
of being unemployed before moving, they are more responsive to oppor
tunities elsewhere. This would counteract the negative effects of the
sex variable and lead to a more ambiguous result. No conclusions can be
drawn without more rigorous analysis.
Homeownership
Owning a home is usually expected to act as a tie to present location.
Time and transactions costs involved in selling a home and perhaps
purchasing another one in a new location would tend to deter migration.
On the other hand, homeownership represents wealth. If a wealth effect
operates, property owners may be more likely to move than renters. In
this sample 64 percent of the migrants owned homes prior to migration
while only 58 percent of the nonmigrants were homeowners. Thus, the
data give some support to the second hypothesis.
Length of Residence
Persons who have lived in their current houses or apartments for
long periods of time are expected to display a lower probability of
moving. Duration of past residence should act as a measure of will
ingness to pull up stakes, leave friends and relatives, and start all
over somewhere else. An individual who remains in the same dwelling
unit for a long time displays an aversion to risk which leads one to
believe that he (or she) would not be likely to do something as filled
with uncertainty as changing states of residence. The data in Table 3.4
do not demonstrate this to be the case. There appears to be very
little difference in length of residence for migrants and nonmigrants.
The average length of residence for both groups is close to 10 years.

50
Length of Employment
As with the last variable, the length of employment would indicate
attachment to present environment. In addition, it can be viewed as a
measure of ability to hold a job. A person who has worked for his (or
her) present employer for only a short period of time is more likely to
have been periodically unemployed. The mean value of this variable is
indeed somewhat lower for migrants than for nonmigrants in the states
under consideration. The average length of employment before moving is
about 21 months for migrants.
Self-Employment
Individuals who are self-employed are expected to migrate less. The
self-employed person is more likely to have an occupation that involves
heavy investment in capital equipment. Dentists and printers are examples.
Although this equipment may be transportable, the costs of moving and
re-establishing at some other location are significant enough to deter
migration. Self-employed workers are also likely to be part of smaller
organizations where job transfer is uncommon. Finally, those who are
employed in professions requiring state licenses, such as medical and
legal fields, are bound to their state to some extent by their licenses.
The mean values for the four sample states show very little difference
in propensity to migrate based upon self-employment, with migrants
showing a slightly higher degree of self-employment. Ten percent of the
migrants from these states were self-employed before moving.

51
Average Age
In most migration studies age is expected to be negatively related
to migration since older people have a shorter work life over which to
realize the gains from migration. In addition, it is felt that older
people have greater ties to present location. That is, they are likely
to have stronger attachments to friends and relatives and are likely to
possess location-specific capital, such as a home. Only upon retirement
are older people expected to show higher migration rates. But since
this study is restricted to members of the labor force, the inverse
relationship is still expected here. It appears, however, that in this
sample, the average age is greater for migrants than for nonmigrants.
Perhaps this occurs because the sample is restricted to Florida-sending
states. Many of the movers to Florida and competing destinations, such
as California, may be anticipating future retirement. For example, a
person may move to Florida and accept a low-paying job knowing that (he
or she) will withdraw from the labor force after a predetermined number
of years. Individuals who have accumulated substantial savings during
their lifetime can afford to do this. In addition, many older migrants
may choose semiretirement rather than retirement, perhaps maintaining
part-time jobs after moving. If the preretirement or semiretirement
thesis is correct, we would expect labor force migration from the sample
states to consist of older migrants than migration from all other states.
In the Panel Study, it is found that the average age of all U.S. labor
force migrants is 36 years old. This compares to a mean age of 38 in
the sample of Florida-sending states. Thus, there is mild evidence in
favor of this hypothesis.

52
Employment Status
People who are unemployed are expected to be more responsive to
opportunities elsewhere than those who are employed. Particularly if
local economic conditions are poor, the process of job search for the
unemployed is likely to include alternative locations. If, however,
employment conditions at other locations are also depressed, the response
of the unemployed is uncertain. The effect of employment status upon
the migration decision seems to depend on local economic conditions
relative to alternative destinations. In the analysis to follow, relative
conditions will be held constant by considering only one state at a
time. The mean value of employment status for all four states combined
is slightly lower for migrants than for nonmigrants. Ninety-six percent
of the migrants in the sample were employed prior to moving.
Marital Status
Marriage is expected to act as a deterrent to migration. As with
the family-size variable, presence of a spouse represents an additional
tie to present location. This is particularly true if the spouse is
working. The data in Table 3.4 show that migrants are less likely to be
married than nonmigrants. Seventy-eight percent of the migrants in the
four-state sample were married, while 81 percent of the nonmigrants were
married.
Race
The expected sign of the race variable is ambiguous. On the one
hand it can be argued that since some blacks are more likely to earn
lower incomes and are more likely to be unemployed than whites, they are

53
more responsive to opportunities elsewhere. On the other hand, there
may also be fewer opportunities elsewhere for blacks because they are
concentrated in low-skill occupations and because of discrimination in
the labor market. In addition, blacks are less likely to be employed in
jobs requiring geographic transfers. For the sample to be considered
here it looks as if migrants are more likely to be white, indicating
that the latter two effects outweigh the first factor.
Average Education
More educated people are expected to move more often. The reason
is twofold. For one, education (in most cases) leads to increased
employment opportunities. Particularly it is believed that increased
education opens up employment possibilities which are more national in
scope resulting in greater migration propensities. In addition, those
who have attained higher education levels are expected to have more and
better information about migration opportunities. The extreme of this
phenomenon occurs in professions that form organizations which aid their
members in locating jobs. The economics profession is an example. The
data in Table 3.4 give preliminary evidence that this thesis is correct.
Migrants in the sample are likely to have at least started college,
while the average nonmigrant posesses only a high school education.
Previous Migration
Migrants are believed to be characterized by frequent mobility
during their lifetimes. It is hypothesized here that a person who has
changed state of residence at least once during his (or her) lifetime is
more likely to move again than an individual who has always lived at the
same location. People who have moved frequently are also less likely

54
to have developed attachments to their present state of residence. In
the four-state sample, 31 percent of the migrants have moved before.
Only 23 percent of the nonmigrants have previously changed their states
of residence. Thus, there is support for the hypothesis set forth
above.
Employment Status of Spouse
The presence of a working spouse is expected to deter migration.
If both partners have substantial earnings, any migration decision must
presumably maximize their combined return. The probability of finding a
location that meets this criterion is lower than the chances of finding
one which maximizes the returns of one family member. It is possible
that only one member of the couple will be properly rewarded upon moving
and the spouse will be tied to this choice. Whether joint utility is
maximized or not, the spouse's employment and income possibilities must
be taken into account in any decision to move. Theoretically, this
should act as a deterrent to migration; however, preliminary evidence in
Table 3.4 suggests that migrants in the sample were more likely to have
working spouses. Thirty-eight percent of the migrants' spouses worked
prior to moving.
Family Income
Migrants are expected to earn higher annual income than nonmigrants.
A few reasons can be given for this belief. First, higher incomes
usually imply greater demand for leisure activity. Changing location is
one way of obtaining more leisure. This is particularly true for this
sample of states which have typically sent large numbers of migrants to
Florida and California, states which possess climates amenable to

55
recreational activities. Higher income people are also more likely to
have visited other states as tourists. Thus, they would have obtained
information about these states which could lead to increased likelihood
of moving. The data in Table 3.4 show significant differences between
family income levels of movers and nonmovers. Migrants earned an average
of $11,240 in 1968. This compares to an average income level of $8,790
for nonmigrants.
Empirical Results
Introduction
The comparison of migrant and nonmigrant characteristics presented
so far serves only to provide a general picture of relationships in the
data. Much can be hidden in such a rudimentary analysis. Since all
four sample states were grouped together, differences in migrant selection
between states are hidden. As an example, migrants from New York may
tend to have larger families than New Yorkers who do not move. There
may, however, be little difference in the family size of movers and non
movers from Ohio, Pennsylvania, and South Carolina. When all four
states are viewed together, very little difference in family size will
show up, even though this is an important determinant in New York. This
problem becomes particularly severe if the variable takes on opposite
signs in each state.
Theoretical reasons exist for expecting differences in the deter
minants of migration between states of origin. The crux of the argument
is that residents of different states face varying relative migration
opportunities. Residents of Ohio, for example, will experience a greater
improvement in climate from moving to Florida than will residents of

56
South Carolina. If older people are more responsive to climate differ
entials than younger people, we might then expect age to be a more impor
tant migration determinant in Ohio than it is in South Carolina. Thus,
variables are significant in one state but not in another because of
differences in initial (origin) conditions. From now on, each state
will be analyzed by itself.
When migrants and nonmigrants are compared only according to the
mean values of a set of variables, other problems arise. First, there
are the usual weaknesses involved in using the mean as a measuring
device (its sensitivity to extreme values is one such weakness). More
importantly, when any variable is considered, the values of remaining
factors are not held constant. Thus, the true effect of each char
acteristic is not being measured. Use of multiple regression techniques
will relieve these problems.
Before the results of the logit analysis of the decision to move
are presented, a discussion of the scaling of some of the qualitative
(dummy) explanatory variables is in order. Particularly, length of
residence, self-employment, and average education are defined in ways
which are somewhat out of the ordinary. Looking back at Table 3.3 it is
observed that these variables are scaled using an ordinal ranking system,
with each value representing a given level or interval. This scaling is
the same as that which appears on the Panel Study tapes. The conventional
method of modeling variables which are coded in such a fashion is to
include separate dummy variables for each category of data. In this
case, however, 21 variables would need to be added to the model. This

57
would result in reduction of degrees of freedom and, more importantly,
would cause logit estimation to become prohibitively expensive.1
A second method of handling the problem would be to redefine the
dummy variables so as to have fewer categories. In the limiting case
only one zero-one dummy is used. Some cutoff point is chosen and a
value of one is assigned to those observations for which the cutoff is
exceeded and a value of zero is assigned otherwise. In this study, for
example, the education variable could be defined so that those individuals
with 12 years or more of schooling were given values of one with the
remaining observations equaling zero. Assumed in such an approach is
that increases in the level of education up until the 12th grade have no
effect on the probability of moving. In addition it is assumed that
increases in education beyond high school have no effect. To validly
use such an approach some prior expectation of the proper cutoff point
is required. Appendix A presents the results of logit estimation when
this approach is adopted. Not having prior knowledge of the correct
cutoff point, the choices used are somewhat arbitrary.
The actual scaling used for the main part of this study (see Table
3.3) also imposes restrictions upon the model. To illustrate these
restrictions, this approach is compared to the conventional method of
defining such data. Suppose an explanatory variable contains three
categories. Following the conventional method, two dummy variables are
introduced. (If three dummies are used, then the constant term must be
eliminated.) Call the dependent variable Y and the two dummies and
1 In fact, given the computer program limitations, logit estimation
would have been impossible.

58
Assuming no other variables in the model, the equation can be written as:
(1)
where a, 6 and 82 are coefficients and E is a random disturbance term.
Y = a + 8 X + B2X2 + E
X
1 and X2
are zero-one
variables we
can further say that:
E
(Y
X1 =
1 and X2 = 0)
= a +
h
(2)
E
(Y
X1 =
0 and X2 = 1)
= a +
b2
(3)
E
(Y
X1 =
0 and X2 = 0)
= a
(4)
Now suppose instead of using the above approach we define a single
variable, X, which takes on the values of 0, 1, and 2 for each of the
three categories. Now the equation can be written as:
Y = a + 6X + E (5)
where a and B are parameters and E is the disturbance. Now we can
say that:
E (Y X = 0) = a (6)
E (Y X = 1) = a + B (7)
E (Y
X = 2) = a + B
(8)
This formulation turns out to be equivalent to estimating the model of
equation (1) with the following restriction attached:
^1 32
(9)
While this restriction should be considered, it may be preferable to the
restrictions imposed when the variable is collapsed into two categories
(as discussed earlier). Using the latter approach, information which
is available about alternative categories goes unused. The chosen
technique uses all of the information, but it imposes rather tight
restrictions on the relationship between categories.

59
Summarizing, it would be optimal to include separate dummy variables
for each category for which there is information. Program and cost limi
tations, however, prevent using this approach. Alternatives include
collapsing the variable into fewer categories or including a single
ordinally ranked categorical variable. While the latter approach is
somewhat unprecedented2, it has been chosen because it utilizes more
information which is available in the data. Both alternatives impose
restrictions which may be considered severe without prior knowledge
of data. In the next four sections, the results of the logit analysis
of the decision to move will be presented and discussed.
Logit Results: Ohio
Table 3.5 presents the regression results for the state of Ohio.
Variables which are significant with at least 95 percent confidence
are sex, marital status, and family income. The sex variable, as
hypothesized, has a positive coefficient. Thus, households with male
heads from Ohio are more likely to migrate than households with female
heads from that state. The most significant variable is marital status.
As expected, unmarried individuals move more often than those who are
married. Thus, marriage does act as a deterrent to migration in Ohio.
Family income is also positive and significant. Higher income families
in Ohio are more likely to change states of residence than lower income
families. This is also the expected result.
2To test whether wage and price controls of the Nixon administration
had any effect upon wage formation, a three-leveled ordinal dummy was used
by Eckstein and Girla (1978). The variable took on the value of .5 for
Phase I of controls and 1 for Phase II of controls. Periods without con
trols took on the value of 0. The variable was found to be insignificant.
The same variable was used in a price equation and was significant.

60
Table 3.5 Results of Logit Analysis: Decision to Move,
Ohio, 1968-1977
Variable Coefficient Asymptotic t-Ratio
Famsz
.147
1.27
Sex
2.802
2.60
Homeown
-3.65
- .64
Lres
.013
.06
Lempl
.052
.31
Emself
-.845
- .73
Avage
.054
1.53
Ernst
-
-
Marst
-4.142
-3.39
Race
.867
1.09
Aveduc
-.024
- .15
Prevmig
.298
.69
Ernst w
.471
.96
Famy
.057
2.06
Constant
-5.000
-2.72
Summary Statistics:
Log of Likelihood Function = -86.0
Likelihood Statistic (at 0) = 133.0
Significance Level (at 0) = .005
Likelihood Statistic (at mean) =27.3
Level of Significance (at mean) = .025
Percentage Correctly Predicted = 85%
Percentage of Migrants Correctly Predicted = 16%

61
The remaining variables are not significant. Family size, average
age, and race are variables which show some importance. Migrant families
in Ohio are more likely to be larger than nonmigrant. This result is
opposite from that expected. A possible explanation is that many
migrant families from Ohio have preschool children. While having
children in school may act as a deterrent to migration, families may
decide to move prior to their children's first enrollment in school.
Indeed, people may move so that their children can go to different
schools. Age shows a positive sign, lending support to the prere
tirement hypothesis set forth earlier. Race is also positive, indi
cating that households headed by white persons are more likely to move
interstate than households headed by nonwhites.
Although insignificant, the homeownership variable has the ex
pected sign. People owning homes in Ohio are less likely to move than
those who rent. They are also more likely to have lived in their homes
and worked for the same employer for longer time periods than non
migrants. These results are contrary to expectations. The migrant is
less likely to be self-employed. Thus, self-employment appears to act
as a migration deterrent. Those who chose to leave Ohio are more
likely to have moved previously, supporting the hypothesis put forth
earlier. Education carries an unexpected negative coefficient. Migrant
household heads are also likely to have spouses who are employed. This
variable has the opposite sign from that which was expected. Employment
status was not included in the analysis because there was no variation
of this characteristic among migrants. Estimation was thus impossible.
When the likelihood value of the decision equation is compared to
the value that results when all coefficients are zeroed out, the equation
is significant at better than the .005 level. When the likelihood

62
value of the equation is tested against that which results when the
right-hand side of the equation is set equal to the mean of the dependent
variable, the equation is significant at the .025 level. Under either
test, the equation appears to produce a good fit.
A correct prediction is defined as one in which the predicted
probability is within 50 percent of the actual value of the dependent
3 p
variable. For migrants this means that move > .50. For nonmigrants
this means ^move < .50. Under this criterion, the equation predicted 85
percent of all (migrant and nonmigrant) individuals in the sample correctly.
Among migrants, 16 percent were correctly predicted.
Logit Results: New York
Table 3.6 presents the decision-to-move results for the state of
New York. Variables which are significant at better than the .05 level
are length of employment, average age, employment status, and spouse's
employment status. Length of employment carries a negative sign, implying
that those who have worked for the same employer for long periods of
time are less likely to move. This result is expected. Migrants from
New York are also older than nonmigrants from that state, providing
further support for the preretirement thesis. Migrant household heads
are less likely to be employed than nonmigrants, but their spouses are
likely to be employed. This last result, which was also found to be
true for Ohio, is contrary to expectations. Perhaps what we are observing
3This definition is admittedly arbitrary. Some cutoff had to be
chosen in order to summarize the results without listing every predicted
value.

63
Variable
Famsz
Sex
Homeown
Lres
Lempl
Emself
Avage
Emst
Marst
Race
Aveduc
Prevmig
Emstw
Famy
Constant
Table 3.6 Results of Logit Analysis: Decision to Move,
New York, 1968-1977
Coefficient
-.165
.127
.773
.374
-.561
-1.277
.074
-2.960
-1.600
-.890
.227
-.448
1.682
.084
Asymptotic t-Ratio
-1.37
.08
1.38
1.51
-2.80
-1.40
2.30
-1.99
-1.03
- .89
1.34
- .71
2.77
1.88
-2.580 -1.39
Summary Statistics:
Log of Likelihood Function = -62.8
Likelihood Statistic (at 0) = 144.8
Level of Significance (at 0) = .005
Likelihood Statistic (at mean) = 38.4
Level of Significance (at mean) = .005
Percentage Correctly Predicted = 83%
Percentage of Migrants Correctly Predicted = 21%

64
are married couples with both spouses being employed in highly mobile,
perhaps professional, occupations. In any case, employed spouses are
not acting as deterrents to migration.
Family income is significant at the 94 percent confidence level.
It carries its expected positive sign. The remaining variables are
insignificant. Sex, self-employment, marital status, and average
education have the signs which were hypothesized earlier. Homeownership
has a positive sign, indicating that migrants are more likely to own a
home than nonmigrants. They are also more likely to have lived in their
homes for longer periods of time. Race, although insignificant, carries
a negative sign. This lends support to the thesis that blacks are more
responsive to opportunities elsewhere. Previous migration also has the
opposite sign from that expected. Migrants from New York are less
likely to have moved before in their lifetimes. This is consistent with
the sign on the length-of-residence variables. Perhaps many of the New
York migrants are middle-aged or older people who have lived in New York
most of their lives and are now moving in anticipation of retirement.
The overall equation is significant at better than the .005 level.
This is true whether the model is compared to one where all coefficients
are zero or compared to a model where the right-hand side is equal to
zero. Eighty-three percent of the individuals in the sample are pre
dicted correctly by the model, while 21 percent of the migrants are
predicted correctly.
Logit Results: South Carolina
The decision-to-move results for South Carolina appear in Table
3.7. Significant variables are family size, average age, marital

65
Variable
Famsz
Sex
Homeown
Lres
Lempl
Emself
Ava ge
Emst
Marst
Race
Aveduc
Prevmig
Emstw
Famy
Constant
Table 3.7 Results of Logit Analysis: Decision to
Move, South Carolina, 1968-1977
Coefficient
.252
1.231
-.180
.286
-.328
-2.713
.093
-.678
-3.494
-.371
.073
.803
-.528
.217
Asymptotic t-Ratio
1.95
1.04
- .33
1.17
-1.76
-1.65
2.51
- 66
-2.46
- .58
.38
1.24
- .89
3.22
-6.333 -2.96
Summary Statistics:
Log of Likelihood Function = -71.7
Likelihood Statistic (at 0) = 164.4
Level of Significance (at 0) = .005
Likelihood Statistic (at mean) = 39.8
Level of Significance (at mean) = .005
Percentage Correctly Predicted = 90%
Percentage of Migrants Correctly Predicted = 31%

66
status, and family income. Marital status and family income show the
expected signs. Contrary to expectations, family size is positive and
significant. Again average age carries a positive sign.
Although insignificant at the 95 percent confidence level, length
of employment and self-employment are important and carry the correct
signs. Of the remaining variables, sex, homeownership, employment
status, average education, previous migration, and spouse's employment
status influence migration decisions in the expected directions. As in
Ohio and New York, length of residence is positively related to migration.
Race displays a negative relationship, indicating that households
headed by blacks are more likely to leave the state.
The South Carolina equation provides a very good fit of the data.
The equation is significant by both criteria (at zero and at the mean)
at better than the .005 level. Ninety percent of the individuals in the
sample were correctly predicted by the model, with 31 percent of the
migrants' choices being predicted accurately.
Logit Results: Pennsylvania
In Table 3.8 are the results of the regression analysis for Penn
sylvania. Variables which are significant at the .05 level are average
education and family income. Average education shows the expected
positive sign. Family income, however, is negatively related to the
probability of moving. This is truly unexpected for a variable that has
been positive and significant for the other three states.
Close to being significant are family size and marital status. As
in Ohio and South Carolina, family size takes on-a -positive sign.
Marital status has the hypothesized negative sign, indicating marriage*

67
Variable
Famsz
Sex
Homeown
Lres
Lempl
Emself
Avage
Ernst
Marst
Race
Aveduc
Prevmig
Ernst w
Famy
Constant
Table 3.8 Results of Logit Analysis: Decision to
Move, Pennsylvania, 1968-1977
Coefficient
.251
1.859
. 105
.022
.253
1.257
.021
Asymptotic t-Ratio
1.85
1.51
. 19
.11
1.34
.88
.59
-2.625
.109
.638
.225
.869
-.219
-1.84
. 17
2.88
.36
1.66
-2.65
-5.650 -2.78
Summary Statistics:
Log of Likelihood Function = -71.6
Likelihood Statistic (at 0) = 57.2
Level of Significance (at 0) = .005
Likelihood Statistic (at mean) = 19.4
Level of Significance (at mean) = .10
Percentage Correctly Predicted = 87%
Percentage of Migrants Corrected Predictly =
7%

68
deters migration. Sex and previous migration are both positively related
to migration, as expected. Homeownership, length of residence, length
of employment, self-employment, and spouse's employment status all show
the wrong sign. Average age has a positive sign, giving further support
to the preretirement thesis, and race is positively related to migration,
indicating that white families leave Pennsylvania more often than do
black families.
The Pennsylvania equation provides a poorer fit than the decision
equations for the other three states. Although significant compared to
the model where all coefficients are zero, the more stringent test
indicates a weaker model. Setting the right-hand side of the equation
equal to the mean value of the dependent variable and comparing the
likelihood value which results with the unconstrained likelihood value
results in a significance level of .10 for the unconstrained model.
Although 87 percent of the sample was predicted correctly by the model,
only 7 percent of the migrants were predicted correctly. Thus, this
model has not been very successful in distinguishing between migrants
and nonmigrants.
Test for Separation of States
In Chapter 2, a theoretical argument was made for estimating
separate equations for each state of origin. In the preceding sections
of this chapter evidence was presented which indicated that there may be
differences between states in the determinants of the migration decision.
In this section, a formal test for equality of coefficients between
states is carried out.

69
The test developed by Chow (1960) for the equality of regression
coefficients between equations is applied to the decision-to-move
equations. First, ordinary least squares (OLS) estimates for each state
are determined. Then the observations from all four states are combined
and an OLS equation is estimated for the combined data.4 The results of
these estimates appear in Table 3.9. The following statistic is calculated:
[SEEq;ota]_ (SEE^y + SEEoH + SEEgc + SEEpA) ] / 3k
[SEENY + SEEQh + SEESC + SEEPA] / n-4k
where SEE stands for the sum of squared residuals, k refers to the
number of coefficients in each equation, and n refers to the size of the
combined sample. The subscript NY refers to the New York equation with
OH standing for Ohio, SC referring to South Carolina, and PA meaning
Pennsylvania. The subscript TOTAL refers to the combined equation. The
statistic is distributed according to an F-distribution. Its value in
this case is 1.67 which compares to a critical F value (14 degrees of
freedom in the numerator, and 779 in the denominator) of 1.5 at the .05
level of significance. Thus, the hypothesis of equality of regression
coefficients between equations is rejected at the 95 percent confidence
level. The belief that there are differences in migration determinants
across states is confirmed.
40LS is chosen for the test because execution of the logit program
is impossible for the combined sample and because no comparable test
exists for comparing coefficients across logit equations.

70
Table 3.9
- Ordinary Least Squares
Results:
Decision to
Move, 1968-
1977
Variable
Ohio
Coefficient
t-Pvatio
New York
Coefficient t-Ratio
Pennsylvania
Coefficient t-Ratio
Famsz
.015
1.02
-.020
-1.51
.023
1.60
Sex
.494
3.32
.073
.50
.323
1.74
Homeown
-.047
- .70
.055
.95
.005
.09
Lres
.002
.05
.060
1.32
.003
.15
Lempl
.005
.21
-.047
-2.63
.024
1.18
Emself
-. 105
- .77
-.061
- .59
.161
1.06
Avage
.006
1.46
.009
2.75
.002
.43
Ernst
.037
.19
-.348
-1.80
.009
.04
Marst
-.626
-4.05
-.198
-1.35
-.402
-2.06
Race
.069
.86
-.136
-1.47
-.012
- .17
Aveduc
-.003
- .14
.020
1.16
.064
2.78
Prevmig
.035
.63
-.041
- .64
.015
.20
Ernst w
.044
.74
. 128
2.21
.106
1.84
Famy
.009
2.48
.009
1.71
-.016
-2.87
Constant
-.210
- .94
.366
1.69
-.210
- .81
R2 =
.13
R2
= .17
R2
= .09

71
Table 3.9 (continued)
South Carolina
Combined
States
Variable
Coefficient
t-Ratio
Coefficient
t-Ratio
Famsz
.018
1.58
.010
1.54
Sex
.174
1.35
.278
3.96
Homeown
-.020
- .40
.005
.18
Lres
.040
1.10
.011
1.00
Lempl
-.042
-2.21
-.014
-1.52
Emself
-.228
-1.76
-.047
- .78
Avage
.009
2.83
.005
3.38
Ernst
-.032
- .28
-.120
-1.47
Marst
-.399
-3.16
-.275
-4.01
Race
-.082
-1.32
-.016
- .52
Aveduc
.006
.29
.018
1.85
Prevmig
.069
.98
.024
.81
Emstw
-.066
-1.27
.041
1.47
Famy
.026
4.03
.005
2.23
Constant
-.045
- .31
CO
rH
l
-1.17
R2 =
.18
R2 =
.06

72
Application to Forecasting
The models for all four states have been re-estimated for the
period 1968 through 1972. The results are presented in Table 3.10. For
the shorter time period there are, of course, fewer migrants. This led
to some problems. First, some of the variables which were used in the
nine-year migration analysis could not be included here. A requirement
for logit estimation is that there be variation within each category of
the dependent variable. Since there are fewer migrants in the sample,
there is less likelihood that some variables will meet this requirement
in the migrant category. Those that do not must be deleted. Second,
with a smaller proportion of migrants in the sample it is more difficult
to distinguish between migrants and nonmigrants. In general, poorer
fitting equations were obtained for the shorter time period. This is
because there were fewer variables that could be included in the model
and because of the smaller proportion of migrants in the sample.
Using the methodology outlined in Chapter 2 a forecast is developed.
The coefficients in Table 3.10 are applied to values of the explanatory
variables for individuals living in the sample states in 1972. Individual
predicted probabilities are calculated and averaged over the number of
individuals in each state. The mean probability for each state is then
multiplied by the total sample size in 1972. The results of this procedure
appear in Table 3.11.
The forecast values are compared to the actual (known) number of
movers during the 1972-76 period. For Ohio, 25 migrants are predicted.
In fact, there were 20 Ohio migrants in the sample between 1972 and
1976. For New York, 13 migrants are predicted compared to 17 actual

Table 3.10 Results of Logit Analysis: Decision to Move, 1968-1972
Ohio
New York
South Carolina
Pennsylvania
Variable
Coefficient
t-Ratio
Coefficient
t-Ratio
Coefficient
t-Ratio
Coefficient
t-Ratio
Famsz
-.813
- .38
-.316
-1.66
.207
1.35
. 164
.87
Sex
-
-
-
-
1.614
1.32
1.087
.89
Homeown
.845
.71
1.118
1.36
-.541
- .69
-.107
- .15
Lres
-
-
-
-
-
-
-
-
Lempl
.570
1.55
-.743
-2.58
-.540
-2.23
.228
.94
Emself
2.885
1.30
-1.940
-1.57
-2.313
-1.41
-
-
Avage
-.062
-1.03
.042
1.07
.074
1.59
-.010
- .23
Ernst
-
-
-
-
-
-
-
-
Mar st
.350
.22
.859
.58
-3.76
-2.40
-1.798
-1.17
Race
-
-
.249
. 19
-1.239
-1.33
.405
.48
Aveduc
-.193
- .59
.388
1.63
.390
1.62
.647
2.35
Prevmig
.204
.26
1.094
1.34
.960
1.18
1.338
1.82
Ernst w
.624
.77
1.617
2.14
-.876
- .97
.711
.96
Famy
.114
2.29
-.045
- .67
.160
2.10
-.212
-1.86
Constant
-4.770
-2.26
-4.949
-2.36
-4.868
-2.88
-5.266
-2.71

Table 3.10 (continued)
Summary Statistics
Ohio
New York
South Carolina
Pennsylvania
Log of Likelihood Function
-38.5
-36.7
-43.9
-45.5
Likelihood Statistic (at 0)
243
246
245
210
Level of Significance (at 0)
.005
.005
.005
.005
Likelihood Statistic (at mean)
11.4
23.4
29.8
18.0
Level of Significance (at mean)
. 10
.025
.01
.10
Percentage Correctly Predicted
95%
95%
94%
93%
Percentage of Migrants Correctly
Predicted
0%
15%
19%
7%
\

75
Table 3.11 Comparison of Forecast to Actual Migration, 1972-1976
State of
Origin
Number of
Migrants-Forecast
Number of
Migrants-Actual
Difference
(Forecast-Actual)
Ohio
25
20
5
New York
13
17
-4
South Carolina
27
15
12
Pennsylvania
10
15
-5
Total
75
67
8

76
movers. Twenty-seven South Carolina migrants are forecast. In fact,
only 15 left that state during the period. The model predicts 10
migrants from Pennsylvania, while 15 actually left the state. In total
75 migrants are forecast to have moved. This compares to 67 people who
actually moved. This is a 12 percent error. It is interesting to note
that the best fitting models in the historical period, 1968-72, do not
provide the most accurate forecasts. The best fitting equation is for
the state of South Carolina, yet the greatest forecast error occurs in
this case.
The decision equations for each state are estimated next for the
period 1972-76. The results appear in Table 3.12. It should first be
observed that these results look very different from the 1968-72 mi
gration equation appearing in Table 3.10. Many variables which are
significant for the earlier period lose their significance in the latter
period and other variables gain significance. For example, for the
state of Ohio, the only significant variable in the 1968-72 equation was
family income. In the 1972-76 period, this variable is no longer
significant and length of employment becomes significant. Significant
variables in the earlier period New York equations are length of em
ployment and employment status of spouse. In the later period, no
variables are significant at the 95 percent confidence level. Length of
employment, marital status, and family income were significant variables
in the first four-year period for the state of South Carolina. Only
family income remained significant in the 1972-76 period. For Penn
sylvania, average education was the most significant 1968-72 variable.
Later, family size, length of employment, average age, and marital
status showed significance. In addition, sign changes frequently

Table 3.12
- Results
of Logit Analysis:
Decision
to Move,
1972-1976
Variable
Ohio
Coefficient
t-Ratio
New York
Coefficient t-
Ratio
South Carolina
Coefficient t-Ratio
Pennsylvania
Coefficient t-Ratio
Famsz
.050
.34
.005
.03
.167
1.09
.307
1.96
Sex
-
-
-
-
.617
.43
1.534
1.69
Homeown
-.015
- .02
-1.212 -1
.66
-.710
-1.03
-.119
- .17
Lres
-
-
-
-
-
-
-
-
Lempl
.610
3.13
-.244 -1
.49
-.398
-1.71
-.380
-2.20
Emself
-.715
- .65
-
-
.921
.84
-
-
Avage
-.019
- .62
.016
.60
.004
.10
.084
2.50
Ernst
-
-
-
-
-.252
- .19
-
-
Marst
-1.108
-1.27
.475
.59
-.681
- .53
-4.011
- 3.15
Race i
-
-
-.576
.71
-.902
- .89
.680
.88
Aveduc
.109
.55
. 148
.78
-. 126
- .44
.378
1.31
Prevmig
.280
.51
-.515
.74
1.387
1.62
.011
.012
Ernst w
.371
.56
-.432
.57
-. 107
- .15
1.516
1.64
Famy
-.029
- .70
.308
.74
.967
2.25
.0004
.005
'Constant
-3.805 -2.23
-2.520 -1.56
-2.148 .83
-7.128
-3.12

Table 3.12 (continued)
Summary Statistics Ohio
Log of Likelihood Function -58.6
Likelihood Statistic (at 0) 218
Level of Significance (at 0) .005
Likelihood Statistic (at mean) 17.4
Level of Significance (at mean) .10
Percentage Correctly Predicted 90%
Percentage of Migrants Correctly
Predicted 5%
New York
South Carolina
Pennsylvania
-53.7
-47.0
-44.2
89
85
182
.005
.005
.005
10.2
11.0
17.4
.10
.10
.10
92%
92%
93%
0%
0%
7%

79
occurred, although this was usually restricted to the insignificant
variables. It should be emphasized that all four-year period equations
are misspecified because some variables could not be included in the
estimation. This makes comparisons of the two periods less meaningful.
There is evidence, however, of instability of the decision-to-move
coefficients between the two time periods. This has negative impli
cations for forecasting based upon previous coefficients. But the
reasonableness of the forecasts which have been obtained provides support
for this methodology. Coefficient changes which occur may be balanced
in such a way as to affect the average probability, the key variable in
the forecast, only slightly.
The final step in the analysis is to apply the 1972-76 coefficients
to the characteristics of people living in the sample states in 1976.
Thus, a 1976-80 forecast is obtained. These predictions are presented
in Table 3.13. The results show that 24 of the Panel Study members who
lived in New York in 1976 will have left by 1980. For both South Carolina
and Pennsylvania 19 migrants are forecast. Ohio is expected to lose 18
migrants during the period.
An aggregate forecast is also presented in Table 3.13. It is
obtained in the following manner. First, the number of persons in the
labor force for each state in 1976 is taken from the Employment and
Training Report of the President. Then, in order to be comparable to
the Panel Study sample, working spouses are removed from the measure.
This is accomplished by multiplying the labor force number by an esti
mate of the proportion of working spouses in the labor force. The
estimates used are the proportions that occur in the Panel Study sample
states in 1976. The resulting measure of labor force household heads is

80
Table 3.13 Forecast: Decision to Move, 1976-1980
State of
Origin
Number of Migrants-
Panel Study
Number of Migrants-
Aggregate Labor Force
Ohio
18
833,000
New York
24
1,867,000
South Carolina
19
309,000
Pennsylvania
19
1,003,000
Total
80
4,012,000

81
then multiplied by average family size to obtain an estimate of the
actual number of potential migrants from each state. The average family
size measure is also estimated from the Panel Study sample.5 Finally,
the potential population at risk is multiplied by the average probability
of migraiton estimated for the Panel Study. The numbers in Table 3.13
result.
Using the procedure put forth in Chapter 2, estimates of aggregate
labor force migration were developed from sample characteristics for the
periods 1968-72 and 1972-76. Average predicted probabilities were
calculated for each period based upon the regression results. The
population at risk for each state was derived using the method discussed
in the last paragraph. The estimates for these two periods and the
forecast are presented together in Table 3.14. It can be seen that for
the 1968-72 period, New York is estimated to have had the largest amount
of out-migration. Pennsylvania is next, followed by Ohio and South
Carolina. For the 1972-76 period, total migration out of the four
states increased. Only New York did not share in this increase. The
largest jump occurred in Ohio which is estimated to have 457,000 more
out-migrants during this period than the number who left in the 1968-72
period. This increase was large enough to put that state in second
place to New York. Pennsylvania and South Carolina showed small in
creases from the previous periods. For the 1976-80 period, total out
migration is expected to decline from the previous period. This decline
will be felt in all states except New York. The greatest decline
5Since this study has found family size to be insignificant in the
determination of the decision to move, the family-size measure is obtained
from the total (migrant and nonmigrant) sample.

82
Table 3.14 Aggregate Labor Force Migration: 1968-1972,
1972-1976, 1976-1980
State of Number of Migrants Number of Migrants Number of Migrants
Origin 1968-1972 1972-1976 1976-1980
Ohio
714,000
1,171,000
833,000
New York
1,708,000
1,643,000
1,867,000
South Carolina
310,000
341,000
309,000
Pennsylvania
1,121,000
1,131,000
1,003,000
Total
3,853,000
4,286,000
4,012,000

83
occurs in Ohio, the same state that was estimated to have an unusually
high level of out-migration in the 1972-76 period. Overall, about 4
million persons are forecast to leave the four states between 1976 and
1980, indicating that there will be a continuing large pool of migrants
from which Florida and other states may draw.
The reasonability of the method used to estimate and forecast
aggregate labor force migration can partially be determined by comparing
the results with measures of out-migration derived from the 1970 Census
of Population. These estimates are not directly comparable, however.
For one, the Census estimates cover the five-year period between 1965
and 1970, while the estimates in this study cover four-year periods.
Since there is some overlap the 1968-72 period is chosen for comparison.
An additional problem in comparing estimates from the two sources is
that the Census figures are not tabulated according to labor force
status in published reports.
Table 3.15 compares the 1965-70 Census estimates to the 1968-72
estimates of out-migration obtained in this study. It can be seen that
with the exception of Ohio, the estimates of migration derived in this
study exceed those obtained from the 1970 Census. Because the Census
estimates include all members of the population (whether they are in the
labor force or not) and because the Census data cover a longer period
the opposite result would be expected. On the other hand, it is generally
believed that there was significant undercount in the 1970 Census. In
addition, because of prosperous economic conditions the early 1970s may
have been years of greater rates of migration. It is known that in
Florida, for example, in-migration accelerated significantly during
this period. Since the origin states chosen for this study are some

84
Table 3.15 Comparison of Panel Study Derived Out-Migration
Estimates to Census Estimates
State of Origin
Panel Study Estimate, 1968-72 Census Estimate, 1965-70
Ohio
714,000 787,546
New York
1,708,000 1,329,432
South Carolina
310,000 248,609
Pennsylvania
1,121,000 781,684
Total
3,853,000 3,147,271

35
of the strongest contributors of Florida migrants, these states are
likely to have experienced higher rates of out-migration. These latter
arguments provide some support for the finding of higher levels of
migration in this study. More detailed evaluation of the estimates is
impossible without having a comparable set of labor force migration
figures for the same period. The next chapter presents the results of
the destination-choice analysis and develops a forecast for the 1976-80
period.

CHAPTER 4
EMPIRICAL RESULTS: DESTINATION CHOICE
Destination Choice: 1963-1977
A summary of the destinations chosen by migrants in the four sample
states appears in Table 4.1. Overall, Florida is the leading destination,
with 20 percent of all out-migrants choosing this state. Florida is the
top location choice for New York, Pennsylvania, and Ohio migrants.
South Carolina migrants showed a preference for moving north, with New
York being the leading destination. Florida, however, followed closely
behind. With the exception of those choosing Florida, California, and
Texas, many migrants chose to move to nearby states. As an example, 14
percent of the Ohio migrants moved to Michigan and 17 percent of New
York's migrants chose to live in New Jersey. Because of technical and
cost limitations, not all destinations which were chosen could be in
cluded in the analysis. The top eight destination states were selected.
These states accounted for 70 percent of the migrants who left the four-
state area of New York, Pennsylvania, Ohio, and South Carolina.
The variables used in the destination-choice equations are presented
in Table 4.2. In the next section the rationale for inclusion of each
variable will be discussed and the results are presented.
Logit Results: Destination Choice
The results of the first destination-choice model estimated appear
in Table 4.3. The sample consists of 88 migrants originating from
86

87
Table 4.1 Destination Choice Summary: 1968-1977
(Number of Migrants)
Destination
State
of Origin
State
New York
Pennsylvania
Ohio South
Carolina
Total
Florida
7
7
6
5
25
California
4
3
5
1
13
New Jersey
5
5
2
1
13
New York
0
2
2
7
11
Illinois
1
1
2
4
8
Virginia
2
2
1
3
8
Michigan
0
1
5
0
6
Texas
0
0
4
0
4
All other
states
11
7
10
12
37
Total
30
28
37
33
125

88
Table 4.2 Variable Definitions: Destination Choice Equations
Dependent Variable
Definition
Choice
Equals 1 for chosen destination,
equals 0 for all other alternatives.
Explanatory Variables
Distance
Straight line mileage between each
origin and each destination.
Tempdiff
Annual average temperature at each
destination minus annual average
temperature at each origin.
Incdiff
Annual per-capita personal income
at each destination minus annual
per-capita personal income at each
origin.
Hdiff
First quarter housing price index at
each destination minus first quarter
housing price index at each origin.
Taxdiff
Annual per-capita state and local
government tax collections at each
destination minus annual per-capita
state and local government tax
collections at each origin.
Unemdiff
Annual average unemployment rate of
each destination minus annual average
unemployment rate at each origin.

89
Table 4.3 Logit Results: Destination Choice,
1968-1977, Model 1
Variable
Coefficient
Asymptotic t-Ratio
Distance
-.703
-2.42
Tempdiff
.229
3.86
Incdiff
.128
1.35
Hdiff
.020
.44
Taxdiff
.010
.22
Unemdiff
.047
. 17
Summary Statistics:
Log of Likelihood Function = -162.5
Likelihood Statistic (at 0) = 23.4
Level of Significance (at 0) = .005
Likelihood Statistic (at mean) = 14.35
Level of Significance (at mean) = .05

90
either New York, Ohio, Pennsylvania, or South Carolina. Each member of
the sample chose one of the eight alternatives put forth in the previous
section.
As in other studies, distance is expected to be negatively related
to destination choice. That is, all other things constant, people are
expected to move to locations which are closer to their initial residence.
Distance is believed to act as a proxy for the cost of moving and the
cost of obtaining information about alternative locations. As an example,
a New Yorker is likely to know more about what it would be like to live
in New Jersey than he would know about the situation in Texas. In fact,
distance turns up negative and significant at the 99 percent confidence
level.
It is expected that individuals are attracted to more temperate
climates. Warmer temperatures, in addition, may represent other amen
ities such as recreational activities and access to the coast. Colder
temperatures at origin locations are expected to act as a further
stimulus to migration. It is hypothesized here that climate is a more
important variable in deciding on a destination to persons originating
in colder climates. Thus, the greater the temperature difference between
origin location and a given destination, the greater is the expected
probability of choosing that destination. This is, in fact, what occurs.
The difference in average temperature variable turns out positive and
significant at the .005 level of significance.
As hypothesized in regional growth models, migration is believed to
be a response to differences in income opportunities. Thus, holding
other variables constant, individuals are believed to choose to move to
that location which promises the maximum possible lifetime gain in

91
income over their present state of residence. Since the interest here
is aggregate flows, average per-capita income differences are used to
represent the opportunities available to the average person. This
variable is positive and significant at only the 90 percent confidence
level. This lends only mild support to the hypothesis that regional
income inequality stimulates migration.
The cost of living at alternative locations is also believed to
influence the location choice. At first, an attempt was made to deflate
the average-income variable by local price indices provided by the
American Chamber of Commerce Association. This, however, caused very
little change in the coefficient of the income difference variable. It
was later thought that just comparing the overall cost-of-living index
may be improper. Rather it may be particular components of the market
basket which are important in location choice. Housing costs and tax
levels were two variables which came to mind. Since moving often involves
selling a home and buying a new home at a chosen location, the differ
ence in state housing price indices between origin and destination
locations was used as a variable. This variable was constructed as an
average of all reporting cities in each state during the first quarter
of the year under consideration (in this case 1968). It was expected to
have a negative sign, meaning that people are expected to move from
areas of high housing costs to locations where housing costs are lower.
At the individual level a person would like to maximize the gain realized
through selling his (or her) present home and purchasing a new one.-
Since the price index also reflects rental costs, renters too are affected
by this variable. When plugged into the model, however, the variable
displayed the wrong sign and showed no significance.

92
Per-capita state and local tax collections are also believed to
make a difference in the location choice. Frequently, migrants are
escaping areas with high state income and local property taxes. Again a
negative sign was expected here, implying that individuals would like to
reduce their tax burden upon moving. But, as before, the variable
showed up to be insignificant.
The final variable in the model of Table 4.3 is one which reflects
employment conditions in sending and receiving areas. It is believed
that individuals will choose to live at locations where there are a
relatively large number of employment opportunities. At the same time,
an individual from an area of high unemployment is expected to be even
more responsive to job availability at other places. This is particularly
the case if he (or she) is currently unemployed. As in other studies,
though, the unemployment-differential variable is of the wrong sign and
insignificant. Other variables representing employment opportunities,
such as the rate of employment growth, were substituted. Still there
was no success.
When the estimated equation is compared to one in which all co
efficients are restricted to zero, the model turns out significant at
the .005 level. When compared to an equation in which the right-hand
side is set equal to the mean value of the dependent variable, the
equation is only significant at the .05 level.
The cost of living and unemployment variables are removed from the
model and the equation is re-estimated. The results are shown in Table
4.4. Distance is now significant at the .005 level. The temperature
difference variable remains extremely significant. The difference in
average annual per-capita income variable is now also significant at the
.005 level.

93
Table 4.4 Logit Results: Destination Choice,
1968-1977, Model 2
Variable
Coefficient
Distance -.741
Tempdiff .231
Incdiff .152
Summary Statistics:
Log of Likelihood Function = -162.8
Likelihood Statistic (at 0) = 22.9
Level of Significance (at 0) = .005
Likelihood Statistic (at mean) = 13.75
Asymptotic t-Ratio
-3.36
4.78
3.55
Level of Significance (at mean) = .005

94
The significance of temperature and income differences lends support
to the Graves (1978) thesis which suggests that migration is just a
response to an increased demand for amenities which are available at
certain locations. The economy is believed to start out in spatial
equilibrium, meaning that amenity-adjusted real income is the same at
all locations. Differences in income across locations are said to exist
because of differences in the availability of location-specific goods,
such as recreation. An exogenous shock, such as an income increase,
causes the demand for location-specific goods to increase. This creates
a temporary disequilibrium which is restored through migration. The
positive and significant coefficient on the income variable says that if
temperature differences are the same for all possible moves, people will
choose places which promise positive income gains. To test whether
individuals will still choose higher income places, even if temperature
differences are allowed to vary, the model is re-estimated with the
climate variable removed. The results of this estimation are presented
in Table 4.5. The income difference variable is no longer significant.
Omitting the climate variable causes a downward specification bias on
the income variable. Thus, the Graves theory receives confirmation. If
the income variable had remained significant, then support would have
been generated for regional growth models which claim that migration is
solely a function of interregional wage differences. But, instead,
income is only significant in conjunction with climate, so that the
trade-off hypothesized by Graves appears to be in operation.
The equation of Table 4.4 is. significant at _the .005 level whether
it is compared to the model where all coefficients are zero or to the

95
Table 4.5 Logit Results: Destination Choice,
1968-1977, Climate Excluded
Variable Coefficient Asymptotic t-Ratio
Distance -.016 -.10
Incdiff .005 .02
Summary Statistics:
Log of Likelihood Function = 174.19
Likelihood Statistic (at 0) = .02
Level of Significance (at 0) = (not significant at any
acceptable level)
Likelihood Statistic (at mean) = -4.5
Level of Significance (at mean) = (not significant at any
acceptable level)

96
model that is evaluated at the mean value of the dependent variable. In
the next section this version of the model of destination choice will be
utilized to obtain a forecast of destination choice.
Application to Forecasting
The procedure used to forecast destination choices follows the
description set forth in Chapter 2. The three-variable version of the
destination-choice equation is first estimated for the 1968-72 period.
The results appear in Table 4.6. Distance and income differences are
still significant variables, with the temperature-differentials var-
riable losing much of its explanatory power. The overall equation is
significant at only the .10 level.
The 1968-72 coefficients were applied to values of the explanatory
variables in order to derive a 1972-76 forecast for each pair of origin
and destination locations. For example, for the New York to Florida
forecast the distance between the two states (constant over time), the
1972 per-capita income difference between them, and the 1972 average
temperature difference are substituted into the 1968-72 equation. The
predicted probability is then applied to the total number of migrants
leaving New York for the eight alternative states between 1972 and 1976.
The total number of available migrants is determined by assuming that
the proportion of migrants who chose alternatives other than the eight
states considered here is the same as it was during the most recent
four-year period. The number choosing alternatives is then subtracted
from the total number of migrants leaving New York for all states. The
resulting number is multiplied by the probability_predicted by the'
model. This process is repeated for each origin and destination pair.

97
Table 4.6 Logit Results: Destination Choice, 1968-1972
Variable
Coefficient Asymptotic t-Ratio
Distance
-.001 -2.15
Tempdiff
.120 1.59
Incdiff
.001 2.20
Summary Statistics
Log of Likelihood
Function = -74.0
Likelihood Statistic (at 0) = 7.4
Level of Significance (at 0) = .10
Likelihood Statistic (at mean) = 7.4
Level of Significance (at mean) = .10

98
Table 4.7 presents the Panel Study forecast for the period 1972-76.
The forecast Is also compared to actual destination choices during that
period. For New York, Pennsylvania, Ohio, and South Carolina combined,
11 migrants are forecast to come to Florida. This compares to 15 who
actually moved to Florida. This is about a 27 percent error. For
California, 3 migrants are predicted. There were actually 8 migrants
from the four states who chose California. Nine migrants are predicted
to go to New Jersey in comparison to 5 who actually did. It is forecast
that 6 movers would go to Illinois, while actually there were none.
Five migrants are expected to go to Virginia. This compares to 8 who
moved there between 1972 and 1976. Four migrants are predicted for
Michigan, compared to 2 people who actually moved there. Finally, 3 are
forecast to move to Texas, the same number that actually moved there.
The overall ability to forecast thus appears to vary from state to
state. Part of the problem, it is believed, results from the small
sample size. A better fitting model was obtained for the 1968-77 time
period. But when the time period is reduced there are smaller numbers
of migrants from which to distinguish. This is believed to contribute
to the reduced significance of the model and the less than adequate
forecasts which were derived from it.
In order to produce a forecast for the 1976-80 period the model was
estimated for the four-year period just prior to it. The results are
presented in Table 4.8. Only the temperature-difference variable
remains significant at the .005 level, with the distance and income-
difference variables falling off to the .10 level. While levels of
significance have changed, the coefficients do not appear to have changed
very much since the 1968-72 period. Only the temperature-difference

99
Table 4.7 Comparison of Forecast to Actual Destination
Choice: 1972-1976
Origin-
Destination
Number of
Migrants-Forecast
Number of
Migrants-Actual
Difference
(Forecast-Actual)
New York-Florida
3
4
-1
New York-California
1
3
-2
New York-New Jersey
4
2
2
New York-Illinois
2
0
2
New York-Virginia
2
3
-1
New York-Michigan
1
0
1
New York-Texas
1
0
1
Pennsylvania-Florida
2
5
-3
Pennsylvania-
California
0
1
-1
Pennsylvania-New
Jersey
2
1
1
Pennsylvania-New
York
2
1
1
Pennsylvania-
Ill inois
1
0
1
Pennsylvania-
Virginia
1
2
-1
Pennsylvania-
Michigan
1
0
1
Pennsylvania-Texas
1
0
1
Ohio-Florida
3
2
1
Ohio-California
1
3
-2
Ohio-New Jersey
2
1
1
Ohio-New York
3
1
2
Ohio-Illinois
2
0
2
Ohio-Virginia
1
1
0
Ohio-Michigan
1
2
-1
Ohio-Texas
1
3
-2
South Carolina-
Florida
3
4
-1
South Carolina-
California
1
1
0
South Carolina-
New Jersey
1
1
0
South Carolina-
New York
2
3
-1
South Carolina-
Illinois
1
0
1
South Carolina-
.
_
Virginia
1
2
-1
South Carolina-
Michigan
1
0
1
South Caro'lina-
Texas
1
0
1

100
variable has taken on added significance in the later period. This
would suggest that migration to Florida, a state for which climate is
the main attraction, would be greater than a forecast based on 1968-72
coefficients would suggest. This is, in fact, the case. The forecast
in Table 4.7 underestimated migration to Florida for all states except
Ohio.
Overall, the equation in Table 4.8 is significant at the .005 level
when compared to the model with coefficients set equal to zero. This is
a better fit than for the previous four-year period. But when the
unconstrained model is compared to one with the right-hand side of the
equation set equal to the mean value of the dependent variable, it is
significant at only the .10 level.
The coefficients for the 1972-76 period were applied to the values
of the explanatory variables in 1976. Predicted probabilities were
obtained and multiplied by estimates of the total number of migrants
leaving each origin state for the eight destination states between 1976
and 1980. The estimates of the total pool of migrants were derived by
taking the number of migrants forecast to leave each state (from Chapter
3) and subtracting out the proportion expected to choose states other
than the eight alternatives. As before, the proportion choosing other
alternatives is assumed constant from the previous period. The destination-
choice forecast is presented in Table 4.9.
The same procedure outlined above was used to distribute the 1976-
80 aggregate forecast for each origin state among the destinations. For
each state, the aggregate forecast obtained from the decision-to-move
analysis is first adjusted by subtracting out the number expected to
choose other alternative states. The remaining number is then mul-

101
Table 4.8 Logit Results: Destination Choice, 1972-1976
Variable
Coefficient Asymptotic t-Ratio
Distance
-.0004 -1.55
Tempdiff
.181 3.67
Incdiff
.001 1.59
Summary Statistics:
Log of Likelihood Function = -85.1
Likelihood Statistic (at 0) = 17.9
Level of Significance (at 0) = .005
Likelihood Statistic (at mean) =4.7
Level of Significance (at mean) = .10

102
Table 4.9 Forecast: Destination Choice, 1976-1980
Origin-
Number of Migrants-
Number of Migrants-
Destination
Panel Study
Aggregate Labor Force
New York-Florida
4
331,000
New York-California
3
252,000
New York-New Jersey
3
199,000
New York-Illinois
2
146,000
New York-Virginia
2
132,000
New York-Michigan
2
119,000
New York-Texas
2
159,000
Pennsylvania-Florida
3
148,000
Pennsylvania-California
2
108,000
Pennsylvania-New Jersey
2
87,000
Pennsylvania-New York
2
87,000
Pennsylvania-Illinois
1
67,000
Pennsylvania-Virginia
1
54,000
Pennsylvania-Michigan
1
54,000
Pennsylvania-Texas
1
67,000
Ohio-Florida
3
114,000
Ohio-California
2
103,000
Ohio-New Jersey
1
60,000
Ohio-New York
1
60,000
Ohio-Illinois
1
60,000
Ohio-Virginia
1
43,000
Ohio-Michigan
1
43,000
Ohio-Texas
1
60,000
South Carolina-Florida
4
59,000
South Carolina-California
3
41,000
South Carolina-New Jersey
2
25,000
South Carolina-New York
1
23,000
South Carolina-Illinois
1
20,000
South Carolina-Virginia
1
18,000
South Carolina-Michigan
1
16,000
South Carolina-Texas
2
25,000

103
tiplied by the 1976-80 destination-choice forecast probabilities. The
predicted distribution of migrants appears in Table 4.9.
The historical aggregate out-migration estimates were also distributed
out among the eight destinations, utilizing predicted probabilities from
the estimated models. These flows, along with the 1976-80 aggregate
forecast, appear in Table 4.10. As was done with the aggregate out
migration forecast from each state, the 1968-72 destination-choice
estimates are compared with 1965-70 Census estimated flows between
states. In Chapter 3 the observation was made that, in total, out
migration based upon Panel Study estimates was higher than Census
estimates despite the facts that the Census is not restricted to labor
force members and that it measures five-year flows. Differences were
partially attributed to Census undercounts and higher rates of migration
in the early 1970s. Remaining differences are attributed to weaknesses
in the methodology employed in the study.
Table 4.11 presents a comparison of Census estimates to Panel Study
based estimates. Since we already know that the absolute number of out-
migrants from each state (with the exception of Ohio) are estimated to
be higher in this study than they are in the 1970 Census, it will be
better to compare the relative proportions of migrants who choose
alternative destinations. Notice that the Census estimates larger
proportions of migrants from all states choosing Florida and California
as destinations than this study estimates. Part of this discrepancy may
arise because of time-period differences. It is believed, however, that
greater reliance should be placed upon the Census estimates. In this
study the small sample size of the destination-choice analysis reduces

Table 4.10 Aggregate Destination Choice
Origin-
Destination
Number of Migrants
1968-1972
New York-Florida
New York-California
New York-New Jersey
New York-Illinois
New York-Virginia
New York-Michigan
New York-Texas
210,000
131,000
355,000
224,000
131,000
171,000
79,000
Pennsylvania-Florida
81,000
Pennsylvania-California
47,000
Pennsylvania-New Jersey
135,000
Pennsylvania-New York
188,000
Pennsylvania-Illinois
81,000
Pennsylvania-Virginia
47,000
Pennsylvania-Michigan
67,000
Pennsylvania-Texas
34,000
Ohio-Florida
63,000
Ohio-California
52,000
Ohio-Ney Jersey
78,000
Ohio-New York
109,000
Ohio-Illinois
89,000
Ohio-Virginia
36,000
Ohio-Michigan
63,000
Ohio-Texas
31,000
South
Carolina-Florida
41,000
South
Carolina-California
19,000
South
Carolina-New Jersey
32,000
South
Carolina-Nex^ York
43,000
South
Carolina-Illinois
30,000
South
Carolina-Virginia
17,000
South
Carolina-Michigan
19,000
South
Carolina-Texas
13,000
1968-1972, 1972-1976, 1976-1980
Number of Migrants
1972-1976
Number of Migrants
1976-1980
443,000
163,000
163,000
82,000
117,000
700,000
140,000
250,000
83,000
91,000
114,000
45,000
61,000
38,000
83,000
244,000
99,000
76,000
99,000
53,000
61,000
38,000
91,000
95,000
30,000
22,000
27,000
15,000
20,000
10,000
30,000
331,000
252,000
199,000
146,000
132,000
119,000
159,000
148,000
108,000
87,000
87,000
67,000
54,000
54,000
67,000
114,000
103,000
60,000
60,000
60,000
43,000
43,000
60,000
59,000
41,000
25,000
23,000
20,000
18,000
16,000
25,000
104

Table 4.11 Comparison of Panel Study Derived Migration Flows to Census Estimates
Origin-Destination
Panel Study Estimate, 1968-1972
Census Estimate
i, 1965'
New York-Florida
210,000
(.16)
182,531
(.29)
New York-California
131,000
(.10)
122,766
(.19)
New York-New Jersey
355,000
(.27)
209,698
(.33)
New York-Illinois
224,000
(.17)
31,041
(.05)
New York-Virginia
131,000
(.10)
36,908
(.06)
New York-Michigan
171,000
(.13)
23,761
(.04)
New York-Texas
79,000
(.06)
30,174
(.04)
Pennsylvania-Florida
81,000
(.12)
58,528
(.15)
Pennsylvania-California
47,000
(.07)
51,612
(.13)
Pennsylvania-New Jersey
135,000
(.20)
100,011
(.28)
Pennsylvania-New York
188,000
(.28)
72,001
(.19)
Pennsylvania-Illinois
81,000
(.12)
21,308
(.06)
Pennsylvania-Virginia
47,000
(.07)
36,450
(.09)
Pennsylvania-Michigan
67,000
(.10)
18,036
(.05)
Pennsylvania-Texas
34,000
(.05)
18,105
(.05)
Ohio-Florida
63,000
(.12)
80,679
(.23)
Ohio-California
52,000
(.10)
70,366
(.20)
Ohio-New Jersey
78,000
(.15)
16,288
(.05)
Ohio-Neib York
109,000
(.21)
33,548
(.10)
Ohio-Illinois
89,000
(.17)
36,910
(.11)
Ohio-Virginia
36,000
(.07)
23,388
(.07)
Ohio-Michigan
63,000
(.12)
58,350
(.17)
Ohio-Texas
31,000
(.06)
25,278
(.07)
South Carolina-Florida
41,000
(.19)
18,951
(.24)
South Carolina-California
19,000
(.09)
11,301
(.15)
South Carolina-New Jersey
32,000
(.15)
6,311
(.08)
South Carolina-New York
43,000
(.20)
13,274
(.17)
South Carolina-Illinois
30,000
(.14)
3,481
(.04)
South Carolina-Virginia
17,000
(.08)
13,316
(.17)
South Carolina-Michigan
19,000
(.09)
3,293
(.04)
South Carolina-Texas
13,000
(.06)
7,808
(.10)
105

106
the viability of the results. The forecasting methodology is based on
the assumption that sample proportions represent population proportions.
When sample size is small the probability of staisfying this assumption
is low. This is particularly true when the number of choices is high,
such as in the destination-choice equations. The number of observations
for each alternative ranged from 25 (Florida) to 4 (Texas).
A final hypothesis is set forth for explaining differences between
Census destination proportions and Panel Study based estimates. The
Census includes the entire population regardless of labor force status.
People who are not labor force members may have higher rates of mi
gration to.states such as Florida or California. Retired people, for
example, have a high propensity to migrate to Sunbelt states. This
group raises the rate of migration to these states. Since the Panel
Study estimates are restricted to labor force members, lower rates of
migration to retirement states are expected.
For the 1976-80 period in-migration to Florida from the four-state
sample is projected to be 652,000. This is in line with current ex
pectations that migration to Florida will remain healthy but will not be
able to maintain recent historical peaks. Floridas major competitor,
California, is forecast to receive 504,000 new migrants between 1976 and
1980. One of its newest competitors, the state of Texas, will gain
311,000 in-migrants during the period.
The poor fits obtained in the shorter time period destination-
choice equations lead to some skepticism about outright acceptance of
the forecast and estimates put forth above. Even with this limitation,
however, these figures appear reasonable and are in line with historical
standards and current expectations of the future.

107
Since climate and income have turned out to be such important
variables in the destination-choice decision, they deserve some special
attention. In the final section of this chapter, a closer look at these
variables over time will be provided.
Income, Climate, and Migration
Earlier it was discovered that climate became a more important
determinant of destination choice during the 1972-76 period than it was
for the 1968-72 period. It was also demonstrated that migration to such
warm weather states as Florida, California, and Texas reached its
maximum during the period 1972-76. Table 4.12 presents the actual
average temperature difference and per-capita income difference data for
the three points in time used in the previous analysis.
It can be observed that temperature differences between sending and
receiving regions are generally at their maximum in 1972. These larger
differences are not the result of lower temperatures at origin states
during that year. Rather, the temperatures in warm weather destination
regions were somewhat higher than usual. Further, some of the northern
receiving regions such as Michigan and Illinois had unusually cold
winters in 1972. Thus, migration to Sunbelt states increased relative
to northern competing regions. This increase in relative attractiveness
combined with the substantially larger number of out-migrants from
sending states, resulted in larger gains for Florida, California, and
Texas. By 1976, temperature differences declined to near normal levels.
This contributed to a forecast of somewhat reduced afcfcrativeness of
Sunbelt states relative to northern receiving regions.

Table 4.12 Climate and Income: 1968, 1972, and 1976
Origin- Average Temperature Difference (C) Per-Capita Income Difference (dollars)
Destination
1968
1972
1976
1968
1972
1976
New York-Florida
9.0
11.4
10.1
-967
-1131
-992
New York-California
5.0
5.3
6.0
-134
-317
64
New York-New Jersey
-1.3
-. 6
-1.0
-160
-193
169
New York-Illinois
-1.5
-2.9
-2.3
-134
-193
332
New York-Virginia
2.2
2.0
2.4
-1065
-1061
-824
New York-Michigan
-1.7
-3.0
-1.7
-424
-502
-106
New York-Texas
5.4
6.8
6.0
-1107
-1274
-857
Pennsylvania-Florida
9.0
11.2
9.7
-253
-259
-358
Pennsylvania-California
5.0
5.1
5.6
580
555
698
Pennsylvania-New Jersey
-1.3
-.8
-1.4
554
679
803
Pennsylvania-New York
0
-.2
-.4
714
872
634
Pennsylvania-Illinois
-1.5
-3.1
-2.7
580
679
966
Pennsylvania-Virginia
2.2
1.8
2.0
-351
-189
-190
Pennsylvania-Michigan
-1.7
-3.2
-2. 1
290
370
528
Pennsylvania-Texas
5.4
6.6
5.6
-393
-402
-223
Ohio-Florida
12.1
14.3
12.8
-341
-324
-324
Ohio-California
8.1
8.2
8.7
492
490
732
Ohio-New Jersey
1.8
2.3
1.7
466
614
837
Ohio-New York
3.1
2.9
2.7
626
807
668
Ohio-Illinois
1.6
0
.4
492
614
1,000
Ohio-Virginia
5.3
4.9
5.1
-439
-254
-156
Ohio-Michigan
1.4
-.10
1.0
202
305
562
Ohio-Texas
8.5
9.7
8.7
-481
-467
-189
South Carolina-Florida
4.6
6.2
5.1
784
740
982
South Carolina-California
. 6
. 1
1.0
1,617
1,554
2,038
South Carolina-New Jersey
-5.7
-5.8
-6.0
1,591
1,678
2,143
South Carolina-New York
-4.4
-5.2
-5.0
1,751
1,879
1,974
South Carolina-Illinois
-5.9
-8.1
-7.3
1,617
1,678
2,306
South Carolina-Virginia
-2.2
-3.2
-2.6
686
810
1,150
South Carolina-Michigan
-6.1
-8.2
-6.7
1,327
1,369
1,868
South Carolina-Texas
1.0
1.6
1.0
644
597
1,117
108

109
The per-capita income data generally paint a picture of persistence
of state income inequality over time. Despite considerable levels of
out-migration from the four origin states and large numbers of in
migrants arriving at the eight destination states, the magnitude of
state income differences has been maintained over time. Thus, the
thesis that migration leads to convergence of regional per-capita income
levels is not supported here.
Despite cold weather, positive income gains available in such
states as Illinois, Michigan, New Jersey, and New York lead to continued
in-migration to these destinations. With the exception of migrants
originating from South Carolina, those arriving in Florida, Texas, and
Virginia can generally expect to experience declines in income but at
the same time see improvements in climate. Thus, support is generated
for the thesis of Graves (1979) that markets adjust to leave utility
constant over space and that migration involves a tradeoff between
income and climate-related amenities. Migration to California, on the
other hand, leads to greater income and better climate, although both
increases are usually moderate. Those leaving South Carolina can expect
to receive substantial income increases, whether moving to a warmer or
colder climate.
Earlier regression results showed that although both income and
climate work together in determining migration flows between states, it
appears that each changes in relative importance at different points in
time. Whenever a factor changes in significance, the distribution of
migrants by destination can be altered. In the final chapter, the
findings of the study will be reviewed and conclusions will be drawn.

CHAPTER 5
SUMMARY AND CONCLUSIONS
Introduction
In this study a model of interstate labor force migration has been
developed and tested. First, decision-to-move models were estimated for
four states which have recently experienced out-migration. Migrants
from each state were found to have been selected by a different mix of
characteristics and circumstances. The out-migrants from all four
states were grouped together and a destination-choice model was es
timated. Seventy percent of the movers were found to have chosen among
eight alternative states. The factors which influenced this decision
were determined. Employment and cost-of-living variables were found to
be of little importance. Distance, income, and climate interacted to
determine location choice. Thus, support was generated for the Graves
(1979) hypothesis of constant spatial utility.
Recent individual models of migration, such as those carried out by
DaVanzo (1977) and Falaris (1978), have analyzed the choice among broad
United States regions. When areas are defined in this way, measures of
regional characteristics reflect averages of all states across the
region. As such, these measures hide many locational differences which
may enter into the migration decision. As an example, Florida and
Georgia are both located in the southeast region of the country, yet
they are certainly distinct enough to be considered as independent
110

Ill
alternatives to a potential migrant. For this reason, the state has
been chosen as the level of locational analysis in this study. An
argument could be made for defining regions at an even smaller level,
such as the urban area, but this would increase the number of alter
natives that would need to be included in a thorough analysis. Tech
nical and cost limitations prohibit this procedure.
Utilizing the results of both the decision-to-move analysis and the
destination-choice model, an aggregate migration forecast was obtained
for each origin and destination combination. Tests of forecasting
accuracy revealed mixed outcomes. The total number of out-migrants from
the four-state sample was predicted reliably. But because of a small
sample and poorly fitting, short time period equation the choice of
destination was not accurately predicted. More specific findings of the
study are reviewed in the following sections.
Review of Findings: Decision to Move
The results of the decision-to-move analysis are summarized by
state in Table 5.1. For the state of Ohio, marriage deters migration.
Mincer's (1978) finding that the presence of a spouse represents a tie
to present location is thus supported. Migrants from this state are
also likely to have higher family incomes. This result supports the
viewpoint that higher income families are more likely to demand goods
which can only be obtained by changing locations. Finally, Ohio migrants
are likely to be members of households headed by males.
As with Ohio migrants, New Yorkers with greater incomes are more
likely to change states of residence. Unemployed persons in New-York
are more likely to move than those who are employed, supporting the

112
Table 5.1 Summary of Findings: Decision to Move
Variable
Ohio
New York
South Carolina
Pennsylv
Famsz
+
-
+*
+
Sex
+*
+
+
+
Homeown
-
+
-
+
Lres
+
+
+
+
Lempl
+
-
+
Emself
-
-
-
+
Avage
+
+*
+
Ernst

JU
-

Marst
-*
-
_*
-
Race
+
-
-
+
Aveduc
-
+
+
+*
Prevmig
+
-
+
+
Ernst w
+
+*
-
+
Famy
+*
+
+*
* Significant at 95 percent confidence level or better.
+ Positive coefficient.
- Negative coefficient.
. Not in equation.

113
thesis that the unemployed are more responsive to migration opportunities.
This, it is recalled, was one of the major conclusions of the DaVanzo
(1977) study. Migrants from New York are also less likely to have held
their most recent job for a long period of time. This variable acts as
an indicator of attachment to present environment and job stability.
Migrants from New York are generally older than nonmigrants. In most
studies, migration rates are found to be lower in older age groups,
except at retirement. The continual presence of positive coefficients
in Florida-sending states equations, however, has led to the development
of an alternative hypothesis. It is believed that many older migrants
are moving in anticipation of retirement at a later date. The positive
and significant sign on family income complements this theory well.
Those with higher incomes are certainly likely to settle at their planned
retirement location at an early age. This is especially true if moving
involves a cut in pay or acceptance of a part-time job. Finally, household
heads from New York with employed spouses are more likely to leave the
state than those whose spouses do not work. Earlier it was theorized
that employed wives deter migration of families. This is because any
new location is less likely to present optimal employment and income
opportunities to both spouses than it is for just one of them. The
positive and significant coefficient on this variable, however, in
dicates that increasing female labor force participation could lead to
more migration from New York. Many spouses who are employed prior to
migration may no longer be employed after the move. Mincer (1978) found
evidence that one family member frequently dominates the migration
decision. Thus, spouses become tied to this decision regardless of the
personal consequences.

114
South Carolina migrants also tend to be older and earn higher
incomes than nonmigrants from that state. Married persons are less
likely to move than those who are not married. Finally, migrant fam
ilies tend to be of larger size than families that remain in South
Carolina. Household size, a priori, was expected to act as a constraint
upon migration.
Education proved to be an important determinant of migration in
Pennsylvania, with more educated people being more likely to leave. But
migrants from the state were also likely to earn lower family income
than those who remained. For New York, Ohio, and South Carolina, the
evidence confirmed the view that increases in income increased the
demand for location-specific goods found in such states as Florida or
California. For Pennsylvania, it appears that greater income increases
the probability of remaining in that state.
The positive income effects found for three of the four states
conflicts with DaVanzo's (1977) finding that the migration process, or
the returns to it, are on balance an inferior good. The difference in
results may arise because this study is only considering states which
have typically sent migrants to Sunbelt states. Sunbelt migrants
typically move in search of increased leisure activities, the demand for
which is expected to increase with income. The additional discovery
that migrants from these states are usually older than nonmigrants adds
further credence to this theory. It is the older migrant planning for
future retirement who is likely to be seeking out more temperate climates.
Those who have in the past earned higher incomes have the greatest
ability to make such long distance, and sometimes income sacrificing.
moves.

115
Over time the migration determinant coefficients did not appear to
be very stable. But, over shorter time periods, there are fewer migrants
available for study. In addition, fewer variables could be considered
in logit estimation, leading to the possibility of misspecified equations.
A decline in the total number of out-migrants is forecast for the 1976
80 period. This is in line with previous expectations that the migration
surge of the early 1970s cannot be maintained.
Review of Findings: Destination Choice
All other things constant, migrants from Ohio, Pennsylvania, South
Carolina, and New York chose to move to states offering greater per-
capita income. This result, at first, seems to offer support to re
gional growth theories, such as that of Smith (1975), which portray
migration solely as a response to income or wage differentials between
regions. But it was discovered that income is only significant when a
variable reflecting differences in climate is included in the model.
Climate turned out to be the most significant determinant of destination
choice. The downward bias on the income coefficient when climate was
excluded from the model indicates a strong negative correlation between
income and climate. Support is thus lent to the Graves (1979) thesis
that long-distance migration involves a trade-off between income and
location-specific, climate-related amenities.
When destination-choice models were estimated for shorter time
periods, coefficient instability was discovered. For the 1972-76
period, climate was found to be a more important determinant of location
choice than during the 1968-72 period. Income,on the other hand,

116
lost significance. Migration to Sunbelt states also showed a large
increase during the later period. A possible conclusion is that the
demand for climate related amenities increased during the whole eight-
year period, creating a disequilibrium situation, which led to greater
levels of migration to warm weather states such as Florida, California,
and Texas.
A forecast of the distribution of migrants from the sample sending
states was developed for the 1976-80 period. This forecast was based
upon the distance, income, and climate relationship during the previous
four-year period. Thus, climate is the primary determinant in this
forecast, promising strong future levels of migration to Sunbelt states.
To the extent, however, that earlier period migration levels were a
response to a disequilibrium situation, the forecast may be overstated.
If migration during the 1972-76 period, in fact, moved the economy
toward a new spatial equilibrium, then that rate of sunbelt migration
would be expected to slow during the adjustment process. New exogenous
shocks could, however, lead to a continuation of current migration
patterns. Given the assumption of the continued importance of climate,
the 1976-80 forecast appears to be reasonable. While maintaining strong
rates of in-migration, Florida is predicted to lose some ground to its
chief competitor, the state of California.
Strengths and Weaknesses of the Study
This has been the first study of individual migration that an
alyzes the decision to move at the state level. Since a different set
of determinants was discovered for each state considered, this approach
has proven fruitful. Studies which have grouped all states together may

117
be hiding these differences. Ambiguous coefficients which appear in
these studies may be the result of aggregation. Destination choice was
also analyzed at the state level. Studies that have combined states
into larger alternative regions are believed to cover up locational
differences which the migrant would be likely to consider in deciding
where to move. This may explain why economic variables rarely turn out
to be significant in these studies.
This has also been the first individual migration model which has
been used for the purpose of forecasting. While tests of accuracy
showed the decision-to-move prediction to be reliable, less success was
obtained in forecasting destination choice. This was partly because of
a small sample size which caused the number of individuals who chose any
given alternative to be low. Poor fits for equations estimated for
short time periods added to this problem. In addition, evidence of
coefficient instability was discovered.
The use of a nine-year time frame for studying the migration
/
decision is the source of potential problems. Migration, in fact, took
place during each year of the period, yet the independent variables
were measured at the initial year. While the choice of a shorter time
period would have relieved this problem, other larger problems would
result. Fewer people migrate during shorter periods of time. This
makes the determination of variables which distinguish migrants from
nonmigrants more difficult. In addition, with fewer numbers of mi
grants, some variables must be excluded for successful logit estimation.
As stated earlier in the paper, it is not believed that the de
cision "whether to move" is always made independently of the decision
"where to move." A methodology which analyzes both decisions simul-

118
taneously would thus seem preferable to the two-stage methodology em
ployed here. But, using existing techniques, trying to estimate such a
model over a large sample while still considering many alternatives is
extremely expensive. In addition, since most individuals do not move
during any reasonably defined period, the observations in such a model
are heavily weighted toward the staying alternative. The current state
of the art in choice theory does not seem to possess the proper tools
for the analysis of decisions where the alternatives are so unbalanced.
Implications of the Study
The influence of age and family income on the decision to move and
the interaction of income and climate in determining destination choice
can be tied together in a consistent fashion. This study presents a
picture of an older than average person with relatively high income who
chooses to move, with his (or her) family to a state with a more tem
perate climate. Possibly this person is moving in anticipation of
retirement. In addition, this move involves a significant probability
of a reduction in income. Another equivalent interpretation is that
older people with higher incomes have a greater demand for location-
specific amenities than do other members of the labor force.
The aging of the "baby boom" generation combined with the recent
decline in the birth rate of the United States implies that an increas
ing absolute number and proportion of the population will be in older
age groups. This occurrence should lead to an increase in the number of
persons who are at risk to the process described above. The implications
for amenity-rich states such as Florida, California, and Texas are
startling. Population growth will continue in these states at perhaps'
even higher rates than have so far occurred. And this increase will
continue to be concentrated in older age groups.

119
The theoretical model and methodology employed in this study can
easily be applied to other states. The results obtained could then be
compared to the findings for the four states considered here. Including
more states in the decision-to-move analysis would also provide a larger
sample for the analysis of destination choice. Better fitting models
might thus be obtained and forecasting accuracy may be improved. Ul
timately, detailed forecasts of labor force migration between all
states could be developed.
Although the focus of this study has been upon labor force mi
gration, the model developed is general enough to be applied to other
demographic groups. The study of, retiree migration would, for instance,
be an important application. Factors that influence the decision to
move and destination choice of the elderly population could be determined.
Ultimately, a forecast of the number of older people who will move to
retirement states such as Florida, California, Arizona, and Texas could
be developed. This forecast could then be combined with labor force
migration predictions for these states. Forecasts of these two components
of population growth should aid in projecting state population figures
and in planning for economic growth. The forecasts can also be used as
input in projecting other economic activities at the state level. State
econometric models, for example, have proven to be very sensitive to the
assumptions made about population growth. This is especially true for
states which are experiencing rapid rates of migration.
It is believed that the findings of this study have added to the
understanding of the migration process. It is hoped that this research
has provided theoretical, methodological, and empirical contributions
that will be of value to others who are interested in the study of
migration. -

APPENDIX A
ALTERNATIVE MODEL FORMULATIONS: DECISION TO MOVE
Combined State Results
Table A.l contains the results of the decision-to-move analysis
when all four origin states are combined into a single sample.
Ordinary least squares estimates appear in the table. A problem
with available storage space prevented successful execution of the
logit program for a sample this large. The variables are defined
as they were in Chapter 3.
Variables that are significant at better than the .05 level are
sex, average age, family income, and marital status. These variables
are the same ones that often turn up significant at the state level.
But other factors, which only occasionally appeared to be important
for separate states, no longer are significant. Examples are average
education (significant in Pennsylvania) and spouse's employment status
(significant in New York). In addition, variables that have opposite
signs for different states become averaged into the combined state
coefficients. For example, in Chapter 3 it was found that Pennsylvania
migrants are negatively selected according to income. The positive
and significant coefficient on combined state family income hides
this discovery.
120

121
Table A.1
- Combined State Results:
Ordinary Least
Decision to Move,
Squares
1968-1977
Variable
Coefficient
t-Ratio
Famsz
.010
1.54
Sex
.278
3.96
Race
-.016
- .52
Homeown
.005
.18
Lempl
-.014
-1.52
Emself
-.047
-.78
Avage
.005
3.38
Aveduc
.018
1.85
Ernst
-.120
1.47
Famy
.005
2.23
Prevmig
.024
.81
Ernst w
.041
-1.47
Marst
-.275
-4.01
Lres
.011
1.00
Constant
-.148
-1.17
R2 =
.06

122.
2
Finally, the low R indicates that only 6 percent of the variation
of the migration decision is being explained by the regression equation.
Thus, the model does not fit well for all four states.
Annual Migration Results
In Table A.2 ordinary least squares results appear for the
decision-to-move model when the sample period is defined as the
1976-77 period. The logit program could not be used here
either. A requirement for logit estimation is that each explanatory
factor display variation within every category. Since there were
so few movers during a one-year period, most of the variables did not
meet this criterion within the migrant category.
The results indicate that very little of the variation in annual
migration is explained by the factors in the model. Only in the Ohio
equation does more than one variable appear to be significant. The
New York and Pennsylvania models contain no significant variables.
One possible conclusion is that the wrong set of explanatory variables
is used in the model and that it should be respecified. Instead, it
is concluded that the proportion of migrants in the sample during a
one-year time period is too small to permit any set of variables to
properly distinguish between migrants and nonmigrants. Longer time
periods should thus be chosen.
Redifinition of Qualitative Explanatory Variables
Table A.3 contains the results of the decision-to-move analysis
when the length of residence, length of employment, and education
variables are redefined to be two-category dummy variables.

Table A.2 Annual Migration Results: Decision to Move, 1976-1977, Ordinary
Variable
Ohio
Coefficient
t-Ratio
Least Squares
New York
Coefficient t-Ratio
South Carolina
Coefficient t-Ratio
Pennsylvania
Coefficient t-Ratio
Famsz
-.001
- .07
.002
.25
.001
.24
-.002
- .45
Sex
-.007
- .13
-.027
- .67
-.009
- .31
-.002
- .07
Race
-.053
-1.56
.017
.56
.021
1.12
-.0001
- .01
Homeown
.002
.05
-.023
- .96
-.041
-2.33
.018
1.04
Lempl
-.028
-2.72
-.004
- .47
.002
.39
.009
1.51
Emself
-.110
-1.68
.011
.24
.006
.16
.019
.51
Avage
.001
1.00
.0002
. 16
-.0001
- .13
-.001
- .76
Aveduc
.012
1.21
-.003
- .47
-.005
- .85
.007
1.16
Ernst
.086
1.12
.029
.64
.017
.39
-.032
- .99
Famy
.004
2.56
. 0005
.39
-.0008
- .63
-.0004
- .42
Prevmig
-.010
- .35
.026
1.10
-.007
- .25
-.008
- .39
Emstw
-.034
-1.10
.007
.31
-.032
-1.71
-.007
- .39
Marst
.051
.90
.017
.42
.057
1.85
.015
.46
Constant
-.085
- .90
-.005
- .08
.010
.19
-.001
- .03
R2 =
.11
R2 =
.02
R2 =
.07
R2 =
.04
123

Table A.3 Results of Logit Analysis With Alternative Dummies, 1968-1977
Variable
Ohio
Coefficient
t-Ratio
New York
Coefficient t-Ratio
South Carolina
Coefficient t-Ratio
Pennsylvania
Coefficient t-Ratio
Famsz
.231
1.93
-.251
-2.03
.205
1.70
.138
1.06
Sex
3.440
2.97
.712
.43
1.352
1.19
2.559
2.06
Race
1.116
1.28
-1.877
-1.70
-.480
- .75
.371
.56
Homeown
-.081
- .13
.903
1.52
-.212
- .37
.671
1.24
Lres
-.538
-1.03
1.096
2.00
.235
.44
-1.289
-2.41
Lempl
-1.036
-1.61
-1.750
-2.67
-1.382
-2.52
.014
.02
Emself
-1.726
-1.79
-.690
- .84
-2.513
-1.57
.692
.52
Avage
.085
2.37
.065
1.91
.084
2.22
.033
1.08
Aveduc
.823
1.51
.621
1.04
.438
.74
.349
.64
Ernst i
-
-
-3.450
-2.38
-.627
- .76
-
-
Famy
.035
1.46
. 134
2.75
.210
3.32
-.113
-1.60
Prevmig
.322
.73
-.510
- .82
.525
.86
.627
.97
Ernst w
.827
1.64
1.253
2.36
-.591
- .99
.538
.95
Marst
l
-5.213
-4.06
-1.452
- .85
-3.401
-2.51
-3.538
-2.54
Constant
-5.862
-4.58
-.094
- .06
-4.362
-3.30
-2.703
-2.53
124

Table A.3 (continued)
Summary Statistics Ohio
Log of Likelihood Function -83.8
Likelihood Statistic (at 0) 137.4
Level of Significance (at 0) .005
Likelihood Statistic (at mean) 31.8
Level of Significance (at mean) .005
Percentage Correctly Predicted 85%
Percentage of Migrants Correctly
Predicted 19%
New York
South Carolina
Pennsylvania
-63.0
-71.1
-73.2
144.4
175.6
54.0
.005
.005
.005
38.0
41.0
16.2
.005
.005
.10
84%
89%
86%
7%
25%
4%
i
125

126
The length of residence variable is assigned a value of 1 for those
living in their current house or apartment for three or more years. It
is assigned a value of 0 for those living in their homes for less than
three years. The variable is significant for the states of New York and
Pennsylvania. When defined as a multi-category variable, length of
residence was insignificant for all states. For the Pennsylvania
observations, the variable now takes on the expected (negative) sign,
while for New York the sign is positive.
The length of employment variable is assigned the value of 1 for
those employed at the same place for 18 months or more. Those at the
job for less than 18 months are given the value of 0. The variable
appears to be negative and significant for New York and South Carolina.
For New York this result is the same as that which occurred when the
variable was alternatively defined. For South Carolina it takes on
greater significance than under the previous formulation. For the
states of Ohio and Pennsylvania, the variable is insignificant under
both regions.
The education variable is assigned the value of 1 for those with 12
or more years of education, while those with less than 12 years of
education are given the value of 0. The variable is insignificant for
all states. When education was defined under the old formulation it was
positive and significant for the state of Pennsylvania and insignificant
for all other states.
For the length of residence and length of employment variables, re
definition has brought the results slightly closer to expectation. The
education variable was brought further from its expected effect as a
result of the revision. Overall, the equations are changed a little

127
by the redefinitions. The Ohio and South Carolina equations provide a
somewhat better fit as a result of the change while the New York and
Pennsylvania equations fit slightly poorer than with the formulation
used in Chapter 3.

APPENDIX B
ALTERNATIVE MODEL FORMULATIONS: DESTINATION CHOICE
Nominal Wage Model
Table B.l presents results obtained when a measure of individual
returns to migration is included as an explanatory variable in the
model of destination choice. Wages at each location are estimated
as a function of age, race, sex, marital status, education, union
membership, experience, and occupation. Using the procedures described
in Chapter 2, values of potential wages at each origin and destination
are estimated for every person in the sample. For each individual,
the variable Wdiff is defined as the difference between his (or her)
potential 1977 hourly wage at each possible destination and the potential
wage at his (or her) origin location. The remaining variables are
defined as they were in Chapter 4.
In this model, only climate (Tempdiff) turns out to be a signi
ficant determinant of destination choice. Both distance and potential
wage gains turn out insignificant. Overall, the model exhibits a
poorer fit than that obtained when the individual returns variable is
replaced with the average aggregate per-capita income gain variable used
in Chapter 4.
Real Wage Model
The wage-difference variable defined in the previous section was
deflated by state cost-of-living indices to obtain a new variable
128

129
Table B.l Nominal Wage Model: Destination Choice, 1968-1977
Variable
Distance
Tempdiff
Wdiff
Coefficient
-.264
.111
.157
Asymptotic t-Ratio
-1.33
3.47
1.67
Summary Statistics:
Log of Likelihood Function = 168.4
Likelihood Statistic (at 0) = 11.6
Level of Significance (at 0) = .01
Likelihood Statistic (at mean) = 2.6
Level of Significance (at mean) = (not significant at any
acceptable level)

130
called Rwdiff. Even when measured in real terms, the individual
returns variable remained insignificant. The overall fit of the
model remained poor. These results appear in Table B.2

131
Table B.2 Real Wage Model: Destination Choice, 1968-1977
Variable
Distance
Tempdiff
RWdiff
Coefficient
-.265
.110
.163
Asymptotic t-Ratio
1.33
3.47
1.64
Summary Statistics:
Log of Likelihood Function = 168.4
Likelihood Statistic (at 0) = 11.6
Level of Significance (at 0) = .01
Likelihood Statistic (at mean) =2.6
Level of Significance (at mean) = (not significant at any
acceptable level)

BIBLIOGRAPHY
Albright, L., Lerman, S.R., and Manski, C.F., "An Introduction to the
Multinominal Probit Model," Cambridge Systematics Inc. Technical
Report, Cambridge Systematics Inc., August 1977.
American Chamber of Commerce Researchers Association, Cost of Living
Indicators: Inter-City Index Report, American Chamber of Com
merce Researchers Association, March 1968.
Borts, G.H. and Stein, J.L., Economic Growth in a Free Market, Columbia
University Press, 1964.
Burns, J.F. and James, M.K., "Migration Into Florida: 1940-1973,"
Work Paper No. 4, Urban and Regional Development Center,
University of Florida, October 1973.
Chow, G.C., "Tests of Equality Between Sets of Coefficients in Two
Linear Regressions," Econometrica, Vol. 28 (1960), pp. 591-605.
Dalton, J.A. and Ford, E.J., Jr., "Concentration and Labor Earnings
in Manufacturing and Utilities," Industrial and Labor Relations
Review, Vol. 31 (October 1977), pp. 45-60.
DaVanzo, J., Why Families Move, R & D Monograph 48, U.S. Department
of Labor, Employment and Training Administration, U.S. Government
Printing Office, 1977.
Dunlevy, J.A. and Gemery, H.A., "The Role of Migrant Stock and Lagged
Migration in the Settlement Patterns of Nineteenth Century In
migrants," The Review of Economics and Statistics, Vol. 59 (May
1977), pp. 137-44.
Eckstein, 0. and Girla, I.A., "Long-Term Properties of the Price-Wage
Mechanisms in the United States, 1891 to 1977," The Review of
Economics and Statistics, Vol. 60 (1978), pp. 323-33.
Eldridge, H.T., "Primary, Secondary, and Return Migration in the
United States, 1955-60," Demography, Vol. 2 (1965), pp. 444-55.
Engler, S.D., "Labor Force Migration to Florida: 1970 to 1976,"
The Florida Outlook, Vol. 2 (September 1978), pp. 25-30.
Falaris, E.M., "Migration: A Study of Choice Among Discrete
Alternatives," Seminar paper presented at the University of
Florida, October 1978.
132

133
Graves, P.E., "A Life-Cycle Empirical Analysis of Migration and
Climate, by Race," Journal of Urban Economics, Vol. 6 (April
1979), pp. 135-47.
Greenwood, M.J., "An Analysis of the Determinants of Geographic Labor
Mobility in the United States," The Review of Economics and
Statistics, Vol. 51 (May 1969), pp. 189-94.
Greenwood, M.J., "Research on Internal Migration in the United States:
A Survey," Journal of Economic Literature, Vol. 13 (June 1975),
pp. 397-433.
Johnston, J., Econometric Methods, McGraw-Hill, Inc., 1963.
Kau, J.B. and Sirmans, C.F., "The Influence of Information Cost and
Uncertainty on Migration: A Comparison of Migrant Types,"
Journal of Regional Science, Vol. 17 (April 1977), pp. 89-96.
Laber, G., "Lagged Response in the Decision to Migrate: A Comment,"
Journal of Regional Science, Vol. 12 (August 1972), pp. 307-10.
Longino, C.F., Jr., "Going Home: Aged Return Migration in the United
States 1965-70," Based on a paper presented at the 31st meeting
of the Gerontological Society, 1979.
Lowry, I.S., Migration and Metropolitan Growth: Two Analytical Models,
Chandler Publishing Company, 1966.
Maddala, G.S. and Roberts, R.B., "Estimation of Econometric Models
Involving Self-Selection," Paper for European Meetings of the
Econometric Society in Vienna, September 1977.
McFadden, D., "Conditional Logit Analysis of Qualitative Choice
Behavior," in Frontiers in Econometrics, Zarembka, P. (ed.),
Academic Press, 1973, pp. 105-42.
McFadden, D., "The Measurement of Urban Travel Demand," Journal of
Public Economics, Vol. 3 (June 1974), pp. 303-28.
Mincer, J., "Family Migration Decisions," Journal of Political Economy,
Vol. 86 (October 1978), pp. 749-73.
Morrison, P.A., "Theoretical Issues in the Design of Population Mobility
Models," Environment and Planning, Vol. 5 (1973), pp. 125-34.
Muth, R.F., "Differential Growth Among U.S. Cities," in Papers in
Quantitative Economics, Quirk, J.P. and Zarley, A.M. (eds.),
The University Press of Kansas, 1968, pp. 311-55.
Nerlove, M. and Press, S.J., "Univariate and Multivariate Log-Linear
and Logistic Models," Rand Corporation Technical Report R-1306-
EDA/NIH, Rand Corporation, December 1973.

134
Orcutt, G.H. Greenberger, M., Korbel, J., N. Rivlin, A.M., Micro
analysis of Socioeconomic Systems: A Simulation Study, Harper
& Brothers, 1961.
Pindyck, R.S. and Rubinfeld, D.L., Econometric Models and Economic
Forecasts, McGraw-Hill, Inc., 1976.
Richardson, H., Regional Growth Theory, John Wiley, 1973.
Rothenberg, J., "On the Microeconomics of Internal Migration," in
Internal Migration: A Comparative Perspective, Brown, A.A.
and Neuberger, E. (eds.), 1977, pp. 183-205.
Smith, D., "Neoclassical Growth Models and Regional Growth in the
U.S.," Journal of Regional Science, Vol. 15 (August 1975),
pp. 165-82.
Schwartz, A., "Interpreting the Effect of Distance on Migration,"
Journal Political Economy, Vol. 81 (September/October 1973),
pp. 1153-69.
U.S. Department of Commerce, Bureau of the Census, Census of Popula
tion: 1970, Subject Reports, Final Report PC(2)-2E, "Migration
Between State Economic Areas," U.S. Government Printing Office,
1972.
U.S. Department of Commerce, Bureau of the Census, Current Population
Reports, Series P-25, No. 701, "Gross Migration by County:
1965-1970," U.S. Government Printing Office, 1977.
U.S. Department of Commerce, Bureau of the Census, Statistical
Abstract of the United States: 1978, U.S. Government Printing
Office, 1978.
U.S. Department of Commerce, Bureau of the Census, Statistical
Abstract of the United States: 1974, U.S. Government Printing
Office, 1974.
U.S. Department of Commerce, Bureau of the Census, Statistical
Abstract of the United States: 1970, U.S. Government Printing
Office, 1970.
U.S. Department of Commerce, National Oceanic and Atmospheric Admin
istration, Climatological Data, National Summary, Annual 1976,
U.S. Government Printing Office, 1976.
U.S. Department of Commerce, National Oceanic and Atmospheric Admin
istration, Climatological Data, National Summary, Annual 1972,
U.S. Government Printing Office, 1972.
U.S. Department of Commerce, National Oceanic and Atmospheric Admin
istration, Climatological Data, National Summary, Annual 1968,
U.S. Government Printing Office, 1968.

135
U.S. Department of Labor, Employment and Training Administration and
U.S. Department of Health, Education and Welfare, Employment and
Training Report of the President, U.S. Government Printing Office,
1978.
U.S. Department of Labor, Manpower Administration, Manpower Report of
the President, U.S. Government Printing Office, 1971.
Vanderkamp, J., "Return Migration: Its Significance and Behavior,"
Western Economic Journal, Vol. 10 (December 1972), pp. 460-65.

BIOGRAPHICAL SKETCH
Sheldon Donald Engler was born on March 30, 1955, in Philadelphia,
Pennsylvania. In 1972, he graduated from North Miami Senior High School
in North Miami, Florida. In 1975, he received a Bachelor of Arts degree
in economics and sociology from the University of South Florida, Tampa,
Florida. In 1978, he received a Master of Arts degree in Economics from
the University of Florida. In August 1979, Mr. Engler accepted employment
as a Research Economist at Louisiana State University, Baton Rouge,
Louisiana.
136

I certify that I have read this study and that in my opinion it
conforms to acceptable standards of scholarly presentation and is
fully adequate, in scope of quality, as a dissertation for the degree
of Doctor of Philosophy.
)
/
Vv r l-1
Je'rome W. Milliman, Chairman
Ptofessor of Economics
I certify that I have read this study and that in my opinion it
conforms to acceptable standards of scholarly presentation and is
fully adequate, in scope of quality, as a dissertation for the degree
of Doctor of Philosophy.
ym. <
*nry fishlcq'nd
Assistant Processor of Economics
I certify that I have read this study and that in my opinion it
conforms to acceptable standards of scholarly presentation and is
fully adequate, in scope of quality, as a dissertation for the degree
of Doctor of Philosophy.
Cy ^ (p /'
/ C-f > A*' £-ft -
David A. Denslow
Associate Professor of Economics

I certify that I have read this study and that in my opinion it
conforms to acceptable standards of scholarly presentation and is
fully adequate, in scope of quality, as a dissertation for the degree
of Doctor of Philosophy.
Stanley K. Smith
Assistant Professor of Economics
I certify that I have read this study and that in my opinion it
conforms to acceptable standards of scholarly presentation and is
fully adequate, in scope of quality, as a dissertation for the degree
of Doctor of Philosophy.
e.
John\f. Henretta
Assistant Professor of
Sociology
This dissertation was submitted to the Graduate Faculty of the Department
of Economics in the College of Business Administration and the Graduate
Council, and was accepted as partial fulfillment of the requirements for
the degree of Doctor of Philosophy.
December 1979
Dean, Graduate School



Table A.2 Annual Migration Results: Decision to Move, 1976-1977, Ordinary
Variable
Ohio
Coefficient
t-Ratio
Least Squares
New York
Coefficient t-Ratio
South Carolina
Coefficient t-Ratio
Pennsylvania
Coefficient t-Ratio
Famsz
-.001
- .07
.002
.25
.001
.24
-.002
- .45
Sex
-.007
- .13
-.027
- .67
-.009
- .31
-.002
- .07
Race
-.053
-1.56
.017
.56
.021
1.12
-.0001
- .01
Homeown
.002
.05
-.023
- .96
-.041
-2.33
.018
1.04
Lempl
-.028
-2.72
-.004
- .47
.002
.39
.009
1.51
Emself
-.110
-1.68
.011
.24
.006
.16
.019
.51
Avage
.001
1.00
.0002
. 16
-.0001
- .13
-.001
- .76
Aveduc
.012
1.21
-.003
- .47
-.005
- .85
.007
1.16
Ernst
.086
1.12
.029
.64
.017
.39
-.032
- .99
Famy
.004
2.56
. 0005
.39
-.0008
- .63
-.0004
- .42
Prevmig
-.010
- .35
.026
1.10
-.007
- .25
-.008
- .39
Emstw
-.034
-1.10
.007
.31
-.032
-1.71
-.007
- .39
Marst
.051
.90
.017
.42
.057
1.85
.015
.46
Constant
-.085
- .90
-.005
- .08
.010
.19
-.001
- .03
R2 =
.11
R2 =
.02
R2 =
.07
R2 =
.04
123


Table 3.10 (continued)
Summary Statistics
Ohio
New York
South Carolina
Pennsylvania
Log of Likelihood Function
-38.5
-36.7
-43.9
-45.5
Likelihood Statistic (at 0)
243
246
245
210
Level of Significance (at 0)
.005
.005
.005
.005
Likelihood Statistic (at mean)
11.4
23.4
29.8
18.0
Level of Significance (at mean)
. 10
.025
.01
.10
Percentage Correctly Predicted
95%
95%
94%
93%
Percentage of Migrants Correctly
Predicted
0%
15%
19%
7%
\


122.
2
Finally, the low R indicates that only 6 percent of the variation
of the migration decision is being explained by the regression equation.
Thus, the model does not fit well for all four states.
Annual Migration Results
In Table A.2 ordinary least squares results appear for the
decision-to-move model when the sample period is defined as the
1976-77 period. The logit program could not be used here
either. A requirement for logit estimation is that each explanatory
factor display variation within every category. Since there were
so few movers during a one-year period, most of the variables did not
meet this criterion within the migrant category.
The results indicate that very little of the variation in annual
migration is explained by the factors in the model. Only in the Ohio
equation does more than one variable appear to be significant. The
New York and Pennsylvania models contain no significant variables.
One possible conclusion is that the wrong set of explanatory variables
is used in the model and that it should be respecified. Instead, it
is concluded that the proportion of migrants in the sample during a
one-year time period is too small to permit any set of variables to
properly distinguish between migrants and nonmigrants. Longer time
periods should thus be chosen.
Redifinition of Qualitative Explanatory Variables
Table A.3 contains the results of the decision-to-move analysis
when the length of residence, length of employment, and education
variables are redefined to be two-category dummy variables.


19
A change in residence will, by itself, cause little change in quantity
of food consumed. Changes in food buying behavior will occur if a move
is accompanied by an income increase or by a change in relative food
prices.
There are also goods contained in the location bundle which can
sometimes only be obtained in varying quantities by moving. An example
is climate. Although typically defined as a characteristic of a place
rather than a good within a location bundle, climate does display
certain characteristics of goods. The quantity of climate can be
represented by measuring factors such as temperature or nearness to the
coast. The price can be measured in terms of moving costs or foregone
earnings involved in moving to a more temperate climate. So although it
is not produced by man, from the consumer's standpoint climate displays
characteristics that other goods possess. The need to move in order to
obtain substantial changes in consumption makes climate an important
determinant in the migration decision. This is true for other fixed-
location amenities such as symphonies, sporting events, and certain
public services.
Another important factor distinguishing locations for the consumer
is the existence of differential employment opportunities across space.
A person may have a potential job available in one place that will earn
him (or her) greater income than jobs at other locations may offer.
Higher income, in turn, allows for greater consumption. These increased
consumption possibilities include leisure activities which are related
to climate. Thus, although many of the goods in the location bundle are
available at all locations, different quantities may be consumed at
various locations because of the uneven spatial distribution of em
ployment and income possibilities.


LIST OF FIGURES
PAGE
Figure 2.1 Illustration of Forecasting Methodology:
Decision to Move 36
Figure 2.2 Illustration of Forecasting Methodology:
Destination Choice 37
xx


96
model that is evaluated at the mean value of the dependent variable. In
the next section this version of the model of destination choice will be
utilized to obtain a forecast of destination choice.
Application to Forecasting
The procedure used to forecast destination choices follows the
description set forth in Chapter 2. The three-variable version of the
destination-choice equation is first estimated for the 1968-72 period.
The results appear in Table 4.6. Distance and income differences are
still significant variables, with the temperature-differentials var-
riable losing much of its explanatory power. The overall equation is
significant at only the .10 level.
The 1968-72 coefficients were applied to values of the explanatory
variables in order to derive a 1972-76 forecast for each pair of origin
and destination locations. For example, for the New York to Florida
forecast the distance between the two states (constant over time), the
1972 per-capita income difference between them, and the 1972 average
temperature difference are substituted into the 1968-72 equation. The
predicted probability is then applied to the total number of migrants
leaving New York for the eight alternative states between 1972 and 1976.
The total number of available migrants is determined by assuming that
the proportion of migrants who chose alternatives other than the eight
states considered here is the same as it was during the most recent
four-year period. The number choosing alternatives is then subtracted
from the total number of migrants leaving New York for all states. The
resulting number is multiplied by the probability_predicted by the'
model. This process is repeated for each origin and destination pair.


27
Florida may exceed the gain resulting from a California to Florida move.
Therefore, more people may migrate to Florida from New York than from
California, despite the fact that Californians and New Yorkers may have
similar characteristics. There is a way, used in this study, to allow
for this occurrence without including a returns variable in the equation.
The model can be estimated for a sample that is restricted to one
origin at a time. In this way, everyone in the sample faces the same
alternative choices even if these choices are viewed relative to their
present location. Differences in migration propensities now occur as a
result of differences in the way individual characteristics are valued
at alternative locations and because of the effects that individual
attributes have upon the way in which people evaluate and respond to
migration opportunities. Estimating the model for various places also
allows for a comparison of the determinants of the migration decision
over space. If the factors important in explaining the New Yorker's
decision to move are significantly different from the factors affecting
the Californian's choice, then there is an additional rationale for
having separate models. Such an approach will be taken in the main body
of this study. Results obtained when origin states are combined are
presented in Appendix A.
Lack of time and resources prevents estimation for all possible
origins. The model will be estimated for four states. These are New
York, Pennsylvania, Ohio, and South Carolina. These states were chosen
because, in the sample to be studied, they are the leading origins- of
migrants to Florida, the destination of primary concern to this study.
In addition, all four are states with relatively large numbers of out-
migrants during the period under study.


81
then multiplied by average family size to obtain an estimate of the
actual number of potential migrants from each state. The average family
size measure is also estimated from the Panel Study sample.5 Finally,
the potential population at risk is multiplied by the average probability
of migraiton estimated for the Panel Study. The numbers in Table 3.13
result.
Using the procedure put forth in Chapter 2, estimates of aggregate
labor force migration were developed from sample characteristics for the
periods 1968-72 and 1972-76. Average predicted probabilities were
calculated for each period based upon the regression results. The
population at risk for each state was derived using the method discussed
in the last paragraph. The estimates for these two periods and the
forecast are presented together in Table 3.14. It can be seen that for
the 1968-72 period, New York is estimated to have had the largest amount
of out-migration. Pennsylvania is next, followed by Ohio and South
Carolina. For the 1972-76 period, total migration out of the four
states increased. Only New York did not share in this increase. The
largest jump occurred in Ohio which is estimated to have 457,000 more
out-migrants during this period than the number who left in the 1968-72
period. This increase was large enough to put that state in second
place to New York. Pennsylvania and South Carolina showed small in
creases from the previous periods. For the 1976-80 period, total out
migration is expected to decline from the previous period. This decline
will be felt in all states except New York. The greatest decline
5Since this study has found family size to be insignificant in the
determination of the decision to move, the family-size measure is obtained
from the total (migrant and nonmigrant) sample.


14
between each destination and the origin, and an interaction term between
the present value of the wage differences and a dummy indicating whether
a person had resided at that destination in the recent past. The
results appear in Table 1.4. None of the variables turn out to be
significant at better than the .10 level. Only the wage variable is
even close to being significant, indicating that families are likely to
move to destinations where the earnings gains are greatest. The sign of
the interaction term indicates that families are more likely to move to
an area where they have lived before than to one where they have never
lived, especially if the family earnings they could receive there are
higher than what they could earn by staying where they are. The unemploy
ment rate is insignificant and shows an unexpected positive sign. The
coefficient of distance shows that migrants are more likely to choose
closer destinations, although it too is insignificant. The model is
estimated using conditional logit, a maximum likelihood technique
designed to analyze multinomial choices among discrete alternatives.
DaVanzo considered the migration decision from the viewpoint of the
married couple. Mincer (1978) notes that there are so far very few
migration studies that consider the effect of family relations on this
decision. He points out that at the individual level a person should
choose to move to that location at which the return is at a maximum.
For a family the optimal move is one which maximizes the combined return
to the family. Whether this last criterion is satisfied or not, frequently
a family will end up at a location which does not reward every family
member with the maximum possible return. The members of the family who
do not reach their private optimum are called tied movers. Mincer
points out that the conflict which could result from tied migration can


I certify that I have read this study and that in my opinion it
conforms to acceptable standards of scholarly presentation and is
fully adequate, in scope of quality, as a dissertation for the degree
of Doctor of Philosophy.
)
/
Vv r l-1
Je'rome W. Milliman, Chairman
Ptofessor of Economics
I certify that I have read this study and that in my opinion it
conforms to acceptable standards of scholarly presentation and is
fully adequate, in scope of quality, as a dissertation for the degree
of Doctor of Philosophy.
ym. <
*nry fishlcq'nd
Assistant Processor of Economics
I certify that I have read this study and that in my opinion it
conforms to acceptable standards of scholarly presentation and is
fully adequate, in scope of quality, as a dissertation for the degree
of Doctor of Philosophy.
Cy ^ (p /'
/ C-f > A*' £-ft -
David A. Denslow
Associate Professor of Economics


TO MY PARENTS


84
Table 3.15 Comparison of Panel Study Derived Out-Migration
Estimates to Census Estimates
State of Origin
Panel Study Estimate, 1968-72 Census Estimate, 1965-70
Ohio
714,000 787,546
New York
1,708,000 1,329,432
South Carolina
310,000 248,609
Pennsylvania
1,121,000 781,684
Total
3,853,000 3,147,271


67
Variable
Famsz
Sex
Homeown
Lres
Lempl
Emself
Avage
Ernst
Marst
Race
Aveduc
Prevmig
Ernst w
Famy
Constant
Table 3.8 Results of Logit Analysis: Decision to
Move, Pennsylvania, 1968-1977
Coefficient
.251
1.859
. 105
.022
.253
1.257
.021
Asymptotic t-Ratio
1.85
1.51
. 19
.11
1.34
.88
.59
-2.625
.109
.638
.225
.869
-.219
-1.84
. 17
2.88
.36
1.66
-2.65
-5.650 -2.78
Summary Statistics:
Log of Likelihood Function = -71.6
Likelihood Statistic (at 0) = 57.2
Level of Significance (at 0) = .005
Likelihood Statistic (at mean) = 19.4
Level of Significance (at mean) = .10
Percentage Correctly Predicted = 87%
Percentage of Migrants Corrected Predictly =
7%


134
Orcutt, G.H. Greenberger, M., Korbel, J., N. Rivlin, A.M., Micro
analysis of Socioeconomic Systems: A Simulation Study, Harper
& Brothers, 1961.
Pindyck, R.S. and Rubinfeld, D.L., Econometric Models and Economic
Forecasts, McGraw-Hill, Inc., 1976.
Richardson, H., Regional Growth Theory, John Wiley, 1973.
Rothenberg, J., "On the Microeconomics of Internal Migration," in
Internal Migration: A Comparative Perspective, Brown, A.A.
and Neuberger, E. (eds.), 1977, pp. 183-205.
Smith, D., "Neoclassical Growth Models and Regional Growth in the
U.S.," Journal of Regional Science, Vol. 15 (August 1975),
pp. 165-82.
Schwartz, A., "Interpreting the Effect of Distance on Migration,"
Journal Political Economy, Vol. 81 (September/October 1973),
pp. 1153-69.
U.S. Department of Commerce, Bureau of the Census, Census of Popula
tion: 1970, Subject Reports, Final Report PC(2)-2E, "Migration
Between State Economic Areas," U.S. Government Printing Office,
1972.
U.S. Department of Commerce, Bureau of the Census, Current Population
Reports, Series P-25, No. 701, "Gross Migration by County:
1965-1970," U.S. Government Printing Office, 1977.
U.S. Department of Commerce, Bureau of the Census, Statistical
Abstract of the United States: 1978, U.S. Government Printing
Office, 1978.
U.S. Department of Commerce, Bureau of the Census, Statistical
Abstract of the United States: 1974, U.S. Government Printing
Office, 1974.
U.S. Department of Commerce, Bureau of the Census, Statistical
Abstract of the United States: 1970, U.S. Government Printing
Office, 1970.
U.S. Department of Commerce, National Oceanic and Atmospheric Admin
istration, Climatological Data, National Summary, Annual 1976,
U.S. Government Printing Office, 1976.
U.S. Department of Commerce, National Oceanic and Atmospheric Admin
istration, Climatological Data, National Summary, Annual 1972,
U.S. Government Printing Office, 1972.
U.S. Department of Commerce, National Oceanic and Atmospheric Admin
istration, Climatological Data, National Summary, Annual 1968,
U.S. Government Printing Office, 1968.


Table 3.12
- Results
of Logit Analysis:
Decision
to Move,
1972-1976
Variable
Ohio
Coefficient
t-Ratio
New York
Coefficient t-
Ratio
South Carolina
Coefficient t-Ratio
Pennsylvania
Coefficient t-Ratio
Famsz
.050
.34
.005
.03
.167
1.09
.307
1.96
Sex
-
-
-
-
.617
.43
1.534
1.69
Homeown
-.015
- .02
-1.212 -1
.66
-.710
-1.03
-.119
- .17
Lres
-
-
-
-
-
-
-
-
Lempl
.610
3.13
-.244 -1
.49
-.398
-1.71
-.380
-2.20
Emself
-.715
- .65
-
-
.921
.84
-
-
Avage
-.019
- .62
.016
.60
.004
.10
.084
2.50
Ernst
-
-
-
-
-.252
- .19
-
-
Marst
-1.108
-1.27
.475
.59
-.681
- .53
-4.011
- 3.15
Race i
-
-
-.576
.71
-.902
- .89
.680
.88
Aveduc
.109
.55
. 148
.78
-. 126
- .44
.378
1.31
Prevmig
.280
.51
-.515
.74
1.387
1.62
.011
.012
Ernst w
.371
.56
-.432
.57
-. 107
- .15
1.516
1.64
Famy
-.029
- .70
.308
.74
.967
2.25
.0004
.005
'Constant
-3.805 -2.23
-2.520 -1.56
-2.148 .83
-7.128
-3.12


127
by the redefinitions. The Ohio and South Carolina equations provide a
somewhat better fit as a result of the change while the New York and
Pennsylvania equations fit slightly poorer than with the formulation
used in Chapter 3.


58
Assuming no other variables in the model, the equation can be written as:
(1)
where a, 6 and 82 are coefficients and E is a random disturbance term.
Y = a + 8 X + B2X2 + E
X
1 and X2
are zero-one
variables we
can further say that:
E
(Y
X1 =
1 and X2 = 0)
= a +
h
(2)
E
(Y
X1 =
0 and X2 = 1)
= a +
b2
(3)
E
(Y
X1 =
0 and X2 = 0)
= a
(4)
Now suppose instead of using the above approach we define a single
variable, X, which takes on the values of 0, 1, and 2 for each of the
three categories. Now the equation can be written as:
Y = a + 6X + E (5)
where a and B are parameters and E is the disturbance. Now we can
say that:
E (Y X = 0) = a (6)
E (Y X = 1) = a + B (7)
E (Y
X = 2) = a + B
(8)
This formulation turns out to be equivalent to estimating the model of
equation (1) with the following restriction attached:
^1 32
(9)
While this restriction should be considered, it may be preferable to the
restrictions imposed when the variable is collapsed into two categories
(as discussed earlier). Using the latter approach, information which
is available about alternative categories goes unused. The chosen
technique uses all of the information, but it imposes rather tight
restrictions on the relationship between categories.


12
In order to test the validity of this hypothesis, a model must be esti
mated using individual data on non-migrants as well as migrants.
Morrison (1973) suggests the use of a theoretical framework which
combines two models of migration. First, he suggests a microeconomic
model which analyzes the decision "whether to move." This model, he
claims, should reflect the idea that individuals have variable decision
thresholds. That is, those with lower thresholds are more likely to
seek out and respond to opportunities elsewhere, and hence are more
likely to move. The level of threshold variability is determined by
such factors as the individual's position in the life cycle, occupa
tionally induced contraints on movements, and prior migration experi
ence. The second type of model suggested by Morrison is one which
allocates those who do move (as a result of the first decision) among
alternative destinations. Thus, this model addresses the question
"where to move." This model, he claims, is more macroeconomic in con
tent .
An empirical study by DaVanzo (1977) divides the migration choice
into the two-stage process suggested above. She looked at a sample of
married couples from the University of Michigan Study of Income Dynamic
First, she estimates "whether to migrate" equations. The dependent
variable in these equations is a zero-one dummy indicating whether or
not a move was made for each couple. A one-year time period (1971-72)
was selected. The explanatory variables can be categorized as employ
ment status, returns to migration, composition of family earnings,
location-specific assets (such as homeownership), previous migration,
age, and education. The basic conclusions can be summarized as follows
Families whose heads are unemployed or are dissatisfied with their jobs


95
Table 4.5 Logit Results: Destination Choice,
1968-1977, Climate Excluded
Variable Coefficient Asymptotic t-Ratio
Distance -.016 -.10
Incdiff .005 .02
Summary Statistics:
Log of Likelihood Function = 174.19
Likelihood Statistic (at 0) = .02
Level of Significance (at 0) = (not significant at any
acceptable level)
Likelihood Statistic (at mean) = -4.5
Level of Significance (at mean) = (not significant at any
acceptable level)


CHAPTER 1
INTRODUCTION
Problem Statement
During recent years experts on migration have begun to focus their
efforts upon understanding the decision to move at the individual level.
A continuous stream of detailed survey data has helped to bring about
this endeavor. In addition, increased availability of sophisticated
computer software packages and advances in econometric techniques have
facilitated the handling and analysis of this data. Until now the
research which has come about with the aid of these tools has been
limited to answering age-old questions in the migration literature.
What are the characteristics which distinguish migrants from nonmi
grants? Do individuals move in the direction of higher wages? How does
the unemployment rate affect the migration decision? What role does
climate play in the choice of location? Does migration lead to regional
income convergence?
Since many of these issues are as yet unsettled, they will be
considered here. The ultimate test of a theory, however, is its ability
to predict the future. With the proportion of population change re
sulting from natural increase on the decline, migration forecasts become
more of a necessity. This is particularly true for states which are
experiencing rapid population growth such as Florida, California, Texas,
and Arizona. In addition, forecasts of other economic events at the
1


116
lost significance. Migration to Sunbelt states also showed a large
increase during the later period. A possible conclusion is that the
demand for climate related amenities increased during the whole eight-
year period, creating a disequilibrium situation, which led to greater
levels of migration to warm weather states such as Florida, California,
and Texas.
A forecast of the distribution of migrants from the sample sending
states was developed for the 1976-80 period. This forecast was based
upon the distance, income, and climate relationship during the previous
four-year period. Thus, climate is the primary determinant in this
forecast, promising strong future levels of migration to Sunbelt states.
To the extent, however, that earlier period migration levels were a
response to a disequilibrium situation, the forecast may be overstated.
If migration during the 1972-76 period, in fact, moved the economy
toward a new spatial equilibrium, then that rate of sunbelt migration
would be expected to slow during the adjustment process. New exogenous
shocks could, however, lead to a continuation of current migration
patterns. Given the assumption of the continued importance of climate,
the 1976-80 forecast appears to be reasonable. While maintaining strong
rates of in-migration, Florida is predicted to lose some ground to its
chief competitor, the state of California.
Strengths and Weaknesses of the Study
This has been the first study of individual migration that an
alyzes the decision to move at the state level. Since a different set
of determinants was discovered for each state considered, this approach
has proven fruitful. Studies which have grouped all states together may


79
occurred, although this was usually restricted to the insignificant
variables. It should be emphasized that all four-year period equations
are misspecified because some variables could not be included in the
estimation. This makes comparisons of the two periods less meaningful.
There is evidence, however, of instability of the decision-to-move
coefficients between the two time periods. This has negative impli
cations for forecasting based upon previous coefficients. But the
reasonableness of the forecasts which have been obtained provides support
for this methodology. Coefficient changes which occur may be balanced
in such a way as to affect the average probability, the key variable in
the forecast, only slightly.
The final step in the analysis is to apply the 1972-76 coefficients
to the characteristics of people living in the sample states in 1976.
Thus, a 1976-80 forecast is obtained. These predictions are presented
in Table 3.13. The results show that 24 of the Panel Study members who
lived in New York in 1976 will have left by 1980. For both South Carolina
and Pennsylvania 19 migrants are forecast. Ohio is expected to lose 18
migrants during the period.
An aggregate forecast is also presented in Table 3.13. It is
obtained in the following manner. First, the number of persons in the
labor force for each state in 1976 is taken from the Employment and
Training Report of the President. Then, in order to be comparable to
the Panel Study sample, working spouses are removed from the measure.
This is accomplished by multiplying the labor force number by an esti
mate of the proportion of working spouses in the labor force. The
estimates used are the proportions that occur in the Panel Study sample
states in 1976. The resulting measure of labor force household heads is


CHAPTER 5
SUMMARY AND CONCLUSIONS
Introduction
In this study a model of interstate labor force migration has been
developed and tested. First, decision-to-move models were estimated for
four states which have recently experienced out-migration. Migrants
from each state were found to have been selected by a different mix of
characteristics and circumstances. The out-migrants from all four
states were grouped together and a destination-choice model was es
timated. Seventy percent of the movers were found to have chosen among
eight alternative states. The factors which influenced this decision
were determined. Employment and cost-of-living variables were found to
be of little importance. Distance, income, and climate interacted to
determine location choice. Thus, support was generated for the Graves
(1979) hypothesis of constant spatial utility.
Recent individual models of migration, such as those carried out by
DaVanzo (1977) and Falaris (1978), have analyzed the choice among broad
United States regions. When areas are defined in this way, measures of
regional characteristics reflect averages of all states across the
region. As such, these measures hide many locational differences which
may enter into the migration decision. As an example, Florida and
Georgia are both located in the southeast region of the country, yet
they are certainly distinct enough to be considered as independent
110


98
Table 4.7 presents the Panel Study forecast for the period 1972-76.
The forecast Is also compared to actual destination choices during that
period. For New York, Pennsylvania, Ohio, and South Carolina combined,
11 migrants are forecast to come to Florida. This compares to 15 who
actually moved to Florida. This is about a 27 percent error. For
California, 3 migrants are predicted. There were actually 8 migrants
from the four states who chose California. Nine migrants are predicted
to go to New Jersey in comparison to 5 who actually did. It is forecast
that 6 movers would go to Illinois, while actually there were none.
Five migrants are expected to go to Virginia. This compares to 8 who
moved there between 1972 and 1976. Four migrants are predicted for
Michigan, compared to 2 people who actually moved there. Finally, 3 are
forecast to move to Texas, the same number that actually moved there.
The overall ability to forecast thus appears to vary from state to
state. Part of the problem, it is believed, results from the small
sample size. A better fitting model was obtained for the 1968-77 time
period. But when the time period is reduced there are smaller numbers
of migrants from which to distinguish. This is believed to contribute
to the reduced significance of the model and the less than adequate
forecasts which were derived from it.
In order to produce a forecast for the 1976-80 period the model was
estimated for the four-year period just prior to it. The results are
presented in Table 4.8. Only the temperature-difference variable
remains significant at the .005 level, with the distance and income-
difference variables falling off to the .10 level. While levels of
significance have changed, the coefficients do not appear to have changed
very much since the 1968-72 period. Only the temperature-difference


66
status, and family income. Marital status and family income show the
expected signs. Contrary to expectations, family size is positive and
significant. Again average age carries a positive sign.
Although insignificant at the 95 percent confidence level, length
of employment and self-employment are important and carry the correct
signs. Of the remaining variables, sex, homeownership, employment
status, average education, previous migration, and spouse's employment
status influence migration decisions in the expected directions. As in
Ohio and New York, length of residence is positively related to migration.
Race displays a negative relationship, indicating that households
headed by blacks are more likely to leave the state.
The South Carolina equation provides a very good fit of the data.
The equation is significant by both criteria (at zero and at the mean)
at better than the .005 level. Ninety percent of the individuals in the
sample were correctly predicted by the model, with 31 percent of the
migrants' choices being predicted accurately.
Logit Results: Pennsylvania
In Table 3.8 are the results of the regression analysis for Penn
sylvania. Variables which are significant at the .05 level are average
education and family income. Average education shows the expected
positive sign. Family income, however, is negatively related to the
probability of moving. This is truly unexpected for a variable that has
been positive and significant for the other three states.
Close to being significant are family size and marital status. As
in Ohio and South Carolina, family size takes on-a -positive sign.
Marital status has the hypothesized negative sign, indicating marriage*


57
would result in reduction of degrees of freedom and, more importantly,
would cause logit estimation to become prohibitively expensive.1
A second method of handling the problem would be to redefine the
dummy variables so as to have fewer categories. In the limiting case
only one zero-one dummy is used. Some cutoff point is chosen and a
value of one is assigned to those observations for which the cutoff is
exceeded and a value of zero is assigned otherwise. In this study, for
example, the education variable could be defined so that those individuals
with 12 years or more of schooling were given values of one with the
remaining observations equaling zero. Assumed in such an approach is
that increases in the level of education up until the 12th grade have no
effect on the probability of moving. In addition it is assumed that
increases in education beyond high school have no effect. To validly
use such an approach some prior expectation of the proper cutoff point
is required. Appendix A presents the results of logit estimation when
this approach is adopted. Not having prior knowledge of the correct
cutoff point, the choices used are somewhat arbitrary.
The actual scaling used for the main part of this study (see Table
3.3) also imposes restrictions upon the model. To illustrate these
restrictions, this approach is compared to the conventional method of
defining such data. Suppose an explanatory variable contains three
categories. Following the conventional method, two dummy variables are
introduced. (If three dummies are used, then the constant term must be
eliminated.) Call the dependent variable Y and the two dummies and
1 In fact, given the computer program limitations, logit estimation
would have been impossible.


45
Pennsylvania, Ohio, or South Carolina. After California, these are also
the leading out-migration states in the sample. These four states were
thus selected as the major sample for this analysis.
Table 3.3 defines the variables to be considered in the decision-
to-move analysis. In the next section, the rationale for including each
variable will be discussed and a first look at the data will be provided.
Comparison of Migrants and Nonmigrants
Family Size
Table 3.4 presents a comparison of migrant and nonmigrant char
acteristics in the four-state sample. A priori it is expected that
persons in larger families are less likely to move since additional
people represent ties to present location. Children in school or a
spouse who is working are both examples of migration ties. Comparison
of the mean values of family size among migrants and nonmigrants, however,
reveals very little difference. Both groups average about five persons
per family. Thus, at first glance, family size does not appear to be a
constraining force.
Sex
Households with male heads are expected to move more often than
households with female heads. One reason is that men are more likely to
be employed in occupations where job transfers are common. In addition,
higher rates of unemployment for women at many locations and lower
income opportunities would deter female migration. The mean value of
the sex variable, however, is only slightly higher for movers than it is
for nonmovers. Eighty-seven percent of the households that moved were
headed by men. It may be that since women are also likely to be earning


126
The length of residence variable is assigned a value of 1 for those
living in their current house or apartment for three or more years. It
is assigned a value of 0 for those living in their homes for less than
three years. The variable is significant for the states of New York and
Pennsylvania. When defined as a multi-category variable, length of
residence was insignificant for all states. For the Pennsylvania
observations, the variable now takes on the expected (negative) sign,
while for New York the sign is positive.
The length of employment variable is assigned the value of 1 for
those employed at the same place for 18 months or more. Those at the
job for less than 18 months are given the value of 0. The variable
appears to be negative and significant for New York and South Carolina.
For New York this result is the same as that which occurred when the
variable was alternatively defined. For South Carolina it takes on
greater significance than under the previous formulation. For the
states of Ohio and Pennsylvania, the variable is insignificant under
both regions.
The education variable is assigned the value of 1 for those with 12
or more years of education, while those with less than 12 years of
education are given the value of 0. The variable is insignificant for
all states. When education was defined under the old formulation it was
positive and significant for the state of Pennsylvania and insignificant
for all other states.
For the length of residence and length of employment variables, re
definition has brought the results slightly closer to expectation. The
education variable was brought further from its expected effect as a
result of the revision. Overall, the equations are changed a little


71
Table 3.9 (continued)
South Carolina
Combined
States
Variable
Coefficient
t-Ratio
Coefficient
t-Ratio
Famsz
.018
1.58
.010
1.54
Sex
.174
1.35
.278
3.96
Homeown
-.020
- .40
.005
.18
Lres
.040
1.10
.011
1.00
Lempl
-.042
-2.21
-.014
-1.52
Emself
-.228
-1.76
-.047
- .78
Avage
.009
2.83
.005
3.38
Ernst
-.032
- .28
-.120
-1.47
Marst
-.399
-3.16
-.275
-4.01
Race
-.082
-1.32
-.016
- .52
Aveduc
.006
.29
.018
1.85
Prevmig
.069
.98
.024
.81
Emstw
-.066
-1.27
.041
1.47
Famy
.026
4.03
.005
2.23
Constant
-.045
- .31
CO
rH
l
-1.17
R2 =
.18
R2 =
.06


50
Length of Employment
As with the last variable, the length of employment would indicate
attachment to present environment. In addition, it can be viewed as a
measure of ability to hold a job. A person who has worked for his (or
her) present employer for only a short period of time is more likely to
have been periodically unemployed. The mean value of this variable is
indeed somewhat lower for migrants than for nonmigrants in the states
under consideration. The average length of employment before moving is
about 21 months for migrants.
Self-Employment
Individuals who are self-employed are expected to migrate less. The
self-employed person is more likely to have an occupation that involves
heavy investment in capital equipment. Dentists and printers are examples.
Although this equipment may be transportable, the costs of moving and
re-establishing at some other location are significant enough to deter
migration. Self-employed workers are also likely to be part of smaller
organizations where job transfer is uncommon. Finally, those who are
employed in professions requiring state licenses, such as medical and
legal fields, are bound to their state to some extent by their licenses.
The mean values for the four sample states show very little difference
in propensity to migrate based upon self-employment, with migrants
showing a slightly higher degree of self-employment. Ten percent of the
migrants from these states were self-employed before moving.


43
Table 3.1 (continued)
State
Washington
West Virginia
Wisconsin
Wyoming
Number of In-
Migrants
6
2
6
2
Number of Out-
Migrants
9
0
2
0
Net Migration
-3
2
4
2


7
likely to migrate) and for residents (making them more likely to re
main) Growth in the resident working-age population is important
because higher growth implies an increased supply of labor occurring
from within the region, thus reducing the number of job opportunities
available to potential in-migrants. Growth in armed services personnel
is included since it is a component of population change that is not
normally explained by the other variables in the equation.
Neoclassical regional growth models theorize that migration is
induced solely by the existence of interregional wage differentials.
That is, migration occurs from low-wage regions (where labor is plen
tiful relative to capital) to high-wage regions (where labor is rela
tively scarce). Increased supply of labor in high-wage regions is
hypothesized to put downward pressure on wages, while decreased supply
of labor in low-wage regions exerts upward pressure on wages. Thus, the
model leads to an equilibrium where wage differentials between regions
are eliminated. Smith (1975) includes a labor sector in his neoclas
sical growth model which posits net migration to be a function of the
difference between the local wage and the national wage. He estimates
his model for various historical periods and generates mild support for
the convergence theory. This is in contrast to other tests of neoclas
sical growth theory, such as that undertaken by Borts and Stein (1964),
in which the theory is rejected. But Smith also discovered a decrease
in responsiveness of potential migrants to income differentials over
time. This he partly attributes to the existence of unemployment which
discourages migration. Neoclassical growth models, whether regional or
national in scope, assume full employment of labor.


117
be hiding these differences. Ambiguous coefficients which appear in
these studies may be the result of aggregation. Destination choice was
also analyzed at the state level. Studies that have combined states
into larger alternative regions are believed to cover up locational
differences which the migrant would be likely to consider in deciding
where to move. This may explain why economic variables rarely turn out
to be significant in these studies.
This has also been the first individual migration model which has
been used for the purpose of forecasting. While tests of accuracy
showed the decision-to-move prediction to be reliable, less success was
obtained in forecasting destination choice. This was partly because of
a small sample size which caused the number of individuals who chose any
given alternative to be low. Poor fits for equations estimated for
short time periods added to this problem. In addition, evidence of
coefficient instability was discovered.
The use of a nine-year time frame for studying the migration
/
decision is the source of potential problems. Migration, in fact, took
place during each year of the period, yet the independent variables
were measured at the initial year. While the choice of a shorter time
period would have relieved this problem, other larger problems would
result. Fewer people migrate during shorter periods of time. This
makes the determination of variables which distinguish migrants from
nonmigrants more difficult. In addition, with fewer numbers of mi
grants, some variables must be excluded for successful logit estimation.
As stated earlier in the paper, it is not believed that the de
cision "whether to move" is always made independently of the decision
"where to move." A methodology which analyzes both decisions simul-


PAGE
Table 5.1 Summary of Findings: Decision
to Move 112
Table A.1 Combined State Results:
Decision to Move, 1968-1977,
Ordinary Least Squares 121
Table A.2 Annual Migration Results:
Decision to Move, 1976-1977,
Ordinary Least Squares 123
Table A.3 Results of Logit Analysis with
Alternative Dummies, 1968-1977 .... 124
Table B.l Nominal Wage Model: Destination
Choice, 1968-1977 129
Table B.2 Real Wage Model: Destination
Choice, 1968-1977 131
viii


Table 3.10 Results of Logit Analysis: Decision to Move, 1968-1972
Ohio
New York
South Carolina
Pennsylvania
Variable
Coefficient
t-Ratio
Coefficient
t-Ratio
Coefficient
t-Ratio
Coefficient
t-Ratio
Famsz
-.813
- .38
-.316
-1.66
.207
1.35
. 164
.87
Sex
-
-
-
-
1.614
1.32
1.087
.89
Homeown
.845
.71
1.118
1.36
-.541
- .69
-.107
- .15
Lres
-
-
-
-
-
-
-
-
Lempl
.570
1.55
-.743
-2.58
-.540
-2.23
.228
.94
Emself
2.885
1.30
-1.940
-1.57
-2.313
-1.41
-
-
Avage
-.062
-1.03
.042
1.07
.074
1.59
-.010
- .23
Ernst
-
-
-
-
-
-
-
-
Mar st
.350
.22
.859
.58
-3.76
-2.40
-1.798
-1.17
Race
-
-
.249
. 19
-1.239
-1.33
.405
.48
Aveduc
-.193
- .59
.388
1.63
.390
1.62
.647
2.35
Prevmig
.204
.26
1.094
1.34
.960
1.18
1.338
1.82
Ernst w
.624
.77
1.617
2.14
-.876
- .97
.711
.96
Famy
.114
2.29
-.045
- .67
.160
2.10
-.212
-1.86
Constant
-4.770
-2.26
-4.949
-2.36
-4.868
-2.88
-5.266
-2.71


64
are married couples with both spouses being employed in highly mobile,
perhaps professional, occupations. In any case, employed spouses are
not acting as deterrents to migration.
Family income is significant at the 94 percent confidence level.
It carries its expected positive sign. The remaining variables are
insignificant. Sex, self-employment, marital status, and average
education have the signs which were hypothesized earlier. Homeownership
has a positive sign, indicating that migrants are more likely to own a
home than nonmigrants. They are also more likely to have lived in their
homes for longer periods of time. Race, although insignificant, carries
a negative sign. This lends support to the thesis that blacks are more
responsive to opportunities elsewhere. Previous migration also has the
opposite sign from that expected. Migrants from New York are less
likely to have moved before in their lifetimes. This is consistent with
the sign on the length-of-residence variables. Perhaps many of the New
York migrants are middle-aged or older people who have lived in New York
most of their lives and are now moving in anticipation of retirement.
The overall equation is significant at better than the .005 level.
This is true whether the model is compared to one where all coefficients
are zero or compared to a model where the right-hand side is equal to
zero. Eighty-three percent of the individuals in the sample are pre
dicted correctly by the model, while 21 percent of the migrants are
predicted correctly.
Logit Results: South Carolina
The decision-to-move results for South Carolina appear in Table
3.7. Significant variables are family size, average age, marital


BIBLIOGRAPHY
Albright, L., Lerman, S.R., and Manski, C.F., "An Introduction to the
Multinominal Probit Model," Cambridge Systematics Inc. Technical
Report, Cambridge Systematics Inc., August 1977.
American Chamber of Commerce Researchers Association, Cost of Living
Indicators: Inter-City Index Report, American Chamber of Com
merce Researchers Association, March 1968.
Borts, G.H. and Stein, J.L., Economic Growth in a Free Market, Columbia
University Press, 1964.
Burns, J.F. and James, M.K., "Migration Into Florida: 1940-1973,"
Work Paper No. 4, Urban and Regional Development Center,
University of Florida, October 1973.
Chow, G.C., "Tests of Equality Between Sets of Coefficients in Two
Linear Regressions," Econometrica, Vol. 28 (1960), pp. 591-605.
Dalton, J.A. and Ford, E.J., Jr., "Concentration and Labor Earnings
in Manufacturing and Utilities," Industrial and Labor Relations
Review, Vol. 31 (October 1977), pp. 45-60.
DaVanzo, J., Why Families Move, R & D Monograph 48, U.S. Department
of Labor, Employment and Training Administration, U.S. Government
Printing Office, 1977.
Dunlevy, J.A. and Gemery, H.A., "The Role of Migrant Stock and Lagged
Migration in the Settlement Patterns of Nineteenth Century In
migrants," The Review of Economics and Statistics, Vol. 59 (May
1977), pp. 137-44.
Eckstein, 0. and Girla, I.A., "Long-Term Properties of the Price-Wage
Mechanisms in the United States, 1891 to 1977," The Review of
Economics and Statistics, Vol. 60 (1978), pp. 323-33.
Eldridge, H.T., "Primary, Secondary, and Return Migration in the
United States, 1955-60," Demography, Vol. 2 (1965), pp. 444-55.
Engler, S.D., "Labor Force Migration to Florida: 1970 to 1976,"
The Florida Outlook, Vol. 2 (September 1978), pp. 25-30.
Falaris, E.M., "Migration: A Study of Choice Among Discrete
Alternatives," Seminar paper presented at the University of
Florida, October 1978.
132


75
Table 3.11 Comparison of Forecast to Actual Migration, 1972-1976
State of
Origin
Number of
Migrants-Forecast
Number of
Migrants-Actual
Difference
(Forecast-Actual)
Ohio
25
20
5
New York
13
17
-4
South Carolina
27
15
12
Pennsylvania
10
15
-5
Total
75
67
8


90
either New York, Ohio, Pennsylvania, or South Carolina. Each member of
the sample chose one of the eight alternatives put forth in the previous
section.
As in other studies, distance is expected to be negatively related
to destination choice. That is, all other things constant, people are
expected to move to locations which are closer to their initial residence.
Distance is believed to act as a proxy for the cost of moving and the
cost of obtaining information about alternative locations. As an example,
a New Yorker is likely to know more about what it would be like to live
in New Jersey than he would know about the situation in Texas. In fact,
distance turns up negative and significant at the 99 percent confidence
level.
It is expected that individuals are attracted to more temperate
climates. Warmer temperatures, in addition, may represent other amen
ities such as recreational activities and access to the coast. Colder
temperatures at origin locations are expected to act as a further
stimulus to migration. It is hypothesized here that climate is a more
important variable in deciding on a destination to persons originating
in colder climates. Thus, the greater the temperature difference between
origin location and a given destination, the greater is the expected
probability of choosing that destination. This is, in fact, what occurs.
The difference in average temperature variable turns out positive and
significant at the .005 level of significance.
As hypothesized in regional growth models, migration is believed to
be a response to differences in income opportunities. Thus, holding
other variables constant, individuals are believed to choose to move to
that location which promises the maximum possible lifetime gain in


5
concerning the destination that is available at the origin. That is,
previous migrants channel information back to potential future migrants.
And, assuming that it is of an encouraging nature, more information
should lead to increased migration.
The interpretation of the migrant stock is subject to some con
troversy. Laber (1972) argued that since the migrant stock is itself a
function of those factors influencing previous migration, it may be
acting as a proxy for lagged explanatory variables. If this is the
case, then a more appropriate specification may be a partial adjustment
model where lagged migration is included as an explanatory variable.
Dunlevy and Gemery (1977) estimate alternative specifications of the
model and conclude that it would be correct to include both migrant
stock and lagged migration variables. They argue that including one
variable while excluding the other results in the included variable
capturing parts of both effects. Thus, it is concluded that both the
information-creating effects of the migrant stock and a lagged migration
adjustment process are operating simultaneously.
A typical model of net migration is developed by Lowry (1966).
Lowry examines a cross section of Standard Metropolitan Statistical
Areas over the period 1950-60. His results are summarized in Table 1.2.
Significant variables in this model are the growth of the resident
working-age population (inversely related to net migration), employment
growth (positively related to net migration), and growth in armed
services personnel (positively related to net migration). Employment
growth is the best variable, the idea simply being that higher growth
reflects greater demand for labor, which translates into greater em
ployment opportunities for potential in-migrants (making them more


113
thesis that the unemployed are more responsive to migration opportunities.
This, it is recalled, was one of the major conclusions of the DaVanzo
(1977) study. Migrants from New York are also less likely to have held
their most recent job for a long period of time. This variable acts as
an indicator of attachment to present environment and job stability.
Migrants from New York are generally older than nonmigrants. In most
studies, migration rates are found to be lower in older age groups,
except at retirement. The continual presence of positive coefficients
in Florida-sending states equations, however, has led to the development
of an alternative hypothesis. It is believed that many older migrants
are moving in anticipation of retirement at a later date. The positive
and significant sign on family income complements this theory well.
Those with higher incomes are certainly likely to settle at their planned
retirement location at an early age. This is especially true if moving
involves a cut in pay or acceptance of a part-time job. Finally, household
heads from New York with employed spouses are more likely to leave the
state than those whose spouses do not work. Earlier it was theorized
that employed wives deter migration of families. This is because any
new location is less likely to present optimal employment and income
opportunities to both spouses than it is for just one of them. The
positive and significant coefficient on this variable, however, in
dicates that increasing female labor force participation could lead to
more migration from New York. Many spouses who are employed prior to
migration may no longer be employed after the move. Mincer (1978) found
evidence that one family member frequently dominates the migration
decision. Thus, spouses become tied to this decision regardless of the
personal consequences.


70
Table 3.9
- Ordinary Least Squares
Results:
Decision to
Move, 1968-
1977
Variable
Ohio
Coefficient
t-Pvatio
New York
Coefficient t-Ratio
Pennsylvania
Coefficient t-Ratio
Famsz
.015
1.02
-.020
-1.51
.023
1.60
Sex
.494
3.32
.073
.50
.323
1.74
Homeown
-.047
- .70
.055
.95
.005
.09
Lres
.002
.05
.060
1.32
.003
.15
Lempl
.005
.21
-.047
-2.63
.024
1.18
Emself
-. 105
- .77
-.061
- .59
.161
1.06
Avage
.006
1.46
.009
2.75
.002
.43
Ernst
.037
.19
-.348
-1.80
.009
.04
Marst
-.626
-4.05
-.198
-1.35
-.402
-2.06
Race
.069
.86
-.136
-1.47
-.012
- .17
Aveduc
-.003
- .14
.020
1.16
.064
2.78
Prevmig
.035
.63
-.041
- .64
.015
.20
Ernst w
.044
.74
. 128
2.21
.106
1.84
Famy
.009
2.48
.009
1.71
-.016
-2.87
Constant
-.210
- .94
.366
1.69
-.210
- .81
R2 =
.13
R2
= .17
R2
= .09


38
migration coefficients are used to forecast later period migrant flows,
coefficient instability implies forecasting problems. Employing energy
shortage coefficients, for example, to project migration for a period in
which we expect no energy problems would be incorrect.
Finally, historical estimates of aggregate labor force migration
can be obtained from the sample by utilizing a methodology similar to
that proposed for forecasting. For the decision-to-move analysis, the
average probability of migration for each state during any given period
can be applied to a measure of the aggregate labor force at the beginning
of that period. The resulting estimates of out-migration can then be
distributed out among destinations by multiplying them times the prob
abilities obtained from the destination-choice equations for the same
period. Thus, estimates of aggregate migration between all possible
origins and destinations are derived from the results of the entire
analysis. In the next section of this chapter, the sources of data will
briefly be described.
Sources of Data
The primary source of data for this study is the University of
Michigan's Panel Study of Income Dynamics. In this survey, 5,862
families have been interviewed each year since 1968. At present, 10
years of data are available on tape. Each year approximately 450
variables are available for each family. Categories under which the
variables can be grouped are family composition information, education,
transportation, housing, employment of head, housework, work for money
by wife, food and clothing, income, intelligence", feelings, and time
use.


99
Table 4.7 Comparison of Forecast to Actual Destination
Choice: 1972-1976
Origin-
Destination
Number of
Migrants-Forecast
Number of
Migrants-Actual
Difference
(Forecast-Actual)
New York-Florida
3
4
-1
New York-California
1
3
-2
New York-New Jersey
4
2
2
New York-Illinois
2
0
2
New York-Virginia
2
3
-1
New York-Michigan
1
0
1
New York-Texas
1
0
1
Pennsylvania-Florida
2
5
-3
Pennsylvania-
California
0
1
-1
Pennsylvania-New
Jersey
2
1
1
Pennsylvania-New
York
2
1
1
Pennsylvania-
Ill inois
1
0
1
Pennsylvania-
Virginia
1
2
-1
Pennsylvania-
Michigan
1
0
1
Pennsylvania-Texas
1
0
1
Ohio-Florida
3
2
1
Ohio-California
1
3
-2
Ohio-New Jersey
2
1
1
Ohio-New York
3
1
2
Ohio-Illinois
2
0
2
Ohio-Virginia
1
1
0
Ohio-Michigan
1
2
-1
Ohio-Texas
1
3
-2
South Carolina-
Florida
3
4
-1
South Carolina-
California
1
1
0
South Carolina-
New Jersey
1
1
0
South Carolina-
New York
2
3
-1
South Carolina-
Illinois
1
0
1
South Carolina-
.
_
Virginia
1
2
-1
South Carolina-
Michigan
1
0
1
South Caro'lina-
Texas
1
0
1


Table 3.12 (continued)
Summary Statistics Ohio
Log of Likelihood Function -58.6
Likelihood Statistic (at 0) 218
Level of Significance (at 0) .005
Likelihood Statistic (at mean) 17.4
Level of Significance (at mean) .10
Percentage Correctly Predicted 90%
Percentage of Migrants Correctly
Predicted 5%
New York
South Carolina
Pennsylvania
-53.7
-47.0
-44.2
89
85
182
.005
.005
.005
10.2
11.0
17.4
.10
.10
.10
92%
92%
93%
0%
0%
7%


LIST OF TABLES
PAGE
Table 1.1 Greenwood Gross Migration Regression
Results 4
Table 1.2 Lowry Net Migration Regression Results 6
Table 1.3 Graves Net Migration Regression Results 10
Table 1.4 DaVanzo Destination Choice Results 15
Table 3.1 Summary of Panel Study Migration, 1968-1977 42
Table 3.2 Florida In-migration by State of Origin,
1968-1977 44
Table 3.3 Variable Definitions for Decision to
Migrate Analysis 46
Table 3.4 Comparison of Migrant and Nonmigrant
Characteristics: New York, Ohio,
Pennsylvania, and South Carolina 48
Table 3.5 Results of Logit Analysis: Decision to
Move, Ohio, 1968-1977 60
Table 3.6 Results of Logit Analysis: Decision to
Move, New York, 1968-1977 63
Table 3.7 Results of Logit Analysis: Decision to
Move, South Carolina, 1968-1977 65
Table 3.8 Results of Logit Analysis: Decision to
Move, Pennsylvania, 1968-1977 67
Table 3.9 Ordinary Least Squares Results: Decision
to Move, 1968-1977 70
Table 3.10 Results of Logit Analysis: Decision to
Move, 1968-1972 73
Table 3.11 Comparison of Forecast to Actual Migration,
1972-1976 75
Table 3.12 Results of Logit Analysis: Decision to
Move, 1972-1976 77
vi


CHAPTER 4
EMPIRICAL RESULTS: DESTINATION CHOICE
Destination Choice: 1963-1977
A summary of the destinations chosen by migrants in the four sample
states appears in Table 4.1. Overall, Florida is the leading destination,
with 20 percent of all out-migrants choosing this state. Florida is the
top location choice for New York, Pennsylvania, and Ohio migrants.
South Carolina migrants showed a preference for moving north, with New
York being the leading destination. Florida, however, followed closely
behind. With the exception of those choosing Florida, California, and
Texas, many migrants chose to move to nearby states. As an example, 14
percent of the Ohio migrants moved to Michigan and 17 percent of New
York's migrants chose to live in New Jersey. Because of technical and
cost limitations, not all destinations which were chosen could be in
cluded in the analysis. The top eight destination states were selected.
These states accounted for 70 percent of the migrants who left the four-
state area of New York, Pennsylvania, Ohio, and South Carolina.
The variables used in the destination-choice equations are presented
in Table 4.2. In the next section the rationale for inclusion of each
variable will be discussed and the results are presented.
Logit Results: Destination Choice
The results of the first destination-choice model estimated appear
in Table 4.3. The sample consists of 88 migrants originating from
86


89
Table 4.3 Logit Results: Destination Choice,
1968-1977, Model 1
Variable
Coefficient
Asymptotic t-Ratio
Distance
-.703
-2.42
Tempdiff
.229
3.86
Incdiff
.128
1.35
Hdiff
.020
.44
Taxdiff
.010
.22
Unemdiff
.047
. 17
Summary Statistics:
Log of Likelihood Function = -162.5
Likelihood Statistic (at 0) = 23.4
Level of Significance (at 0) = .005
Likelihood Statistic (at mean) = 14.35
Level of Significance (at mean) = .05


Abstract of Dissertation Presented to the Graduate Council
of the University of Florida in Partial Fulfillment of the
Requirements for the Degree of Doctor of Philosophy
AN ECONOMETRIC MODEL OF INTERSTATE LABOR FORCE MIGRATION
By
Sheldon Donald Engler
December 1979
Chairman: Jerome W. Milliman
Major Department: Economics
The purpose of this study is to analyze and forecast migration
decisions of labor force members. Location choice is viewed within
a theoretical framework which assumes that the individual is a utility
maximizer. Location bundles which include some goods available at all
locations and other goods which are location-specific can be varied by
moving. Each person reaches his (or her) optimum by first considering
the opportunities which are available at all potential locations.
Differences in individual characteristics and circumstances are
hypothesized to cause people to evaluate these opportunities in
alternative ways.
A two-step methodology is adopted for the empirical section of
this study. First, the "whether to move" decision is modeled for all
sample members. Then, the choice of "where to move" is analyzed for
the proportion of the sample who migrated. The first decision is
considered for four separate states. States were chosen based upon
historical rates of out-migration, particularly to Florida. The
x


8
Richardson (1973) criticizes neoclassical models for assuming that
individuals are simply maximizing their income in the migration decision.
He proposes a model (which is as yet untested) in which net migration is
a function of wage differentials, agglomeration economies, and locational
preferences. Agglomeration economies are divided into two types.
Household agglomeration economies refer to the benefits and costs of
life in large cities for households. They include benefits accruing
from larger labor markets, availability of leisure and cultural facil
ities, the quality of public service, and environmental amenities.
Business agglomeration economies refer to advantages which urban areas
offer businesses. They attract firms which lead to jobs which in turn
attract migrants. The level of agglomeration economies is thus hypoth
esized to influence the rate of in-migration to a region. Locational
preferences are meant to explain why individuals may remain in low-
income regions despite the existence of improved income opportunities
elsewhere. Thus, they measure the retentive power of a region and are
related to the rate of out-migration. Examples of factors which in
fluence locational preferences are community ties, sociocultural traditions,
and length of settlement.
In a recent study Graves (1979) rejects the theory that migration
is a response to interregional income differences. Rather, he advances
the view that any income differences which may exist between regions are
compensated for by differences in the amounts of other amenity-oriented
goods which are available. For example, an area which offers an income
advantage over other places is believed to offer a more objectionable
climate, allowing for less involvement in leisure or recreational
activities. Moving to a location with a better climate, under this


33
chooses California. Micro type variables, such as individual potential
wages, which take into account differences among people were tested.
Their failure to yield meaningful results led to the adoption of a more
macro approach.
Equation (6) was estimated for the period 1968-77. For a forecasting
test, it was estimated for the periods 1968-72 and 1972-76. The sample
was defined as it was for the decision-to-move analysis, except that
only migrants were considered. Migrants from New York, Ohio, Pennsylvania,
and South Carolina were then grouped together. The top eight destination
states were selected as potential alternatives. These included Florida,
California, New York, New Jersey, Michigan, Illinois, Texas, and Virginia.
Migrants choosing other states were excluded from the analysis.
The model is estimated using a multinomial "conditional" logit
program. Estimating equation (6) using this technique allows us to
derive predicted probability values. These probabilities can be rep
resented as follows:
P.
i
P.0
l tp
J P.0
i tp
£ e
i=l
(10)
where i indexes alternatives and J is the total number of choices
available to the individual. The variables in the vector P. will be
x
defined as differences between the value at location i and the value at
the individual's origin of residence. Probabilities will be calculated
for migration between each possible origin and destination combination.
Since starting from any origin these values will-not vary across "indi
viduals, there will be a total of 32 probabilities calculated (8 alter
natives x 4 origins).


ACKNOWLEDGEMENTS
The author wishes to thank Professor Jerome Milliman for his guid
ance and support throughout the course of study. Special thanks are
given to Professor Henry Fishkind for his intellectual contributions
and his continued friendship. Acknowledgements are also extended to
Professors Stanley Smith, David Denslow, John Henretta, Angela O'Rand,
and William Tyler. Finally, the author would like to thank Alee
Williams and Doreen Willmeroth, whose typing and editing skills were
invaluable to the production of the finished paper.
iii


Table A.3 (continued)
Summary Statistics Ohio
Log of Likelihood Function -83.8
Likelihood Statistic (at 0) 137.4
Level of Significance (at 0) .005
Likelihood Statistic (at mean) 31.8
Level of Significance (at mean) .005
Percentage Correctly Predicted 85%
Percentage of Migrants Correctly
Predicted 19%
New York
South Carolina
Pennsylvania
-63.0
-71.1
-73.2
144.4
175.6
54.0
.005
.005
.005
38.0
41.0
16.2
.005
.005
.10
84%
89%
86%
7%
25%
4%
i
125


131
Table B.2 Real Wage Model: Destination Choice, 1968-1977
Variable
Distance
Tempdiff
RWdiff
Coefficient
-.265
.110
.163
Asymptotic t-Ratio
1.33
3.47
1.64
Summary Statistics:
Log of Likelihood Function = 168.4
Likelihood Statistic (at 0) = 11.6
Level of Significance (at 0) = .01
Likelihood Statistic (at mean) =2.6
Level of Significance (at mean) = (not significant at any
acceptable level)


68
deters migration. Sex and previous migration are both positively related
to migration, as expected. Homeownership, length of residence, length
of employment, self-employment, and spouse's employment status all show
the wrong sign. Average age has a positive sign, giving further support
to the preretirement thesis, and race is positively related to migration,
indicating that white families leave Pennsylvania more often than do
black families.
The Pennsylvania equation provides a poorer fit than the decision
equations for the other three states. Although significant compared to
the model where all coefficients are zero, the more stringent test
indicates a weaker model. Setting the right-hand side of the equation
equal to the mean value of the dependent variable and comparing the
likelihood value which results with the unconstrained likelihood value
results in a significance level of .10 for the unconstrained model.
Although 87 percent of the sample was predicted correctly by the model,
only 7 percent of the migrants were predicted correctly. Thus, this
model has not been very successful in distinguishing between migrants
and nonmigrants.
Test for Separation of States
In Chapter 2, a theoretical argument was made for estimating
separate equations for each state of origin. In the preceding sections
of this chapter evidence was presented which indicated that there may be
differences between states in the determinants of the migration decision.
In this section, a formal test for equality of coefficients between
states is carried out.


15
Table 1.4 DaVanzo Destination Choice Results
Variable
Coefficient
Asymptotic t-Ratio
fam
PV. .
11
.00548
1.53
fam
PV.. Here Before .
il 1
.0139
1.46
Unemployment Rate
.00909
.05
Ln Distance ..
il
-.322
-.81
Variable Definitions:
f am
PV,, -present value of the difference between what the
family could earn at destination j and what it
could earn if it stayed in its 1971 location, i
Here Before -dummy that indicates whether the family
^ resided in area j recently (between 1968
and 1970)
Unemployment Rate -unemployment rate in 1971 at destination j
Ln Distance .. -natural logarithm of the distance between
origin i and destination j


AN ECONOMETRIC MODEL OF INTERSTATE
LABOR FORCE MIGRATION
BY
SHELDON DONALD ENGLER
A DISSERTATION PRESENTED TO THE GRADUATE COUNCIL
OF THE UNIVERSITY OF FLORIDA IN
PARTIAL FULFILLMENT OF THE REQUIREMENTS
FOR THE DEGREE OF DOCTOR OF PHILOSOPHY
UNIVERSITY OF FLORIDA
1979


129
Table B.l Nominal Wage Model: Destination Choice, 1968-1977
Variable
Distance
Tempdiff
Wdiff
Coefficient
-.264
.111
.157
Asymptotic t-Ratio
-1.33
3.47
1.67
Summary Statistics:
Log of Likelihood Function = 168.4
Likelihood Statistic (at 0) = 11.6
Level of Significance (at 0) = .01
Likelihood Statistic (at mean) = 2.6
Level of Significance (at mean) = (not significant at any
acceptable level)