Citation
Understanding education

Material Information

Title:
Understanding education : three essays analyzing unintended outcomes of school policies
Creator:
Luallen, Jeremy Clayton ( Dissertant )
Kenny, Lawrence W. ( Thesis advisor )
Place of Publication:
Gainesville, Fla.
Publisher:
University of Florida
Publication Date:
Copyright Date:
2005
Language:
English
Physical Description:
xi, 107 p.

Subjects

Subjects / Keywords:
Crime in schools ( jstor )
Criminals ( jstor )
Juvenile delinquency ( jstor )
Juveniles ( jstor )
Property crimes ( jstor )
Schools ( jstor )
Teachers ( jstor )
Urban crime ( jstor )
Violent crimes ( jstor )
ZIP codes ( jstor )
Economics thesis, Ph. D. ( local )
Dissertations, Academic -- UF -- Economics ( local )

Notes

Abstract:
The goal of my study is to examine specific school policies to determine if unintentional consequences result from these policies. Specifically, I focus on two main issues as they relate to student and teachers outcomes. I begin by looking at the effect of incapacitating juveniles in school as a force influencing juvenile crime. I exploit teacher strikes as a measure of unexpected student absence from school to measure the effect of school in preventing juvenile crime. My data set consists of information on every juvenile arrest made in Washington State over a 22-year period. I show that previous estimates of the effect of school incapacitation are systematically underestimated, that criminal activity increases as students continue to remain out of school. I also show that these increases in crime reflect an increase in overall crime, not a displacement. Lastly, I show that repeat juvenile offenders are more likely to have committed their first crime on a strike day, relative to a normal school day. Chapters 2 and 3 of the study focus on the role of teacher networks in influencing teacher mobility. Specifically, my study develops a model of teacher networks that describes how teachers assemble networks through professional development activities (PDAs) and how these networks provide an effective sorting mechanism for public school teachers. I empirically test the existence of teacher networks with 2 distinct datasets. The dataset in Chapter 2 is comprised of various reports covering all 67 Florida school districts. Besides examining how professional development affects teacher movement, I am able to exploit the macro nature of the data to compare district characteristics (such as differences in compensation levels and school district density) to examine how these factors also influence teacher mobility. The dataset in Chapter 3 uses survey data from the ''Schools and Staffing Survey'' and includes over 17,000 teachers. The high-powered nature of this dataset allows me to identify specific details, such as teacher salary incentives, individual network strength and union membership. Ultimately I conclude that teacher networks are an integral part of a teacher's transfer decision and have a sizable impact on intra-district teacher mobility.
Subject:
binomial, crime, development, economics, education, incapacitation, inservice, intradistrict, juvenile, Luallen, mobility, negative, networks, professional, property, schools, strikes, transfer, violent, Washington
General Note:
Title from title page of source document.
General Note:
Document formatted into pages; contains 118 pages.
General Note:
Includes vita.
Thesis:
Thesis (Ph. D.)--University of Florida, 2005.
Bibliography:
Includes bibliographical references.
General Note:
Text (Electronic thesis) in PDF format.

Record Information

Source Institution:
University of Florida
Holding Location:
University of Florida
Rights Management:
Copyright Luallen, Jeremy Clayton. 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.
Embargo Date:
7/30/2007
Resource Identifier:
71638324 ( OCLC )

Downloads

This item has the following downloads:

PDF ( .pdf )

luallen_j_Page_022.txt

luallen_j_Page_036.txt

luallen_j_Page_093.txt

luallen_j_Page_011.txt

luallen_j_Page_079.txt

luallen_j_Page_020.txt

luallen_j_Page_039.txt

luallen_j_Page_042.txt

luallen_j_Page_109.txt

luallen_j_Page_073.txt

luallen_j_Page_057.txt

luallen_j_Page_061.txt

luallen_j_Page_055.txt

luallen_j_Page_108.txt

luallen_j_Page_080.txt

luallen_j_Page_029.txt

luallen_j_Page_060.txt

luallen_j_Page_030.txt

luallen_j_Page_026.txt

luallen_j_Page_050.txt

luallen_j_Page_027.txt

luallen_j_Page_024.txt

luallen_j_Page_077.txt

luallen_j_Page_099.txt

luallen_j_Page_018.txt

luallen_j_Page_004.txt

luallen_j_Page_006.txt

luallen_j_Page_038.txt

luallen_j_Page_014.txt

luallen_j_Page_037.txt

luallen_j_Page_046.txt

luallen_j_Page_045.txt

luallen_j_Page_083.txt

luallen_j_Page_043.txt

luallen_j_Page_101.txt

luallen_j_Page_054.txt

luallen_j_Page_116.txt

luallen_j_Page_034.txt

luallen_j_Page_074.txt

luallen_j_Page_007.txt

luallen_j_Page_058.txt

luallen_j_Page_095.txt

luallen_j_Page_097.txt

luallen_j_Page_063.txt

luallen_j_Page_056.txt

luallen_j_Page_010.txt

luallen_j_Page_044.txt

luallen_j_Page_113.txt

luallen_j_Page_100.txt

luallen_j_Page_062.txt

luallen_j_Page_085.txt

luallen_j_Page_086.txt

luallen_j_Page_012.txt

luallen_j_Page_084.txt

luallen_j_Page_078.txt

luallen_j_Page_052.txt

luallen_j_Page_017.txt

luallen_j_Page_105.txt

luallen_j_Page_098.txt

luallen_j_Page_092.txt

luallen_j_Page_069.txt

luallen_j_Page_005.txt

luallen_j_Page_111.txt

luallen_j_Page_068.txt

luallen_j_Page_072.txt

luallen_j_Page_048.txt

luallen_j_Page_019.txt

luallen_j_Page_035.txt

luallen_j_Page_065.txt

luallen_j_Page_066.txt

luallen_j_Page_040.txt

luallen_j_Page_021.txt

luallen_j_Page_106.txt

luallen_j_Page_009.txt

luallen_j_Page_070.txt

luallen_j_Page_102.txt

luallen_j_Page_032.txt

luallen_j_Page_008.txt

luallen_j_Page_001.txt

luallen_j_Page_082.txt

luallen_j_Page_096.txt

luallen_j_Page_041.txt

luallen_j_Page_015.txt

luallen_j_Page_107.txt

luallen_j_Page_013.txt

luallen_j_Page_104.txt

luallen_j_Page_094.txt

luallen_j_Page_081.txt

luallen_j_Page_059.txt

luallen_j_Page_118.txt

luallen_j_Page_016.txt

luallen_j_Page_090.txt

luallen_j_Page_003.txt

luallen_j_Page_051.txt

luallen_j_Page_023.txt

luallen_j_Page_025.txt

luallen_j_Page_049.txt

luallen_j_pdf.txt

luallen_j_Page_053.txt

luallen_j_Page_071.txt

luallen_j_Page_110.txt

luallen_j_Page_089.txt

luallen_j_Page_087.txt

luallen_j_Page_033.txt

luallen_j_Page_076.txt

luallen_j_Page_002.txt

luallen_j_Page_117.txt

luallen_j_Page_088.txt

luallen_j_Page_031.txt

luallen_j_Page_028.txt

luallen_j_Page_075.txt

luallen_j_Page_047.txt

luallen_j_Page_115.txt

luallen_j_Page_064.txt

luallen_j_Page_091.txt

luallen_j_Page_112.txt

luallen_j_Page_103.txt

luallen_j_Page_067.txt

luallen_j_Page_114.txt


Full Text












UNDERSTANDING EDUCATION: THREE ESSAYS ANALYZING UNINTENDED
OUTCOMES OF SCHOOL POLICIES
















By

JEREMY CLAYTON LUALLEN


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

UNIVERSITY OF FLORIDA


2005

































Copyright 2005

by

Jeremy Clayton Luallen















ACKNOWLEDGMENTS

Throughout the course of my graduate and dissertation work here at the University

of Florida, there have been many friends, relatives and colleagues whom I have relied on

for much professional and emotional support. It is for this reason that I wish to

acknowledge the contributions of all these important individuals, who were key to the

success of this dissertation. In addition to the academic and financial support these

individuals have provided, they have also made outstanding sacrifices to better my

growth, not only as a professional, but also personally.

First and foremost I would like to recognize my parents Dean and Nancy for their

amazing support of my efforts. Their guidance through the years has bettered me in ways

that I cannot even begin to measure. I would also like to thank my fiancee Jaclyn for

never-ending support of me and my work over the past several years. I would like to

specially acknowledge Dr. Lawrence Kenny for his wonderful friendship and guidance as

a mentor over the years. I would also like to thank my sister Jessica who has inspired me

in many ways. There are also many others I would like to acknowledge for their

assistance in making this dissertation a success. I would like to thank Steven Slutsky,

David Figlio, David Denslow, David Hedge, Randall Reback, Richard Romano, Sarah

Hamersma, David Brasington, David Sappington, Bill Bomberger, Chunrong Ai, Scott

Hankins, Mark Hoekstra and Jim Dewey for their many comments, incisive criticism, and

overall enthusiasm for my research.









I would like to thank Anne Stahl at the National Juvenile Court Data Archive for

her guidance and patience in helping me to acquire data for this project. Thanks also go

to Kathy Budge, Dennis Small, and many others at the Superintendent of Public

Instruction Office in Washington State for providing supplementary data that has been

crucial to the success of my research. I would also like to thank Martha Haynes and the

Staff of the Education Information and Accountability Services division of the Florida

Department of Education, who supplied various reports and data utilized in this

dissertation. I would also like to thank Kerry Gruber and the entire staff at the National

Center of Education Statistics that supplies and distributes the Schools and Staffing

Survey. I am also especially grateful to all of the teachers and administrators throughout

the State of Florida who shared their time, their thoughts and their experiences with me.
















TABLE OF CONTENTS

page

A C K N O W L E D G M E N T S ......... .................................................................................... iii

LIST OF TABLES ............................... .............. ................. vii

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

A B STR A C T ................................................. ..................................... .. x

CHAPTER

1 SCHOOL'S OUT...FOREVER: A STUDY OF JUVENILE CRIME, AT-
RISK YOUTHS AND TEACHER STRIKES.............. .... ................... 1

1.1 Introdu action ................... ......... .......... ................. ............... 1
1.2 The Identification Strategy ........................................ ....... ............... 3
1.2.1 Zip-Code M atching................. ......... ...... .................5
1.2.2 The Basic M odel ........................ .. ........................ .8
1 .3 D a ta ................................................... ................ ................ 9
1.3.1 M ethodology .. ..................... .. .......... ...... .............. .... ............ ..10
1.3.2 Defining the Param eters............... ......................... .................11
1.4 R egression A nalysis........................................................ ............... 13
1.4.1 Community Differences...................................... .................. 16
1.4.2 Differences in Offense Types ................................................18
1.4.3 Types of Offenders .............. .... ............................ ........... 19

2 THE ROLE OF TEACHER NETWORKING IN TEACHER TRANSFER
DECISIONS AND TEACHER MOBILITY ....................................................... 42

2 .1 Introdu action ................................................................................ 42
2.2 Discussion of Networking...................... .... .......................... 43
2.2.1 H ow Teachers N etw ork ........................................ .....................43
2.2.2 How Teachers Use Their Networks.............................................45
2.3 The Basic Model .......................................... ...................46
2.4 D ata and Em pirical Study ........................................ ...................... 47
2.4.1 D efining the Param eters............................................................... 48
2.4.2 E m pirical T testing ....................................... ......... ............... 52
2.5 R obustness C hecks.......................................................... ............... 55
2.5.1 R e-exam ining the D ata ...................................... ............... 55









2.5.2 U observed C orrelation ........... ...................... ............ ........57
2 .6 C o n c lu sio n ........................................................................................... 5 8

3 WITH A LITTLE HELP FROM MY FRIENDS: EVIDENCE OF TEACHER
NETWORKS USING MICRO DATA......................................................65

3.1 Introduction .................. .................................................. 65
3.2 The Identification Strategy .............. ...................... ............... ... 67
3.2.1 What Are Networks And How Are They Useful...........................67
3.2.2 How Do Teachers Build Networks? ...........................................69
3.2.3 M easuring Teacher Networks ...................................................70
3.2.4 The B asic M odel ........................................ ........................ 72
3 .3 D ata ................................................................................. 7 6
3 .4 E m pirical E v iden ce ................................................... ..............................77
3.5 R obustness C hecks.......................................................... ............... 80
3 .6 C o n c lu sio n ........................................................................................... 8 3

APPENDIX

A SUMMARY STATISTICS AND SUPPLEMENTARY REGRESSION
A N A L Y S IS ................................................... ................ 9 6

B COMPLETE RESULTS OF SPECIFIED REGRESSIONS ..............................98

L IST O F R E F E R E N C E S ...................................................................... ..................... 103

BIOGRAPHICAL SKETCH .............................................................. ............... 107
















LIST OF TABLES


Table pge

1-1 Negative Binomial Regression onto Total Crime with Full Data Set....................30

1-2 Urban Subsample Negative Binomial Regression onto Total Crime ..................31

1-3 Suburban Subsample Negative Binomial Regression onto Total Crime ..............32

1-4 Rural Subsample Negative Binomial Regression onto Total Crime ...................33

1-5 Effects of Strike Days by Crime Type and by Community Type..........................34

1-6 Effects of Strike Days by Crime Type and by Offender Type ...........................35

1-7 Temporal Displacement of Crime by Crime Type and by Offender Type............36

1-8 Temporal Displacement of Crime through Lagged Strike Variable....................37

1-9 Changes in Total Juvenile Crime, by Days Elapsed since the Start of a Strike.....38

1-10 Effects of Strike Days with Specific Years Dropped from the Sample.................40

1-11 Average Effect of Randomly Generated Strike Days on Total Crime...................41

1-12 Effects of Strike Days by Offender Type with the Last 4 Years of the Sample
D ro p p e d ................................. ....................................................... ............... 4 1

2-1 Ordinary Least Squares Regression onto teacher transfers using full sample
with and without robust standard errors...................................... ............... 60

2-2 Percent change teacher transfers per school for a one standard deviation
increase in each significant variable with full sample ....................................... 61

2-3 Ordinary Least Squares Regression onto teacher transfers using sample with
biasing outlier om itted ....................... .. ...................... .... ........ ...... ............62

2-4 Outlier excluded percent change teacher transfers per school for a one standard
deviation increase in each significant variable .............. ........................ ......... 63

2-5 Ordinary Least Squares Regression of other types of teacher movements with
outlier ex clu ded ............................................................................ ............... 64









3-1 Professional Development Activities Summary Statistics................................84

3-2 Summary Statistics for Various Teacher Groups........................................85

3-3 Probit Regression of Intra-District Movement ............................................... 86

3-4 Probit Regression of Intra-District Movement ............................................... 87

3-5 Within State Probit Regression of Interdistrict Movement............................... 93

3-6 Out Of State Probit Regression of Interdistrict Movement .............................. 94

3-7 Probit Regression of Intradistrict Movement with District Fixed Effects.............95

A -1 Strike Sum m ary Statistics ......................................................... .............. 96

A-2 Regressions of Full Data Set excluding Zip Codes with less than 500 students ...97

B R ep orted on T ab le 3-4 ................................................................ .....................99

B -2 R reported on Table 3-5 ................................................ ............................. 100

B -3 R reported on T able 3-6 ........................................... ........................................ 10 1

B -4 R reported on T able 3-7 ........................................... ........................................ 102
















LIST OF FIGURES


Figure


1-1 Average Percent of Total Arrests as Function of Days between Offense and
C ap tu re ...................................................................................3 9

3-1 Comparisons of Teacher Characteristics Along the Distribution of PDA ..............88


page















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

UNDERSTANDING EDUCATION: THREE ESSAYS ANALYZING UNINTENDED
OUTCOMES OF SCHOOL POLICIES

By

Jeremy Clayton Luallen

August 2005

Chair: Lawrence Kenny
Major Department: Economics

The goal of my study is to examine specific school policies to determine if

unintentional consequences result from these policies. Specifically, I focus on two main

issues as they relate to student and teachers outcomes. I begin by looking at the effect of

incapacitating juveniles in school as a force influencing juvenile crime. I exploit teacher

strikes as a measure of unexpected student absence from school to measure the effect of

school in preventing juvenile crime. My data set consists of information on every

juvenile arrest made in Washington State over a 22-year period. I show that previous

estimates of the effect of school incapacitation are systematically underestimated, that

criminal activity increases as students continue to remain out of school. I also show that

these increases in crime reflect an increase in overall crime, not a displacement. Lastly, I

show that repeat juvenile offenders are more likely to have committed their first crime on

a strike day, relative to a normal school day.









Chapters 2 and 3 of the study focus on the role of teacher networks in influencing

teacher mobility. Specifically, my study develops a model of teacher networks that

describes how teachers assemble networks through professional development activities

(PDAs) and how these networks provide an effective sorting mechanism for public

school teachers. I empirically test the existence of teacher networks with 2 distinct

datasets. The dataset in Chapter 2 is comprised of various reports covering all 67 Florida

school districts. Besides examining how professional development affects teacher

movement, I am able to exploit the macro nature of the data to compare district

characteristics (such as differences in compensation levels and school district density) to

examine how these factors also influence teacher mobility. The dataset in Chapter 3 uses

survey data from the "Schools and Staffing Survey" and includes over 17,000 teachers.

The high-powered nature of this dataset allows me to identify specific details, such as

teacher salary incentives, individual network strength and union membership. Ultimately

I conclude that teacher networks are an integral part of a teacher's transfer decision and

have a sizable impact on intra-district teacher mobility.














CHAPTER 1
SCHOOL'S OUT...FOREVER: A STUDY OF JUVENILE CRIME, AT-RISK
YOUTHS AND TEACHER STRIKES

1.1 Introduction

Although we have taken important strides in understanding the economics of crime,

there remains a great deal which we have yet to fully understand. The focus of this

chapter is juvenile crime in particular. Because juvenile crime can be especially hard to

study, due to general limitations on access to data, economists still have a lot to explore

on this front.

Over the past 20 years, economists and social scientists have attacked the problem

of crime in four distinct ways. Specifically they have analyzed the potential effects of

deterrence, retribution, rehabilitation and incapacitation as forces for reducing crime

(Ehrlich 1981).1 Deterrence, a widely explored topic, stresses the importance of

imposing penalties as an effective means of preventing crime, because the perceived

costs to criminals of criminal activity increases (Mocan and Rees 1999, Levitt 1998,

Freeman 1996, Ehrlich and Gibbons 1977, Lochner 2003). Retribution addresses the

actual punishment of criminals. It suggests that a criminal experiences an increased cost

to his crime when a punishment is imposed on him and/or other criminals (Levitt 1998).

Rehabilitation stresses the importance of reforming criminals, through treatment and

rehabilitative programs, in preventing future crime (Cuellar, Markowitz and Libby 2003).

Finally, incapacitation deals with the notion that the physical lock-up and detaining of


SI am not arguing that these four ideas are mutually exclusive ways of preventing or reducing crime.









criminals may be an effective means of preventing crime (Freeman 1996, Lochner 2004).

This paper will closely examine the role of incapacitation in preventing juvenile crime.

Specifically I deal with the effect of incapacitating juveniles in school on juvenile

crime. Brian Jacob and Lars Lefgren (2003) were the first to address this topic in their

paper, "Are Idle Hands the Devil's Workshop? Incapacitation, Concentration and

Juvenile Crime". They examine the effect of school attendance on juvenile crime by

using teacher in-service days, days where teaching professionals are required to attend

work when children are not required to be at school, as a source of variation in student

attendance. While their use of in-service days is very clever, I utilize teacher strikes as a

source of variation of student school attendance to gain additional insights into the

relationship between juvenile crime and school incapacitation.

One primary advantage to using teacher strikes over in-service days as a source of

variation in school attendance is that strikes often occur in blocks larger than one

individual day. This allows me to observe the effect of prolonged absence from school

on juvenile crime. If juvenile criminal habits do change as absence from school

increases, then a one day absence cannot reflect the average effect of school

incapacitation.

Another advantage of utilizing teacher strikes is that they are relatively

unpredictable. Since in-service days are planned at the beginning of the regular school

year, parents have adequate time to make arrangements for their children during these

off-days. In contrast, school closures caused by teacher strikes are often reported in local

newspapers only a couple of days before they are likely to take place, and even then,

there is no guarantee that these reported strikes will actually occur. Additionally, a









parent's ability to plan for their children's activities during strike days is complicated by

the fact that details revealed in newspapers and other information sources may prove to

be inaccurate or unjustified.2 The late-breaking and incomplete nature of information

leaves parents little time to plan their children's activities during the days of a teacher

strike. Thus teacher strike days are likely to result in more unsupervised students, and

more juvenile crime.

The main drawback to this use of strike as variation is that teacher strikes are often

very sparse, and mostly occur in a only a few school districts. In order to overcome this

difficulty, I rely on data from Washington State. Washington provides an ideal

environment for this type of study because it has an extensive strike history as well as a

detailed juvenile arrest data set that dates back as far as 1980.3

1.2 The Identification Strategy

As stated earlier, the objective of this analysis is to measure the impact of school

incapacitation on juvenile crime rates by using teacher strike days as variation in student

absence from school on an ordinary school day. Arguably, teacher strikes are a source of

variation that is exogenous to variations in juvenile crime. Issues that lead teachers to

strike include pay, class size and teacher planning time, none of which are likely to

influence daily variations in juvenile crime.4


2 For instance, a school district may report that it is sure teachers will return to work when a court
injunction is issued, and subsequently the teachers may defy the injunction.

3 Unlike most states in the U.S., teacher strike activity in Washington remains very much active to this day.
The most recent strike event took place in 2003 in Marysville school district. This strike lasted nearly two
months (approximately 50 total days). A summary table describing teacher strikes in Washington can be
seen in Table A-1.

4 In fact, areas where juvenile crime may be an increasing problem are arguably less-likely to see a teacher
strike, because greater crime has been shown to be positively correlated with higher teacher pay (Grogger
1997).









In order to draw direct comparisons between crime on strike vs. nonstrike days, I

must be able to control for days in which schools are closed for other reasons. This

would include days that coincide with spring, summer and winter vacations, weekends,

teacher in-service and half-days, national holidays, etc. However since the data cover a

large span of time (22 years), and since each Washington school district determines their

own school calendar, it is impossible to know where all of these nontypical school days

occur.5 Further, a prolonged teacher strike sometimes results in a reworking of the

existing school calendar (largely unobservable) so that a school district can meet the

requirement set by the state that the school year last for 180 days.6 Since I cannot control

for these school holidays with complete accuracy, I must include only those days which I

am strongly confident students are regularly scheduled to be in school.

To minimize any possibility that school holidays could bias the results, I take extra

measures when eliminating questionable days. I begin by eliminating weekends and

national holidays. If a national holiday falls on a Saturday or Sunday, I eliminate the

preceding Friday or subsequent Monday, respectively. The first two Fridays of October

are dropped because they are both traditional days for the state mandated teacher in-

service day. The entire months of June and July were also excluded, and only the last

three school days in August are allowed in the sample. Finally I eliminate the first week

5Each school district may vary on the scheduling of breaks, half-days, inservice days, etc. To the best of
my knowledge, there is no state or federal agency that has archived these calendars over the 21 years.
Since the advent of increased technological integration with school systems, some districts are beginning to
electronically archive these calendars, however these calendars generally only go back a couple of years.
The Washington State Superintendent of Public Instruction was able supply me with sufficient calendars
for most districts, for each year from Aug. 2000 to June 2004.

6 The shortening of spring and winter breaks is common, as well as the lengthening of the school year into
summer vacation.

7 Thanksgiving falls on Thursday, however both Thursday and Friday are dropped from the sample because
students are given both of those days off from school.









in January, the last two full weeks (at least) of March, the first two weeks of April (at

least 8 school days), and the last two full weeks (at least) of December. Based on these

criteria, I can be reasonably certain that each observation in the analysis is a typical

school day.8 Days which are included in the sample but are not ordinary school days will

serve to downwardly bias the results, because more juvenile crime should be naturally

occurring on those days. This is only true because this measurement of strikes as a

treatment is limited to the days where students have missed an expected day of school.9

1.2.1 Zip-Code Matching

My data set consists of individual-arrest data, where each juvenile is matched to his

home address zip code. Since zip codes are set by the Address Information division of

the U.S. Post Office, they look very different from other boundaries like county borders,

congressional districts, and most importantly school districts. In order to match each

juvenile (and his/her arrest) to his/her appropriate school district, I must address two

problems. The first problem is that zip codes often spill over multiple school districts.

Therefore it is often uncertain which school district a juvenile is in, given their home zip

code. The second problem is that zip codes are redefined over time. First I will focus on

how zip codes change, and how this problem is resolved so that identification strategy is

preserved.

As cities and towns develop both inside and outside city limits, the Post Office

creates new zip codes and new zip-code offices to handle increasing mail traffic. When




8 This is not a perfect matching process because half days and minor differences between districts are
impossible to pinpoint.

9 The timing of strikes is very particular and seldom overlaps with nonschool days, except weekends.









these new zip codes are created, they are drawn strictly as subsets of existing zip codes.10

The creation of these new zip codes makes zip-code fixed effects useless from one year

to the next. In addition, it is also very difficult to pinpoint when new zip codes are

created. However it is possible to see which zip codes were created and what preexisting

zip code they were originally a part of. By using zip-code maps of Washington State

from 1984 and 2000, I am able to map each subset zip code back into its original (1984)

zip code.1 Essentially I treat zip codes in Washington as if they are never partitioned,

beginning in 1984. This method leaves me with 510 zip codes in total and a stable zip

code definition over a 22 year time frame.

The second problem is matching these zip codes to their corresponding school

district. Zip-code overlap with school districts make this matching process difficult,

however the approach I take is sound, and fairly straight forward. I began by totaling the

number of schools in each zip code, noting which district they serve.12 I then divided the

total number of schools in that zip code for a given district by the total number of schools

in that zip code, regardless of their district. Essentially what I am left with is a

probability, that I will call p*, which reflects the probability that a child living in a

particular zip code is also part of a particular school district. Thisp* measure allows me

to assign a proper strike treatment to zip codes.


10 So far I have not been able to find an instance in Washington State since 1984 where a zip code was
created from two or more existing zip codes

1 Four maps of Washington State zip codes were provided by the Western Economics Research Co., Inc.
courtesy of the Suzzallo Library at the Washington Library. They encompass they following years: 1984,
1989, 1992 and 1994. Zip code maps of Washington State for the year 2000 are made available by the
Census.

12 I did this using a master list of schools (as of the 2003-04 school year) that includes each school's
address, grade level served, school code and district name. This list was provided by the Office of
Superintendent of Public Instruction in Washington State. It is also publicly available.









There are several features of zip codes that simplify this matching process. One

helpful factor is that many zip codes do not overlap school district boundaries at all. This

leaves out any guess work. Another helpful point is that zip codes and school districts

often share a significant number of common boundaries. These shared boundaries occur

where divisions seem logical, such as along a county boundary, or a major river or

highway. These convenient features lessen the confusion of the matching process. It is

also worth noting that not all zip codes contain schools.13 In my sample there were four

zip codes without any schools; two were in the heart of Seattle, and two were located in

rural areas. Conveniently, none of these four zip codes overlapped school district

boundaries, so assigning them to the appropriate school district was simple.

The last concern I am left with deals with how school districts change over time, if

at all. If school district boundaries are changing frequently and dramatically over time,

then the matching process breaks down. However, what I find is that school district

boundaries are stable over time. In order for a district to change its boundaries, it must

engage in costly and time-consuming legal processes, however that does not imply that

such adjustments never occur.14 In fact, the two most frequent changes to school district

boundaries seem to be district consolidation and district dissolution.15 Neither changes

turn out not to be problematic. These changes only occur a total of four times over the




13 The zip codes I use in the sample are zip codes with residences. Many zip codes like business zips, P.O.
Box zips, and other nonresidential zips, do not contain schools. These kinds of zips are not included in the
data sample.

14 For an outline of regulations and requirements surrounding district boundary changes, please refer to the
publication, "Changing School District Boundaries: A Lay Person's Guide" published by the Washington
State Board of Education in conjunction with the Office of Superintendent of Public Instruction
15 Consolidation is when two or more districts form a new superdistrict. Dissolution is when a one or more
districts are absorbed into existing districts.









course of the 22 years and involve districts which do not experience any teacher strikes. I

treat these integrated districts as if they had never been separate.

The school district zip code matching process I invoke does not provide perfectly

accurate matches in all cases, but I can be certain of how any mismatching will bias the

results. If I say there was a strike in an unaffected zip code, the effect of the strike will be

biased toward zero because juvenile crime should be unchanged. If I say that there was

not a strike in an affected group, then the strike effect will fail to pick up any increase in

juvenile crime rates. Any mismatching that arises from the imperfect matching process

will downwardly bias the results.

1.2.2 The Basic Model

In this basic model I am attempting to explain changes in juvenile crime for

ordinary school days as a function of teacher strikes. I can start by expressing a simple

regression model in the following form:

Juvenile School Day Crime = a + P(p*)(Teacher Strike)

Again, all of the information from the crime data is reported at the zip-code level,

however the teacher strikes occur at the school-district level. To adjust I introduce p*

where* represents the probability that the student population of a zip code is treated by

the teacher strike. Rather than have fixed effects to take into account differences across

zip codes, I want to initially include specific zip-code characteristics to make sure the

data are well-behaved. I consider income levels, welfare status, parental education,

juvenile work status, juvenile gender, single parent households and community

characteristics in this model. Also I need time-fixed effects to control for temporal

changes over 22 years. I therefore include year, month and day fixed effects in the

model.









The regression model now takes the form:

Juvenile School Day Crimemyd = a + 11(p*)(Teacher Strike)myd +...
+...2(Median Income)myd + P3(Welfare)myd + P4(Parent Education)myd
+... 5(Student Employment)myd + 36(Juvenile Gender)myd +...
+.. 07(Urban)myd + 8(Single Parent)myd 61(Year) + 82(Month) + 63(Day)

Once I am satisfied that the data exhibit all of the normal properties one would expect

from specifying a model of juvenile crime, I can then shift the analysis to include zip

code fixed effects take into account differences which I cannot control for. When I move

to this specification, the final regression model looks like the following:

Juvenile School Day Crimemydz = a + 11(p*)(Teacher Strike) mydz ...
+... 6i(Year) + 62(Month) + 63(Day) + 64(Zip)

1.3 Data

My juvenile-arrest data comes from "Washington Juvenile Court Case Records"

made available by the National Juvenile Court Data Archive. It provides juvenile arrest

data over 22 years, spanning from 1980 to 2001. This data set is both lengthy as well as

highly detailed at the individual level. It provides the home address zip codes of the

offenders, the nature of the crimes committed, the dates of the offenses (as well as the

arrests), as well as many other important characteristics surrounding the reported arrests.

In total when I include only Washington juveniles who were arrested for crimes that

occurred during the ordinary school days preselected for my study, I have 401,864 arrest

cases starting in 1980.

In addition to the juvenile arrest data for Washington State, I use data provided by

the "Census 2000 Summary File 3" and "1990 Summary Tape File 3" to derive annual

zip-code characteristics. These Census files give zip-code level information on

population and school enrollment numbers, as well adult educational attainment, the

number of households who are welfare recipients, type of community, median household









income and student employment. Because these characteristics are only available for

these two periods (1990 and 2000) I trend them over the 22 years using an exponential

function rather than a linear function to describe the path of the trend.16

Finally, teacher strike information came from the report "Public Employee Strikes

in Washington" published in 2003 by the Public Employees Relations Commission in

Washington State. It covers all Washington public employee strikes since 196717;

however since this publication lacks some important details about these strikes, it has

been supplemented with specific information from news articles provided by the

Associated Press and the Seattle Times to strengthen the accuracy of the treatment effect.

These articles are useful in pinpointing specific details surrounding each strike such as

whether the school remained opened with emergency substitutes, and what information

was distributed to parents in districts where strikes occurred.

1.3.1 Methodology

Choosing a proper methodology for this analysis requires some extra care. Because

the analysis is done at the daily level, a huge percentage of observations for any given

day will be zero. The distribution of the dependent variable is skewed towards zero,

making an OLS regression analysis inappropriate. This could be solved by using a

poisson regression analysis, however the fit for a poisson regression for this dataset is

poor.18 Also the poisson analysis does not account for strangely behaved standard errors


16 I also trended these variables linearly, however when I do this I end up with some numbers that are not
feasible do to the imprecise nature of this process. When I impose minimum and maximum constraints on
these variables, I find that these results are insensitive to whether these variables are defined linearly or
exponentially.

17 This report began with the enactment of Public Employees Collective Bargaining Act in 1967.

18 A preliminary test of the data shows that the dependent variable generally has a standard deviation
roughly 3 times larger than the mean. In addition, I observe large values of the Chi squared terms in the









and overdispersion. In this case, overdispersion may take the form of a single juvenile

committing several crimes on a single strike day, or a gang of juveniles committing what

could be considered one crime. Ultimately the chosen method of regression analysis is a

negative binomial regression analysis. This deals with problems of overdispersion that a

poisson does not correct for.19 The dependent variable is aggregate crime per day (crime

count) in a zip code, and the exposure is total student population in said zip code.

1.3.2 Defining the Parameters

As described earlier, I begin by controlling for differences among zip codes directly

rather than using zip-code fixed effects. Each variable is derived from data taken from

the 1990 and 2000 Census. Since I have only two observations for each variable, I

interpolate annual values using an exponential path function. I started by solving for the

growth rate (Gi, zip) of each variable izip:

Gi, zip = ((Censusi, zip 2000/ Censusi, zip 1990)(110)) 1

Then I interpolated each year's observation according to the equation:

Variable izip in year t = Censusi, zip 1990*(1 + Gi, zip)(n), where n = t 1990

Overall this analysis uses 7 variables to capture differences in communities. Income is

measured using Median Income, which comes directly from the Census data. The

Welfare variable is a percentage of households in a zip code that currently receive public

assistance income. Of course I expect that local poverty is a serious factor in juvenile

crime, so both of these variables seem to be relevant (Mocan and Rees 1999). Juveniles

from poor neighborhoods are more likely to be less educated, and therefore more likely to

poisson "goodness-of-fit" tests. Both of these characteristics suggest that the poisson model in not the
correct model for these regressions.

19 The overdispersion parameter alpha is shown to be significantly different from zero. Thus I can conclude
that the negative binomial regression significantly different from the poisson regression.









be criminally active overall (Freeman 1992). Since juveniles from poor households have

a smaller opportunity cost of committing crime, I expect Median Income to be negatively

correlated with juvenile crime, and naturally the converse should be true for Welfare.

The Poor Parental Education variable is defined as the percent of adults (25 years+)

who have not completed the equivalent of a high school education. My prediction here is

that as adult dropout rates increase, juvenile crime increases. Because poor adult/parental

education equates to less adult human capital, parents are likely to have lower demand for

child quality (Becker and Tomes 1976, 1986). Therefore juveniles in these

neighborhoods should also have a lower opportunity cost of crime. In addition, one can

argue that poor adult/parental education may also imply that parents have fewer resources

to provide monitoring for juveniles, less overall time to provide supervision of their own

children (parents may be working long hours or more than one job), or possibly

uninformed/undesirable parenting techniques.

Our Student Employment variable describes the percent of 16-19 year olds who are

both employed and also enrolled in school. The prediction of this variable is not clear.

On one hand, students who are employed in after-school jobs may develop a sense of

responsibility from working a job, or that they may fear losing future wages should they

be caught committing a crime. This suggests that Student Employment should negatively

influence juvenile crime. On the other hand, juveniles who work may have greater access

to reliable transportation, or less restrictive parents, both of which could contribute to a

greater propensity to commit crime. In this case, Student Employment would positively

influence juvenile crime.









The Juvenile Male variable measures the percent of the juvenile population in a zip

code who are male. Since juvenile males traditionally make up the majority of juvenile

crime, juvenile crime for a zip code should increase as a juvenile population there

becomes proportionately more male.20

The Single parent variable is the percent of single parent households in that zip

code. One expects that this variable should be positively correlated with juvenile crime

so that more single parent households lead to more potentially unsupervised juveniles,

and potentially greater juvenile crime.

The Urban variable is the percent of the total zip code population who live in an

urban community. This should pick up basic differences among community types,

however I also interact this variable with the strike measure to see if strikes have

differential impacts in different types of communities. For several reasons, I expect

crime in Urban areas to differ from that of Rural or Suburban areas. Urban areas feature

many characteristics, such as increased criminal opportunity, higher pecuniary benefit,

and increased criminal anonymity, which makes criminal activity in these areas more

desirable (Glaeser and Sacerdote 1999). As such this Urban variable should be positively

correlated with teacher strikes. Lastly, the Strike variable is a dummy variable, with a

value of 1 describing a teacher strike event, and zero otherwise.

1.4 Regression Analysis

The initial results of the negative binomial regression analysis seem to confirm

previous findings about the nature of school incapacitation. Column 1 in Table 1-1



20 From 1993 to 2002, males made up about 71-75% of total juvenile arrests. For more information about
juvenile crime by gender, please reference the Crime in the U.S. annual report published by the Federal
Bureau of Investigation.









shows that the presence of teacher strikes seems to have a positive and significant effect

on juvenile criminal behavior. But before I begin to quantify the effect of school

incapacitation on juvenile crime, I want to make sure that the data are behaving properly.

The signs of the covariates in Column 1 of Table 1-1 seem to support the

predictions made on the effects of income, welfare, parent's education and urban status

with respect to juvenile crime. First, we see that median income and welfare are

negatively and positively significant, respectively. This implies that children in lower

income households and welfare receiving households are more likely to engage in

juvenile crime. These results support the predictions associated with poorer living

conditions. In addition, lower parental education leads to higher crime as well. Again,

for several reasons this result makes sense. The Urban variable is positive and

significant, showing that urban communities experience more overall juvenile crime.

Our Student Employment variable is positive and statistically significant. This may

imply after-school jobs provide students with greater resources which they may use to

commit crime. Further, it may reflect the fact that as more students get after-school jobs,

they are spending less time increasing their human capital (through study, after-school

activities, etc.)

One problem is that the effect of the Juvenile Male variable is negative and

significant, which contradicts the predicted effect. It says that across juvenile

populations, those that are comprised of relatively more females experience more

juvenile crime. One possible explanation for this result is that it may be reflecting

problems in variable measurement for small communities. For rural communities there

are two basic problems. The first is that with a small population of juveniles, the









variation of the Juvenile Male variable increases significantly.21 The second is that this

unusual variation makes interpolating this variable equally unreliable. Since a significant

portion of zip codes have very small populations of juveniles, it is possible that the

negative result I observe is being driven by rural zip codes that have atypical juvenile

gender characteristics.22

To check the sensitivity of the initial results I begin by eliminating months in which

strikes did not occur from the sample: January, March, and May (Specification II).23 I

then went a step further and eliminated every month other than September, October and

April from the sample (Specification III). September and October are preserved because

they are the most common months for strikes, and April was included because in 1991

there was a "state-wide" teacher strike that involved at least 41 school districts. The

results of Specifications II & III are reported in Columns 2 and 3 of Table 1-1

respectively. These results are not significantly different from Specification I.

Recalling the zip code matching process described earlier in the paper, I am able to

accurately match back to 1984 zip codes. However the crime data dates as far back as

1980. To deal with potential zip code mismatches before 1984, I drop the first four years

of the sample (Specification IV).24 This fourth specification shows the strike variable


21 For zip codes with juvenile populations less than 500, the standard deviation of the Juvenile Male
variable is approximately 3 times larger than the standard deviation of those zip codes with 500 or more
juveniles.
22 In fact when I look at only urban zip codes, I see that this variable seems to correct itself, which gives
credence to this argument. Further, Table A-2 shows that when I drop the smallest zip codes (population
less than 500 juveniles) I see that the negative significance of the coefficient is destroyed. Thus it seems
plausible that poor interpolation of this variable for sparsely populated areas has tainted the coefficients.

23 Over the 22 year period, there was one school district that orchestrated a strike in January, but it only
lasted for one day.
24 In doing this I eliminate 9 strike events, 3 of which lasted 9 or more days.









remains significant as the sample is modified (Column 4 of Table 1-1). These results

seem robust to many different representations of the data set.

It is also important to verify that these results are not being driven by zip code

characteristics being specified inaccurately. The most immediate check of this is to use

zip code fixed effects in lieu of explicit variables to capture the difference. Column 5

shows the addition of zip code fixed effects does not significantly alter the results.

The coefficient of the strike variable Column 5 of Table 1-1 shows that strikes have

a positive effect on juvenile crime that is statistically significant. To quantify the impact

of this increase, I take the partial derivative of the dependant variable with respect to the

strike variable (aTotalcrime/aStrike). When I divide this marginal effect by the average

of the dependent variable, I am left with a change that reflects a proportion of the mean of

the dependent variable. In the full data set, the marginal effect of the strike variable

suggests that total juvenile crime increases by 56.71% on days when strikes occur. This

is more than a modest change in total crime.

1.4.1 Community Differences

Since the full data set includes all communities, some of which may look and

behave very differently from one another, I want to test whether juvenile crime in

different communities is differentially affected on strike days. To capture community

differences, I run separate regressions for each community type, where I restrict the

sample to only those communities in which 51% or more of the population in a zip code

live in one of three types of communities: urban, suburban or rural.25 The results of these


25 The reader should be aware that this specification implies that zip codes which are considered one-third
urban, one-third suburban, and one-third rural, will therefore be excluded from this analysis. However
because of the way in which the Census defines these communities, these kinds of zip codes make up a
very small portion of the total sample.









regressions can be viewed in the Columns of Tables 1-2, 1-3 and 1-4. These results show

that unexpected school absences significantly influence juvenile crime only in urban

communities.

This result is not at all unreasonable, given there are many differences among these

types of communities that no doubt influence criminal behavior. It seems understandable

that rural communities experience no change in juvenile crime when a strike occurs.

Rural communities may provide less opportunity for crime, as well as a small town

atmosphere that makes anonymity difficult. The lack of a significant effect of strikes in

suburban communities is less believable since they arguably provide greater opportunity

and anonymity, however there may yet be other, (possibly unobservable), characteristics

about suburban characteristics that naturally deter juvenile crime on these strike days.

For instance, suburban communities may be populated by more involved parents, may

have more effective emergency resources, etc. The strike effects in urban communities

however are positive and significant. In urban communities juvenile crime increases

19.68% on strike days.26

Again I run the same specification tests to test the robustness of these results. The

signs of the covariates in every urban subsample are the same as in Table 1-1, except for

Juvenile Male. The Juvenile Male covariate does take on a positive and significant

coefficient in the urban subsample, however in the suburban and rural subsamples it

remains negative and significant.





26 The huge decrease in the magnitude of the strike coefficient from the full sample to urban subsample
indicates that small populations in rural communities are exaggerating the effects of the strikes in the full
sample.









1.4.2 Differences in Offense Types

Since school incapacitation may be differentially affecting the types of crimes that

juveniles are committing, I partition the dependent variable measure of total crime into 5

specific crime types. These types of crimes include: drug and alcohol related crimes,

mischievous crimes, property crimes, violent crimes, and finally weapons and

endangerment crimes. I still consider all three community classifications when doing this

analysis because I may yet find significant results in types of crimes for rural and/or

suburban communities.27 Table 1-5 shows the results of this extended analysis.

Rural and suburban communities continue to experience no change in juvenile

crime resulting from strike days regardless of the type of crime that is being committed.

Urban communities though do experience an array of changes in juvenile criminal

behavior. Both mischievous crimes and property crimes seem to increase in the presence

of a teacher strike. I estimate that mischievous crime increases by 48% on average for

days when strikes occur.28 Property crime increases by an average of 28.81%. This

change in property crime is nearly double previous estimates in the literature. Violent

crime decreases by approximately 31.53%.29 Drug and alcohol crimes, and weapons and

endangerment crimes are unaffected by school incapacitation.

These changes in criminal activity seem to make sense. It seems logical that

mischievous crime would be most affected by school incapacitation. After all, these are

27 This could arise if increases in certain crimes were offsetting decreases in other types of crimes.

28 It is difficult to compare my measure of mischievous crime to Jacob and Lefgren's measure of minor
offenses. They define minor offenses as "NIBRS Group B offenses", however my measure of mischievous
crime is a select subset of these crimes.

29 Jacob and Lefgren estimate that school attendance decreases property crime by approximately 14%.
They also estimate that violent crime increases on school days by about 28%. Therefore, my estimate of
31.53% for violent crime is approximately 10% larger than previous estimates.









the types of crimes that often result from boredom rather than calculated criminal

thought. On the other hand, the decrease in violent crime that occurs is harder to explain.

I use, as Jacob and Lefgren do, social interaction theory to explain this result. Jacob and

Lefgren argue that juvenile violence against other juveniles arises, at least in part, from

disputes formed in and around the classroom. When these juveniles are not forced to be

at school, there is a decrease in the amount of overall violent crime that reflects a

decrease in juvenile violence against other juveniles. This explanation seems credible.30

Given the limits of the data set, I cannot test this theory further.31

1.4.3 Types of Offenders

In addition to the topics already spoken to, the detailed level of the data also allows

me to also explore the nature of the criminal, beyond superficial characteristics. By

observing criminal record, I attempt to answer whether these additional crimes are caused

by normal delinquents facing greater criminal opportunity, or by juveniles who do not

normally engage in criminal activity. I again partition the dependent variable into two

measures: 1) total crime by repeat offenders and 2) total crime by one-time offenders. I

define repeat offenders not just as those juveniles who have committed previous crimes,

but also those juveniles who will commit future crimes. Thus a juvenile who commits

multiple crimes is considered to be a repeat offender juvenile even on their first crime.

The results of these regressions can be seen in Table 1-6.

Additional overall crime is induced for both offender types, however the average

increase in crime for one-time offenders as a percent of the mean crime level for that



30 In 1998 and 1997, 62% of victims of juvenile violence were 17 or younger.

31 I have no information concerning the victims of the crimes in my data set.









group is much higher than for the repeat offenders. On strike days there is a 33.45%

increase in average total crime for one-time offenders, and a 13.89% increase in average

total crime for repeat offenders. Essentially, one-time offenders as a group seem to be

much more affected by the lack of school incapacitation then the repeat offenders. This

evidence suggests that the increased crime on strike days is less about criminal motive,

and is more a result of boredom.

If these two groups are motivated differently to commit crimes, these effects may

not be expressing differences across groups in the same types of crimes. As a society we

may be more concerned if the lack of school incapacitation were inducing more property

crime by one-time offenders rather than repeat offenders (where the crime may have

eventually happened anyway). Therefore I am interested in examining whether these two

groups differ in offense types on these strike days. To do this I create a dependent

variable that equals total crime conditional on offender type and offense type. The results

are expressed in Table 1-6. The decrease in violent crime comes mainly from the repeat

offenders. Violent crimes committed by repeat offenders drop by 36.72% on strike days.

In addition, both groups are contributing to the overall increases in mischievous crimes

and property crimes. Property crimes by one-time offenders and repeat offenders

increase by 56.75% and 25.40% respectively. Likewise mischievous crimes increase by

83.11% for one-time offenders and by 30.82% for repeat offenders.

1.4.4 Strikes Days as Gateway Crimes

Since the increase in crimes committed on strike days is substantially larger for

first-time offenders, it may also be true that these juveniles may carry this behavior

forward after their first arrest. To see if this is the case, I examine whether or not a









current day repeat offender was significantly more likely to haven gotten his/her start

(commit his/her very first crime) on a strike day versus a normal school day. To do this, I

generate a daily crime total that expresses the total of"first time" crimes by subsequent

repeat offenders. The results of this regression can be also viewed in Table 1-6. A future

repeat offender is found to be significantly more likely have committed his first crime on

a strike day rather than an ordinary school day. This suggests that not only is school

incapacitation preventing the creation of crimes, but it may also be preventing the

creation of criminals.

1.4.5 Displacement of Crime

Thus far it seems the evidence suggests juvenile crime increases when students are

not incapacitated in school. However what is not clear is whether these changes describe

an overall increase in the amount of juvenile crime, or a displacement of crime from one

day to another. Jacob and Lefgren try to speak to this question of temporal displacement

in their paper, however their use of in-service days again limits the power of their

analysis. Jacob and Lefgren find that there is no temporal displacement of crime to in-

service days, however because in-service days are completely known, this result may

simply be showing that juvenile criminals will plan to commit more crime when facing

more criminal opportunity. What we are really interested in is whether a juvenile will

temporally displace a "planned" crime given an unexpected opportunity to do so, or

whether an unexpected opportunity to commit crime results in more "unplanned" crime.

Again teacher strikes provide a more ideal measure of unexpected criminal opportunity,

so I revisit this question of temporal displacement.









To test whether crime is being displaced, I begin by aggregating Total Crime to a

weekly measurement rather than daily32. If juvenile crime is simply being displaced from

weekend crimes to weekday crimes, then I should not observe an effect for the strike

variable. To capture the weekly aggregated effect of strikes, I proportion the fraction of

the week where strikes occurred as a new measure of the strike treatment. Despite this

aggregation, I find that strikes are still a significant factor in contributing to additional

juvenile crime (Table 1-7).

Table 1-7 shows a positive and significant strike coefficient on Total Crime in the

full data set. This is a strong indication that overall juvenile crime is increasing, rather

than being displaced across time. However, if displacement of crime varies across crime

type, or criminal type, we cannot conclude that no crimes are being displaced. To see if

the type of crime or the characteristics of the perpetrator influences whether displacement

of crime occurs, I partition weekly total crime by crime type and offender as before. The

results of these regressions are also reported in Table 1-7. These strike coefficients on

Table 1-7 show that every specification of crime type and criminal type results in an

increase of overall juvenile crime, and not temporal displacement of crime. I confirm the

lack of displacement in total crime by aggregating further to a monthly crime total.

Weekly and Monthly aggregations of juvenile crime provide a good test of

displacement of crime. I perform another test by lagging the strike effect, in order to pick

up acute changes in crime shortly after the strike ends. If fewer crimes are committed on

days following the strike, then the coefficient of the lagged strike variable should be



32 This weekly measurement includes weekends and holidays, since strike day crime could displaced from
usual weekend crime. In addition, week fixed effects are used in place of day fixed effects.









negative and significant.33 I run seven separate regressions, shown in Table 1-8, labeled

"1 day" through "7 days." For a strike that ends on day t, the first treatment of the lagged

strike variable under "1 day" begins on day (t + 1), under "2 days" begins on day (t + 2),

and so on. Table 1-8 shows that there is no significant effect from the lagged strike

treatment. These results provide further evidence that juvenile crime occurring on strike

days is additional crime, rather than displaced crime.

1.4.6 Duration Effects on Criminal Activity

One last advantage I gain from using teacher strikes as an instrument, in

conjunction with the strength of the daily arrest data, is that I can track how juvenile

crime changes over time with prolonged absence. If it were true that an increase in

juvenile crime is at least partially caused by boredom and inactivity on the part of

juveniles, then I should expect that the longer juveniles remain "inactive", the more crime

is likely to result. Arguably there is just as much uncertainty over when teacher strikes

will end as when they will begin. As such, trends in juvenile crimes up until the start of

the strike and over the period of the strike should not be biased by student or parent

foresight. In addition, the longer a strike continues, the less able a parent or guardian is

to provide adequate supervision to students. For example, a parent may be able to take

one or two days away from work to supervise their child, but the cost of taking 10

consecutive days off work is substantially larger.

To examine if juvenile crime is changing with the length of a teacher strike I

partition the strike variable into 4 groups of new independent variables. Three of these

variables are clustered groups of 3 consecutive strike days, and the last variable pertains


33 This sample period also includes weekends. I also revert back to day fixed effects rather than week FE.










to a strike in its 10th day or higher.34 These results can be viewed in Table 1-9. Table 1-9

shows that for the first 3 days, there is no significant change in overall juvenile crime.35

After the 5th day or so, the strike coefficient becomes positive and significant, reflecting

an increase in overall crime. Further, after the 5th or so strike day there is a gradual

increase in the level of juvenile crime as strike length continues. In unreported

regressions I find that as time passes, property and mischievous crime trends upward,

while the decrease in violent crime remains relatively constant across days. The positive

coefficient shows that these increases in property and mischievous crime overpower the

decreases in violent crime. Column 2 of Table 1-9 repeats the regression in Column 1,

only with the strike variable clustered at the two-day level rather than at the three-day

level. Clearly the evidence suggests that the longer juveniles remain absent from school,

the more overall daily juvenile crime will be observed.36

1.5 Robustness Checks

The results of the analysis up to this point seem to point to a very clear story about

the relationship between juvenile crime and school incapacitation. In this section I

perform some robustness checks to ensure that these results are not spurious. I begin by



34 do this because if I were to partition the strike variable for each separate day (1st day, 2nd day, etc.), then
the number of treatments would be very small for the later strike days and the coefficients would be
unreliable.

35 When I break up the dependent variable (total crime) by crime type as before, violent, property and
mischievous crime types are individually significant. However in the first 3 days, the opposite signs and
magnitudes of the coefficients wash out aggregate effects in total crime. These results are consistent with
Jacob and Lefgren's findings (2003).
36 It is worth keeping in mind that these strike day counts are also limited to school days. That is to say that
when a strike has reached its 10th day, a student has missed exactly 2 full weeks of school. It is also very
likely that these criminal behaviors will change after a strike has gone on for an extensive amount of time.
At that point, juvenile crime may level-off or even decrease substantially. Of course given the sparse and
limited nature of extremely lengthy teacher strikes, an analysis of that nature is inherently difficult not to
mention highly unreliable.






25


testing whether the strike variable is correlated with some unobservable characteristic,

and that perhaps the results are picking up some alternate relationship. For instance, if a

strike event is in fact very foreseeable, then one might speculate that the increase in

juvenile crime predicted by the strike variable is really picking up an increased police

presence on days that strike occur. That is to say that the results are not reflecting a true

increase in juvenile crime, but that there are simply more police on the streets capturing a

larger portion of the same total amount of juvenile crime.

If this were true, then one would expect that the number of days between an offense

and an arrest would be shorter on strike days. The graphs on Figure 1-1 show that this is

not the case. When I consider only zip codes where strikes did occur, I see that the

percent of quick arrests37 is significantly higher on regular school days than on strike

days. I explain this by suggesting that on regular school days, police have a better idea

who the likely criminals are (repeat offenders), as opposed to strike days where offenders

are more likely to be (previously unidentified) first time, or one-time offenders.

1.5.1 Importance of Singular Strikes

Given the uniqueness of strikes, I must confirm that these results are not being

solely driven by a small subsets of strikes. To do this I systematically eliminate each

individual year from my sample starting with 1984. Table 1-10 shows that strikes are a

significant determinant of juvenile crime in every year with the possible exception of

1985 (z = 1.70). When I drop 1985 from the sample I find a much weaker significance

than when I eliminate any other year. In September of 1985, the Seattle school district

engaged in a teacher strike that lasted 20 school days. Given that the evidence has


37 I define these as arrests where the associated offense occurred less than 48 hours before the arrest.









suggested that juvenile crime in urban districts is most affected by teacher strikes, it is

possible that these results are being solely driven by the circumstances surrounding this

one district.

To test this condition I restrict the sample twice, once to look at the sample without

the Seattle school district, and then again to look at only the Seattle school district. If

Seattle is driving all of the results, then the strike variable should be insignificant when I

exclude Seattle, and extremely significant when I exclude all other districts except for

Seattle. Table 1-10 shows that strikes still have a positive and significant effect when I

drop Seattle from the analysis. In addition the strike variable in the Seattle subsample is

marginally significant (z = 1.82). Therefore I can be confident that while this one event

obviously contributes to the result, (as I would expect the second longest strike in the

most urban school district in Washington State to do), it is not the sole force driving the

results.38

1.5.2 Reproducibility of Results

Another robustness check I perform is to make sure that the results cannot be easily

reproduced with random generations of the strike variable. To check if this is the case, I

use a uniformly generated random variable on the interval from 0 to 1 to re-specify a

random dummy strike variable. Since teachers are on strike 2,351 days out of the

1,703,910 day sample, strike days occur 0.13797 % of the time. I assign a strike value of

1 to the random variable, when the value of the random observation generated is less than

or equal to 0.0013797. If the randomly generated observation is greater than 0.0013797,

then it is reassigned a value of zero. The new random strike variable should approximate

38 The Seattle strike of 1985 is actually tied with the Mukilteo strike of 1990 (at 20 days) for the second
longest strike in Washington State history from 1980-2001.









the number of strikes in the original sample. The results in Table 1-11 show the mean of

the random strike coefficient over 25 trial regressions, as well as the average number of

random strike days generated. Of the 25 trials, only 2 random strike variable coefficients

were significant. One of the statistically significant coefficients was positive, while the

other was negative.

1.5.3 Further Tests

One final test of the data is to make sure that my finding of new criminal

participation on strike days (when the sample is broken up into one-time and repeat

offenders) is not a byproduct of mislabeling offenders. Since the crime data ends in

2001, it could be that in the years leading up to 2001 many future career criminals are

being mislabeled as one-time offenders due to the fact that future arrests past 2001 are

unknown. To test this possibility, I drop the last 4 years of the data sample and repeat the

regressions using offender type as the dependent variable. Table 1-12 shows that this

mislabeling does not affect the results. The effects of a strike on repeat offenders and

one-time offenders are nearly identical for either specification.

1.6 Conclusion

The results presented in this chapter provide a clearer picture concerning the

relationship between schooling and urban juvenile crime. My main findings are the

following: 1) The effects of a lack of school incapacitation are larger than those found in

Jacob and Lefgren's study. The evidence suggests that property crime rises by as much

as 29%, almost twice as much as what is predicted in the existing literature. Likewise,

violent juvenile crime decreases by as little as 31.53%, and as much as 36.72%. These

estimates of increased violent crime represent a 10%-25% increase in previous estimates.







28

2) I confirm that these changes in daily crime reflect changes in total crime, and not a

displacement of crime from one day to the next, regardless of crime type and criminal

type. 3) Different types of juveniles are affected differently by a lack school

incapacitation. A significant proportion of the decrease in the level of violent crime can

be attributed to repeat juvenile offenders, while one-time offenders contribute more to the 28

increases in property crime and crimes of mischief. 4). A failure to incapacitate juveniles

will result in significantly more crime by those juveniles who might not have otherwise

engaged in criminal acts or who rarely engage in such acts. In fact, juveniles who

become repeat offenders are more likely to have gotten their criminal start when

incapacitation was expected but not implemented, as opposed to ordinary school days.

Incapacitation seems to have the greatest implications for new and seemingly

preventable urban crime. I find that on strike days, new criminals are engaging in new

criminal acts. Incapacitating these juveniles seems to be effective at preventing at-risk

urban youths from engaging in new and risky behavior. I do not find that school

incapacitation has any significant effect in suburban and rural communities. It may be

true that the characteristics that differentiate these community types from urban

communities, as well as each other, inherently reduce juvenile criminal behaviors,

however such characteristics may be largely unobservable to us.

If the evidence presented in this paper is accurate, then what we are learning is that

how we manage children's time outside of school is very important. If parents are

strained to provide adequate supervision for their children, juveniles who we might

consider to be at-risk may be the ones who are affected most, though at-risk juveniles are

not the only ones affected. Families and businesses also bear the burden of increased









juvenile delinquency. These results have profound implications for many urban school

policies and programs. How school districts budget their students' days off is no longer a

trivial matter. For example, school districts with long breaks (like in a traditional school

calendar) may have very different juvenile criminal behaviors than a school district with

frequent but shorter breaks (like in a year-round school calendar). Differences in how

school calendars overlap within and among districts should have an effect on the nature

of juvenile crime. The length of the school day or school year in a district, school district

policies regarding student attendance requirements, and even the nature of after school

programs within a district are all policy issues that determine more than just a child's

educational outcome. It is my goal to continue to explore this data set and scrutinize

these kinds of school policies, so that I can better understand the impact that these kinds

of policies may have.











Table 1-1: Negative Binomial Regression onto Total Crime with Full Data Set
Variables (I) (II) (III) (IV) (V)
Strike 0.195** 0.187** 0.210** 0.260** 0.220**


(4.06)**


(3.86)**


(4.16)**


(5.06)**


(4.38)**


Median Income



Welfare


Urban



Poor Parental Educ.



Juvenile Maleness



Student Employment


Single Parent House


Alpha


-0.000058**
(102.47)**

0.475**
(7.71)**

0.132**
(26.96)**

0.329**
(10.40)**

-0.330**
(28.06)**

0.291**
(17.49)**

0.500**
(9.52)**

0.989**
(156.56)**


-0.000059**
(83.82)**

0.452**
(5.96)**


-0.000059**
(60.76)**

0.520**
(4.97)**


0.112** 0.097**
(18.45)** (11.54)**


0.346**
(8.88)**

-0.341**
(26.24)**


0.398**
(7.45)**

-0.355**
(22.32)**


0.297** 0.290**
(14.43)** (10.23)**


0.495**
(7.64)**

1.009**
(126.04)**


0.466**
(5.20)**

0.997**
(91.12)**


Number of obs. 1,703,910 1,128,630 583,440 479,400 479,400
Time Fixed Effects Y Y Y Y Y
Zip Fixed Effects N N N N Y
R-Squared 0.0367 0.0364 0.0367 0.0279 0.0559
In Columns I V, the sample is re-specified as follows (Type I through Type V):
(I) Sample includes all ordinary school days from 1980-2001
(II) January, March and May are excluded from the sample 1980-2001
(III) Includes only April, September and October from 1980 2001
(IV) Only April, September and October from 1984 2001
(V) Only April, September and October from 1984 2001, with zip code fixed effects
Z-statistics are given in the parentheses for every table. Indicates significance at the 5% level,
** is significant at the 1% level or smaller. To interpret these regressions, I have to calculate the marginal
effect of these variables. The marginal effect of the strike variable is 0.149, which represents a 56.71%
increase in average total juvenile crime on strike days.


-0.000055**
(52.80)**

0.821**
(6.28)**


0.055**
(6.14)**

0.482**
(6.67)**

-0.396**
(22.72)**

0.296**
(9.16)**

0.985**
(9.67)**

0.956**
(87.74)**


0.759**
(79.06)**










Table 1-2: Urban Subsample Negative Binomial Regression onto Total Crime
Variables (I) (II) (III) (IV) (V)
Strike 0.151** 0.129* 0.178** 0.211** 0.214**
(3.07)** (2.60)* (3.50)** (4.11)** (4.26)**

Median Income -0.000061** -0.000062** -0.000062** -0.000057** -
(83.07)** (67.39)** (48.42)** (42.17)** -

Welfare 1.394** 1.369** 1.418** 1.703** -
(16.40)** (13.01)** (9.79)** (9.46)** -

Poor Parental Educ. 0.398** 0.403** 0.449** 0.462** -
(9.49)** (7.79)** (6.31)** (4.65)** -

Juvenile Maleness 0.602** 0.577** 0.482** 0.397** -
(11.13)** (8.60)** (5.16)** (3.62)** -

Student Employment 0.214** 0.224** 0.214** 0.157** -
(9.10)** (7.67)** (5.34)** (3.47)** -

Single Parent House 0.739** 0.790** 0.945** 1.276** -
(8.28)** (7.12)** (6.24)** (7.55)** -

Alpha 0.676** 0.685** 0.666** 0.638** 0.494**
(114.06)** (91.00)** (65.18)** (62.65)** (54.87)**
Number of obs. 454,953 301,312 155,601 129,593 129,593
Time Fixed Effects Y Y Y Y Y
Zip Fixed Effects N N N N Y
R-Squared 0.0484 0.0484 0.0487 0.0400 0.0635
Table 1-2 limits the sample to include only zip codes where 51% or more of the population lives in an area
considered by the Census to be Urban
In Columns I V, the sample is re-specified as follows (Type I through Type V):
(I) Sample includes all ordinary school days from 1980-2001
(II) January, March and May are excluded from the sample 1980-2001
(III) Includes only April, September and October from 1980 2001
(IV) Only April, September and October from 1984 2001
(V) Only April, September and October from 1984 2001, with zip code fixed effects
Z-statistics are given in the parentheses for every table. Indicates significance at the 5% level,
** is significant at the 1% level or smaller. To interpret these regressions, I have to calculate the marginal
effect of these variables. The marginal effect of the strike variable is 0.121, which represents a 19.68%
increase in average total juvenile crime on strike days.










Table 1-3: Suburban Subsample Negative Binomial Regression onto Total Crime
Variables (I) (II) (III) (IV) (V)
Strike -0.027 -0.064 -0.159 -0.162 -0.449
(0.12) (0.29) (0.61) (0.62) (1.74)

Median Income -0.000015** -0.000016** -0.000013** -0.000012** -
(8.01)** (6.54)** (4.74)** (5.66)** -

Welfare -1.929** -2.146** -1.483** -2.005** -
(9.21)** (8.31)** (4.20)** (4.43)** -

Poor Parental Educ. 1.639** 1.856** 1.683** 2.393** -
(13.33)** (12.29)** (8.09)** (8.55)** -

Juvenile Maleness -1.432** -1.973** -2.151** -2.899** -
(5.33)** (5.94)** (4.74)** (5.66)** -

Student Employment -0.121* -0.198** -0.172 -0.064 -
(2.23)* (2.96)** (1.87) (0.59) -

Single Parent House 2.285** 2.369** 0.217** 3.062** -
(14.13)** (11.97)** (7.83)** (8.82)** -

Alpha 1.326** 1.303** 1.240** 1.203** 0.969**
(59.08)** (54.84)** (39.01)** (37.78)** (34.42)**
Number of obs. 134,100 88,841 45,794 37,992 37,992
Time Fixed Effects Y Y Y Y Y
Zip Fixed Effects N N N N Y
R-Squared 0.0294 0.0300 0.0302 0.0227 0.051
Table 1-3 limits the sample to include only zip codes where 51% or more of the population lives in an area
considered by the Census to be Suburban
In Columns I V, the sample is re-specified as follows (Type I through Type V):
(I) Sample includes all ordinary school days from 1980-2001
(II) January, March and May are excluded from the sample 1980-2001
(III) Includes only April, September and October from 1980 2001
(IV) Only April, September and October from 1984 2001
(V) Only April, September and October from 1984 2001, with zip code fixed effects
Z-statistics are given in the parentheses for every table. Indicates significance at the 5% level,
** is significant at the 1% level or smaller










Table 1-4: Rural Subsample Negative Binomial Regression onto Total Crime

Variables (I) (II) (III) (IV) (V)
Strike 0.017 0.027 -0.343 -0.285 -0.392
(0.09) (0.14) (1.45) (1.20) (1.65)

Median Income -0.000034** -0.000036** -0.000036** -0.000033** -
(26.68)** (22.50)** (16.62)** (13.59)** -


Welfare -0.077 -0.063 -0.108 0.117 -
(0.68) (0.45) (0.56) (0.48) -


Poor Parental Educ. -0.282** -0.332** -0.218 -0.200
(3.81)** (3.63)** (1.77) (1.32) -


Juvenile Maleness -0.357** -0.362** -0.370** -0.421** -
(37.69)** (32.83)** (25.81)** (25.64)** -


Student Employment 0.142** 0.156** 0.180** 0.267** -
(4.34)** (3.86)** (3.25)** (4.15)** -

Single Parent House 0.379** 0.401** 0.317* 0.864** -
(4.41)** (3.81)** (2.15)* (5.21)** -

Alpha 2.679** 2.767** 2.810** 2.712** 2.197**
(80.71)** (66.36)** (49.08)** (46.97)** (44.48)**
Number of obs. 1,101,359 729,469 377,244 304,491 304,491
Time Fixed Effects Y Y Y Y Y
Zip Fixed Effects N N N N Y
R-Squared 0.0313 0.0308 0.0316 0.0218 0.0551
Table 1-4 limits the sample to include only zip codes where 51% or more of the population lives in an area
considered by the Census to be Rural
In Columns I V, the sample is re-specified as follows (Type I through Type V):
(I) Sample includes all ordinary school days from 1980-2001
(II) January, March and May are excluded from the sample 1980-2001
(III) Includes only April, September and October from 1980 2001
(IV) Only April, September and October from 1984 2001
(V) Only April, September and October from 1984 2001, with zip code fixed effects
Z-statistics are given in the parentheses for every table. Indicates significance at the 5% level,
** is significant at the 1% level or smaller










Table 1-5: Effects of Strike Days by Crime Type and by Community Type
Strike Strike Strike
Dependant Variable Coefficient for Coefficient for Coefficient for
Urban Sub- Suburban Sub- Rural Sub-
Sample Sample Sample
Drugs and Alcohol 0.200 -1.188 -0.113
(1.36) (1.48) (0.20)

Mischievous 0.536** -0.813 -0.241
(3.88)** (0.73) (0.34)

Property 0.257** -0.128 -0.276
(3.79)** (0.34) (0.80)

Violent -0.364* -1.103 -0.266
(2.50)* (1.44) (0.52)

Weapon and -0.014 -53.096
Endangerment (0.05) (0.02) -
Number of obs. 129,593 37,992 304,491
Time Fixed Effects Y Y Y
Zip Fixed Effects y y Y
Ave. R-Squared 0.0506 0.0423 0.0579
Each row represents a new regression with the listed crime type as the dependant variable. The reported
number is the strike coefficient for the regression with the corresponding dependant variable. Each column
is a different subsample group: urban, suburban and rural. All 14 regressions are defined as a Type V
regression, including data from only April, September and October from 1984 2001, and zip code fixed
effects. The Pseudo R-squared reported at the bottom of each column is the average R-squared over the 5
regressions. Z-statistics are given in the parentheses for every table. Indicates significance at the 5%
level, ** is significant at the 1% level or smaller. The marginal effects of these variables show that violent
crime decreases by 31.53% on strike days. Property crime and mischievous crime increase by 28.81% and
48% respectively. Again, these effects reflect changes in juvenile crime as a proportion of the mean. The
coefficients of the marginal effects are (-0.032), (0.072) and (0.025) for violent, property and mischievous
crime respectively.










Table 1-6: Effects of Strike Days by Crime Type and by Offender Type
Strike
Strike Coefficient Stri
Coefficient for
Dependant Variable for One-Time
Offenders Repeat
Offenders
Total Crime 0.382** 0.156**
(4.05)** (2.77)**

Drugs and Alcohol 0.337 0.153
(1.18) (0.93)

Mischievous 1.024** 0.597**
(3.66)** (3.99)**

Property 0.534** 0.220**
(4.44)** (2.81)**

Violent -0.567 -0.408**
(1.48) (2.59)**

Weapon and -1.330 -0.009
Endangerment (1.20) (0.03)

Gateway Crime -0.286*
-- (2.00)*
Number of obs. 129,593 129,593
Time Fixed Effects Y Y
Zip Fixed Effects Y Y

Table 1-7 includes only urban zip codes. All 13 regressions are defined as a Type V regression, including
data from only April, September and October from 1984 2001, and zip code fixed effects.
Z-statistics are given in the parentheses for every table. Indicates significance at the 5% level,
** is significant at the 1% level or smaller. The marginal effects of these variables show that for repeat
offenders, violent crime decreases by 36.72%, property crime increases by 25.4%, and mischievous crime
increases by 30.82% on strike days. The coefficients of these marginal effects are (-0.032), (0.046) and
(0.012). For one-time offenders property crime increases by 56.75% and mischievous crime increases by
83.11% on strike days. The coefficients of these marginal effects are (0.030) and (0.008). Again, these
effects reflect changes in juvenile crime as a proportion of the mean.











Table 1-7: Temporal Displacement of Crime by Crime Type and by Offender Type
Strike Strike Strike Strike
Weekly Aggregated Dependant Coefficient Coefficient Coefficient for Coefficient
Variable for Full for Urban One-Time for Repeat
Data Set Subsample Offenders Offenders


Total Crime



Mischievous


0.235**
(4.50)**


Property


Violent


0.243**
(4.67)**

0.570**
(4.39)**

0.312**
(4.73)**

-0.300*
(2.32)*


0.418**
(4.51)**

0.820**
(3.09)**

0.492**
(4.27)**

-0.223
(0.73)


0.176**
(3.11)**

0.501**
(3.55)**

0.225**
(3.06)**

-0.317*
(2.27)*


Monthly Aggregated Dependant
Variable


Total Crime


0.534** 0.485**
(3.86)** (3.73)**


Number of obs. 440,640 117,491 117,491 117,491
Time Fixed Effects Y Y Y Y
Y Y Y Y
Zip Fixed Effects

Each row represents a new regression with the listed crime type aggregated at the weekly level as the
dependant variable. The reported number is the strike coefficient for the regression with the corresponding
dependant variable. Each column is a different sample group. Column 1 is the full sample. The remaining
three Columns are from the urban sample. The top 10 regressions are defined as a Type I regressions
including zip code fixed effects. The bottom 2 regression coefficients represent the strike effect on Monthly
Total Crime. Z-statistics are given in the parentheses for every table. Indicates significance at the 5%
level, ** is significant at the 1% level or smaller










Table 1-8: Temporal Displacement of Crime through Lagged Strike Variable
Lagged Strike Variable Strike Coefficient
For Urban Sub-
Sample
1 Day -0.033
(0.62)

2 Days -0.028
(0.54)

3 Days -0.030
(0.57)

4 Days -0.046
(0.87)

5 Days -0.041
(0.78)

6 Days -0.093
(1.75)

7 Days -0.070
(1.33)
Number of obs. 293,112
Time Fixed Effects Y
Zip Fixed Effects Y
Ave. R-Squared 0.0597
Each row represents a new regression with strike variable lagged the by the corresponding row. Each
regression is defined as a Type IV regression including zip code fixed effects of the urban subsample. The
notable exception is that weekends are included to allow for the correct number of strike treatments.
Therefore, days where lagged strike treatments overlap breaks, holidays and other nontypical school days
are included in the sample. Z-statistics are given in the parentheses for every table. Indicates significance
at the 5% level, ** is significant at the 1% level or smaller











Table 1-9: Changes in Total Juvenile Crime, by Days Elapsed since the Start of a Strike
Partitioned Strike Partitioned Strike
Variable Strike Coefficient Variable Strike Coefficient


Days 1 3



Days 4 6



Days 7 9



Day 10+


Alpha


-0.005
(0.06)

0.185
(1.91)

0.261*
(2.54)*

0.313**
(2.93)**


Days 1 2



Days 3 4



Days 5 6



Days 7 8



Days 9 10



Day 11+


0.533**
(101.05)**


Alpha


0.058
(0.58)

0.006
(0.07)

0.262*
(2.37)*

0.284**
(2.62)**

0.228
(0.98)

0.305**
(2.67)**


0.533**
(101.04)**


Number of obs. 454,953 Number of obs. 454,953
Time Fixed Effects Y Time Fixed Effects Y
Zip Fixed Effects Y Zip Fixed Effects Y
R-Squared 0.0713 R-Squared 0.0713

Table 1-8 includes only urban zip codes. These two regression are defined as a Type I regressions, using
data including all ordinary school days from 1980-2001, and zip code fixed effects. Both regressions use
Total Crime as their dependant variable. Z-statistics are given in the parentheses for every table. *
Indicates significance at the 5% level, ** is significant at the 1% level or smaller





























C-
0)




LO











LO











q-




0-


0 1 2 3 4 5 6 7 8 9 10 11 12 13 14
DaysToCapture


Figure 1-1: Average Percent of Total Arrests as Function of Days between Offense and
Capture A) NonStrike Days for Zip Codes Where Strikes Occur, B) Strike
Days for Zip Codes Where Strikes Occur


I I I I I I I I I I I I I I I
0 1 2 3 4 5 6 7 8 9 10 11 12 13 14
DaysToCapture











B











Table 1-10: Effects of Strike Days with Specific Years Dropped from the Sample
Year Dropped Strike Coefficient Year Dropped Strike Coefficient


1984


1985


1986


1987


1988


1989


1990


1991


1992


0.164**
(3.39)**

0.093
(1.70)

0.166**
(3.43)**

0.160**
(3.16)**

0.163**
(3.37)**

0.148**
(3.02)**

0.189**
(3.77)**

0.229**
(3.18)**

0.157**
(3.24)**


1993


1994


1995


1996


1997


1998


1999


2000


2001


0.166**
(3.41)**

0.163**
(3.28)**

0.166**
(3.40)**

0.157**
(3.23)**

0.153**
(3.16)**

0.156**
(3.23)**

0.157**
(3.26)**

0.144**
(2.99)**

0.150**
(3.11)**


Without Seattle 0.133* Seattle Only 0.153
(2.04)* (1.82)


Ave.Number of obs. 358274 358274
Time Fixed Effects Y Y
Zip Fixed Effects Y Y
Ave. R-Squared 0.0603 0.0603
The top half of this table is a regression of Strikes on Total Crime with each year systematically dropped
from the sample. Each regression includes only urban zip codes. All 18 regressions are defined as a Type I
regressions, using data including all ordinary school days from 1984-2001, and zip code fixed effects. The
bottom half of the table reports the effect of strike days, for the same Type I sample, with Seattle dropped
from the sample, and with Seattle making up the entire sample. The bottom of the table shows the average
number of observations for the top 18 regressions, as well as the average R-squared term. Z-statistics are
given in the parentheses for every table. Indicates significance at the 5% level,** is significant at the 1%
level or smaller.










Table 1-11: Average Effect of Randomly Generated Strike Days on Total Crime
Dependant Variable: Total Random Generations of
Crime Strike Days

Ave. Random Strike 0.002
Coefficient (0.05)

Average Number 2,355
of Strike Days

Number of Trials 25

Number of obs. 1,703,910
Time Fixed Effects Y
Zip Fixed Effects N
Ave. R-Squared 0.0380

Table 1-12: Effects of Strike Days by Offender Type with the Last 4 Years of the Sample
Dropped
Strike
Strike Coefficient Stri
Coefficient for
Dependant Variable for One-Time
Repeat
Offenders Repeat
Offenders

Total Crime 0.382** 0.156**
(4.05)** (2.77)**

Total Crime without 0.383** 0.128*
1998-2001 (4.08)** (2.31)*

Number of obs. 129,593 129,593
Time Fixed Effects Y Y
Zip Fixed Effects Y Y

Table 1-11 is the average strike coefficient over 25 trials using a randomly generated strike variable. It
includes the full data set. Table 1-12 is a Type V regression using just the urban subsample, and includes
zip code fixed effects. Z-statistics are given in the parentheses for every table. Indicates significance at
the 5% level, ** is significant at the 1% level or smaller















CHAPTER 2
THE ROLE OF TEACHER NETWORKING IN TEACHER TRANSFER DECISIONS
AND TEACHER MOBILITY

2.1 Introduction

In recent years, there has been growing research to support the notion that teacher

movement is an important topic, and not one to be overlooked. Thus far, research on

teacher sorting behavior has shown that student characteristics seem to influence a

teacher's transfer decision (Hanushek, Kain, Rivkin 2001). More specifically, teachers

sort away from relatively low-income, low-achieving schools. As a result, less

experienced teachers are more typically found in poor, nonwhite, low-performing

schools, particularly those in urban areas (Lankford, Loeb, Wyckoff 2002). This is most

important because teacher experience seems to be directly linked to student achievement

(Hanushek, Kain, Rivkin 1998, 1999)1.

In addition to student characteristics, other characteristics (such as salary

differences) have been studied in an effort to better understand the factors that influence

teacher movement. Research aimed at identifying specific characteristics and conditions

that promote teacher mobility have given more insight into the questions of when and

why teachers move, however researchers still do not have the complete picture on how

teachers move and why.




1 Hanuchek, Kain and Rivkin (1998) show that variations in teacher quality account for 7.5% of the total
variation in student achievement. In addition, teacher quality differences tend to outweigh school quality
differences considerably.









The focus of this paper is to better understand the nature of teacher transfers, with

special attention given to teacher intradistrict transfers from school to school.

Specifically this paper will identify the use of Master Inservice workshops aimed at

professional development as a powerful means of building teacher networks. Ultimately

I test the hypothesis that Inservice workshops give teachers greater opportunities to

network with teachers from other schools in their district, and therefore result in a higher

rate of teacher transfer. The evidence will show that a one standard deviation increase in

workshop attendance contributes to a 12.5% increase in teacher transfer rates per school.

Thus teacher networks are a crucial element in influencing teacher transfer rates.

Section 2.2 of this paper will provide a context for the basic assumptions outlying

the arguments set forth, Section 2.3 will specify the model therein, Section 2.4 reports the

preliminary results, Section 2.5 will check the validity of the results, and Section 2.6 will

make the concluding remarks.

2.2 Discussion of Networking

2.2.1 How Teachers Network

The idea that professionals develop networks is not a new concept in the social

sciences. The term "networking" spans across many professions and takes many

different forms. Like many professionals, teachers also engage in networking for various

reasons I will address later in this section. First I will address the issue of how teachers

build networks. Clearly some teacher networking takes place in private. Teachers may

spend recreational time together getting to know one another, or they may talk over

lunch. However these examples of the networking process, though valid, are not feasibly

quantified or qualified. In order to quantify teacher networks, I rely on teacher Inservice

activity to determine if teacher networks aid in mobility. Specifically I focus on the









number of hours teachers spend in Inservice activity as my measure of network

size/strength. I argue that Master Inservice workshops, a part of the teacher re-

certification process, provide an effective forum for networking. To understand why this

is the case, it is necessary to first review how the re-certification process works.

Every teacher in the state of Florida is required to have an up-to-date license that

enables him or her to teach in Florida public schools. This license is earned by every

teacher at the beginning of his or her career and must be re-certified every five years.

Teachers earn their re-certification through any combination of three things: they may

take state-issued Florida Subject Area Exam in appropriate fields to prove their level of

competency, they may enroll in and complete required coursework (6 credits minimum)

at an approved university or college, and/or they may choose to complete up to 120 hours

of Master Inservice courses/workshops offered by the school district.

This state mandated re-certification process provides an ideal environment for

testing the effects of teacher networking. Teachers who take the Florida Subject Area

Exam in their appropriate fields or who decide to take classes at a College or University

have little to no interaction their colleagues in their school districts. However, the Master

Inservice option is quite different from other re-certification options. First, Inservice

courses/workshops are almost exclusively made up of other teachers, all within the same

district and from differing schools.2 Thus teachers interact with others who share a

common profession, and often common interests. Secondly, the environment of most

workshops is designed in such a way that encourages interaction among participants.


2 It is possible that a marginal number of administrators, county officials, etc. attend these teacher
components for other reasons (professional or personal). They are not included in the Inservice data in this
paper.









Workshop assemblies are long in duration (often several hours per session) and can span

weeks or months. In addition, individuals are often called upon to share experiences with

all the participants, or present material regarding how the workshop has influenced their

classroom activity. Thirdly, Inservice is an attendance based program, as opposed to the

graded college coursework or graded test format. One could argue that the relatively

stress-free nature of the Inservice allows teachers to spend more time getting to know

each other, free from the distraction of grade requirements.

2.2.2 How Teachers Use Their Networks

There are many reasons why any professional or businessperson would network

with others in their field of work. They yearn to share ideas, understand their

competition, and exchange information that can benefit their personal interests,

sometimes in a mutual way. Why teachers network with each other is not so different.

They exchange classroom experiences and teaching methods, information on workplace

environments, as well as build personal bonds that enhance their feelings towards the

profession, or even their personal lives. As a result they could possibly use their networks

to plan intradistrict school events, find work in the summer, or even to socialize.

However, there are specific applications of strong teacher networks that would

enhance mobility within a district. First, teachers use their networks to gain information

about the various student and administrative characteristics of schools in their district.

Greater access to this information lowers the cost of job searching and enhances mobility.

Secondly, teachers use their networks to reveal their teacher quality in a less costly way

for the purpose of being hired at other schools within their own district. If teachers do

use their networks to sort into preferred schools in a more efficient, less costly way, then









there are obvious implications for student outcomes, since others have successfully linked

student outcomes to teacher mobility.

2.3 The Basic Model

A teacher's desire to transfer to school n (new school) is based on the utility they

receive at school n minus the utility they receive at school o (old school). The utility at

the old school is given by:

Vo(so, qo, ao, do, o)

where o indexes the old school, So is the salary of the old school, qo is a measure of

student/classroom quality, ao (administration quality), do (distance from school to home),

and o (opportunity cost of teaching). Similarly the utility at the new school is given by:

Vn(sn, qn, an, dn, o)

where n indexes the new school, sn is the salary of the new school, qn is a measure of

student/classroom quality, an (administration quality), dn (distance from school to home),

and o (opportunity cost of teaching). Teachers incur explicit costs to movement. These

costs arise from the search and application process. More specifically, teachers may have

to submit paperwork, spend some time interviewing, or generally gather information

about the school they are transferring to. Thus I write this teacher decision as:

Transfer when: Vn Vo C(in, rn, nj*) > 0

Do not transfer when: Vn Vo C(i, rn, nj*) < 0

In this case C(.) represents the explicit cost borne by the teacher, where in is the amount

of relevant school information a teacher desires to gather, rn represents district and school

requirements for transfers (paperwork, interviews, etc.), and nj* representing the network









size of the jth individual. I further note the relationship of each parameter to this explicit

cost function.

It is reasonable to assume that the less a teacher knows about a prospective school,

the more time he/she spends gathering relevant information. Therefore we expect an

increase in necessary or relevant information (in) to translate to a direct increase in C(.),

such that 8C/8i, > 0. In addition an increase in district and/or school requirements for

transfer will increase time and resources spent towards transfer, such that CC/Or, > 0.

When looking at network size it becomes apparent that a larger network size will lend

itself to better flow of school information, possibly fewer or shorter interviews, and more

inside information that helps one to eliminate unnecessary measures. Therefore it seems

logical that a larger network will decrease external cost, such that 8C/cnj* < 0. Lastly we

must qualify the network to external cost relationship as not being strict. This allows for

the possibility that for school n, one's network is totally ineffective. That is to say that

one's network is not the relevant network, i.e., a network in another state. From these

equations it is clear why teachers may be more likely to transfer as their relevant network

size increases

2.4 Data and Empirical Study

My data for my study comes from a variety of sources. I utilize individual school

data taken from the "Florida School Indicators Report" (FSIR) and the "School Advisory

Council Report" (SACR). Countywide data was also supplied in various forms from the

Florida State Department of Education (FLDOE). These forms include the "FLDOE

State Survey #2", "Teacher Exit Interview Information", an "Inservice Hours Report",

"The Profiles of Florida District Schools (Students and Staff Data)", and "Teacher Salary,









Experience and Degree Level". The sample includes 67 observations, one for each

Florida school district.

Because several variables introduced in the following section are highly collinear to

district size, such variables have been defined on a per school basis.3

2.4.1 Defining the Parameters

The dependent variable for my study (Transfers) is the number of transfers within

a district per school. In order to approximate the closeness of different school job

options, or capture the density of school job choice in a district, I use the total area,

expressed in sq. miles, divided by the number of schools in the district (School

Distance). Since a denser school district gives a mobile teacher more viable, less costly

job options, I expect this variable to be positively correlated with the total number of

transfers per school.

Student/classroom quality is measured by taking the standard deviation of the

percentage of free lunch eligible students among all schools in a given district. This

measure (School Difference) should capture the variation of the percentage of the student

body that is of low socio-economic status (SES) across schools. In essence, this variable

measures how schools in a district differ along family backgrounds. Since greater

variation in work environments should give teachers a greater incentive to sort, this

variable should be positively correlated with transfers per school.

Wage differences may also play an important role in teacher transfer decisions.

There are no wage differences among public schools in a given district, but wage

differences among districts still exist and are important to account for. As such, Salary


3Transfers, Network, Dislike Supervisor, Principal Moves, School Changes.









Difference is defined as the average salary of the highest paying district directly adjacent

to the current district minus the average salary of the district in which a teacher is

currently employed. If a teacher considers salary difference in a move between districts,

it is plausible that they may consider the first-best alternative salary difference above all

else. If higher salaries provide incentives for teachers to leave their district rather than

move within district, then there should be less movement between schools within district.

Therefore this variable should be negatively correlated with within-district movement.

In addition to the interdistrict salary difference, a wage concern for teachers should

also be the wage forgone from remaining a teacher. To capture the opportunity wage

forgone, I construct a variable (Wage Ratio) using a county-level wage index. This

wage index reflects the average wages across a predetermined set of occupations for each

county. By dividing average salary for a district by its county wage index, I measure how

well teachers in a county are paid relative to other occupations. The larger the Wage

Ratio, the less appealing otherjobs are. As a result this variable should be positively

correlated with teacher transfers.

My measure of network size (Network) equals total hours spent in Inservice per

school.4 Since more hours spent in Inservice should lead to larger and more effective

networks, this variable should be positively correlated with teacher transfers per school.

It is also true that the Network variable is positively correlated with larger school

size. This is a potential problem since school size, measured in average number of

teacher faculty per school, is positively correlated with district size itself. That is to say



4Component hours geared towards non-instructional staff (i.e., food service employees, transportation staff,
administrative staff, substitute teachers, etc.) are not included in this measurement. Since each district
adheres to a statewide numbering standard for component identification, I can be sure that I only include
components which were teacher-oriented.









that as school districts grow, the size of the schools as measured in total faculty also

grows. Since teachers in larger districts (with larger schools) have more positions to

choose from, they should have greater transfer opportunities. To prevent this relationship

from biasing any estimates of the Network variable, I create a variable to control for the

effect of larger school size. This variable, School Size, measures the average teaching

staff per school in the a district.

New public schools opening and old public schools closing also present teachers

with new choices in making transfer decisions. New schools that open in a district may

provide up-to-date facilities and technology in the classroom, making teaching there more

desirable. School closures force teachers to reorganize into other schools or leave

teaching entirely. The variable (School Changes) captures these elements of school

turnover. This variable is measured by assigning a value of positive one for every school

gained and/or lost from the 2000-01 school year to the 2001-02 school year, and then

summing the totals for each county and dividing by the number of schools. Since school

districts where schools open and/or close should result in more teacher transfers, this

variable should be positively correlated with transfers.

There is another variable emphasized in this model that is based on administrative

movement, specifically movement between schools by existing principals. When

principals transfer from school to school within a district, positive teacher/principal

relationships may induce principals to recruit teachers to work at a new assignment. In

essence principals may bring their favorite teachers with them to another school under the

expectation/promise of a positive work environment. This variable (Principal Moves) is

measured as the total number of principals who moved from one school to another









school, divided by the number of schools. If principals take teachers with them when

they move between schools, then principal transfers should be positively correlated with

teacher transfers.

Another variable that should affect teacher movement within a district is the

number of relevant alternatives. It is arguably difficult for teachers to move between

high schools and elementary/middle schools. Obviously high school teachers who work

in districts where there is only one high school cannot move to another high school in

their district. As a result, one would expect that school districts with only one high

school will have an inherently lower rate of teacher transfer compared to similar districts

with multiple high schools. To account for this disparity I include a dummy variable

(One HS) that receives a value of 1 when a school district has only one high school (zero

otherwise). This variable should be negatively correlated with teacher transfers.

The last two variables I include in the regression help to describe the state of a

district's preferential treatment towards transferring existing teachers to fill new

vacancies when they open, rather than hiring new teachers. Some districts require that

open positions are filled with teachers already employed before a new teacher may be

hired; while some other districts give no special consideration to existing teachers over

new applicants. Since more relaxed or favorable policies towards teacher transfers may

enhance teacher mobility, I introduce two dummy variables that describe the priority

given to teacher transfers. The first variable (Transfer PC) describes a district that

affords partial consideration to existing teachers. This variable takes on a value of 1

when a district gives preferential treatment, but not exclusive privilege, to its teacher

transfers (zero otherwise). The second dummy variable (Transfer CC) describes a









district's policy to necessarily fill a new position with a teacher transfer request before

considering a new applicant. This variable takes on a value of 1 when a district offers

complete consideration to existing teachers before all others (zero otherwise). In both

cases, these variables should be positively correlated with transfers per school to reflect

the enhanced opportunity for mobility given by preference.

2.4.2 Empirical Testing

The analysis begins with a simple OLS regression onto the dependent variable

Transfers. The results are expressed in Table 2-1, listed by coefficient with the relevant

t-statistic listed in parenthesis.

Columns 1 and 2 express the OLS regression with our 10 major variables of

interest. Some of these variables that do not have the predicted sign, in this case School

Distance, Transfer PC and Transfer CC, and are not significantly different from zero.

Of the variables that are significant, (Network, School Difference, School Changes and

School Size), each has the predicted sign.

The positive and significant coefficient of the Network variable in Column 2

seems to suggest that networks exist, and that in fact they do serve to enhance teacher

intradistrict mobility. As a teacher networks grows, teacher mobility rises. This provides

evidence that teachers are using their networks to gain information and enhance the

transfer process. To determine the magnitude of the impact of networking on transfer

rates, I look at the percent change in the average number of teacher transfers per school

when the amount of networking increases by one standard deviation. I find that an

increase in the number of inservice hours per school by approximately 1000 hours

increases the teacher transfer rate per school by approximately 11.49%. These results can

be viewed in Table 2-2.









The results in Table 2-1 also suggest that student characteristics, school size, and

schools opening and closing all influence teacher movement. Teachers seem to sort more

frequently in districts where variation in student quality across schools is high. The

variable School Difference has a significant impact in teacher movement and shows that

teachers are more likely to move around, or sort, when there are sizeable differences

among student populations at different schools. According to the regression, a one

standard deviation increase in the measure of school heterogeneity implies a nearly

15.25% increase in teacher transfers per school

Teachers also seem to be more mobile in districts where schools are larger. The

School Size variable is positive and significant, suggesting teachers face greater

opportunity to sort when there are more total positions in a school. The coefficient of the

School Size variable implies that a one standard deviation increase in the average size of

a school leads to a roughly 17.85% increase in transfer rates per school.

Besides student and school characteristics, teachers also have greater mobility in

districts where schools are opening and closing with greater frequency. The School

Changes variable suggests that as schools open and close in a district, teacher mobility

rises. One would expect that as schools open in a district, teachers seek new facilities,

new classrooms, etc. Further, as schools close teachers are forced to shuffle and sort to

other schools so that they may continue teaching in their district. The evidence suggests

that when 5 schools are added in a district with 100 schools, teacher transfers increase by

nearly 30% for every school in the district.

Column 2 runs the same regression using robust standard errors. Using these

robust standard errors helps to ease worries about heteroskedasticity in the regression. If









it were true that smaller districts have smaller measurement error than larger districts

(which seems plausible), then robust standard errors help to control for this problem. We

see roughly the same results as the regression in Column 1. School Size and School

Difference have the predicted sign and are significant at the 1% level. School Changes

also has the predicted sign and is significant at the 0.05% level.

Column 3 reports the same regression as in Column 1 except that one additional

variable is included. This variable (Dislike Supervisor) is a rough measure of teacher

dissatisfaction with administrative oversight in one's own school. It is calculated for

each district as the number of teachers per school who cited "dissatisfaction with

supervisor" as a reason for voluntarily terminating their employment on the Teacher Exit

Interview from the 2000-2001 school year.5 If teacher discontent with administration

leads to greater mobility, then this variable should be positively correlated with teacher

transfers.

In Column 3, the variable Dislike Supervisor has the predicted sign but is not

significant. It suggests that job environments and teacher attitudes are not necessarily an

important part of the movement decision. This result seems to defy intuition, however in

the next section I will show later that this result is marginal, and that Dislike Supervisor

is sensitive to model specification. Column 4 reports the same regression as in Column 3

using robust standard errors. It is also worth noting that the Dislike Supervisor variable

is not an ideal measure of overall teacher attitudes towards administration because it does

not include teachers who continued to teach in their own districts and only reflects a self-


5 This survey includes teachers who continued to teach in other districts or other states, teachers who stayed
employed in education departments in and outside of their district, teachers who left the education
profession altogether, those who retired, and teachers who left for private schools in and out of their
district.









selecting subgroup of teachers. It does however have some benefits. First, since there is

no descriptive teacher evaluation of administration formally collected by the FLDOE or

school districts, nor is there useable county data of the same nature, this variable offers at

least some approximation of what that data might yield. Secondly, because the teachers

in these interviews have already committed themselves to alternative employment, the

responses given in these interviews are assured to be candid and honest. It is unlikely

that a teacher respondent fears having their identity becoming known to others in these

interviews.

Overall, these results seem at least superficially consistent with Hanushek, Kain

and Rivkin findings (1999, 2001) that student characteristics influence mobility, and that

salary changes at best have only a modest impact. However none of these variables seem

to have a powerful an impact as the School Changes variable. There seems to be a large

disparity between the predictive power of this variable, and other variables. To ensure

that our results are valid, and to understand why each variable induces the kind of change

it does, I check the robustness of the data.

2.5 Robustness Checks

2.5.1 Re-examining the Data

With respect to the School Changes variable, one observation in the dataset

experienced an abnormally high rate of school change between the years of 2000-2001

and 2001-2002. This county in particular closed two school facilities to open one, much

larger school, that would incorporate all of the existing student population of the other

schools, however the small size of this district seems to exacerbate the measurement of

the School Changes variable. It is possible that this one observation alone is driving the

results of the School Changes variable. If this observation is more of an outlier, then it









may be overstating the overall importance of the variable it influences. An OLS

regression analysis without this outlier is reported in Table 2-3.

Once the sample is modified to eliminate the possible outlier, the magnitude of the

coefficient of the School Changes variable is reduced by almost half, though the variable

remains significant. This does not necessarily mean that this variable is not important,

however it does show that perhaps this outlier should be dropped to maintain the integrity

of these results. These new results suggest that a one standard deviation increase in

school openings/closings leads to a 14.55% increase in teacher transfers per school. This

change is considerably lower than the previous estimate.

All other variables in the regression have the predicted sign and all of the formerly

significant variables are still significant. Network and School Size are significant at the

1% level, while School Difference becomes marginally significant at the 10% level.

Column 2 reports the same regression with robust standard errors. Column 3 includes the

Dislike Supervisor variable, which is significant at the 10% level. This result suggests

that there may be some value to job environment and administrative quality that teachers

consider in their movement decisions. Finally Column 4 presents the same regression as

in Column 3, including robust standard errors. I again assess the impact of each variable

independently, by determining the effect of an increase of one standard deviation of each

variable. These results can be viewed in Table 2-4. With the exception of School

Changes, there are no substantial differences in the magnitudes of the coefficients of the

other variables.

The key variable of interest, Network, is positive and significant in each regression

specification on Table 2-3, suggesting that networks are an important factor contributing









to teacher movement. However, to confirm that the results are credible, I further test the

robustness of these results. Specifically I address whether the variable itself is

endogenous to the system. In the following section I test whether the measurement of

networking is correlated with some unobserved teacher characteristicss.

2.5.2 Unobserved Correlation

If my measure of teacher networks is correlated with some unobserved teacher

characteristicss, then it may be true that there is no real networking taking place.

Instead, teachers who complete high levels of Inservice may simply be more likely to

transfer for other reasons. If that were true then one should expect that teachers who

complete Inservice would also be just as likely to move in ways other than just

intradistrict transfers. To test whether the Network variable is reflecting true networking

or is just reflecting self-selection of highly mobile teachers into Inservice, I utilize new

dependent variables that depict other types of teacher movement.

Private Move equals the number of teachers who claim a move into private

schools per school. The second variable, District Move, measures the number of

teachers moving into other Florida districts per school. The last variable is teacher

movement to other states to teach. This variable, State Move, equals the number of

teachers moving to other states per school in the district. Regression results with these

new dependent variables are reported in Table 2-5 using robust standard errors.

Columns 1 and 2 report the regressions with Private Move as the new dependent

variable, and using robust standard errors. Columns 3 and 4 do the same with District

Move as the new dependent variable, and Columns 5 and 6 have State Move as the new

dependent variable. In each of these specifications, network size does not seem to be a

significant factor influencing teacher movement. Since Inservice is not significant to any









of these different kinds of movement, it seems plausible that teachers who tend to move

around more generally, are not self-select into attending a greater amount of Inservice.

Thus I can be more confident that the positive effect of networking is not simply a

reflection of the preferences of more mobile teachers.

2.6 Conclusion

The evidence set forth in this paper supports the argument that Master Inservice

components provide an effective environment for teachers to network with each other.

These networks seem to provide a basis for teachers to gather important information

about schools and allow teachers access to transfer options that might have previously

been closed to them. Overall, after controlling for determinants of mobility and

eliminating biasing outlier effects, I show that an increase of one standard deviation in

Inservice hours causes roughly a 12.74% increase in teacher transfers per school. The

results suggest that teachers exploit these networks to sort into schools that they find

desirable. If teachers can more effectively sort away from less desirable schools, such as

low-income or failing schools, into more desirable schools, then it is easy to conceive that

worse schools will be worse off as teacher networks improve.

Of course the goal of the Inservice plans set forth by school districts is to improve

teacher quality. However, the unintended consequences of such staff development

policies that encourage, or at least facilitate, teacher networking seems to exist as well.

Further, the consequence of this particular policy seems to be quite substantial in

promoting a potentially harmful action such as teacher transfer. This result may have

considerable meaning for the future of public education policy. State officials must begin

to weigh the possible negative effects of increased teacher networking with the potential






59


(and realized) benefits of Inservice, so that such policy analysis can be conducted in the

future with a complete understanding of social welfare consequence.












Table 2-1: Ordinary Least Squares Regression onto teacher transfers using full sample
with and without robust standard errors


Variable


Dependent

Network


School Distance


Principal Moves


School Changes


School Difference


Salary Difference


Wage Ratio


School Size


Dislike Supervisor


One HS


Transfer Policy
PC

Transfer Policy
CC


Constant


Column 1

Transfers

0.0003261
(1.56)

0.0065532
(1.12)

0.3216886
(0.12)

17.86031
(4.88)**

0.0584747
(2.10)**

-0.0000891
(1.61)

-0.0018573
(0.27)

0.0501213
(2.46)**


-0.7262709
(1.32)

-0.2608151
(0.65)

-0.2381146
(0.57)

0.2509715
(0.09)


Column 2

Transfers

0.0003261
(2.00)**

0.0065532
(1.16)

0.3216886
(0.12)

17.86031
(3.42)**

0.0584747
(2.34)**

-0.0000891
(1.50)

-0.0018573
(0.22)

0.0501213
(2.80)**


-0.7262709
(1.27)

-0.2608151
(0.64)

-0.2381146
(0.67)

0.2509715
(0.08)


Column 3

Transfers

0.0003474
(1.67)

0.0078015
(1.33)

0.9270341
(0.34)

17.55322
(4.84)**

0.0557304
(2.01)**

-0.0000766
(1.38)

0.0001112
(0.02)

0.0448708
(2.19)**

4.17856
(1.43)

-0.7582897
(1.39)

-0.2026472
(0.51)

-0.1001645
(0.24)

-0.5541842
(0.20)


Column 4

Transfers

0.0003474
(2.19)**

0.0078015
(1.33)

0.9270341
(0.33)

17.55322
(3.20)**

0.0557304
(2.16)**

-0.0000766
(1.21)

0.0001112
(0.02)

0.0448708
(2.96)**

4.17856
(1.17)

-0.7582897
(1.31)

-0.2026472
(0.48)

-0.1001645
(0.25)

-0.5541842
(0.19)


R-Squared 0.6154 0.6154 0.6295 0.6295
Number of Obs. 67 67 67 67
Robust SE N Y N Y
denotes significance at the 10% level
**denotes significance at the 5% level






61


Table 2-2: Percent change teacher transfers per school for a one standard deviation
increase in each significant variable with full sample
Variable Mean Standard Deviation Range of % Change
School Changes 0.027 0.05 30.06 to 30.58 %
School Size 33.418 11.61 17.85 to 19.94 %
School Difference 15.945 7.99 15.25 to 16.00 %
Network 1758.317 1047.62 11.49 to 12.46 %













Table 2-3: Ordinary Least Squares Regression onto teacher transfers using sample with


biasin outlier omitted
Variable Column 1


Dependent

Network


School Distance


Principal Moves


School Changes


School Difference


Salary Difference


Wage Ratio


School Size


Dislike Supervisor


One HS


Transfer Policy PC


Transfer Policy CC


Constant

R-Squared
Number of Obs.
Robust SE


Transfers
0.0003563
(1.79)*

0.0010389
(0.18)

0.3430332
(0.13)

9.11579
(1.90)*

0.0433008
(1.60)

-0.0000671
(1.26)

-0.0000751
(0.01)

0.0502127
(2.60)**




-0.4594238
(0.86)

-0.2247365
(0.59)

-0.3207089
(0.81)

0.0560173
(0.02)

0.6033
66
N


Column 2
Transfers
0.0003563
(2.22)**

0.0010389
(0.20)

0.3430332
(0.12)

9.11579
(1.91)*

0.0433008
(1.74)*

-0.0000671
(1.06)

-0.0000751
(0.01)

0.0502127
(2.89)***




-0.4594238
(0.94)

-0.2247365
(0.55)

-0.3207089
(0.89)

0.0560173
(0.02)

0.6033
66
Y


Column 3
Transfers
0.0003822
(1.95)*

0.0021892
(0.37)

1.147493
(0.46)

8.322631
(1.76)*

0.0393933
(1.48)

-0.0000516
(0.97)

0.0020135
(0.30)

0.0442077
(2.29)**

4.78275
(1.74)*

-0.4825945
(0.92)

-0.1563356
(0.41)

-0.1669839
(0.42)

-0.8754046
(0.33)

0.6247
66
N


Column 4
Transfers
0.0003822
(2.46)**

0.0021892
(0.41)

1.147493
(0.36)

8.322631
(1.73)*

0.0393933
(1.55)

-0.0000516
(0.79)

0.0020135
(0.30)

0.0442077
(2.95)***

4.78275
(1.52)

-0.4825945
(0.96)

-0.1563356
(0.37)

-0.1669839
(0.42)

-0.8754046
(0.33)

0.6247
66
Y






63


Table 2-4: Outlier excluded percent change teacher transfers per school for a one
standard deviation increase in each significant variable
Variable Mean Standard Deviation Range of % Change
School Size 33.696 11.48 17.40 to 19.73 %
Network 1777.904 1043.22 12.73 to 13.65 %
School Difference 15.881 8.03 10.84 to 11.91 %
School Changes 0.023 0.04 0.37 to 11.36 %
Dislike Supervisor 0.027 0.05 8.78 %







64


Table 2-5: Ordinary Least Squares Regression of other types of teacher movements with
outlier excluded


Variable


Dependent


Network


School Distance


Principal Moves


School Changes


School
Difference


Salary Difference


Wage Ratio


School Size


Dislike
Supervisor


One HS


Transfer Policy
PC

Transfer Policy
CC


Constant

R-Squared

Number of Obs.

Robust SE


Column 1 Column 2 Column 3 Column 4


Private
Move


Private District District
Move Move Move


Column 5 Column 6

State Move State Move


0.0000101 0.0000104 0.0000342 0.0000305 0.0000014 0.0000034
(1.21) (1.22) (0.49) (0.43) (0.06) (0.01)

-0.000214 -0.000196 0.0024853 0.0021902 -0.000891 -0.000813
(1.09) (1.02) (0.95) (0.85) (1.36) (1.22)

-0.001503 0.0073589 -1.829443 -1.930217 -0.190106 -0.142873
(0.02) (0.07) (2.09)** (2.01)** (0.53) (0.39)

0.125987 0.1214916 4.121108 4.236288 0.6497598 0.5957743
(1.11) (1.07) (1.44) (1.46) (0.90) (0.78)

-0.001018 -0.001058 0.0119132 0.0124806 0.0015131 0.0012472
(0.73) (0.74) (0.99) (0.99) (0.44) (0.35)

-0.000001 -0.000001 -0.000001 -0.000003 -0.000002 -0.000001
(0.43) (0.31) (0.06) (0.18) (0.37) (0.20)

0.0002086 0.0002341 -0.006608 -0.006912 -0.002273 -0.002131
(0.96) (1.02) (2.47)** (2.35)** (2.41)** (2.18)**

-0.000328 -0.000405 0.0019503 0.0028223 0.0017899 0.0013812
(0.42) (0.51) (0.32) (0.47) (0.76) (0.58)


0.0611693
(0.66)


-0.694535
(0.70)


0.3255332
(1.10)


-0.014332 -0.014801 0.2299846 0.2333494 0.0774991 0.0759221
(1.04) (1.08) (0.87) (0.88) (0.95) (0.91)

0.0199595 0.0197389 -0.111869 -0.121802 -0.007048 -0.002392
(1.06) (1.05) (0.73) (0.81) (0.17) (0.06)

0.0022509 0.0042704 -0.174603 -0.196926 -0.050714 -0.040251
(0.12) (0.21) (1.20) (1.29) (1.04) (0.78)


-0.041984
(0.50)

0.1931
66


-0.053770
(0.59)

0.1977
66


2.621647
(2.52)**

0.3416
66

Y
N


2.756905
(2.38)**

0.3461
66


1.011499
(2.84)**

0.2398
66


0.9481022
(2.54)**

0.2511
66


N N


Outlier














CHAPTER 3
WITH A LITTLE HELP FROM MY FRIENDS: EVIDENCE OF TEACHER
NETWORKS USING MICRO DATA

3.1 Introduction

In the last several years, the issue of teacher mobility has come to the foreground of

discussions about the public education. Parents and families who value teacher quality

have come to realize that teacher mobility is a topic that has potentially far reaching

consequences, not just for education labor markets, but also in terms of levels of

education and student achievement. Schools that struggle with teacher retention often

find job vacancies difficult to fill, and may compensate by hiring less than fully qualified

teachers, expanding class sizes, canceling course offerings and assigning teachers from

other subject areas (NCES 1997). Of course these kinds of actions may have adverse

effects on student learning and achievement. As a result, economists and policy makers

have done more to explore the issues of teacher attrition and movement within school

districts. The goals of such recent studies have been to shed light on the questions of

which teachers are the most mobile, how do these teachers move, when do these

movements take place, and what are the potential implications for students.

The most recent studies focused on teacher transfer behavior and sorting within

education have provided evidence that supports the notion that student characteristics and

school quality are important factors in a teacher's transfer decision (Hanushek, Kain, and

Rivkin 2001). This also seems consistent with teacher self-reported information. In the

2000-2001 Teacher Follow-up Survey, administered by the National Center of Education









Statistics, 32% of teachers reported poor workplace conditions as a primary reason for the

movement from one school to another. Research shows that teachers sort away from

low-income, low-achieving schools. Consequently, less experienced teachers are more

typically found in poor, nonwhite, low-performing schools, particularly those in urban

areas (Lankford, Loeb, and Wyckoff 2002).

For policymakers, this can be a troubling result. If more experienced teachers tend

to sort away from poorly performing schools, then it is the students at these schools who

stand to lose the most, because experienced teachers have been found to be more

effective than novice teachers in terms of higher student achievement (Hanushek, Kain,

and Rivkin 1998, 1999).1 In addition, urban schools and highly urban school districts

face high rates of teacher turnover (Imazeki 2003). If teachers are sorting away from

these areas, then these urban districts may have the most difficult time filling vacancies

with replacement staff.

Aside from student characteristics, economists have also studied the potential

effects of salary changes on teacher mobility. So far researchers have found that salary

levels have only a modest impact in terms of student achievement and almost no impact

on teacher mobility (Hanushek, Kain, and Rivkin 1999). However, others have also

found that salary can influence a teacher's movement choice in ways other than strict

transfer, such as length of stay in teaching (Murnane and Olsen 1990). Although much

has been learned about teacher mobility, researchers still do not have the complete picture

on how teachers move and why.



1 Hanushek, Kain and Rivkin (1998) show that variations in teacher quality account for 7.5% of the total
variation in student achievement. In addition, teacher quality differences tend to outweigh school quality
differences considerably.









The focus of this paper is to better understand the nature of teacher transfers from

school to school by introducing the idea of teacher networks as an effective sorting

mechanism for public school teachers. To study this issue I rely on information provided

by the 1999-2000 Schools and Staffing Survey (SASS) put out by the National Center for

Education Statistics. The individually detailed and specific nature of the dataset not only

allows me to consider how individual networks affect mobility but also allows me to

control for a well-defined set of variables, giving my study a more complete and detailed

look at networks. This paper will begin discussion on the notion of teacher networks, and

will address how these networks influence teacher movement and to what extent.

Specifically this paper describes how professional development activities (PDAs) provide

opportunity for teachers to network with other teachers. PDAs will be shown to have a

sizable impact on intradistrict teacher mobility.

3.2 The Identification Strategy

Before I describe how I identify teacher networks, let me first discuss what teacher

networks are, how they are useful, and how teachers build these networks.

3.2.1 What Are Networks And How Are They Useful

The term network in this paper is defined as group or "family" of co-workers or

colleagues within one's field of work. Specifically, in this paper a network refers to a

teacher's associations with other teachers who work within their own district. Although

the context of a network here is limited to the field of education, people from all

professions form networks for various reasons.

There are many reasons why any professional or businessperson would network

with others in their field of work. They yearn to share ideas, understand their

competition, and exchange information that can benefit their personal interests,









sometimes in a mutual way. Why teachers network with each other is not so different.

They exchange classroom experiences and teaching methods as well as information on

workplace environments, as well as build personal bonds that enhance their feelings

towards the profession, or even their personal lives. As a result, teachers may be able to

use their networks to plan intradistrict school events more efficiently, find work in the

summer time, or sometimes just to socialize with others in their field.

However, there are two key uses of teacher networks that are the focus of this

paper. First I argue that teachers use their networks to gain greater information about

school characteristics. Because there are many characteristics about schools that may be

difficult to observe (existing job vacancies, quality of administrative support,

departmental environments, etc.), networks may allow teachers to evaluate other schools

in their district more efficiently. Greater information about schools should allow teachers

to sort into their preferred school more easily.

Secondly, I argue that teachers use their networks to reveal their quality as a job

candidate in a less costly way. Because it can be costly (or sometimes impossible) to

reveal one's aptitude for a job vacancy, teachers may call on their networks to provide

important or relevant information to potential employers. For example, a teacher may

ask another teacher to put in a favorable word to a hiring principal. In this way, networks

provide a valuable advantage to those seeking employment, while at the same time being

a more credible or reliable source of information for potential employers. If teachers do

use their networks to foster more efficient, less costly sorting behavior, then policies that

encourage network building may have unintended outcomes for teacher mobility and

possibly student achievement.









3.2.2 How Do Teachers Build Networks?

Clearly some teacher networking takes place in private forums. Teachers may

spend recreational time together getting to know one another, or they may talk over

lunch. However these examples of the networking process are not feasibly quantified.

To find a more measurable environment where networking takes place, I focus on

professional development activities. Professional Development Activities (PDAs) consist

of an array of different activities geared towards maintaining and enhancing teacher

competency and knowledge of various issues (including such issues as the use of

technology in the classroom, student assessment and state education standards, teaching

methods, etc.).

PDAs can include many different types of activities such as workshops, mentoring

and/or peer observation, collaborative research, University coursework, etc. Table 3-1

reports a summary of self-reported professional development activity from the 1999-2000

Schools and Staffing Survey.2 Though each of these activities may be available to most

teachers in the sample, attending workshops/conferences/training clearly seems to be the

most popular form of PDA in the sample, with nearly 94% of teachers reporting they

have engaged in this form of PDA in the past year (prior to the 1999-2000 school year).

PDAs provide an ideal environment for teachers to network with one another.

Teachers who participate in workshops or conferences have greater opportunity to meet

teachers from other schools within their district and enhance their networks. In addition

most PDA is geared to a specific employment group within the school system. This



2 The 1999-2000 Schools and Staffing Survey (SASS) is a survey of involving approximately 56,000 public
school teachers from all over the U.S., and is produced by the National Center of Education Statistics. It is
the main data source for this chapter and is described further in Section 3.3.









means that administrative personnel do not usually engage in the same PDA as teachers.

Likewise teachers do not usually engage in the same PDA as other support staff (such as

bus drivers, cafeteria workers, etc.). Thus I can be somewhat confident that if teachers

are building networks through PDA, they are generally relevant networks.

Besides simply meeting other teachers through PDA, these activities often provide

an environment conducive to promoting teacher interaction. Workshops especially are

designed in such a way that encourages interaction among participants. Workshop

assemblies are usually long in duration (often several hours per session) and can span

weeks or months. In addition, individuals are often called upon to share experiences with

all the participants, or present material regarding how the workshop has influenced their

classroom activity.

3.2.3 Measuring Teacher Networks

In order to measure the size of a teacher's network, I utilize national survey data

that provides information on professional development activity. The 1999-2000 Schools

and Staffing Survey (SASS), the survey from which the overall dataset is derived, asks

respondents to report the number of hours they spent in each of six major categories3 of

PDA within the last school year (1998-99). Rather than report the specific number of

hours, respondents are asked to select a (predetermined) range of hours that includes the

actual number of hours they have spent in each category of PDA. Using the midpoint of

the reported range as my best estimate of reported hours, I create a continuous variable of





3 These six categories are: PDAs with focus on 1) in-depth study of main assignment, 2) content and
performance standards of main assignment, 3) methods of teaching, 4) uses of computers for instruction,
5) student assessment/methods of testing, and 6) student management in the classroom.









PDA Hours (Hours), by summing the reported hours across all six PDA categories for

each individual.4

In general the distribution of PDA Hours across my sample is skewed to the left

with approximately 3.1 % of all teachers reporting that they have not engaged in any

form of PDA within the past year. Both the level of skewness of the sample

(approximately 1.30) and percent of teachers reporting no PDA within the past year are

roughly equivalent across movement categories as well.5 In addition, the distribution of

PDA Hours across these movement categories is also roughly identical. Of those

teachers who did not move, 24.8 % (or 332) are in the upper quartile of the entire

distribution. Likewise 25.2 % of teachers (or 88) who transferred within district were in

the upper quartile of the entire distribution of PDA Hours.6 It is also worth noting that in

each individual PDA category, only a small fraction of teachers are top-coded (reporting

the maximum number of hours allowed in the survey).7 Although for each PDA category

the percent of teachers who are top-coded ranges from 16.4 % to 2.1 %, these differences

do not vary significantly across movement types. Overall the data seems to indicate that





4 It is important to note that the highest range of hours that teachers could report is "33 or more". Where
this happens, I use 60 hours as the maximum value in this range. 60 seems to be a reasonable upper limit
of hours, however I have also tested maximum values at 48 and 64. These changes do not significantly
impact the results.

5 The PDA distribution of teachers who move intradistrict is slightly more skewed (1.37) than the same
distribution for those who did not move (1.25). Additionally, the kurtosis of the PDA distribution for all
movement categories is nearly equivalent (4.34 for non-movers, 4.64 for intradistrict movers, and 4.61 for
those who leave their district).

6 Of those teachers who moved out of their district, 21.7 % (or 98) are in the upper quartile of the entire
distribution. This includes both interdistrict and interstate movement.

7 There is only one teacher in my sample who is top-coded for every category of PDA (and has thus taken
the maximum amount of PDA measured in the survey)









along the PDA measurement, teachers in various movement categories do not look

different enough from one another to warrant any concern these groups are incomparable.

3.2.4 The Basic Model

In this basic model I am attempting to explain how a teacher's decision to transfer

to another school (in year t) within their own district is affected by their network. My

measure of a teacher network here is the number of hours a teacher spent in professional

development activities in the previous year (in year t -1). I can start by expressing a

simple regression model in the following form:

Transfer) = a + P(PDA Hours(t 1)) + 6X

Here the dependent variable is a dummy that expresses a teacher's transfer within

his own district. The dependent variable (Transfer) is assigned a value of 1 for a within

district move, and 0 otherwise. In addition to the amount of networking a teacher does,

the basic characteristics of the most mobile teachers are somewhat different from those

who are less mobile. To capture these differences I include variable X, a vector of

individual and school characteristics that influence a transfer decision. Table 3-2 reports

some descriptive statistics about the characteristics of teachers who do and do not move

around in various ways. The most notable difference is that teachers who move within

district tend on average to be about 3-6 years less experienced (and younger) then those

who do not move. Since inexperienced teachers are generally not as far along the career

path as more experienced teachers, this difference seems intuitive. In addition, I must

also consider marital status and job status in the model. Those teachers who are not

married may find a move to a new school to be more costly since spousal relocation can

complicate matters. Those teachers who are employed part-time by their district should









have a greater overall benefit to searching for job vacancies since they stand to benefit

from becoming a full-time employee.

Aside from the natural differences that exist among teachers, a teacher's transfer

decision should also depend on the working conditions he/she faced the previous year.

Specifically, a teacher's decision to transfer may be influenced by student

quality/characteristics, administrative quality/characteristics, or compensation levels. As

a measure of student quality (L), I construct a measure:

%Free Lunch Eligible = (%free-lunch eligibleshool,(t -1) %free-lunch eligibledistrict,(t 1))

This measure is the difference between the percent of the student population who

are free-lunch eligible at a teacher's school and the average percent of the student

population who are free-lunch eligible at the district level. It is designed to capture the

socio-economic status (SES) of the student population at a teacher's initial school relative

to the SES of the district as a whole. Given that teachers desire to sort away from low

SES schools, this variable should have a positive impact on transfers.

To measure administrative quality (Adm), I utilize the variable:

Poor Administration = a dummy variable, receives a value of 1 is a teacher reports
that administrative support is poor within the past year (t 1)

This variable captures a teacher's perception of how supportive/encouraging the

administration at their school is towards its staff. The predicted sign of this variable

should be positive, since in schools where administrative quality is perceived to be poor,

teacher mobility should be higher.8 Some additional measures of the working conditions

in my regressions specification include:


8 I have also measured poor administrative quality using the average teacher response for each school. My
results do not change with this different measure.









Student Threat = a dummy variable, equals 1 if a teacher reports being threatened
by a student within the past year (t 1) (Thr).

Again because receiving a student threat indicates poor working conditions, the sign of

the coefficient should be positive to express teachers' preferences to sort away from

undesirable schools.9

Bonus Pay = a dummy variable, equals 1 if a teacher reports receiving bonus pay,
separate from salary and extracurricular pay within the past year (BP).

Bonus pay should give teachers an incentive to stay at their existing school, and should

therefore be negatively correlated with transfer rates.

I also use the self-reported data to construct a measure of job satisfaction (Job) that

reflects a teacher's general attitude toward their workplace:

Job Dissatisfaction = a dummy variable, equals 1 if a teacher reports being strongly
dissatisfied with being a teacher at their school in year (t 1) (Job)

This variable is intended to capture not just the quality of the students, but also a

teacher's perception about the quality of their co-workers, their administrative superiors,

school resource availability, and general issues surrounding their employment. Since

greater general job dissatisfaction should lead to greater mobility, I expect the sign of this

variable to be positive.

While it is true that salary differences between schools seem to affect teacher

transfer decisions, this analysis is limited to those teachers whose movement is within

their district. Given that salaries schedules are generally set at the district level and do





9 I have also measured student threat using the average teacher response for each school. My results do not
change with this different measure.









not vary across schools in the same district, base salary differences between schools are

negligible10. As a result, I do not include base salary as a variable in my analysis.1

I also control for whether a teacher belongs to a union in this model for two main

reasons. The first is that union membership may provide existing teachers with

preferential treatment in filling ajob vacancy. Another reason is that districts may find

unionized teachers more difficult to fire. As a result, problem teachers may be shuffled

around more often. These two explanations of how unions affect transfer rates are very

different, however in both scenarios union membership should positively affect the

probability of transferring.

Finally I add state fixed-effects into the regression model to control for basic

institutional and cultural differences across states, and a dummy = 1 for missing data in

the % Free-Lunch (L) variable.12 It would be ideal to use school district fixed-effects,

however, doing this results in a dramatic reduction of my sample size.13 Once this is

done, the regression takes the following form:

Transfer(,) = a + PiHours(t i) + p2L(t 1-) + 33Thr(t 1) + 04BP(t 1) + P5PT(t 1) + 36Exp(t 1)
+ 37Adm(t 1) + PsMar(t 1) + 390Job(t 1)+ 3ioAge + PiiUnion + p12LDummy
+ P13StateFE

where PTis a dummy for job type (1 if part-time, 0 otherwise), Exp is the total years of

full-time experience, Age is teacher age, Union is a dummy for union membership (1 is

10 In the SASS, over 98% of school districts reported having a set salary schedule for teachers in their
district. This means that virtually every teacher in the sample who moved within district did not face salary
differences between the schools in their district

1 Although other types of compensation such as incentive pay are an important component to one's
decision to transfer, those types of compensation could be endogenously determined and therefore
inappropriate to include.
12 The Free-Lunch variable is the only variable in the regression where observations were missing.
Roughly 157 (or 9%) of observations were missing this data.

13 Never the less I show on Table 3-7 that the inclusion of district fixed-effects does not alter the results.









unionized, 0 otherwise) and Mar is a dummy for marriage status (1 is married, 0

otherwise).

3.3 Data

The data for this model comes from the 1999 2000 Schools and Staffing Survey

(SASS) and the 2000 2001 Teacher Follow-up Survey (TFS) produced by the National

Center for Education Statistics (NCES).14 The data are comprised of individual (teacher)

level survey information that can be matched across surveys, as well as relevant survey

information for the relevant schools, principals and school districts. In the original 1999-

2000 SASS survey there were over 52,000 respondents, but the TFS consists of only a

subset of those original respondents.

In aggregate the TFS includes 6,758 teachers, however not all of the teachers in the

survey are included in this analysis.15 Since this analysis focuses on public school

teachers who move within district relative to those who do not change schools, I

eliminate those respondents who either leave their district, or leave teaching altogether.

Later in the paper I will use other types of movement (interdistrict and interstate

movements) to show that PDA hours do not influence other mobility decisions,

particularly those decisions where networks should not matter.








14 The datasets utilized in this paper contain identifiable and sensitive information. While there are
versions of these data that are available to the general public, the datasets used in this paper are only
available at the discretion of the NCES, and with their proper approval.

15 Those respondents in the TFS who left the teaching profession after the 1999-2000 school year (2,374
respondents) are given a version of the TFS that is different from the survey given to those who did not
leave teaching.









In addition, I eliminate those teachers who initially teach at private, Indian, or

charter schools, teachers who reported their move was mostly involuntary16, and those

teachers who were not considered full or part time regular teachers in 1999 (e.g.,

specialists, substitutes, student teachers and itinerant teachers). Once these criteria are

imposed on the data set, 1,688 observations remain. Of the 1,688 observations in the

dataset, 349 (or approximately 20.7%) of those were teachers who transferred within their

district.

3.4 Empirical Evidence

Because the dependent variable is a binary choice variable with a successful

transfer equal to 1, probit estimation is appropriate. Column 1 of Table 3-3 reports a

preliminary regression of PDA Hours onto the dependent variable, with the specified

covariates also included.17 At first glance it seems that networking does not have a

significant impact on the likelihood of transfer. Networks though are a story of collegial

kinship, and like any friendship, benefits may not be realized until the parties involved

are familiar and comfortable enough with one another to begin a relationship. Thus

networks may be difficult to establish initially. Further, networks may require a level of

maintenance. If that is the case, then teachers who engage in PDA frequently may have

an increasing advantage over those who do not. This may imply that there are increasing

returns to hours spent in PDA in terms of networking. A simple test of this assumption is

to include a squared term of PDA Hours.


16 It is also worth noting that any involuntary teacher movements who may be unavoidably included in the
sample (thus creating noisy measures of intradistrict movement), are likely to bias any positive estimates of
the effect of networking towards zero.

1 All regression estimates are clustered at the district level to control for possible standard error correlation
at the district level. In addition, all regression estimates include robust standard errors, to account for
possible heteroskedasticity.









In Column 2 of Table 3-3, the addition of a squared measure of PDA Hours in the

regression seems to suggest that there may be some value to networking. Specifically,

the coefficient on the squared term is positive and significant at the 0.025 level, which

may indicate that while having only a few hours networking may have no impact on

transferring, positive returns to networking may exist for large values of PDA Hours. To

further test whether this is the case, I use a spline to allow the effect of an additional hour

of PDA to be greater at high levels of PDA hours then at low levels of PDA hours.

The spline estimates a continuous, piecewise-linear relationship between PDA Hours and

transfer probability, and is defined as follows:

PDA Hours (H) : from [0 H*] hours:
=H ifH < H*
= H* ifH > H*

PDA Hours (H) : from [H* Hmax] hours:
= 0 ifH < H*
=(H-H*) ifH>H*

Table 3-4 reports the coefficients of several regressions using the spline. I test

various cutoffs of H*, where H* is chosen at decile intervals along the distribution of

PDA Hours. Table 3-4 illustrates that the 2nd segment of the spline is positive and

significant for most decile cutoffs18, but the best fit for the regression is where the cutoff

is at the 80th percentile. This table confirms that additional PDA Hours, particularly for

those teachers in the top 20% of the overall distribution of PDA Hours, seem to have a

positive and significant impact on a teacher's transfer decision. Interpretation of the

coefficients suggests that a one standard deviation increase in the total number if PDA



18 For every cutoff chosen at or above the 20th percentile of the distribution, the 2nd segment of the spline
estimation is positive and significant.









Hours (approximately 46.6) can lead to a 0.020 to 0.078 increase in the probability of

transferring within one's own school district.

Other covariates also seem to have a significant impact of the transfer decision.

The evidence seems to indicate that teachers with less experience and those who work

only part-time are significantly more likely to move between schools. Intuitively this

seems valid since teachers with little experience have had arguably less time to sort into a

desirable school and should therefore be more likely to transfer than teachers who are

further along the career path. In addition, part-time employees stand to secure higher

wages if they search for full-time employment, and have lower time-cost to job

searching.

Besides the natural sorting taking place for younger/less experienced teachers,

teachers who work in low SES schools and who face undesirable working conditions

clearly have a greater incentive to sort into better schools, especially given that

additional/incentive pay is rarely available for teachers to work in less desirable schools.

The evidence in Tables 3-3 and 3-4 seem to support this claim. Teachers at schools

where student SES as a whole is lower than the district average, schools where their

overall job satisfaction is generally low, and teachers that feel administrative quality is

generally poor, seem to have a higher propensity to sort.

Lastly, Tables 3-3 and 3-4 show that there is also some advantage to a teacher

being part of a teacher union in terms of increased mobility. This could be because

teachers who are part of a union receive priority in filling a position vacancy. Another

interpretation of this union effect could also be that unionized teachers are difficult to

fire, and therefore get shuffled around more often. Overall the results shown in Tables 3-









3 and 3-4 are consistent with previous findings in the literature, and not at all surprising

given the economics going on. These regression specifications provides some valuable

insight into the teacher transfer decision, and also tells us that teacher networks are not an

integral component to teacher mobility.

3.5 Robustness Checks

The evidence in the previous section suggests that there is a significant relationship

between teacher networks and transfer behavior. However it is important to test the

robustness of these results. If the time a teacher spends in PDA is simply correlated with

some other (possibly unobservable) characteristicss, then the explanation provided by

networking may be inaccurate. If it is true that teachers who spend a great deal of time

doing professional development activities are simply more likely to transfer, then the

results are spurious in a networking framework.

Before I test the robustness of the results in the previous section, I look to see if

those teachers who engage in the most professional development look substantially

different from the average teacher. I compare the characteristics of teachers in every

decile of the distribution of PDA hours to the average characteristics of the entire sample

of full and part-time teachers (2,280 observations). The results of these comparisons are

reported in Figure 3-1. Each graph on Figure 3-1 shows the mean of the given

characteristic for each interval (0% 10%, 10% 20%, etc) of the PDA distribution,

centered about the mean for the entire sample. I use the standard errors at each interval to

construct a 95% confidence interval on which the mean lies, and test whether the sample

average falls in the range of the confidence interval.

There are only a few characteristics where teachers who engage in the most PDA

look significantly different from the sample average. The graphs of Age and Full-Time









Experience illustrate that teachers in the top end of the distribution of PDA Hours are

significantly older and more experienced. However, previous evidence suggests that

more experienced teachers, who are also generally older, are also less likely to transfer

schools. Therefore, if older and more experienced teachers are taking the most hours of

professional development, then we should expect PDA Hours to be negatively correlated

with the probability of transferring, which does not seem to be the case.

The graphs in Figure 3-1 may also suggest that those teachers at the very bottom of

the distribution of PDA Hours may be disproportionately white, male part-time teachers,

with respect to the entire distribution of hours in PDA. However it is not clear that this

presents a problem of selection bias in who decides to take PDA as it relates to transfer

rates. Currently there is no evidence to suggest that qualities of teacher race or gender

are effective at predicting/explaining teacher movement. In addition, Table 3-4 showed

that part-time teachers are more likely to move than full-time teachers. If part-time

teachers are more likely to spend only a small amount of time in PDA, this should also

downwardly bias the estimated effect of PDA Hours on transfers. These results are

promising, however they are not conclusive. Teachers may still be sorting into the

distribution of PDA Hours based on some characteristic that is not observable.

To test whether teachers are sorting into high or low levels of professional

development activity, I begin by looking at other types of teacher movement. If it is true

that teachers with high levels of PDA participation are simply more likely to move, then

one might expect these same teachers are also more likely to move in ways other than just

intradistrict transfer. However, since most PDAs are activities limited to one's own

school district, then networks should be largely ineffective for movement outside the









scope of the immediate district. So if PDA hours are a strong predictor of other types of

movement, then that may be an indication that the network relationship is unjustified.

Tables 3-5 and 3-6 report similar regressions similar to those reported in Table 3-4.

Table 3-5 shows the effect of a change in PDA Hours on the probability of

movement to another school district within one's state. Table 3-6 shows the effect of a

change in PDA Hours on the probability of movement to another school district in

another state. In both cases, the effect of increased PDA Hours is not significantly

different from zero. For Tables 3-5 and 3-6, neither spline segment is significant, which

suggests that the amount of time a teacher spends in PDAs has no significant impact on

how likely they are to move out-of-district (within the state), or even out-of-state. This

helps to quell some suspicion that teachers with high levels of PDA are simply more

mobile.

It may also be possible that the relationship between PDA and mobility is a

function of district characteristics or behavior. For example, if districts with "naturally"

high levels of teacher mobility also demand an unusually high number of PDA hours for

their faculty (arguably this may be the case with some urban school districts), then the

positive effect of networking is simply a reflection of high levels of mobility in a subset

of districts. To account for this possibility, I repeat the regressions in Table 3-4 with

district fixed effects. Because the use of district-fixed effects limits the sample to

districts where both movement and nonmovement occur, the number of observations

drops to only 398. Naturally this limits the power of the analysis, and is the main

problem with using district fixed effects throughout. Table 3-7 reports the results of these

regressions. According to the data, high levels of PDA have a positive effect on the









transfer decision and low levels of PDA have a negative effect on the transfer decision.

The results in Table 3-7 confirm that district specific behavior does not fully explain the

relationship hours in PDA and mobility.

3.6 Conclusion

The evidence presented in this paper clearly seems to indicate that time spent in

professional development activities leads to networks that enhance teacher mobility

within their own district. These PDAs arguably provide a forum in which teachers can

build networks to gather important information about school characteristics, work

environments and job availability from which they can sort in a less costly, more efficient

way. It also seems plausible that teachers can use their networks with favored colleagues

to reveal their aptitude for a job vacancy, which may allow teachers access to transfer

options that might have previously been closed to them.

Results show that a one standard deviation increase in the amount of networking

leads to a 0.020 to 0.078 increase in the probability of transferring within one's own

school district. If teachers use these networks to more effectively sort away from less

desirable schools, such as low-income or failing schools, then networks arising from

PDAs may contribute to the disparities in teacher quality across school types. It is also

possible that if networks make it more difficult for low achieving schools to retain

qualified teachers, then student achievement at these schools may also be diminished by

networks. Oddly enough, this unintended consequence of staff development policies

would be subversive to the goal of PDAs, which is to improve education through greater

teacher quality.











Table 3-1: Professional Development Activities Summary Statistics
Number of Teachers Who Reported

Percent Who
Yes Nocip
Participated


785 1495 34.4 %


University Course in your Main
Teaching Field

University Course NOT in your Main
Teaching Field

Observational Visits to Other Schools

Individual or Collaborative Reseach on
a Topic of Interest

Regularly Scheduled Collaboration with
other Teachers on Issues of Instruction

Mentoring or Peer Observation

Participating in a Network of Teachers
(e.g. Internet Organization)

Attending Workshops, Conferences
or Training

Presenter at a Workshop, Conference
or Training


26.9 %


33.6 %

45.3 %


68.5 %


44.4 %

24.5 %


93.9 %


19.8 %


614 1666


766 1514

1032 1248



1561 719


1014 1266

559 1721


2141 139


452 1828










Table 3-2: Summary Statistics for Various Teacher Groups
No Move Move

Inter-
Intra-District Inter-State
District

484 (21.2
Number of Teachers 1400 (61.4%) 370(16.2%) 484(22 96(3.9%)


Mean Std. Dev. Mean Mean Mean

PDA Hours 55.728 46.662 59.000 49.636 48.220

% Free-Lunch Eligible 0.002 0.225 0.045 -0.014 -0.021

Bonus Pay 0.137 0.344 0.113 0.101 0.146

Student Threat 0.095 0.294 0.127 0.130 0.083

Job Dissatisfaction 0.021 0.144 0.048 0.066 0.072

Full-Time Experience 10.915 9.856 7.959 5.134 4.583

Part-time employment 0.040 0.197 0.081 0.041 0.010

Poor Administration 0.069 0.254 0.102 0.101 0.104

Union Member 0.783 0.411 0.778 0.652 0.667

Age 40.484 11.340 37.956 33.634 32.030

Gender (Male) 0.282 0.450 0.235 0.299 0.292

Race (White) 0.902 0.297 0.862 0.902 0.927

Marital Status 0.695 0.460 0.678 0.665 0.667










Table 3-3: Probit Regression of Intra-District Movement
Variables 1 2


PDA Hours


(PDA Hours)2


% Free-Lunch Eligible


Bonus Pay


Student Threat


Job Dissatisfaction


Full-Time Experience


Part-time employment


Poor Administration


Union Member


Age


Marital Status

Dummy for Missing
% Free Lunch Eligible

Constant


0.0010626 0.0006944
(1.27) (0.62)

0.00003
(2.30)**

0.301017 0.3069657
(1.84)* (1.87)*


0.1669206
(1.40)


0.1629131
(1.37)


0.1009573 0.1002069
(0.87) (0.86)

0.4224374 0.4353552
(1.85)* (1.92)*


0.0143332
(2.34)**


0.0131213
(2.13)**


0.5452476 0.5394231
(3.43)** (3.41)**

0.3542717 0.3600298
(2.55)** (2.59)**

0.2372197 0.2386774
(2.36)** (2.37)**


-0.005178 0.0055949
(1.01) (1.09)

0.0237581 0.0261624
(0.29) (0.32)

0.1770793 0.176266
(1.20) (1.20)


0.2702219
(1.37)


0.2997908
(1.50)


Pseudo R-squared 0.0955 0.0984
Number of Obs. 1627 1627
Robust SE, clustered Y Y
State FE Y Y
* denotes significance at the 0.05 level
** denotes significance at the 0.025 level







87


Table 3-4: Probit Regression of Intra-District Movement
Regression of PDA Hours with H* Chosen at Decile Intervals Along the Distribution of PDA Hours
H* = H* = H* = H H= H* = H* =
H* = 9 H* = 17 H* = 68
22.5 29.5 40.5 51.5 87.5 122.5
Variables (10%) (20%) (30%) (40%) (50%) (60%) (70%) (80%) (90%)


PDA Hours
0 H*


PDA Hours
>H*

% Free-Lunch
Eligible

Job
Dissatisfaction

Full-Time
Experience

Part-time
employment

Poor
Administration

Union
Member

Student Threat

Bonus Pay

Age

Dummy for
Missing
% Free Lunch
Eligible
Marital Status

Constant


Pseudo R-
Squared
Number of
Obs.
State FE
Robust SE,
clustered


-0.0262 -0.0170 -0.0117 -0.0074 -0.0047 -0.0033 -0.0022 -0.0018 -0.0006
(1.12) (1.68) (1.63) (1.43) (1.34) (1.27) (1.12) (1.15) (0.55)
0.0013 0.0017 0.0018 0.0020 0.0022 0.0026 0.0031 0.0043 0.0061

(1.52) (1.85)* (1.94)* (1.94)* (2.00)** (2.08)** (2.16)** (2.45)** (2.51)**
0.3041 0.3066 0.3077 0.3074 0.3067 0.3086 0.3117 0.3081 0.3010

(1.86)* (1.87)* (1.88)* (1.88)* (1.88)* (1.89)* (1.91)* (1.88)* (1.83)*
0.4214 0.4254 0.4282 0.4285 0.4306 0.4338 0.4377 0.4385 0.4288

(1.86)* (1.87)* (1.88)* (1.88)* (1.89)* (1.90)* (1.92)* (1.93)* (1.89)*
-0.0141 -0.0136 -0.0135 -0.0134 -0.0132 -0.0130 -0.0130 -0.0130 -0.0134

(2.29)** (2.22)** (2.19)** (2.17)** (2.13)** (2.10)** (2.10)** (2.10)** (2.18)**
0.5383 0.5362 0.5351 0.5373 0.5417 0.5458 0.5479 0.5477 0.5381

(3.37)** (3.36)** (3.36)** (3.38)** (3.41)** (3.44)** (3.45)** (3.46)** (3.41)**
0.3575 0.3616 0.3636 0.3641 0.3638 0.3621 0.3595 0.3596 0.3593

(2.57)** (2.59)** (2.61)** (2.61)** (2.61)** (2.60)** (2.58)** (2.58)** (2.59)**
0.2344 0.2406 0.2419 0.2411 0.2415 0.2416 0.2399 0.2387 0.2373

(2.32)** (2.38)** (2.40)** (2.39)** (2.39)** (2.39)** (2.37)** (2.37)** (2.36)**
0.1011 0.1016 0.1022 0.1025 0.1016 0.1000 0.0984 0.0987 0.0991
(0.87) (0.87) (0.87) (0.88) (0.87) (0.86) (0.85) (0.85) (0.85)
-0.1674 -0.1639 -0.1639 -0.1657 -0.1658 -0.1650 -0.1635 -0.1633 -0.1658
(1.41) (1.38) (1.38) (1.39) (1.40) (1.39) (1.38) (1.38) (1.40)
-0.0053 -0.0054 -0.0054 -0.0055 -0.0055 -0.0056 -0.0057 -0.0058 -0.0056
(1.02) (1.05) (1.06) (1.07) (1.08) (1.09) (1.10) (1.12) (1.09)
0.1789 0.1779 0.1798 0.1796 0.1790 0.1777 0.1772 0.1807 0.1776

(1.22) (1.21) (1.22) (1.22) (1.22) (1.21) (1.21) (1.23) (1.21)
0.0232 0.0224 0.0232 0.0245 0.0256 0.0265 0.0266 0.0277 0.0264
(0.28) (0.27) (0.28) (0.30) (0.31) (0.32) (0.33) (0.34) (0.32)
-0.9070 0.5313 0.4935 0.4260 -0.9811 0.3674 0.3596 -0.9753 0.3402
(1.29) (2.15)** (2.09)** (1.93)* (1.45) (1.78)* (1.75)* (1.44) (1.69)
0.0993 0.1004 0.1004 0.1002 0.1002 0.1003 0.1005 0.1012 0.1009


1627
Y
Y


1627
Y
Y


1627
Y
Y


1627
Y
Y


1627
Y
Y


1627
Y
Y


1627
Y
Y


1627
Y
Y


1627
Y
Y


* denoted significance at the 0.05 level
** denoted significance at the 0.025 level







88




W A











0-
0 20 40 60 80 100
PDA Distribution



B








8-0


0 20 40 60 80 100
FDA Distribution


8C




. A I I T I T T I I


i I I 1 I 1 4
O-1

0 20 40 60 80 100
PDA distribution

A) Percent of teachers receiving poor administrative support
B) Percent of teachers receiving bonus pay
C) Percent of teachers reporting overall job dissatisfaction

Figure 3-1. Comparisons of Teacher Characteristics Along the Distribution of PDA.





























0 20 40 60 80 100
PDA Distribution


E






ai)




o t
0-


L..


0 20 40 60 80 100
FDA Distribution

F



t-











0 20 40 60 80 100
PDA Distribution
D) Percent of teachers with a master's degree
E) Percent of free-lunch eligible relative to the district average
F) Percent of teachers who belong to a union


Figure 3-1. Continued.




Full Text

PAGE 1

UNDERSTANDING EDUCATION: THREE ESSAYS ANALYZING UNINTENDED OUTCOMES OF SCHOOL POLICIES By JEREMY CLAYTON LUALLEN A DISSERTATION PRESENTED TO THE GRADUATE SCHOOL OF THE UNIVERSITY OF FLOR IDA IN PARTIAL FULFILLMENT OF THE REQUIREMENTS FOR THE DEGREE OF DOCTOR OF PHILOSOPHY UNIVERSITY OF FLORIDA 2005

PAGE 2

Copyright 2005 by Jeremy Clayton Luallen

PAGE 3

iii ACKNOWLEDGMENTS Throughout the course of my graduate and di ssertation work here at the University of Florida, there have been many friends, re latives and colleagues wh om I have relied on for much professional and emotional support. It is for this reason that I wish to acknowledge the contributions of all these important individu als, who were key to the success of this dissertation. In addition to the academic and financial support these individuals have provided, they have also made outstandi ng sacrifices to better my growth, not only as a professi onal, but also personally. First and foremost I would like to recogni ze my parents Dean and Nancy for their amazing support of my efforts. Their guidance through the years has bettered me in ways that I cannot even begin to measure. I w ould also like to thank my fiance Jaclyn for never-ending support of me and my work over the past several year s. I would like to specially acknowledge Dr. Lawrence Kenny for his wonderful friendship and guidance as a mentor over the years. I would also like to thank my sister Jessica who has inspired me in many ways. There are also many othe rs I would like to acknowledge for their assistance in making this disse rtation a success. I would lik e to thank Steven Slutsky, David Figlio, David Denslow, David He dge, Randall Reback, Richard Romano, Sarah Hamersma, David Brasington, David Sappingt on, Bill Bomberger, Chunrong Ai, Scott Hankins, Mark Hoekstra and Jim Dewey for th eir many comments, incisive criticism, and overall enthusiasm for my research.

PAGE 4

iv I would like to thank Anne Stahl at the National Juvenile Court Data Archive for her guidance and patience in help ing me to acquire data for th is project. Thanks also go to Kathy Budge, Dennis Small, and many othe rs at the Superintendent of Public Instruction Office in Washington State for pr oviding supplementary data that has been crucial to the success of my research. I woul d also like to thank Martha Haynes and the Staff of the Education Information and Acc ountability Services di vision of the Florida Department of Education, who supplied vari ous reports and data utilized in this dissertation. I would also like to thank Kerry Gruber and the entire staff at the National Center of Education Statistics that supp lies and distributes the Schools and Staffing Survey. I am also especially grateful to a ll of the teachers and administrators throughout the State of Florida who shared their time, their thoughts and their experiences with me.

PAGE 5

v TABLE OF CONTENTS page ACKNOWLEDGMENTS.................................................................................................iii LIST OF TABLES............................................................................................................vii LIST OF FIGURES...........................................................................................................ix ABSTRACT....................................................................................................................... ..x CHAPTER 1 SCHOOLÂ’S OUTÂ…FOREVER: A ST UDY OF JUVENILE CRIME, ATRISK YOUTHS AND TEACHER STRIKES.........................................................1 1.1 Introduction..................................................................................................1 1.2 The Identification Strategy..........................................................................3 1.2.1 Zip-Code Matching..........................................................................5 1.2.2 The Basic Model..............................................................................8 1.3 Data..............................................................................................................9 1.3.1 Methodology..................................................................................10 1.3.2 Defining the Parameters.................................................................11 1.4 Regression Analysis...................................................................................13 1.4.1 Community Differences.................................................................16 1.4.2 Differences in Offense Types........................................................18 1.4.3 Types of Offenders........................................................................19 2 THE ROLE OF TEACHER NETWOR KING IN TEACHER TRANSFER DECISIONS AND TE ACHER MOBILITY.........................................................42 2.1 Introduction................................................................................................42 2.2 Discussion of Networking..........................................................................43 2.2.1 How Teachers Network.................................................................43 2.2.2 How Teachers Use Their Networks...............................................45 2.3 The Basic Model........................................................................................46 2.4 Data and Empirical Study..........................................................................47 2.4.1 Defining the Parameters.................................................................48 2.4.2 Empirical Testing...........................................................................52 2.5 Robustness Checks.....................................................................................55 2.5.1 Re-examining the Data..................................................................55

PAGE 6

vi 2.5.2 Unobserved Correlation.................................................................57 2.6 Conclusion.................................................................................................58 3 WITH A LITTLE HELP FROM MY FRIENDS: EVIDENCE OF TEACHER NETWORKS USING MICRO DATA..................................................................65 3.1 Introduction................................................................................................65 3.2 The Identification Strategy........................................................................67 3.2.1 What Are Networks And How Are They Useful...........................67 3.2.2 How Do Teachers Build Networks?..............................................69 3.2.3 Measuring Teacher Networks........................................................70 3.2.4 The Basic Model............................................................................72 3.3 Data............................................................................................................76 3.4 Empirical Evidence....................................................................................77 3.5 Robustness Checks.....................................................................................80 3.6 Conclusion.................................................................................................83 APPENDIX A SUMMARY STATISTICS AND SUPPLEMENTARY REGRESSION ANALYSIS............................................................................................................96 B COMPLETE RESULTS OF SPE CIFIED REGRESSIONS.................................98 LIST OF REFERENCES.................................................................................................103 BIOGRAPHICAL SKETCH...........................................................................................107

PAGE 7

vii LIST OF TABLES Table page 1-1 Negative Binomial Regression onto Total Crime with Full Data Set....................30 1-2 Urban Subsample Negative Binomia l Regression onto Total Crime....................31 1-3 Suburban Subsample Negative Bino mial Regression onto Total Crime...............32 1-4 Rural Subsample Negative Binom ial Regression onto Total Crime.....................33 1-5 Effects of Strike Days by Crime Type and by Community Type..........................34 1-6 Effects of Strike Days by Crime Type and by Offender Type..............................35 1-7 Temporal Displacement of Crime by Crime Type and by Offender Type............36 1-8 Temporal Displacement of Crime through Lagged Strike Variable......................37 1-9 Changes in Total Juvenile Crime, by Da ys Elapsed since the Start of a Strike.....38 1-10 Effects of Strike Days with Sp ecific Years Dropped from the Sample.................40 1-11 Average Effect of Randomly Gene rated Strike Days on Total Crime...................41 1-12 Effects of Strike Days by Offender T ype with the Last 4 Years of the Sample Dropped..................................................................................................................41 2-1 Ordinary Least Squares Regression ont o teacher transfers using full sample with and without robust standard errors.................................................................60 2-2 Percent change teacher transfers pe r school for a one standard deviation increase in each significan t variable with full sample...........................................61 2-3 Ordinary Least Squares Regression ont o teacher transfers using sample with biasing outlier omitted...........................................................................................62 2-4 Outlier excluded percent change teacher transfers per school for a one standard deviation increase in e ach significant variable......................................................63 2-5 Ordinary Least Squares Regression of other types of teacher movements with outlier excluded......................................................................................................64

PAGE 8

viii 3-1 Professional Development Ac tivities Summary Statistics.....................................84 3-2 Summary Statistics fo r Various Teacher Groups...................................................85 3-3 Probit Regression of Intra-District Movement......................................................86 3-4 Probit Regression of Intra-District Movement......................................................87 3-5 Within State Probit Regression of Interdistrict Movement....................................93 3-6 Out Of State Probit Regressi on of Interdistrict Movement...................................94 3-7 Probit Regression of Intradistrict Mo vement with District Fixed Effects.............95 A-1 Strike Summary Statistics......................................................................................96 A-2 Regressions of Full Data Set excludin g Zip Codes with less than 500 students...97 B-1 Reported on Table 3-4...........................................................................................99 B-2 Reported on Table 3-5.........................................................................................100 B-3 Reported on Table 3-6.........................................................................................101 B-4 Reported on Table 3-7.........................................................................................102

PAGE 9

ix LIST OF FIGURES Figure page 1-1 Average Percent of Total Arrests as Function of Days between Offense and Capture ...................................................................................................................39 3-1 Comparisons of Teacher Characteris tics Along the Distribution of PDA...............88

PAGE 10

x Abstract of Dissertation Pres ented to the Graduate School of the University of Florida in Partial Fulfillment of the Requirements for the Degree of Doctor of Philosophy UNDERSTANDING EDUCATION: THREE ESSAYS ANALYZING UNINTENDED OUTCOMES OF SCHOOL POLICIES By Jeremy Clayton Luallen August 2005 Chair: Lawrence Kenny Major Department: Economics The goal of my study is to examine speci fic school policies to determine if unintentional consequences resu lt from these policies. Specifically, I focus on two main issues as they relate to student and teachers outcomes. I begin by looki ng at the effect of incapacitating juveniles in school as a force influencing juvenile crime. I exploit teacher strikes as a measure of unexpect ed student absence from school to measure the effect of school in preventing juvenile crime. My data set consists of information on every juvenile arrest made in Washington State ove r a 22-year period. I show that previous estimates of the effect of school incapacitation are systema tically underestimated, that criminal activity increa ses as students continue to remain out of school. I also show that these increases in crime reflect an increase in overall crime, not a di splacement. Lastly, I show that repeat juvenile offenders are more likely to have committed their first crime on a strike day, relative to a normal school day.

PAGE 11

xi Chapters 2 and 3 of the study focus on the role of teacher networks in influencing teacher mobility. Specifically, my study deve lops a model of teacher networks that describes how teachers assemble networks th rough professional development activities (PDAs) and how these networks provide an effective sorting mechanism for public school teachers. I empirically test the exis tence of teacher networks with 2 distinct datasets. The dataset in Chapter 2 is compri sed of various reports covering all 67 Florida school districts. Besides examining how professional development affects teacher movement, I am able to exploit the macro nature of the data to compare district characteristics (such as differences in compen sation levels and school district density) to examine how these factors also influence teach er mobility. The dataset in Chapter 3 uses survey data from the “Schools and Staffing Survey” and includes over 17,000 teachers. The high-powered nature of this dataset allows me to identify specific details, such as teacher salary incentives, individual networ k strength and union membership. Ultimately I conclude that teacher networks are an integr al part of a teacher’s transfer decision and have a sizable impact on intra-district teacher mobility.

PAGE 12

1 CHAPTER 1 SCHOOLÂ’S OUTÂ…FOREVER: A STUD Y OF JUVENILE CRIME, AT-RISK YOUTHS AND TEACHER STRIKES 1.1 Introduction Although we have taken important strides in understanding the economics of crime, there remains a great deal which we have yet to fully understand. The focus of this chapter is juvenile crime in pa rticular. Because juvenile crim e can be especially hard to study, due to general limitations on access to da ta, economists still have a lot to explore on this front. Over the past 20 years, economists and soci al scientists have attacked the problem of crime in four distinct ways. Specifically they have analyzed the potential effects of deterrence, retribution, rehabilitation and incapacitation as forces for reducing crime (Ehrlich 1981).1 Deterrence, a widely explored topic, stresses the importance of imposing penalties as an effective means of preventing crime, because the perceived costs to criminals of crim inal activity increases (Mocan and Rees 1999, Levitt 1998, Freeman 1996, Ehrlich and Gibbons 1977, Loc hner 2003). Retribu tion addresses the actual punishment of criminals. It suggests th at a criminal experiences an increased cost to his crime when a punishment is imposed on him and/or other criminals (Levitt 1998). Rehabilitation stresses the importance of reforming criminals, through treatment and rehabilitative programs, in preventing future crime (Cuellar, Markowitz and Libby 2003). Finally, incapacitation deals with the notion that the physical lock-up and detaining of 1 I am not arguing that these four ideas are mutually exclusive ways of preventing or reducing crime.

PAGE 13

2 criminals may be an effective means of pr eventing crime (Freeman 1996, Lochner 2004). This paper will closely examine the role of incapacitation in preventing juvenile crime. Specifically I deal with the effect of in capacitating juveniles in school on juvenile crime. Brian Jacob and Lars Lefgren (2003) we re the first to addre ss this topic in their paper, “Are Idle Hands the Devil’s Wo rkshop? Incapacitation, Concentration and Juvenile Crime”. They examine the effect of school attendance on juvenile crime by using teacher in-service days, days where teac hing professionals are required to attend work when children are not requ ired to be at school, as a source of variation in student attendance. While their use of in-service days is very clever, I utilize teache r strikes as a source of variation of studen t school attendance to gain additional insights into the relationship between juvenile crime and school incapacitation. One primary advantage to using teacher stri kes over in-service days as a source of variation in school attendance is that strikes often occur in blocks larger than one individual day. This allows me to observe the effect of prolonge d absence from school on juvenile crime. If juvenile criminal habits do change as absence from school increases, then a one day absence cannot reflect the averag e effect of school incapacitation. Another advantage of utilizing teacher strikes is that they are relatively unpredictable. Since in-service days are pl anned at the beginning of the regular school year, parents have adequate time to make arrangements for thei r children during these off-days. In contrast, school closures caused by teacher strikes are often reported in local newspapers only a couple of days before they are likely to take place, and even then, there is no guarantee that these reported strikes will actually occur. Additionally, a

PAGE 14

3 parentÂ’s ability to plan for their childrenÂ’s activities during strike days is complicated by the fact that details revealed in newspapers and other information sources may prove to be inaccurate or unjustified.2 The late-breaking and incomplete nature of information leaves parents little time to plan their childrenÂ’s activities during the days of a teacher strike. Thus teacher strike days are likely to result in more unsupervised students, and more juvenile crime. The main drawback to this use of strike as variation is that t eacher strikes are often very sparse, and mostly occur in a only a few school districts. In order to overcome this difficulty, I rely on data from Washington State. Washington provides an ideal environment for this type of study because it has an extensive strike history as well as a detailed juvenile arrest data se t that dates back as far as 1980.3 1.2 The Identification Strategy As stated earlier, the objective of this anal ysis is to measure the impact of school incapacitation on juvenile crime rates by using t eacher strike days as variation in student absence from school on an ordinary school da y. Arguably, teacher stri kes are a source of variation that is exogenous to variations in j uvenile crime. Issues that lead teachers to strike include pay, class size and teacher planning time, none of which are likely to influence daily variations in juvenile crime.4 2 For instance, a school district may report that it is sure teachers will return to work when a court injunction is issued, and subsequently the teachers may defy the injunction. 3 Unlike most states in the U.S., teach er strike activity in Washington rema ins very much active to this day. The most recent strike event took place in 2003 in Marysville school distri ct. This strike lasted nearly two months (approximately 50 total days). A summary table describing teacher strikes in Washington can be seen in Table A-1. 4 In fact, areas where juvenile crim e may be an increasing problem are ar guably less-likely to see a teacher strike, because greater crime has been shown to be positively correlated with higher teacher pay (Grogger 1997).

PAGE 15

4 In order to draw direct comparisons between crime on strike vs. nonstrike days, I must be able to control for days in which schools are closed for other reasons. This would include days that coincide with sp ring, summer and winter vacations, weekends, teacher in-service and half-days, national holid ays, etc. However since the data cover a large span of time (22 years), and since each Washington school district determines their own school calendar, it is impossible to know where all of these nontypical school days occur.5 Further, a prolonged teacher strike sometimes results in a reworking of the existing school calendar (largely unobservable) so that a school di strict can meet the requirement set by the state that the school year last for 180 days.6 Since I cannot control for these school holidays with complete accu racy, I must include only those days which I am strongly confident students are regul arly scheduled to be in school. To minimize any possibility that school ho lidays could bias the results, I take extra measures when eliminating questionable da ys. I begin by eliminating weekends and national holidays.7 If a national holiday falls on a Saturday or Sunday, I eliminate the preceding Friday or subsequent Monday, respec tively. The first two Fridays of October are dropped because they are both traditiona l days for the state mandated teacher inservice day. The entire months of June and July were al so excluded, and only the last three school days in August are allowed in the sample. Finally I elim inate the first week 5Each school district may vary on the scheduling of breaks, half-days, inservice days, etc. To the best of my knowledge, there is no state or federal agency th at has archived these calendars over the 21 years. Since the advent of increased techno logical integration with school systems, some districts are beginning to electronically archive these calendars, however these cale ndars generally only go back a couple of years. The Washington State Superintendent of Public Instruction was able supply me with sufficient calendars for most districts, for each year from Aug. 2000 to June 2004. 6 The shortening of spring and winter breaks is comm on, as well as the lengthening of the school year into summer vacation. 7 Thanksgiving falls on Thursday, however both Thursd ay and Friday are dropped from the sample because students are given both of those days off from school.

PAGE 16

5 in January, the last two full weeks (at least) of March, the first tw o weeks of April (at least 8 school days), and the last two full w eeks (at least) of December. Based on these criteria, I can be reasonably certain that each observation in the analysis is a typical school day.8 Days which are included in the sample but are not ordinary school days will serve to downwardly bias the results, because more juvenile crime should be naturally occurring on those days. This is only true because this measurement of strikes as a treatment is limited to the days where students have missed an expected day of school.9 1.2.1 Zip-Code Matching My data set consists of individual-arrest da ta, where each juvenile is matched to his home address zip code. Since zip codes are set by the Addr ess Information division of the U.S. Post Office, they look very differe nt from other boundari es like county borders, congressional districts, and most importantly school districts. In order to match each juvenile (and his/her arrest) to his/her a ppropriate school district, I must address two problems. The first problem is that zip code s often spill over multiple school districts. Therefore it is often uncertain which school district a juveni le is in, given their home zip code. The second problem is that zip codes ar e redefined over time. First I will focus on how zip codes change, and how this problem is resolved so that identification strategy is preserved. As cities and towns develop both inside a nd outside city limits, the Post Office creates new zip codes and new zi p-code offices to handle increasing mail traffic. When 8 This is not a perfect matching process because half days and minor differences between districts are impossible to pinpoint. 9 The timing of strikes is very particular and seldom overlaps with nonschool days, except weekends.

PAGE 17

6 these new zip codes are created, they are draw n strictly as subsets of existing zip codes.10 The creation of these new zip codes makes zipcode fixed effects useless from one year to the next. In addition, it is also very difficult to pinpoint when new zip codes are created. However it is possible to see which zip codes were create d and what preexisting zip code they were originally a part of. By using zip-co de maps of Washington State from 1984 and 2000, I am able to map each subset zip code back into its original (1984) zip code.11 Essentially I treat zip codes in Washington as if they are never partitioned, beginning in 1984. This method leaves me with 510 zip codes in total and a stable zip code definition over a 22 year time frame. The second problem is matching these zip codes to their corresponding school district. Zip-code overlap with school dist ricts make this matching process difficult, however the approach I take is sound, and fair ly straight forward. I began by totaling the number of schools in each zip code noting which district they serve.12 I then divided the total number of schools in that zip code for a given district by the to tal number of schools in that zip code, regardless of their district Essentially what I am left with is a probability, that I will call p *, which reflects the probability that a child living in a particular zip code is also part of a particular school district. This p measure allows me to assign a proper strike tr eatment to zip codes. 10 So far I have not been able to find an instance in Washington State since 1984 where a zip code was created from two or more existing zip codes 11 Four maps of Washington State zip codes were provided by the Western Economics Research Co., Inc. courtesy of the Suzzallo Library at the Washington Library. They encompass they following years: 1984, 1989, 1992 and 1994. Zip code maps of Washington State for the year 2000 are made available by the Census. 12 I did this using a master list of schools (as of the 2003-04 school year) that includes each schoolÂ’s address, grade level served, school code and district name. This list was provided by the Office of Superintendent of Public Instruction in Washington State. It is also publicly available.

PAGE 18

7 There are several features of zip codes th at simplify this matching process. One helpful factor is that many zip codes do not ove rlap school district bound aries at all. This leaves out any guess work. Another helpful poi nt is that zip codes and school districts often share a significant numb er of common boundaries. Th ese shared boundaries occur where divisions seem logical, such as along a county boundary, or a major river or highway. These convenient features lessen the confusion of the matching process. It is also worth noting that not all zip codes contain schools.13 In my sample there were four zip codes without any schools; two were in th e heart of Seattle, and two were located in rural areas. Conveniently, none of these four zip codes overla pped school district boundaries, so assigning them to the appr opriate school distri ct was simple. The last concern I am le ft with deals with how school districts change over time, if at all. If school district boundaries are changing frequent ly and dramatically over time, then the matching process breaks down. Howe ver, what I find is that school district boundaries are stable over time. In order for a district to change its boundaries, it must engage in costly and time-consuming legal pr ocesses, however that does not imply that such adjustments never occur.14 In fact, the two most freque nt changes to school district boundaries seem to be district cons olidation and district dissolution.15 Neither changes turn out not to be pr oblematic. These changes only occu r a total of four times over the 13 The zip codes I use in the sample are zip codes with residences. Many zip codes like business zips, P.O. Box zips, and other nonresidential zips, do not contain schools. These kinds of zips are not included in the data sample. 14 For an outline of regulations and requirements surrou nding district boundary changes, please refer to the publication, “Changing School District Boundaries: A Lay Person’s Guide” published by the Washington State Board of Education in conjunction with the Office of Superintendent of Public Instruction 15 Consolidation is when two or more districts form a new superdistrict. Dissolution is when a one or more districts are absorbed in to existing districts.

PAGE 19

8 course of the 22 years and involve districts wh ich do not experience any teacher strikes. I treat these integrated districts as if they had never been separate. The school district zip code matching pr ocess I invoke does not provide perfectly accurate matches in all cases, but I can be certain of how any mism atching will bias the results. If I say there was a st rike in an unaffected zip code, th e effect of the strike will be biased toward zero because juvenile crime s hould be unchanged. If I say that there was not a strike in an affected group, then the stri ke effect will fail to pick up any increase in juvenile crime rates. Any mismatching that arises from the imperfect matching process will downwardly bias the results. 1.2.2 The Basic Model In this basic model I am attempting to explain changes in juvenile crime for ordinary school days as a function of teacher strikes. I can start by expressing a simple regression model in the following form: Juvenile School Day Crime = + ( p *)(Teacher Strike) Again, all of the information from the crime data is reported at the zip-code level, however the teacher strikes occur at the schoo l-district level. To adjust I introduce p where p* represents the probability that the student population of a zip code is treated by the teacher strike. Rather than have fixed effects to take into account differences across zip codes, I want to initially include specifi c zip-code characteristics to make sure the data are well-behaved. I consider income levels, welfare status, parental education, juvenile work status, juvenile gender, single parent households and community characteristics in this model. Also I need time-fixed effects to control for temporal changes over 22 years. I ther efore include year, month a nd day fixed effects in the model.

PAGE 20

9 The regression model now takes the form: Juvenile School Day Crimemyd = + 1( p *)(Teacher Strike)myd +… +… 2(Median Income)myd + 3(Welfare)myd + 4(Parent Education)myd +… 5(Student Employment)myd + 6(Juvenile Gender)myd +… +… 7(Urban)myd + 8(Single Parent)myd 1(Year) + 2(Month) + 3(Day) Once I am satisfied that the data exhibit a ll of the normal properties one would expect from specifying a model of juvenile crime, I can then shift the analysis to include zip code fixed effects take into account differences which I cannot control for. When I move to this specification, the final regr ession model looks like the following: Juvenile School Day Crimemydz = + 1( p *)(Teacher Strike) mydz +… +… 1(Year) + 2(Month) + 3(Day) + 4(Zip) 1.3 Data My juvenile-arrest data comes from “Was hington Juvenile Court Case Records” made available by the National Juvenile Court Da ta Archive. It prov ides juvenile arrest data over 22 years, spanning from 1980 to 2001. This data set is both lengthy as well as highly detailed at the individual level. It provides the home address zip codes of the offenders, the nature of the crimes committed, the dates of the offenses (as well as the arrests), as well as many other important ch aracteristics surrounding th e reported arrests. In total when I include only Washington juve niles who were arrested for crimes that occurred during the ordinary school days preselected for my study, I have 401,864 arrest cases starting in 1980. In addition to the juvenile arrest data for Washington State, I use data provided by the “Census 2000 Summary File 3” and “1990 Su mmary Tape File 3” to derive annual zip-code characteristics. These Census files give zip-code level information on population and school enrollment numbers, as well adult educationa l attainment, the number of households who are welfare reci pients, type of comm unity, median household

PAGE 21

10 income and student employment. Because th ese characteristics are only available for these two periods (1990 and 2000) I trend them over the 22 years using an exponential function rather than a li near function to describe the path of the trend.16 Finally, teacher strike information came from the report “Public Employee Strikes in Washington” published in 2003 by the Publ ic Employees Relations Commission in Washington State. It covers all Wa shington public employee strikes since 196717; however since this publication lacks some im portant details about these strikes, it has been supplemented with specific informa tion from news articles provided by the Associated Press and the Seattle Times to streng then the accuracy of the treatment effect. These articles are useful in pinpointing specific details surro unding each strike such as whether the school remained opened with emer gency substitutes, and what information was distributed to parents in di stricts where strikes occurred. 1.3.1 Methodology Choosing a proper methodology for this analys is requires some extra care. Because the analysis is done at the daily level, a huge percentage of observations for any given day will be zero. The distribution of the de pendent variable is skewed towards zero, making an OLS regression analysis inappropr iate. This could be solved by using a poisson regression analysis, however the fit fo r a poisson regression for this dataset is poor.18 Also the poisson analysis does not accoun t for strangely behave d standard errors 16 I also trended these variables linearly, however when I do this I end up with some numbers that are not feasible do to the imprecise nature of this process. When I impose minimum and maximum constraints on these variables, I find that these results are insens itive to whether these variables are defined linearly or exponentially. 17 This report began with the enactment of Public Employees Collective Bargaining Act in 1967. 18 A preliminary test of the data shows that the dependent variable generally has a standard deviation roughly 3 times larger than the mean. In addition, I observe large values of the Chi squared terms in the

PAGE 22

11 and overdispersion. In this case, overdisper sion may take the form of a single juvenile committing several crimes on a single strike day, or a gang of juveniles committing what could be considered one crime. Ultimately the chosen method of regression analysis is a negative binomial regression analysis. This deals with problems of overdispersion that a poisson does not correct for.19 The dependent variable is aggregate crime per day (crime count) in a zip code, and the exposure is total student population in said zip code. 1.3.2 Defining the Parameters As described earlier, I begin by controlli ng for differences among zip codes directly rather than using zip-code fixed effects. E ach variable is derived from data taken from the 1990 and 2000 Census. Since I have only two observations for each variable, I interpolate annual values usi ng an exponential path function. I started by solving for the growth rate (Gi, zip) of each variable izip: Gi, zip = ((Censusi, zip 2000/ Censusi, zip 1990)(1/10)) 1 Then I interpolated each year’s obse rvation according to the equation: Variable izip in year t = Censusi, zip 1990*(1 + Gi, zip)(n), where n = t – 1990 Overall this analysis uses 7 variables to cap ture differences in communities. Income is measured using Median Income, which comes directly from the Census data. The Welfare variable is a percentage of households in a zip code that currently receive public assistance income. Of course I expect that local poverty is a serious factor in juvenile crime, so both of these variables seem to be relevant (Mocan and Rees 1999). Juveniles from poor neighborhoods are more likely to be less educated, and ther efore more likely to poisson “goodness-of-fit” tests. Both of these char acteristics suggest that the poisson model in not the correct model for these regressions. 19 The overdispersion parameter alpha is shown to be significantly different from zero. Thus I can conclude that the negative binomial regression significantly different from the poisson regression.

PAGE 23

12 be criminally active overall (Freeman 1992). Since juveniles from poor households have a smaller opportunity cost of committing crime, I expect Median Income to be negatively correlated with juvenile crime, and naturally the converse should be true for Welfare. The Poor Parental Education variable is de fined as the percent of adults (25 years+) who have not completed the equivalent of a high school education. My prediction here is that as adult dropout rates incr ease, juvenile crime increases Because poor adult/parental education equates to less adu lt human capital, parents are likely to have lower demand for child quality (Becker and Tomes 1976, 1986) Therefore juveniles in these neighborhoods should also have a lower opportunity cost of crime. In addition, one can argue that poor adult/parental education may also imply that parents have fewer resources to provide monitoring for juveniles, less overa ll time to provide supervision of their own children (parents may be working long hour s or more than one job), or possibly uninformed/undesirable parenting techniques. Our Student Employment variable describe s the percent of 16-19 year olds who are both employed and also enrolled in school. The prediction of this variable is not clear. On one hand, students who are employed in af ter-school jobs may develop a sense of responsibility from working a j ob, or that they may fear losi ng future wages should they be caught committing a crime. This suggests that Student Employment should negatively influence juvenile crime. On the other ha nd, juveniles who work may have greater access to reliable transportation, or less restrictive parents, both of which could contribute to a greater propensity to commit crime. In th is case, Student Employ ment would positively influence juvenile crime.

PAGE 24

13 The Juvenile Male variable measures the pe rcent of the juvenile population in a zip code who are male. Since juvenile males trad itionally make up the majority of juvenile crime, juvenile crime for a zip code shoul d increase as a juvenile population there becomes proportionately more male.20 The Single parent variable is the percent of single parent house holds in that zip code. One expects that this variable should be positively co rrelated with juvenile crime so that more single parent households lead to more potentially unsupervised juveniles, and potentially greater juvenile crime. The Urban variable is the percent of th e total zip code popula tion who live in an urban community. This should pick up ba sic differences among community types, however I also interact this variable with the strike measure to see if strikes have differential impacts in different types of co mmunities. For several reasons, I expect crime in Urban areas to differ from that of Rural or Suburban areas. Urban areas feature many characteristics, such as increased crim inal opportunity, highe r pecuniary benefit, and increased criminal anonymity, which make s criminal activity in these areas more desirable (Glaeser and Sacerdote 1999). As su ch this Urban variable should be positively correlated with teacher strikes. Lastly, the Strike variable is a dummy variable, with a value of 1 describing a teacher stri ke event, and zero otherwise. 1.4 Regression Analysis The initial results of the negative binomi al regression analysis seem to confirm previous findings about the nature of sc hool incapacitation. Column 1 in Table 1-1 20 From 1993 to 2002, males made up about 71-75% of total juvenile arrests. For more information about juvenile crime by gender, please reference the Crime in the U.S. annual report published by the Federal Bureau of Investigation.

PAGE 25

14 shows that the presence of teacher strikes seems to have a positive and significant effect on juvenile criminal behavior. But before I begin to quantify the effect of school incapacitation on juvenile crime, I want to make sure that the data are behaving properly. The signs of the covariates in Column 1 of Table 1-1 seem to support the predictions made on the effects of income, we lfare, parentÂ’s educat ion and urban status with respect to juvenile crime. First, we see that median income and welfare are negatively and positively significant, respectively. This implies that children in lower income households and welfare receiving house holds are more likely to engage in juvenile crime. These results support th e predictions associated with poorer living conditions. In addition, lower pa rental education leads to high er crime as well. Again, for several reasons this result makes sens e. The Urban variable is positive and significant, showing that urba n communities experience more overall juvenile crime. Our Student Employment variable is positive and statistically significant. This may imply after-school jobs provide students with greater resources which they may use to commit crime. Further, it may reflect the fact that as more student s get after-school jobs, they are spending less time increasing their human capit al (through study, after-school activities, etc.) One problem is that the effect of the Juvenile Male variab le is negative and significant, which contradicts the predicted effect. It says that across juvenile populations, those that are comprised of re latively more females experience more juvenile crime. One possible explanation for this result is that it may be reflecting problems in variable measurement for sma ll communities. For rural communities there are two basic problems. The first is that with a small populati on of juveniles, the

PAGE 26

15 variation of the Juvenile Male variable increases significantly.21 The second is that this unusual variation makes interpolating this variab le equally unreliable. Since a significant portion of zip codes have very small populati ons of juveniles, it is possible that the negative result I observe is being driven by ru ral zip codes that ha ve atypical juvenile gender characteristics.22 To check the sensitivity of the initial resu lts I begin by eliminating months in which strikes did not occur from the sample: January, March, and May (Specification II).23 I then went a step further and eliminated ev ery month other than September, October and April from the sample (Specification III). September and October are preserved because they are the most common months for strike s, and April was included because in 1991 there was a “state-wide” teacher strike that involved at l east 41 school districts. The results of Specifications II & III are repor ted in Columns 2 and 3 of Table 1-1 respectively. These results are not sign ificantly different from Specification I. Recalling the zip code matching process descri bed earlier in the paper, I am able to accurately match back to 1984 zip codes. Howe ver the crime data dates as far back as 1980. To deal with potential zip code mismat ches before 1984, I drop the first four years of the sample (Specification IV).24 This fourth specificati on shows the strike variable 21 For zip codes with juvenile populations less than 500, the standard deviation of the Juvenile Male variable is approximately 3 times larger than the standard deviation of those zip codes with 500 or more juveniles. 22 In fact when I look at only urban zip codes, I see th at this variable seems to correct itself, which gives credence to this argument. Further, Table A-2 shows that when I drop the smallest zip codes (population less than 500 juveniles) I see that the negative significance of the coefficient is destroyed. Thus it seems plausible that poor interpolation of this variable for sparsely populated areas has tainted the coefficients. 23 Over the 22 year period, there wa s one school district that orchestrated a strike in January, but it only lasted for one day. 24 In doing this I eliminate 9 strike events, 3 of which lasted 9 or more days.

PAGE 27

16 remains significant as the sample is modifi ed (Column 4 of Table 1-1). These results seem robust to many different repr esentations of the data set. It is also important to verify that thes e results are not being driven by zip code characteristics being specified inaccurately. The most immediate check of this is to use zip code fixed effects in lieu of explicit vari ables to capture the difference. Column 5 shows the addition of zip code fixed effects does not significantly alter the results. The coefficient of the strike variable Colu mn 5 of Table 1-1 shows that strikes have a positive effect on juvenile crime that is sta tistically significant. To quantify the impact of this increase, I take the partial derivative of the dependant variable with respect to the strike variable ( Totalcrime/ Strike). When I divide this marginal effect by the average of the dependent variable, I am left with a ch ange that reflects a pr oportion of the mean of the dependent variable. In the full data set, the marginal effect of the strike variable suggests that total juvenile crime increases by 56.71% on days when strikes occur. This is more than a modest change in total crime. 1.4.1 Community Differences Since the full data set includes all co mmunities, some of which may look and behave very differently from one another, I want to test whether juvenile crime in different communities is differentially affected on strike days. To capture community differences, I run separate regressions for each community type, where I restrict the sample to only those communities in which 51% or more of the population in a zip code live in one of three types of co mmunities: urban, suburban or rural.25 The results of these 25 The reader should be aware that this specification implies that zip codes which are considered one-third urban, one-third suburban, and one-third rural, will th erefore be excluded from this analysis. However because of the way in which the Census defines thes e communities, these kinds of zip codes make up a very small portion of the total sample.

PAGE 28

17 regressions can be viewed in the Columns of Tables 1-2, 1-3 and 1-4. These results show that unexpected school absences significantly influence juvenile crime only in urban communities. This result is not at all unreasonable, given there are many differences among these types of communities that no doubt influence cr iminal behavior. It seems understandable that rural communities experience no change in juvenile crime when a strike occurs. Rural communities may provide less opportunity for crime, as well as a small town atmosphere that makes anonymity difficult. The lack of a significant effect of strikes in suburban communities is less believable sin ce they arguably provide greater opportunity and anonymity, however there may yet be other, (possibly unobservabl e), characteristics about suburban characteristics th at naturally deter juvenile crime on these strike days. For instance, suburban communities may be pop ulated by more involved parents, may have more effective emergency resources, etc. The strike effects in urban communities however are positive and significant. In ur ban communities juvenile crime increases 19.68% on strike days.26 Again I run the same specification tests to test the robustness of these results. The signs of the covariates in ev ery urban subsample are the same as in Table 1-1, except for Juvenile Male. The Juvenile Male covariate does take on a positive and significant coefficient in the urban subsample, however in the suburban and rural subsamples it remains negative and significant. 26 The huge decrease in the magnitude of the strike coefficient from the full sample to urban subsample indicates that small populations in rural communities are exaggerating the effects of the strikes in the full sample.

PAGE 29

18 1.4.2 Differences in Offense Types Since school incapacitation may be differentially affecting the types of crimes that juveniles are committing, I partition the dependent variable measure of total crime into 5 specific crime types. These types of crim es include: drug and alcohol related crimes, mischievous crimes, property crimes, vi olent crimes, and finally weapons and endangerment crimes. I still consider all th ree community classifica tions when doing this analysis because I may yet find significant re sults in types of crimes for rural and/or suburban communities.27 Table 1-5 shows the results of this extended analysis. Rural and suburban communities continue to experience no change in juvenile crime resulting from strike days regardless of the type of crime that is being committed. Urban communities though do experience an a rray of changes in juvenile criminal behavior. Both mischievous crimes and propert y crimes seem to increase in the presence of a teacher strike. I estimate that misc hievous crime increases by 48% on average for days when strikes occur.28 Property crime increases by an average of 28.81%. This change in property crime is nearly double prev ious estimates in the literature. Violent crime decreases by approximately 31.53%.29 Drug and alcohol crimes, and weapons and endangerment crimes are unaffected by school incapacitation. These changes in criminal activity seem to make sense. It seems logical that mischievous crime would be most affected by school incapacitation. After all, these are 27 This could arise if increases in certain crimes were offsetting decreases in other types of crimes. 28 It is difficult to compare my measure of mischiev ous crime to Jacob and Lefgren’s measure of minor offenses. They define minor offenses as “NIBRS Group B offenses”, however my measure of mischievous crime is a select subset of these crimes. 29 Jacob and Lefgren estimate that school attendance decreases property crime by approximately 14%. They also estimate that violent crim e increases on school days by about 28%. Therefore, my estimate of 31.53% for violent crime is approximatel y 10% larger than previous estimates.

PAGE 30

19 the types of crimes that often result from boredom rather than calculated criminal thought. On the other hand, the decrease in violen t crime that occurs is harder to explain. I use, as Jacob and Lefgren do, social interacti on theory to explain th is result. Jacob and Lefgren argue that juvenile violence against other juveniles arises, at least in part, from disputes formed in and around the classroom. When these juveniles are not forced to be at school, there is a decreas e in the amount of overall vi olent crime that reflects a decrease in juvenile violence against other j uveniles. This explan ation seems credible.30 Given the limits of the data set, I cannot test this theory further.31 1.4.3 Types of Offenders In addition to the topics already spoken to, the detailed level of the data also allows me to also explore the nature of the crim inal, beyond superficial characteristics. By observing criminal record, I attempt to answer whether these additional crimes are caused by normal delinquents facing greater criminal opportunity, or by juveniles who do not normally engage in criminal activity. I agai n partition the dependent variable into two measures: 1) total crime by repeat offenders and 2) total crime by one-time offenders. I define repeat offenders not just as those juveniles who have committed previous crimes, but also those juveniles who will commit future crimes. Thus a juvenile who commits multiple crimes is considered to be a repeat offender juvenile even on their first crime. The results of these regressions can be seen in Table 1-6. Additional overall crime is induced for both offender types, however the average increase in crime for one-time offenders as a percent of the mean crime level for that 30 In 1998 and 1997, 62% of victims of juvenile violence were 17 or younger. 31 I have no information concerning the victims of the crimes in my data set.

PAGE 31

20 group is much higher than for the repeat o ffenders. On strike days there is a 33.45% increase in average total crime for one-time offenders, and a 13.89% increase in average total crime for repeat offenders. Essentiall y, one-time offenders as a group seem to be much more affected by the lack of school in capacitation then the repeat offenders. This evidence suggests that the incr eased crime on strike days is less about criminal motive, and is more a result of boredom. If these two groups are motivated differen tly to commit crimes, these effects may not be expressing differences ac ross groups in the same types of crimes. As a society we may be more concerned if the lack of school incapacitation were i nducing more property crime by one-time offenders rather than repe at offenders (where the crime may have eventually happened anyway). Therefore I am interested in examining whether these two groups differ in offense types on these strike days. To do this I create a dependent variable that equals total crime conditional on offender type and offense type. The results are expressed in Table 1-6. The decrease in violent crime comes mainly from the repeat offenders. Violent crimes committed by repe at offenders drop by 36.72% on strike days. In addition, both groups are cont ributing to the overall increas es in mischievous crimes and property crimes. Property crimes by one-time offenders and repeat offenders increase by 56.75% and 25.40% re spectively. Likewise mischi evous crimes increase by 83.11% for one-time offenders and by 30.82% for repeat offenders. 1.4.4 Strikes Days as Gateway Crimes Since the increase in crimes committed on strike days is substantially larger for first-time offenders, it may also be true that these juveniles may carry this behavior forward after their first arrest. To see if this is the case, I examine whether or not a

PAGE 32

21 current day repeat offender was significantl y more likely to haven gotten his/her start (commit his/her very first crime) on a strike day versus a normal school day. To do this, I generate a daily crime total that expresses the total of “first time” crimes by subsequent repeat offenders. The results of this regression can be also vi ewed in Table 1-6. A future repeat offender is found to be significantly more likely have committed his first crime on a strike day rather than an ordinary school day. This sugge sts that not only is school incapacitation preventing the creation of crimes, but it may also be preventing the creation of criminals. 1.4.5 Displacement of Crime Thus far it seems the evidence suggests juve nile crime increases when students are not incapacitated in school. However what is not clear is whether th ese changes describe an overall increase in the amount of juvenile crime, or a di splacement of crime from one day to another. Jacob and Lefgren try to sp eak to this question of temporal displacement in their paper, however their use of in-s ervice days again limits the power of their analysis. Jacob and Lefgren find that there is no temporal displacement of crime to inservice days, however because in-service da ys are completely known, this result may simply be showing that juvenile criminals will plan to commit more crime when facing more criminal opportunity. What we are really interested in is wh ether a juvenile will temporally displace a “planned” crime given an unexpected opportunity to do so, or whether an unexpected opportunity to commit cr ime results in more “unplanned” crime. Again teacher strikes provide a more ideal measure of unexpected criminal opportunity, so I revisit this question of temporal displacement.

PAGE 33

22 To test whether crime is being displace d, I begin by aggregating Total Crime to a weekly measurement rather than daily32. If juvenile crime is simply being displaced from weekend crimes to weekday crimes, then I s hould not observe an effect for the strike variable. To capture the week ly aggregated effect of stri kes, I proportion the fraction of the week where strikes occurred as a new meas ure of the strike treatment. Despite this aggregation, I find that strike s are still a significant factor in contributing to additional juvenile crime (Table 1-7). Table 1-7 shows a positive and significant st rike coefficient on Total Crime in the full data set. This is a strong indication th at overall juvenile crime is increasing, rather than being displaced across time. However, if displacement of crim e varies across crime type, or criminal type, we cannot conclude th at no crimes are being displaced. To see if the type of crime or the characteristics of the perpetrator influences whether displacement of crime occurs, I partition weekly total crim e by crime type and offender as before. The results of these regressions ar e also reported in Table 1-7. These strike coefficients on Table 1-7 show that every specification of cr ime type and criminal type results in an increase of overall juvenile crime, and not temporal displacement of crime. I confirm the lack of displacement in total crime by aggr egating further to a monthly crime total. Weekly and Monthly aggregations of juvenile crime provide a good test of displacement of crime. I perform another test by lagging the strike effe ct, in order to pick up acute changes in crime shortly after the strike ends. If fewer crimes are committed on days following the strike, then the coefficient of the lagge d strike variable should be 32 This weekly measurement includes weekends and ho lidays, since strike day crime could displaced from usual weekend crime. In addition, week fixed effects are used in place of day fixed effects.

PAGE 34

23 negative and significant.33 I run seven separate regre ssions, shown in Table 1-8, labeled “1 day” through “7 days.” For a strike that ends on day t the first treatment of the lagged strike variable under “1 day” begins on day ( t + 1), under “2 days” begins on day ( t + 2), and so on. Table 1-8 shows that there is no significant effect from the lagged strike treatment. These results provide further evid ence that juvenile crime occurring on strike days is additional crime, rather than displaced crime. 1.4.6 Duration Effects on Criminal Activity One last advantage I gain from using teacher strikes as an instrument, in conjunction with the strength of the daily arre st data, is that I can track how juvenile crime changes over time with prolonged absen ce. If it were true that an increase in juvenile crime is at least partially cause d by boredom and inactivity on the part of juveniles, then I should expect that the longer juveniles rema in “inactive”, the more crime is likely to result. Arguably there is just as much uncertainty over when teacher strikes will end as when they will begin. As such, tr ends in juvenile crimes up until the start of the strike and over the period of the strike should not be biased by student or parent foresight. In addition, the longe r a strike continues, the less able a parent or guardian is to provide adequate supervision to students. For example, a parent may be able to take one or two days away from work to supe rvise their child, but the cost of taking 10 consecutive days off work is substantially larger. To examine if juvenile crime is changi ng with the length of a teacher strike I partition the strike variable in to 4 groups of new independent variables. Three of these variables are clustered groups of 3 consecutive strike days, and the la st variable pertains 33 This sample period also includes w eekends. I also revert back to day fixed effects rather than week FE.

PAGE 35

24 to a strike in its 10th day or higher.34 These results can be view ed in Table 1-9. Table 1-9 shows that for the first 3 days, there is no si gnificant change in ove rall juvenile crime.35 After the 5th day or so, the strike coefficient becomes positive and significant, reflecting an increase in overall crime. Further, after the 5th or so strike day there is a gradual increase in the level of juvenile crime as strike length continues. In unreported regressions I find that as time passes, property and misc hievous crime trends upward, while the decrease in violent crime remains re latively constant across days. The positive coefficient shows that these increases in property and mischievous crime overpower the decreases in violent crime. Column 2 of Table 1-9 repeats the regression in Column 1, only with the strike variable clustered at the twoday level rather than at the three-day level. Clearly the evidence sugge sts that the longer juveniles remain absent from school, the more overall daily juvenile crime will be observed.36 1.5 Robustness Checks The results of the analysis up to this point seem to point to a very clear story about the relationship between juveni le crime and school incapaci tation. In this section I perform some robustness checks to ensure that these results are not spurious. I begin by 34 I do this because if I were to partition the strike variable for each separate day (1st day, 2nd day, etc.), then the number of treatments would be very small for the later strike days and the coefficients would be unreliable. 35 When I break up the dependent variable (total crim e) by crime type as before, violent, property and mischievous crime types are individually significant. However in the first 3 days, the opposite signs and magnitudes of the coefficients wash out aggregate effects in total crime. These results are consistent with Jacob and LefgrenÂ’s findings (2003). 36 It is worth keeping in mind that these strike day counts are also limited to school days. That is to say that when a strike has reached its 10th day, a student has missed exactly 2 full weeks of school. It is also very likely that these criminal behaviors will change after a strike has gone on for an extensive amount of time. At that point, juvenile crime may level-off or even decrease substantially. Of course given the sparse and limited nature of extremely lengthy teach er strikes, an analysis of that na ture is inherently difficult not to mention highly unreliable.

PAGE 36

25 testing whether the strike variable is corre lated with some unobservable characteristic, and that perhaps the results are picking up some alternate relationship. For instance, if a strike event is in fact very foreseeable, th en one might speculate that the increase in juvenile crime predicted by the strike variab le is really picking up an increased police presence on days that strike occur. That is to say that the results are not reflecting a true increase in juvenile crime, but that there ar e simply more police on the streets capturing a larger portion of the same total amount of juvenile crime. If this were true, then one would expect th at the number of days between an offense and an arrest would be shorter on strike days. The graphs on Figure 1-1 show that this is not the case. When I consider only zip c odes where strikes did occur, I see that the percent of quick arrests37 is significantly higher on regular school days than on strike days. I explain this by sugges ting that on regular school days police have a better idea who the likely criminals are (repeat offenders), as opposed to strike days where offenders are more likely to be (previously unidentifie d) first time, or one-time offenders. 1.5.1 Importance of Singular Strikes Given the uniqueness of strikes, I must confirm that these results are not being solely driven by a small subsets of strikes. To do this I systematically eliminate each individual year from my sample starting with 1984. Table 1-10 shows that strikes are a significant determinant of juvenile crime in every year with the possible exception of 1985 (z = 1.70). When I drop 1985 from the sample I find a much weaker significance than when I eliminate any other year. In September of 1985, the Seattle school district engaged in a teacher strike that lasted 20 school days. Given that the evidence has 37 I define these as arrests where the associated offe nse occurred less than 48 hours before the arrest. 25

PAGE 37

26 suggested that juvenile crime in urban districts is most affect ed by teacher strikes, it is possible that these results are being solely driven by the circumstances surrounding this one district. To test this condition I restri ct the sample twice, once to look at the sample without the Seattle school district, and then again to look at only the Seattle school district. If Seattle is driving all of the results, then the strike variable should be insignificant when I exclude Seattle, and extremely significant when I exclude all other districts except for Seattle. Table 1-10 shows that strikes still have a positive and signi ficant effect when I drop Seattle from the analysis. In addition the strike variable in the Seattle subsample is marginally significant (z = 1.82). Therefore I can be confident that while this one event obviously contributes to the re sult, (as I would expect the second longest strike in the most urban school district in Washington State to do), it is not the sole force driving the results.38 1.5.2 Reproducibility of Results Another robustness check I perform is to make sure that the results cannot be easily reproduced with random generations of the strike variable. To check if this is the case, I use a uniformly generated random variable on the interval from 0 to 1 to re-specify a random dummy strike variable. Since teach ers are on strike 2,351 days out of the 1,703,910 day sample, strike days occur 0.13797 % of the time. I assign a strike value of 1 to the random variable, when the value of the random observation generated is less than or equal to 0.0013797. If the randomly gene rated observation is greater than 0.0013797, then it is reassigned a value of zero. The ne w random strike variable should approximate 38 The Seattle strike of 1985 is actually tied with the Mukilteo strike of 1990 (at 20 days) for the second longest strike in Washington State history from 1980-2001.

PAGE 38

27 the number of strikes in the original sample. The results in Table 1-11 show the mean of the random strike coefficient over 25 trial re gressions, as well as the average number of random strike days generated. Of the 25 trials only 2 random strike variable coefficients were significant. One of the statistically si gnificant coefficients was positive, while the other was negative. 1.5.3 Further Tests One final test of the data is to make sure that my finding of new criminal participation on strike days (when the samp le is broken up into one-time and repeat offenders) is not a byproduct of mislabeling offenders. Since the crime data ends in 2001, it could be that in the years leading up to 2001 many future career criminals are being mislabeled as one-time offenders due to the fact that future arrests past 2001 are unknown. To test this possibilit y, I drop the last 4 years of th e data sample and repeat the regressions using offender type as the dependent variable. Table 1-12 shows that this mislabeling does not affect the results. The effects of a strike on repeat offenders and one-time offenders are nearly iden tical for either specification. 1.6 Conclusion The results presented in this chapter provide a clearer pi cture concerning the relationship between schooling and urban juve nile crime. My main findings are the following: 1) The effects of a lack of school incapacitation are larger than those found in Jacob and LefgrenÂ’s study. The evidence sugge sts that property crime rises by as much as 29%, almost twice as much as what is pr edicted in the existing literature. Likewise, violent juvenile crime decrea ses by as little as 31.53%, a nd as much as 36.72%. These estimates of increased violent crime represent a 10%-25% increase in previous estimates.

PAGE 39

28 2) I confirm that these change s in daily crime reflect change s in total crime, and not a displacement of crime from one day to the next, regardless of crime type and criminal type. 3) Different types of juveniles are affected differently by a lack school incapacitation. A significant propo rtion of the decrease in the level of violent crime can be attributed to repeat juvenile offenders, while one-time offenders contribute more to the increases in property crime and crimes of mischi ef. 4). A failure to incapacitate juveniles will result in significantly more crime by those juveniles who might not have otherwise engaged in criminal acts or who rarely enga ge in such acts. In fact, juveniles who become repeat offenders are more likely to have gotten their criminal start when incapacitation was expected but not implement ed, as opposed to or dinary school days. Incapacitation seems to have the greatest implications for new and seemingly preventable urban crime. I find that on strike days, new criminals are engaging in new criminal acts. Incapacitating these juveniles seems to be e ffective at preventing at-risk urban youths from engaging in new and ris ky behavior. I do not find that school incapacitation has any significant effect in suburban and rural communities. It may be true that the characteristics that differe ntiate these community types from urban communities, as well as each other, inherent ly reduce juvenile criminal behaviors, however such characteristics may be largely unobservable to us. If the evidence presented in this paper is ac curate, then what we are learning is that how we manage childrenÂ’s time outside of sc hool is very important. If parents are strained to provide adequate supervision for their children, juveniles who we might consider to be at-risk may be the ones who are affected most, though at-risk juveniles are not the only ones affected. Families and bus inesses also bear the burden of increased 28 28

PAGE 40

29 juvenile delinquency. These results have profound implications for many urban school policies and programs. How school districts bud get their studentsÂ’ days off is no longer a trivial matter. For example, school districts with long breaks (like in a traditional school calendar) may have very different juvenile crim inal behaviors than a school district with frequent but shorter breaks (like in a year -round school calendar). Differences in how school calendars overlap within and among dist ricts should have an effect on the nature of juvenile crime. The length of the school da y or school year in a di strict, school district policies regarding student atte ndance requirements, and even the nature of after school programs within a district are all policy issu es that determine more than just a childÂ’s educational outcome. It is my goal to conti nue to explore this da ta set and scrutinize these kinds of school policies, so that I can better understand the impact that these kinds of policies may have.

PAGE 41

30 Table 1-1: Negative Bino mial Regression onto Total Crime with Full Data Set Variables ( ) ( ) ( ) ( V) (V) Strike Median Income Welfare Urban Poor Parental Educ. Juvenile Maleness Student Employment Single Parent House Alpha 0.195** 0.187** 0.210** 0.260** 0.220** (4.06)** (3.86)** (4.16)** (5.06)** (4.38)** -0.000058** -0.000059** -0.000059** -0.000055** — (102.47)** (83.82)** (60.76)** (52.80)** — 0.475** 0.452** 0.520** 0.821** — (7.71)** (5.96)** (4.97)** (6.28)** — 0.132** 0.112** 0.097** 0.055** — (26.96)** (18.45)** (11.54)** (6.14)** — 0.329** 0.346** 0.398** 0.482** — (10.40)** (8.88)** (7.45)** (6.67)** — -0.330** -0.341** -0.355** -0.396** — (28.06)** (26.24)** (22.32)** (22.72)** — 0.291** 0.297** 0.290** 0.296** — (17.49)** (14.43)** (10.23)** (9.16)** — 0.500** 0.495** 0.466** 0.985** — (9.52)** (7.64)** (5.20)** (9.67)** — 0.989** 1.009** 0.997** 0.956** 0.759** (156.56)** (126.04)** (91.12)** (87.74)** (79.06)** Number of obs. Time Fixed Effects Zip Fixed Effects R-Squared 1,703,910 1,128,630 583,440 479,400 479,400 Y Y Y Y Y N N N N Y 0.0367 0.0364 0.0367 0.0279 0.0559 In Columns – V, the sample is re-specified as follows (Type through Type V): (I) – Sample includes all ordinary school days from 1980-2001 (II) – January, March and May are excluded from the sample 1980-2001 (III) – Includes only April, September and October from 1980 2001 (IV) – Only April, September and October from 1984 – 2001 (V) – Only April, September and October from 1984 – 2001, with zip code fixed effects Z-statistics are given in the parentheses for every table. Indicates significance at the 5% level, ** is significant at the 1% level or smaller. To interpret these regressions, I have to calculate the marginal effect of these variables. The marginal effect of th e strike variable is 0.149, which represents a 56.71% increase in average total juvenile crime on strike days.

PAGE 42

31 Table 1-2: Urban Subsample Negative Bi nomial Regression onto Total Crime Variables ( ) ( ) ( ) ( V) (V) Strike Median Income Welfare Poor Parental Educ. Juvenile Maleness Student Employment Single Parent House Alpha 0.151** 0.129* 0.178** 0.211** 0.214** (3.07)** (2.60)* (3.50)** (4.11)** (4.26)** -0.000061** -0.000062** -0.000062** -0.000057** — (83.07)** (67.39)** (48.42)** (42.17)** — 1.394** 1.369** 1.418** 1.703** — (16.40)** (13.01)** (9.79)** (9.46)** — 0.398** 0.403** 0.449** 0.462** — (9.49)** (7.79)** (6.31)** (4.65)** — 0.602** 0.577** 0.482** 0.397** — (11.13)** (8.60)** (5.16)** (3.62)** — 0.214** 0.224** 0.214** 0.157** — (9.10)** (7.67)** (5.34)** (3.47)** — 0.739** 0.790** 0.945** 1.276** — (8.28)** (7.12)** (6.24)** (7.55)** — 0.676** 0.685** 0.666** 0.638** 0.494** (114.06)** (91.00)** (65.18)** (62.65)** (54.87)** Number of obs. Time Fixed Effects Zip Fixed Effects R-Squared 454,953 301,312 155,601 129,593 129,593 Y Y Y Y Y N N N N Y 0.0484 0.0484 0.0487 0.0400 0.0635 Table 1-2 limits the sample to include only zip codes where 51% or more of the population lives in an area considered by the Census to be Urban In Columns – V, the sample is re-specified as follows (Type through Type V): (I) – Sample includes all ordinary school days from 1980-2001 (II) – January, March and May are excluded from the sample 1980-2001 (III) – Includes only April, September and October from 1980 2001 (IV) – Only April, September and October from 1984 – 2001 (V) – Only April, September and October from 1984 – 2001, with zip code fixed effects Z-statistics are given in the parentheses for every table. Indicates significance at the 5% level, ** is significant at the 1% level or smaller. To inte rpret these regressions, I have to calculate the marginal effect of these variables. The marginal effect of th e strike variable is 0.121, which represents a 19.68% increase in average total juvenile crime on strike days.

PAGE 43

32 Table 1-3: Suburban Subsample Negative Binomial Regression onto Total Crime Variables ( ) ( ) ( ) ( V) (V) Strike Median Income Welfare Poor Parental Educ. Juvenile Maleness Student Employment Single Parent House Alpha -0.027 -0.064 -0.159 -0.162 -0.449 (0.12) (0.29) (0.61) (0.62) (1.74) -0.000015** -0.000016** -0.000013** -0.000012** — (8.01)** (6.54)** (4.74)** (5.66)** — -1.929** -2.146** -1.483** -2.005** — (9.21)** (8.31)** (4.20)** (4.43)** — 1.639** 1.856** 1.683** 2.393** — (13.33)** (12.29)** (8.09)** (8.55)** — -1.432** -1.973** -2.151** -2.899** — (5.33)** (5.94)** (4.74)** (5.66)** — -0.121* -0.198** -0.172 -0.064 — (2.23)* (2.96)** (1.87) (0.59) — 2.285** 2.369** 0.217** 3.062** — (14.13)** (11.97)** (7.83)** (8.82)** — 1.326** 1.303** 1.240** 1.203** 0.969** (59.08)** (54.84)** (39.01)** (37.78)** (34.42)** Number of obs. Time Fixed Effects Zip Fixed Effects R-Squared 134,100 88,841 45,794 37,992 37,992 Y Y Y Y Y N N N N Y 0.0294 0.0300 0.0302 0.0227 0.051 Table 1-3 limits the sample to include only zip codes where 51% or more of the population lives in an area considered by the Census to be Suburban In Columns – V, the sample is re-specified as follows (Type through Type V): (I) – Sample includes all ordinary school days from 1980-2001 (II) – January, March and May are excluded from the sample 1980-2001 (III) – Includes only April, September and October from 1980 2001 (IV) – Only April, September and October from 1984 – 2001 (V) – Only April, September and October from 1984 – 2001, with zip code fixed effects Z-statistics are given in the parentheses for every table. Indicates significance at the 5% level, ** is significant at the 1% level or smaller

PAGE 44

33 Table 1-4: Rural Subsample Negative Binomial Regression onto Total Crime Table 1-4 limits the sample to include only zip codes where 51% or more of the population lives in an area considered by the Census to be Rural In Columns – V, the sample is re-specified as follows (Type through Type V): (I) – Sample includes all ordinary school days from 1980-2001 (II) – January, March and May are excluded from the sample 1980-2001 (III) – Includes only April, September and October from 1980 2001 (IV) – Only April, September and October from 1984 – 2001 (V) – Only April, September and October from 1984 – 2001, with zip code fixed effects Z-statistics are given in the parentheses for every table. Indicates significance at the 5% level, ** is significant at the 1% level or smaller Variables ( ) ( ) ( ) ( V) (V) Strike Median Income Welfare Poor Parental Educ. Juvenile Maleness Student Employment Single Parent House Alpha 0.017 0.027 -0.343 -0.285 -0.392 (0.09) (0.14) (1.45) (1.20) (1.65) -0.000034** -0.000036** -0.000036** -0.000033** — (26.68)** (22.50)** (16.62)** (13.59)** — -0.077 -0.063 -0.108 0.117 — (0.68) (0.45) (0.56) (0.48) — -0.282** -0.332** -0.218 -0.200 — (3.81)** (3.63)** (1.77) (1.32) — -0.357** -0.362** -0.370** -0.421** — (37.69)** (32.83)** (25.81)** (25.64)** — 0.142** 0.156** 0.180** 0.267** — (4.34)** (3.86)** (3.25)** (4.15)** — 0.379** 0.401** 0.317* 0.864** — (4.41)** (3.81)** (2.15)* (5.21)** — 2.679** 2.767** 2.810** 2.712** 2.197** (80.71)** (66.36)** (49.08)** (46.97)** (44.48)** Number of obs. Time Fixed Effects Zip Fixed Effects R-Squared 1,101,359 729,469 377,244 304,491 304,491 Y Y Y Y Y N N N N Y 0.0313 0.0308 0.0316 0.0218 0.0551

PAGE 45

34 Table 1-5: Effects of Strike Days by Crime Type and by Community Type Dependant Variable Strike Coefficient for Urban SubSample Strike Coefficient for Suburban SubSample Strike Coefficient for Rural SubSample Drugs and Alcohol Mischievous Property Violent Weapon and Endangerment 0.200 -1.188 -0.113 (1.36) (1.48) (0.20) 0.536** -0.813 -0.241 (3.88)** (0.73) (0.34) 0.257** -0.128 -0.276 (3.79)** (0.34) (0.80) -0.364* -1.103 -0.266 (2.50)* (1.44) (0.52) -0.014 -53.096 — (0.05) (0.02) — Number of obs. Time Fixed Effects Zip Fixed Effects Ave. R-Squared 129,593 37,992 304,491 Y Y Y Y Y Y 0.0506 0.0423 0.0579 Each row represents a new regression with the listed crime type as the dependant variable. The reported number is the strike coefficient for the regression with the corresponding dependant variable. Each column is a different subsample group: urban, suburban and rural. All 14 regressions are defined as a Type V regression, including data from only April, September and October from 1984 – 2001, and zip code fixed effects. The Pseudo R-squared reported at the bottom of each column is the average R-squared over the 5 regressions. Z-statistics are given in the parenthese s for every table. Indicates significance at the 5% level, ** is significant at the 1% level or smaller. The marginal effects of these variables show that violent crime decreases by 31.53% on strike days. Property crime and mischievous crime increase by 28.81% and 48% respectively. Again, these eff ects reflect changes in juvenile crim e as a proportion of the mean. The coefficients of the marginal effects are (–0.032), (0.072) and (0.025) for violent, property and mischievous crime respectively.

PAGE 46

35 Table 1-6: Effects of Strike Days by Crime Type and by Offender Type Dependant Variable Strike Coefficient for One-Time Offenders Strike Coefficient for Repeat Offenders Total Crime Drugs and Alcohol Mischievous Property Violent Weapon and Endangerment Gateway Crime 0.382** 0.156** (4.05)** (2.77)** 0.337 0.153 (1.18) (0.93) 1.024** 0.597** (3.66)** (3.99)** 0.534** 0.220** (4.44)** (2.81)** -0.567 -0.408** (1.48) (2.59)** -1.330 -0.009 (1.20) (0.03) — 0.286* — (2.00)* Number of obs. Time Fixed Effects Zip Fixed Effects 129,593 129,593 Y Y Y Y Table 1-7 includes only urban zip codes. All 13 regressions are defined as a Type V regression, including data from only April, September and October from 1984 – 2001, and zip code fixed effects. Z-statistics are given in the parentheses for every table. Indicates significance at the 5% level, ** is significant at the 1% level or smaller. The ma rginal effects of these vari ables show that for repeat offenders, violent crime decreases by 36.72%, prope rty crime increases by 25.4%, and mischievous crime increases by 30.82% on strike days. The coefficients of these marginal effect s are (-0.032), (0.046) and (0.012). For one-time offenders pr operty crime increases by 56.75% an d mischievous crime increases by 83.11% on strike days. The coefficients of these ma rginal effects are (0.030) an d (0.008). Again, these effects reflect changes in juvenile cr ime as a proportion of the mean.

PAGE 47

36 Table 1-7: Temporal Displacement of Cr ime by Crime Type and by Offender Type Weekly Aggregated Dependant Variable Strike Coefficient for Full Data Set Strike Coefficient for Urban Subsample Strike Coefficient for One-Time Offenders Strike Coefficient for Repeat Offenders Total Crime Mischievous Property Violent Monthly Aggregated Dependant Variable Total Crime 0.235** 0.243** 0.418** 0.176** (4.50)** (4.67)** (4.51)** (3.11)** — 0.570** 0.820** 0.501** — (4.39)** (3.09)** (3.55)** — 0.312** 0.492** 0.225** — (4.73)** (4.27)** (3.06)** — -0.300* -0.223 -0.317* — (2.32)* (0.73) (2.27)* 0.534** 0.485** — — (3.86)** (3.73)** — — Number of obs. Time Fixed Effects Zip Fixed Effects 440,640 117,491 117,491 117,491 Y Y Y Y Y Y Y Y Each row represents a new regression with the listed crime type aggregated at the weekly level as the dependant variable. The reported number is the strike coefficient for the regressi on with the corresponding dependant variable. Each column is a different sample group. Column 1 is the full sample. The remaining three Columns are from the urban sample. The top 10 regressions are defined as a Type I regressions including zip code fixed effects. The bottom 2 regression coefficients represent the strike effect on Monthly Total Crime. Z-statistics are given in the parenthese s for every table. Indicates significance at the 5% level, ** is significant at the 1% level or smaller

PAGE 48

37 Table 1-8: Temporal Displacement of Crime through Lagged Strike Variable Lagged Strike Variable Strike Coefficient For Urban SubSample 1 Day -0.033 (0.62) 2 Days -0.028 (0.54) 3 Days -0.030 (0.57) 4 Days -0.046 (0.87) 5 Days -0.041 (0.78) 6 Days -0.093 (1.75) 7 Days -0.070 (1.33) Number of obs. Time Fixed Effects Zip Fixed Effects Ave. R-Squared 293,112 Y Y 0.0597 Each row represents a new regression with strike variable lagged the by the corresponding row. Each regression is defined as a Type IV regression including zip code fixed effects of the urban subsample. The notable exception is that weekends are included to allow for the correct number of strike treatments. Therefore, days where lagged strike treatments overlap breaks, holidays and other nontypical school days are included in the sample. Z-statistics are given in th e parentheses for every table. Indicates significance at the 5% level, ** is significant at the 1% level or smaller

PAGE 49

38 Table 1-9: Changes in Total Juvenile Crime, by Days Elapse d since the Start of a Strike Partitioned Strike Variable Strike Coefficient Partitioned Strike Variable Strike Coefficient Days 1 3 Days 4 6 Days 7 9 Day 10+ — — Alpha -0.005 Days 1 2 0.058 (0.06) (0.58) 0.185 Days 3 4 0.006 (1.91) (0.07) 0.261* Days 5 6 0.262* (2.54)* (2.37)* 0.313** Days 7 8 0.284** (2.93)** (2.62)** — Days 9 10 0.228 — (0.98) — Day 11+ 0.305** — (2.67)** 0.533** Alpha 0.533** (101.05)** (101.04)** Number of obs. Time Fixed Effects Zip Fixed Effects R-Squared 454,953 Number of obs. 454,953 Y Time Fixed Effects Y Y Zip Fixed Effects Y 0.0713 R-Squared 0.0713 Table 1-8 includes only urban zip codes. These two regression are defined as a Type I regressions, using data including all ordinary school days from 1980-2001, and zip code fixed effects. Both regressions use Total Crime as their dependant variable. Z-statistic s are given in the parentheses for every table. Indicates significance at the 5% level, ** is significant at the 1% level or smaller

PAGE 50

39 0 .05 .1 .15 .2 PctArrests 0 1 2 3 4 5 6 7 8 9 10 11 12 13 14 DaysToCapture 0 .05 .1 .15 .2 PctArrests 0 1 2 3 4 5 6 7 8 9 10 11 12 13 14 DaysToCapture Figure 1-1: Average Percent of Total Arrests as Function of Days between Offense and Capture A) NonStrike Days for Zip C odes Where Strikes Occur, B) Strike Days for Zip Codes Where Strikes Occur A B

PAGE 51

40 Table 1-10: Effects of Strike Days with Specific Years Dropped from the Sample Year Dropped Strike Coefficient Year Dropped Strike Coefficient 1984 0.164** 1993 0.166** (3.39)** (3.41)** 1985 0.093 1994 0.163** (1.70) (3.28)** 1986 0.166** 1995 0.166** (3.43)** (3.40)** 1987 0.160** 1996 0.157** (3.16)** (3.23)** 1988 0.163** 1997 0.153** (3.37)** (3.16)** 1989 0.148** 1998 0.156** (3.02)** (3.23)** 1990 0.189** 1999 0.157** (3.77)** (3.26)** 1991 0.229** 2000 0.144** (3.18)** (2.99)** 1992 0.157** 2001 0.150** (3.24)** (3.11)** Without Seattle 0.133* Seattle Only 0.153 (2.04)* (1.82) Ave.Number of obs. Time Fixed Effects Zip Fixed Effects Ave. R-Squared 358274 Y Y 0.0603 358274 Y Y 0.0603 The top half of this table is a regression of Strike s on Total Crime with each year systematically dropped from the sample. Each regression includes only urban zip codes. All 18 regressions are defined as a Type I regressions, using data including all ordinary school days from 1984-2001, and zip code fixed effects. The bottom half of the table reports the effect of strike days, for the same Type I sample, with Seattle dropped from the sample, and with Seattle making up the entir e sample. The bottom of the table shows the average number of observations for the top 18 regressions, as well as the average R-squared term. Z-statistics are given in the parentheses for every table. Indicates sign ificance at the 5% level,** is significant at the 1% level or smaller.

PAGE 52

41 Table 1-11: Average Effect of Randomly Generated Strike Days on Total Crime Dependant Variable: Total Crime Random Generations of Strike Days Ave. Random Strike Coefficient Average Number of Strike Days Number of Trials 0.002 (0.05) 2,355 25 Number of obs. Time Fixed Effects Zip Fixed Effects Ave. R-Squared 1,703,910 Y N 0.0380 Table 1-12: Effects of Strike Days by Offender Type with the Last 4 Years of the Sample Dropped Dependant Variable Strike Coefficient for One-Time Offenders Strike Coefficient for Repeat Offenders Total Crime Total Crime without 1998-2001 0.382** 0.156** (4.05)** (2.77)** 0.383** 0.128* (4.08)** (2.31)* Number of obs. Time Fixed Effects Zip Fixed Effects 129,593 129,593 Y Y Y Y Table 1-11 is the average strike coefficient over 25 trials using a randomly generated strike variable. It includes the full data set. Table 1-12 is a Type V regression using just the urban subsample, and includes zip code fixed effects. Z-statistic s are given in the parentheses for ever y table. Indicates significance at the 5% level, ** is significant at the 1% level or smaller

PAGE 53

42 CHAPTER 2 THE ROLE OF TEACHER NETWORKING IN TEACHER TRANSFER DECISIONS AND TEACHER MOBILITY 2.1 Introduction In recent years, there has been growing research to support the notion that teacher movement is an important topic, and not one to be overlooked. Thus far, research on teacher sorting behavior has shown that st udent characteristics seem to influence a teacherÂ’s transfer decision (Hanushek, Kai n, Rivkin 2001). More specifically, teachers sort away from relatively low-income, low-achieving schools. As a result, less experienced teachers are more typicall y found in poor, nonwhite, low-performing schools, particularly those in urban areas (Lankford, Loeb, Wyckoff 2002). This is most important because teacher experience seems to be directly linked to student achievement (Hanushek, Kain, Rivkin 1998, 1999)1. In addition to student char acteristics, other characteristics (such as salary differences) have been studied in an effort to better understand the factors that influence teacher movement. Research aimed at identif ying specific characteristics and conditions that promote teacher mobility have given mo re insight into the questions of when and why teachers move, however researchers stil l do not have the complete picture on how teachers move and why. 1 Hanuchek, Kain and Rivkin (1998) show that variatio ns in teacher quality accoun t for 7.5% of the total variation in student achievement. In addition, teache r quality differences tend to outweigh school quality differences considerably.

PAGE 54

43 The focus of this paper is to better unders tand the nature of teac her transfers, with special attention given to teacher intradis trict transfers from school to school. Specifically this paper will identify the us e of Master Inservice workshops aimed at professional development as a powerful means of building teacher networks. Ultimately I test the hypothesis that Inservice work shops give teachers greater opportunities to network with teachers from other schools in thei r district, and theref ore result in a higher rate of teacher transfer. The evidence will s how that a one standard deviation increase in workshop attendance contributes to a 12.5% increas e in teacher transfer rates per school. Thus teacher networks are a crucial element in influencing teacher transfer rates. Section 2.2 of this paper will provide a c ontext for the basic assumptions outlying the arguments set forth, Section 2.3 will specif y the model therein, Section 2.4 reports the preliminary results, Section 2.5 will check the validity of the results, and Section 2.6 will make the concluding remarks. 2.2 Discussion of Networking 2.2.1 How Teachers Network The idea that professionals develop networks is not a new concept in the social sciences. The term “networking” span s across many professions and takes many different forms. Like many professionals, t eachers also engage in networking for various reasons I will address later in this section. First I will address th e issue of how teachers build networks. Clearly some teacher networking takes plac e in private. Teachers may spend recreational time together getting to know one another, or they may talk over lunch. However these examples of the netw orking process, though valid, are not feasibly quantified or qualified. In order to quantify teacher networks, I rely on teacher Inservice activity to determine if teacher networks aid in mobility. Specifically I focus on the

PAGE 55

44 number of hours teachers spend in Inse rvice activity as my measure of network size/strength. I argue that Master Inse rvice workshops, a part of the teacher recertification process, provide an effective forum for networking. To understand why this is the case, it is necessary to first revi ew how the re-certifi cation process works. Every teacher in the state of Florida is required to have an up-to-date license that enables him or her to teach in Florida public schools. This license is earned by every teacher at the beginning of his or her career and must be re-c ertified every five years. Teachers earn their re-c ertification through any combinati on of three things: they may take state-issued Florida Subject Area Exam in appropriate fields to prove their level of competency, they may enroll in and complete required coursework (6 credits minimum) at an approved university or college, and/ or they may choose to complete up to 120 hours of Master Inservice courses/workshop s offered by the school district. This state mandated re-certi fication process provides an ideal environment for testing the effects of teacher networking. Teachers who take the Florida Subject Area Exam in their appropriate fields or who decide to take classes at a College or University have little to no interaction th eir colleagues in their school dist ricts. However, the Master Inservice option is quit e different from other re-certifi cation options. First, Inservice courses/workshops are almost exclusively made up of other teachers, all within the same district and from differing schools.2 Thus teachers interact with others who share a common profession, and often common interest s. Secondly, the environment of most workshops is designed in such a way that encourages interacti on among participants. 2 It is possible that a marginal num ber of administrators, county officials, etc. attend these teacher components for other reasons (professional or personal). They are not included in the Inservice data in this paper.

PAGE 56

45 Workshop assemblies are long in duration (often several hours per session) and can span weeks or months. In addition, individuals ar e often called upon to sh are experiences with all the participants, or present material re garding how the workshop has influenced their classroom activity. Thirdly, Inservice is an attendance based program, as opposed to the graded college coursework or graded test format. One could argue that the relatively stress-free nature of the Inservice allows teachers to spend more time getting to know each other, free from the distr action of grade requirements. 2.2.2 How Teachers Use Their Networks There are many reasons why any professi onal or businessperson would network with others in their field of work. Th ey yearn to share ideas, understand their competition, and exchange information that can benefit their personal interests, sometimes in a mutual way. Why teachers netw ork with each other is not so different. They exchange classroom experiences and teaching methods, information on workplace environments, as well as build personal bonds that enhance their feelings towards the profession, or even their persona l lives. As a result they could possibly use their networks to plan intradistrict school events, find work in the summer, or even to socialize. However, there are specific applications of strong teacher networks that would enhance mobility within a district. First, te achers use their networks to gain information about the various student and administrative ch aracteristics of schools in their district. Greater access to this information lowers the cost of job searching and enhances mobility. Secondly, teachers use their networ ks to reveal their teacher quality in a less costly way for the purpose of being hired at other schools within their own distri ct. If teachers do use their networks to sort into preferred school s in a more efficient, less costly way, then

PAGE 57

46 there are obvious implications for student out comes, since others have successfully linked student outcomes to teacher mobility. 2.3 The Basic Model A teacher’s desire to transfer to school n (new school) is based on the utility they receive at school n minus the utility they receive at school o (old school). The utility at the old school is given by: Vo(so, qo, ao, do, o) where o indexes the old school, so is the salary of the old school qo is a measure of student/classroom quality, ao (administration quality), do (distance from school to home), and o (opportunity cost of teaching). Similarly the utility at the new school is given by: Vn(sn, qn, an, dn, o) where n indexes the new school, sn is the salary of the new school qn is a measure of student/classroom quality, an (administration quality), dn (distance from school to home), and o (opportunity cost of teaching). Teachers in cur explicit costs to movement. These costs arise from the search and application pro cess. More specifically, teachers may have to submit paperwork, spend some time interv iewing, or generally gather information about the school they are transferring to. Thus I write this teacher decision as: Transfer when: Vn – Vo – C(in, rn, nj*) 0 Do not transfer when: Vn – Vo – C(in, rn, nj*) < 0 In this case C( ) represents the explicit cost borne by the teacher, where in is the amount of relevant school information a teacher desires to gather, rn represents district and school requirements for transfers (paper work, interviews, etc.), and nj* representing the network

PAGE 58

47 size of the jth individual. I further note the relationship of each parameter to this explicit cost function. It is reasonable to assume that the less a teacher knows about a prospective school, the more time he/she spends gathering relevant informati on. Therefore we expect an increase in necessary or relevant information (in) to translate to a direct increase in C( ), such that C/ in > 0. In addition an increase in dist rict and/or school requirements for transfer will increase time and resour ces spent towards transfer, such that C/ rn > 0. When looking at network size it becomes appare nt that a larger network size will lend itself to better flow of school information, pos sibly fewer or shorter interviews, and more inside information that helps one to eliminat e unnecessary measures. Therefore it seems logical that a larger network will decrease external cost, such that C/ nj* 0. Lastly we must qualify the network to external cost relatio nship as not being strict. This allows for the possibility that for school n one’s network is totally ineffective. That is to say that one’s network is not the releva nt network, i.e., a network in another state. From these equations it is clear why teachers may be more likely to transfer as their relevant network size increases 2.4 Data and Empirical Study My data for my study comes from a variet y of sources. I utilize individual school data taken from the “Florida School Indicators Repor t” (FSIR) and the “School Advisory Council Report” (SACR). Countywide data was also supplied in various forms from the Florida State Department of Education (F LDOE). These forms include the “FLDOE State Survey #2”, “Teacher Exit Interview In formation”, an “Inservice Hours Report”, “The Profiles of Florida Distri ct Schools (Students and Sta ff Data)”, and “Teacher Salary,

PAGE 59

48 Experience and Degree Level”. The samp le includes 67 observations, one for each Florida school district. Because several variables introduced in th e following section are highly collinear to district size, such variables have been defined on a per school basis.3 2.4.1 Defining the Parameters The dependent variable for my study ( Transfers) is the number of transfers within a district per school. In or der to approximate the closen ess of different school job options, or capture the density of school job choice in a di strict, I use the total area, expressed in sq. miles, divided by the number of schools in the district ( School Distance ). Since a denser school di strict gives a mobile teacher more viable, less costly job options, I expect this variable to be positively correlated with the total number of transfers per school. Student/classroom quality is measured by taking the standard deviation of the percentage of free lunch eligible students among all schools in a given district. This measure ( School Difference ) should capture the variation of the percentage of the student body that is of low socio-economic status (SES) across schools. In essence, this variable measures how schools in a district differ along family backgrounds. Since greater variation in work environments should give teachers a greater incentive to sort, this variable should be positively corr elated with transfers per school. Wage differences may also play an importa nt role in teacher transfer decisions. There are no wage differences among public schools in a given district, but wage differences among districts still exist and ar e important to account for. As such, Salary 3Transfers Network Dislike Supervisor Principal Moves School Changes

PAGE 60

49 Difference is defined as the average salary of the highest paying district directly adjacent to the current district minus the average sala ry of the district in which a teacher is currently employed. If a teacher considers sa lary difference in a move between districts, it is plausible that they may consider the firs t-best alternative salary difference above all else. If higher salaries provide incentives fo r teachers to leave thei r district rather than move within district, then there should be less movement between schools within district. Therefore this variable should be negativel y correlated with within-district movement. In addition to the interdistrict salary di fference, a wage concern for teachers should also be the wage forgone from remaining a teacher. To capture the opportunity wage forgone, I construct a variable ( Wage Ratio ) using a county-level wage index. This wage index reflects the average wages across a predetermined set of occupations for each county. By dividing average salary for a dist rict by its county wage index, I measure how well teachers in a county are paid relative to other occupations. The larger the Wage Ratio the less appealing other jobs are. As a result this variable should be positively correlated with teacher transfers. My measure of network size ( Network ) equals total hours spent in Inservice per school.4 Since more hours spent in Inservice sh ould lead to larger and more effective networks, this variable should be positively co rrelated with teacher tr ansfers per school. It is also true that the Network variable is positively corr elated with larger school size. This is a potential problem since sc hool size, measured in average number of teacher faculty per school, is positively correlated with district size itself. That is to say 4Component hours geared towards non-instructional staff (i.e., food service employees, transportation staff, administrative staff, substitute teacher s, etc.) are not included in this measurement. Since each district adheres to a statewide numbering standard for component identification, I can be sure that I only include components which were teacher-oriented.

PAGE 61

50 that as school districts grow, the size of th e schools as measured in total faculty also grows. Since teachers in larger districts (with larger schools) have more positions to choose from, they should have greater transfer opportunities. To prev ent this relationship from biasing any estimates of the Network variable, I create a vari able to control for the effect of larger school size. This variable, School Size, measures the average teaching staff per school in the a district. New public schools opening and old public sc hools closing also present teachers with new choices in making transfer decisions New schools that open in a district may provide up-to-date facilities and technology in the classroom making teaching there more desirable. School closures force teachers to reorganize into other schools or leave teaching entirely. The variable ( School Changes ) captures these el ements of school turnover. This variable is measured by assigning a value of positive one for every school gained and/or lost from the 2000-01 school year to the 2001-02 school year, and then summing the totals for each county and dividi ng by the number of schools. Since school districts where schools open a nd/or close should result in mo re teacher transfers, this variable should be positively co rrelated with transfers. There is another variable emphasized in th is model that is based on administrative movement, specifically movement between schools by existing principals. When principals transfer from school to school w ithin a district, positive teacher/principal relationships may induce principa ls to recruit teachers to wo rk at a new assignment. In essence principals may bring their favorite te achers with them to another school under the expectation/promise of a positive work environment. This variable ( Principal Moves ) is measured as the total number of principa ls who moved from one school to another

PAGE 62

51 school, divided by the number of schools. If principals take teachers with them when they move between schools, then principal transfers should be posit ively correlated with teacher transfers. Another variable that should affect teach er movement within a district is the number of relevant alternatives. It is ar guably difficult for teachers to move between high schools and elementary/middle schools. Obviously high school teachers who work in districts where there is only one high school cannot move to another high school in their district. As a result, one would exp ect that school district s with only one high school will have an inherently lower rate of teacher transfer compared to similar districts with multiple high schools. To account for this disparity I include a dummy variable ( One HS ) that receives a value of 1 when a school district ha s only one high school (zero otherwise). This variable should be nega tively correlated with teacher transfers. The last two variables I include in the re gression help to desc ribe the state of a districtÂ’s preferential treatment towards transferring existing teachers to fill new vacancies when they open, rather than hiring new teachers. Some districts require that open positions are filled with teachers al ready employed before a new teacher may be hired; while some other distri cts give no special consideratio n to existing teachers over new applicants. Since more relaxed or favor able policies towards teacher transfers may enhance teacher mobility, I introduce two dummy variables that describe the priority given to teacher transfers. The first variable ( Transfer PC ) describes a district that affords partial consideration to existing teach ers. This variable takes on a value of 1 when a district gives preferential treatment, but not exclusive privilege, to its teacher transfers (zero otherwise). The second dummy variable ( Transfer CC ) describes a

PAGE 63

52 districtÂ’s policy to necessarily fill a new position with a teacher transfer request before considering a new applicant. This variable takes on a value of 1 when a district offers complete consideration to existing teachers befo re all others (zero otherwise). In both cases, these variables should be positively correlated with tr ansfers per school to reflect the enhanced opportunity for mobility given by preference. 2.4.2 Empirical Testing The analysis begins with a simple OLS regression onto the dependent variable Transfers The results are expressed in Table 21, listed by coefficient with the relevant t-statistic listed in parenthesis. Columns 1 and 2 express the OLS regre ssion with our 10 major variables of interest. Some of these va riables that do not have the predicted sign, in this case School Distance Transfer PC and Transfer CC and are not significantly different from zero. Of the variables that are significant, ( Network School Difference School Changes and School Size ), each has the predicted sign. The positive and significant coefficient of the Network variable in Column 2 seems to suggest that networks exist, and th at in fact they do se rve to enhance teacher intradistrict mobility. As a teacher networks gr ows, teacher mobility rises. This provides evidence that teachers are using their netw orks to gain information and enhance the transfer process. To determine the magnit ude of the impact of networking on transfer rates, I look at the percent change in the av erage number of teacher transfers per school when the amount of networking increases by one standard deviation. I find that an increase in the number of inservice h ours per school by approximately 1000 hours increases the teacher transfer rate per sc hool by approximately 11.49%. These results can be viewed in Table 2-2.

PAGE 64

53 The results in Table 2-1 also suggest that student characteris tics, school size, and schools opening and closing all influence teacher movement. Teachers seem to sort more frequently in districts wher e variation in student quality across schools is high. The variable School Difference has a significant impact in t eacher movement and shows that teachers are more likely to move around, or so rt, when there are sizeable differences among student populations at different school s. According to the regression, a one standard deviation increase in the measure of school heterogene ity implies a nearly 15.25% increase in teacher transfers per school Teachers also seem to be more mobile in districts where schools are larger. The School Size variable is positive and significant, suggesting teachers face greater opportunity to sort when there are more total positions in a school. The coefficient of the School Size variable implies that a one standard deviation incr ease in the average size of a school leads to a roughly 17.85% increa se in transfer rates per school. Besides student and school characteristics, teachers also have greater mobility in districts where schools are opening and closing with greater frequency. The School Changes variable suggests that as schools open a nd close in a district, teacher mobility rises. One would expect that as schools ope n in a district, teachers seek new facilities, new classrooms, etc. Further, as schools clos e teachers are forced to shuffle and sort to other schools so that they may continue teach ing in their district. The evidence suggests that when 5 schools are added in a district with 100 schools, teacher transfers increase by nearly 30% for every school in the district. Column 2 runs the same regression usi ng robust standard errors. Using these robust standard errors helps to ease worries ab out heteroskedasticity in the regression. If

PAGE 65

54 it were true that smaller districts have sma ller measurement error than larger districts (which seems plausible), then r obust standard errors help to control for this problem. We see roughly the same results as the regression in Column 1. School Size and School Difference have the predicted sign and are significant at the 1% level. School Changes also has the predicted sign and is significant at the 0.05% level. Column 3 reports the same regression as in Column 1 except that one additional variable is included. This variable ( Dislike Supervisor ) is a rough measure of teacher dissatisfaction with administra tive oversight in one’s own sc hool. It is calculated for each district as the number of teachers pe r school who cited “dissatisfaction with supervisor” as a reason for voluntarily termin ating their employment on the Teacher Exit Interview from the 2000-2001 school year.5 If teacher discontent with administration leads to greater mobility, then this variable should be positively correlated with teacher transfers. In Column 3, the variable Dislike Supervisor has the predicted sign but is not significant. It suggests that job environments and teacher attitudes are not necessarily an important part of the movement decision. This result seems to defy intuition, however in the next section I will show later that this result is marginal, and that Dislike Supervisor is sensitive to model specification. Column 4 reports the same regression as in Column 3 using robust standard errors. It is also worth noting that the Dislike Supervisor variable is not an ideal measure of overall teacher a ttitudes towards administ ration because it does not include teachers who continued to teach in their own districts and only reflects a self5 This survey includes teachers who co ntinued to teach in other districts or other states, teachers who stayed employed in education departments in and outside of their district, teacher s who left the education profession altogether, those who retired, and teachers who left for private scho ols in and out of their district.

PAGE 66

55 selecting subgroup of teachers. It does however have some benefits. First, since there is no descriptive teacher evaluation of administ ration formally collected by the FLDOE or school districts, nor is there us eable county data of the same na ture, this variable offers at least some approximation of what that data might yield. Secondly, because the teachers in these interviews have already committed themselves to alternative employment, the responses given in these interviews are assu red to be candid and honest. It is unlikely that a teacher respondent fears having their identity becoming known to others in these interviews. Overall, these results seem at least superficially consistent with Hanushek, Kain and Rivkin findings (1999, 2001) th at student characteristics in fluence mobility, and that salary changes at best have only a modest im pact. However none of these variables seem to have a powerful an impact as the School Changes variable. There seems to be a large disparity between the predictive power of this variable, and other variables. To ensure that our results are valid, and to understand w hy each variable induces the kind of change it does, I check the robustness of the data. 2.5 Robustness Checks 2.5.1 Re-examining the Data With respect to the School Changes variable, one observa tion in the dataset experienced an abnormally high rate of school change between the years of 2000-2001 and 2001-2002. This county in particular closed two school facilities to open one, much larger school, that would incorporate all of the existing student population of the other schools, however the small size of this distri ct seems to exacerbate the measurement of the School Changes variable. It is possible that this one observation alone is driving the results of the School Changes variable. If this observation is more of an outlier, then it

PAGE 67

56 may be overstating the overall importance of the variable it influences. An OLS regression analysis without this outlier is reported in Table 2-3. Once the sample is modified to eliminate the possible outlier, the magnitude of the coefficient of the School Changes variable is reduced by almo st half, though the variable remains significant. This doe s not necessarily mean that th is variable is not important, however it does show that perhaps this outlier should be dropped to maintain the integrity of these results. These new results suggest that a one standard deviation increase in school openings/closings leads to a 14.55% increa se in teacher transfers per school. This change is considerably lower than the previous estimate. All other variables in the re gression have the predicted si gn and all of the formerly significant variables ar e still significant. Network and School Size are significant at the 1% level, while School Difference becomes marginally significant at the 10% level. Column 2 reports the same regression with r obust standard errors. Column 3 includes the Dislike Supervisor variable, which is significant at the 10% level. This result suggests that there may be some value to job envir onment and administrative quality that teachers consider in their movement decisions. Finally Column 4 presents the same regression as in Column 3, including robust standard errors. I again assess the impact of each variable independently, by determining the effect of an increase of one standard deviation of each variable. These results ca n be viewed in Table 2-4. With the exception of School Changes there are no substantial differences in th e magnitudes of the coefficients of the other variables. The key variable of interest, Network is positive and signif icant in each regression specification on Table 2-3, suggesting that netw orks are an important factor contributing

PAGE 68

57 to teacher movement. However, to confirm that the results are credible, I further test the robustness of these results. Specifically I address whether the variable itself is endogenous to the system. In the following se ction I test whether the measurement of networking is correlated with some unobserved teacher ch aracteristic(s). 2.5.2 Unobserved Correlation If my measure of teacher networks is correlated with some unobserved teacher characteristic(s), then it may be true that there is no real netw orking taking place. Instead, teachers who complete high levels of Inservice may simply be more likely to transfer for other reasons. If that were true then one should expect that teachers who complete Inservice would also be just as likely to move in ways other than just intradistrict transfers. To test whether the Network variable is reflec ting true networking or is just reflecting self-selection of highly mobile teachers into Inservice, I utilize new dependent variables that depict other types of teacher movement. Private Move equals the number of teachers who claim a move into private schools per school. The second variable, District Move measures the number of teachers moving into other Florida districts per school. The last variable is teacher movement to other states to teach. This variable, State Move equals the number of teachers moving to other states per school in the district. Regression results with these new dependent variables are reported in Table 2-5 using robust standard errors. Columns 1 and 2 report the regressions with Private Move as the new dependent variable, and using robust standard erro rs. Columns 3 and 4 do the same with District Move as the new dependent variab le, and Columns 5 and 6 have State Move as the new dependent variable. In each of these specifications, networ k size does not seem to be a significant factor influencing te acher movement. Since Inservi ce is not significant to any

PAGE 69

58 of these different kinds of movement, it seem s plausible that teachers who tend to move around more generally, are not se lf-select into attending a grea ter amount of Inservice. Thus I can be more confident that the positive effect of networking is not simply a reflection of the preferences of more mobile teachers. 2.6 Conclusion The evidence set forth in this paper s upports the argument that Master Inservice components provide an effective environment for teachers to network with each other. These networks seem to provide a basis for teachers to gather important information about schools and allow teachers access to tran sfer options that mi ght have previously been closed to them. Overall, after c ontrolling for determinants of mobility and eliminating biasing outlier effect s, I show that an increase of one standard deviation in Inservice hours causes roughly a 12.74% increase in teacher transfers per school. The results suggest that teachers exploit these networks to sort into schools that they find desirable. If teachers can more effectively so rt away from less desirable schools, such as low-income or failing schools, into more desirabl e schools, then it is easy to conceive that worse schools will be worse off as teacher networks improve. Of course the goal of the Inservice plans set forth by school dist ricts is to improve teacher quality. However, the unintended consequences of such staff development policies that encourage, or at least facilitate, teacher networking seems to exist as well. Further, the consequence of this particular policy seems to be quite substantial in promoting a potentially harmful action such as teacher transfer. This result may have considerable meaning for the future of public education policy. State officials must begin to weigh the possible negative effects of incr eased teacher networking with the potential

PAGE 70

59 (and realized) benefits of Inservice, so that such policy analysis can be conducted in the future with a complete understanding of social welfare consequence.

PAGE 71

60 Table 2-1: Ordinary Least Squares Regressi on onto teacher transfers using full sample with and without robust standard errors Variable Column 1 Column 2 Column 3 Column 4 Dependent Transfers Transf ers Transfers Transfers Network 0.0003261 (1.56) 0.0003261 (2.00)** 0.0003474 (1.67) 0.0003474 (2.19)** School Distance 0.0065532 (1.12) 0.0065532 (1.16) 0.0078015 (1.33) 0.0078015 (1.33) Principal Moves 0.3216886 (0.12) 0.3216886 (0.12) 0.9270341 (0.34) 0.9270341 (0.33) School Changes 17.86031 (4.88)** 17.86031 (3.42)** 17.55322 (4.84)** 17.55322 (3.20)** School Difference 0.0584747 (2.10)** 0.0584747 (2.34)** 0.0557304 (2.01)** 0.0557304 (2.16)** Salary Difference -0.0000891 (1.61) -0.0000891 (1.50) -0.0000766 (1.38) -0.0000766 (1.21) Wage Ratio -0.0018573 (0.27) -0.0018573 (0.22) 0.0001112 (0.02) 0.0001112 (0.02) School Size 0.0501213 (2.46)** 0.0501213 (2.80)** 0.0448708 (2.19)** 0.0448708 (2.96)** Dislike Supervisor — — — — 4.17856 (1.43) 4.17856 (1.17) One HS -0.7262709 (1.32) -0.7262709 (1.27) -0.7582897 (1.39) -0.7582897 (1.31) Transfer Policy PC -0.2608151 (0.65) -0.2608151 (0.64) -0.2026472 (0.51) -0.2026472 (0.48) Transfer Policy CC -0.2381146 (0.57) -0.2381146 (0.67) -0.1001645 (0.24) -0.1001645 (0.25) Constant 0.2509715 (0.09) 0.2509715 (0.08) -0.5541842 (0.20) -0.5541842 (0.19) R-Squared 0.6154 0.6154 0.6295 0.6295 Number of Obs. 67 67 67 67 Robust SE N Y N Y denotes significance at the 10% level **denotes significance at the 5% level

PAGE 72

61 Table 2-2: Percent change teacher transfer s per school for a one standard deviation increase in each significant variable with full sample Variable Mean Standard Deviation Range of % Change School Changes 0.027 0.05 30.06 to 30.58 % School Size 33.418 11.61 17.85 to 19.94 % School Difference 15.945 7.99 15.25 to 16.00 % Network 1758.317 1047.62 11.49 to 12.46 %

PAGE 73

62 Table 2-3: Ordinary Least Squares Regression onto teacher transfers using sample with biasing outlier omitted Variable Column 1 Column 2 Column 3 Column 4 Dependent Transfers Transf ers Transfers Transfers Network 0.0003563 (1.79)* 0.0003563 (2.22)** 0.0003822 (1.95)* 0.0003822 (2.46)** School Distance 0.0010389 (0.18) 0.0010389 (0.20) 0.0021892 (0.37) 0.0021892 (0.41) Principal Moves 0.3430332 (0.13) 0.3430332 (0.12) 1.147493 (0.46) 1.147493 (0.36) School Changes 9.11579 (1.90)* 9.11579 (1.91)* 8.322631 (1.76)* 8.322631 (1.73)* School Difference 0.0433008 (1.60) 0.0433008 (1.74)* 0.0393933 (1.48) 0.0393933 (1.55) Salary Difference -0.0000671 (1.26) -0.0000671 (1.06) -0.0000516 (0.97) -0.0000516 (0.79) Wage Ratio -0.0000751 (0.01) -0.0000751 (0.01) 0.0020135 (0.30) 0.0020135 (0.30) School Size 0.0502127 (2.60)** 0.0502127 (2.89)*** 0.0442077 (2.29)** 0.0442077 (2.95)*** Dislike Supervisor — — — — 4.78275 (1.74)* 4.78275 (1.52) One HS -0.4594238 (0.86) -0.4594238 (0.94) -0.4825945 (0.92) -0.4825945 (0.96) Transfer Policy PC -0.2247365 (0.59) -0.2247365 (0.55) -0.1563356 (0.41) -0.1563356 (0.37) Transfer Policy CC -0.3207089 (0.81) -0.3207089 (0.89) -0.1669839 (0.42) -0.1669839 (0.42) Constant 0.0560173 (0.02) 0.0560173 (0.02) -0.8754046 (0.33) -0.8754046 (0.33) R-Squared 0.6033 0.6033 0.6247 0.6247 Number of Obs. 66 66 66 66 Robust SE N Y N Y

PAGE 74

63 Table 2-4: Outlier excluded percent cha nge teacher transfers per school for a one standard deviation increase in each significant variable Variable Mean Standard Deviation Range of % Change School Size 33.696 11.48 17.40 to 19.73 % Network 1777.904 1043.22 12.73 to 13.65 % School Difference 15.881 8.03 10.84 to 11.91 % School Changes 0.023 0.04 0.37 to 11.36 % Dislike Supervisor 0.027 0.05 8.78 %

PAGE 75

64 Table 2-5: Ordinary Least Squares Regression of other types of teacher movements with outlier excluded Variable Column 1 Column 2 Column 3 Column 4 Column 5 Column 6 Dependent Private Move Private Move District Move District Move State Move State Move Network 0.0000101 (1.21) 0.0000104 (1.22) 0.0000342 (0.49) 0.0000305 (0.43) 0.0000014 (0.06) 0.0000034 (0.01) School Distance -0.000214 (1.09) -0.000196 (1.02) 0.0024853 (0.95) 0.0021902 (0.85) -0.000891 (1.36) -0.000813 (1.22) Principal Moves -0.001503 (0.02) 0.0073589 (0.07) -1.829443 (2.09)** -1.930217 (2.01)** -0.190106 (0.53) -0.142873 (0.39) School Changes 0.125987 (1.11) 0.1214916 (1.07) 4.121108 (1.44) 4.236288 (1.46) 0.6497598 (0.90) 0.5957743 (0.78) School Difference -0.001018 (0.73) -0.001058 (0.74) 0.0119132 (0.99) 0.0124806 (0.99) 0.0015131 (0.44) 0.0012472 (0.35) Salary Difference -0.000001 (0.43) -0.000001 (0.31) -0.000001 (0.06) -0.000003 (0.18) -0.000002 (0.37) -0.000001 (0.20) Wage Ratio 0.0002086 (0.96) 0.0002341 (1.02) -0.006608 (2.47)** -0.006912 (2.35)** -0.002273 (2.41)** -0.002131 (2.18)** School Size -0.000328 (0.42) -0.000405 (0.51) 0.0019503 (0.32) 0.0028223 (0.47) 0.0017899 (0.76) 0.0013812 (0.58) Dislike Supervisor — — 0.0611693 (0.66) — — -0.694535 (0.70) — — 0.3255332 (1.10) One HS -0.014332 (1.04) -0.014801 (1.08) 0.2299846 (0.87) 0.2333494 (0.88) 0.0774991 (0.95) 0.0759221 (0.91) Transfer Policy PC 0.0199595 (1.06) 0.0197389 (1.05) -0.111869 (0.73) -0.121802 (0.81) -0.007048 (0.17) -0.002392 (0.06) Transfer Policy CC 0.0022509 (0.12) 0.0042704 (0.21) -0.174603 (1.20) -0.196926 (1.29) -0.050714 (1.04) -0.040251 (0.78) Constant -0.041984 (0.50) -0.053770 (0.59) 2.621647 (2.52)** 2.756905 (2.38)** 1.011499 (2.84)** 0.9481022 (2.54)** R-Squared 0.1931 0.1977 0.3416 0.3461 0.2398 0.2511 Number of Obs. 66 66 66 66 66 66 Robust SE Y Y Y Y Y Y Outlier N N N N N N

PAGE 76

65 CHAPTER 3 WITH A LITTLE HELP FROM MY FRIENDS: EVIDENCE OF TEACHER NETWORKS USING MICRO DATA 3.1 Introduction In the last several years, the issue of teacher mobility has come to the foreground of discussions about the public education. Pa rents and families who value teacher quality have come to realize that teacher mobility is a topic that has pot entially far reaching consequences, not just for education labor ma rkets, but also in terms of levels of education and student achievement. Schools that struggle with teacher retention often find job vacancies difficult to fill, and may compensate by hiring less than fully qualified teachers, expanding class sizes, canceling cour se offerings and assi gning teachers from other subject areas (NCES 1997). Of course these kinds of actions may have adverse effects on student learning and achievement. As a result, economists and policy makers have done more to explore the issues of t eacher attrition and movement within school districts. The goals of such recent studies have been to shed light on the questions of which teachers are the most mobile, how do these teachers move, when do these movements take place, and what are the potential implications for students. The most recent studies focused on teacher transfer behavior and sorting within education have provided evidence that supports the notion that stude nt characteristics and school quality are important factors in a teach erÂ’s transfer decision (Hanushek, Kain, and Rivkin 2001). This also seems consistent wi th teacher self-reported information. In the 2000-2001 Teacher Follow-up Survey, administered by the National Center of Education

PAGE 77

66 Statistics, 32% of teachers reported poor wor kplace conditions as a primary reason for the movement from one school to another. Rese arch shows that teachers sort away from low-income, low-achieving schools. Conseque ntly, less experienced teachers are more typically found in poor, nonwh ite, low-performing schools, pa rticularly those in urban areas (Lankford, Loeb, and Wyckoff 2002). For policymakers, this can be a troubling re sult. If more experienced teachers tend to sort away from poorly performing schools, then it is the students at these schools who stand to lose the most, because experien ced teachers have been found to be more effective than novice teachers in terms of higher student achievement (Hanushek, Kain, and Rivkin 1998, 1999).1 In addition, urban schools a nd highly urban school districts face high rates of teacher tur nover (Imazeki 2003). If teache rs are sorting away from these areas, then these urban districts may ha ve the most difficult time filling vacancies with replacement staff. Aside from student characteristics, econom ists have also studied the potential effects of salary changes on t eacher mobility. So far researchers have found that salary levels have only a modest impact in terms of student achievement and almost no impact on teacher mobility (Hanushek, Kain, and Rivk in 1999). However, others have also found that salary can influence a teacherÂ’s m ovement choice in ways other than strict transfer, such as length of stay in te aching (Murnane and Olsen 1990). Although much has been learned about teacher mobility, resear chers still do not have the complete picture on how teachers move and why. 1 Hanushek, Kain and Rivkin (1998) s how that variations in teacher qu ality account for 7.5% of the total variation in student achievement. In addition, teache r quality differences tend to outweigh school quality differences considerably.

PAGE 78

67 The focus of this paper is to better unde rstand the nature of teacher transfers from school to school by introducing the idea of t eacher networks as an effective sorting mechanism for public school teachers. To st udy this issue I rely on information provided by the 1999-2000 Schools and Staffing Survey (SA SS) put out by the National Center for Education Statistics. The individually detailed and specific nature of the dataset not only allows me to consider how individual networ ks affect mobility but also allows me to control for a well-defined set of variables, gi ving my study a more complete and detailed look at networks. This paper will begin disc ussion on the notion of teacher networks, and will address how these networks influence teacher movement and to what extent. Specifically this paper descri bes how professional developm ent activities (PDAs) provide opportunity for teachers to netw ork with other teachers. PD As will be shown to have a sizable impact on intradistrict teacher mobility. 3.2 The Identification Strategy Before I describe how I identify teacher ne tworks, let me first discuss what teacher networks are, how they are useful, a nd how teachers build these networks. 3.2.1 What Are Networks And How Are They Useful The term network in this paper is define d as group or “family” of co-workers or colleagues within one’s field of work. Specifically, in this paper a network refers to a teacher’s associations with other teachers w ho work within their own district. Although the context of a network here is limited to the field of education, people from all professions form networks for various reasons. There are many reasons why any professi onal or businessperson would network with others in their field of work. Th ey yearn to share ideas, understand their competition, and exchange information that can benefit their personal interests,

PAGE 79

68 sometimes in a mutual way. Why teachers netw ork with each other is not so different. They exchange classroom experiences and teaching methods as well as information on workplace environments, as well as build pe rsonal bonds that enhance their feelings towards the profession, or even their personal lives. As a result, teachers may be able to use their networks to plan intradistrict school events more efficiently, find work in the summer time, or sometimes just to socialize w ith others in their field. However, there are two key uses of teacher networks that are the focus of this paper. First I argue that teachers use their networks to gain greater information about school characteristics. Because there are many characteristics about schools that may be difficult to observe (existing job vacanci es, quality of administrative support, departmental environments, etc.), networks may allow teachers to evaluate other schools in their district more efficiently. Greater information about schools should allow teachers to sort into their pref erred school more easily. Secondly, I argue that teachers use their networks to reve al their quality as a job candidate in a less costly wa y. Because it can be costly (or sometimes impossible) to reveal oneÂ’s aptitude for a job vacancy, t eachers may call on their networks to provide important or relevant information to potential employers. For example, a teacher may ask another teacher to put in a favorable word to a hiring principal. In this way, networks provide a valuable advantage to those seeking employment, while at the same time being a more credible or reliable source of info rmation for potential employers. If teachers do use their networks to foster more efficient, less costly sorting behavi or, then policies that encourage network building may have unint ended outcomes for teacher mobility and possibly student achievement.

PAGE 80

69 3.2.2 How Do Teachers Build Networks? Clearly some teacher networking takes pl ace in private forums. Teachers may spend recreational time together getting to know one another, or they may talk over lunch. However these examples of the networ king process are not f easibly quantified. To find a more measurable environment where networking takes place, I focus on professional development activities. Profe ssional Development Activities (PDAs) consist of an array of different activities geared towards maintaining and enhancing teacher competency and knowledge of various issues (including such issu es as the use of technology in the classroom, st udent assessment and state ed ucation standards, teaching methods, etc.). PDAs can include many different types of activities such as workshops, mentoring and/or peer observation, collabo rative research, University coursework, etc. Table 3-1 reports a summary of self-reported profe ssional development activity from the 1999-2000 Schools and Staffing Survey.2 Though each of these activities may be available to most teachers in the sample, attending workshops/c onferences/training clearly seems to be the most popular form of PDA in the sample, w ith nearly 94% of teachers reporting they have engaged in this form of PDA in the pa st year (prior to the 1999-2000 school year). PDAs provide an ideal environment for t eachers to network with one another. Teachers who participate in workshops or c onferences have greater opportunity to meet teachers from other schools within their distri ct and enhance their networks. In addition most PDA is geared to a specific employm ent group within the school system. This 2 The 1999-2000 Schools and Staffing Survey (SASS) is a survey of involving approximately 56,000 public school teachers from all over the U.S., and is produced by the National Center of Education Statistics. It is the main data source for this chapter and is described further in Section 3.3.

PAGE 81

70 means that administrative personnel do not usuall y engage in the same PDA as teachers. Likewise teachers do not usually engage in th e same PDA as other s upport staff (such as bus drivers, cafeteria workers, etc.). Thus I can be somewhat confident that if teachers are building networks through PDA, they are generally relevant networks. Besides simply meeting other teachers th rough PDA, these activities often provide an environment conducive to promoting teache r interaction. Workshops especially are designed in such a way that encourages interaction among participants. Workshop assemblies are usually long in duration (often several hours per session) and can span weeks or months. In addition, individuals ar e often called upon to sh are experiences with all the participants, or present material re garding how the workshop has influenced their classroom activity. 3.2.3 Measuring Teacher Networks In order to measure the size of a teacherÂ’s network, I utilize national survey data that provides information on professional development activity. The 1999-2000 Schools and Staffing Survey (SASS), the survey from which the overall data set is derived, asks respondents to report the number of hours they spent in e ach of six major categories3 of PDA within the last school y ear (1998-99). Rather than re port the specific number of hours, respondents are asked to select a (pre determined) range of hours that includes the actual number of hours they have spent in each category of PDA. Using the midpoint of the reported range as my best estimate of repor ted hours, I create a co ntinuous variable of 3 These six categories are: PDAs w ith focus on 1) in-depth study of main assignment, 2) content and performance standards of main assignmen t, 3) methods of teaching, 4) uses of computers for instruction, 5) student assessment/methods of testing, and 6) student management in the classroom.

PAGE 82

71 PDA Hours ( Hours) by summing the reported hours acro ss all six PDA categories for each individual.4 In general the distribution of PDA Hours across my sample is skewed to the left with approximately 3.1 % of all teachers repo rting that they have not engaged in any form of PDA within the past year. Bo th the level of skewness of the sample (approximately1.30) and percent of teachers re porting no PDA within the past year are roughly equivalent across m ovement categories as well.5 In addition, the distribution of PDA Hours across these movement categories is also roughly identical. Of those teachers who did not move, 24.8 % (or 332) are in the upper quartile of the entire distribution. Likewise 25.2 % of teachers (or 88) who transfe rred within district were in the upper quartile of the enti re distribution of PDA Hours.6 It is also worth noting that in each individual PDA category, only a small fr action of teachers are top-coded (reporting the maximum number of hours allowed in the survey).7 Although for each PDA category the percent of teachers who ar e top-coded ranges from 16.4 % to 2.1 %, these differences do not vary significantly across movement types. Overall the data seems to indicate that 4 It is important to note that the highest range of hour s that teachers could report is “33 or more”. Where this happens, I use 60 hours as the maximum value in this range. 60 seems to be a reasonable upper limit of hours, however I have also tested maximum values at 48 and 64. These chan ges do not significantly impact the results. 5 The PDA distribution of t eachers who move intradistrict is slightly more skewed (1.37) than the same distribution for those who did not move (1.25). Additionally, the kurtosis of the PDA distribution for all movement categories is nearly equivalent (4.34 for non-movers, 4.64 for intradistrict movers, and 4.61 for those who leave thei r district). 6 Of those teachers who moved out of their district, 21.7 % (or 98) are in the upper quartile of the entire distribution. This includes both interdistrict and interstate movement. 7 There is only one teacher in my sample who is to p-coded for every category of PDA (and has thus taken the maximum amount of PDA measured in the survey)

PAGE 83

72 along the PDA measurement, teachers in va rious movement categories do not look different enough from one another to warrant an y concern these groups are incomparable. 3.2.4 The Basic Model In this basic model I am attempting to explain how a teacher’s decision to transfer to another school (in year t ) within their own district is affected by their network. My measure of a teacher network here is the num ber of hours a teacher spent in professional development activities in th e previous year (in year t – 1). I can start by expressing a simple regression model in the following form: Transfer(t) = + (PDA Hours(t – 1)) + X Here the dependent variable is a dummy th at expresses a teacher’s transfer within his own district. The dependent variable (Tra nsfer) is assigned a value of 1 for a within district move, and 0 otherwise. In additi on to the amount of networking a teacher does, the basic characteristics of th e most mobile teachers are somewhat different from those who are less mobile. To capture these diffe rences I include variable X, a vector of individual and school characteri stics that influence a transfer decision. Table 3-2 reports some descriptive statistics about the charac teristics of teachers who do and do not move around in various ways. The most notable di fference is that teach ers who move within district tend on average to be about 3-6 years less experi enced (and younger) then those who do not move. Since inexperienced teachers are generally not as far along the career path as more experienced teachers, this di fference seems intuitive. In addition, I must also consider marital status and job status in the model. Those teachers who are not married may find a move to a new school to be more costly since spousal relocation can complicate matters. Those teachers who are employed part-time by their district should

PAGE 84

73 have a greater overall benefit to searching fo r job vacancies since they stand to benefit from becoming a full-time employee. Aside from the natural differences that exist among teachers, a teacher’s transfer decision should also depend on the working co nditions he/she faced the previous year. Specifically, a teacher’s decision to tr ansfer may be influenced by student quality/characteristics, administrative quality/c haracteristics, or comp ensation levels. As a measure of student quality ( L ), I construct a measure: %Free Lunch Eligible = (%free-lunch eligibleschool,(t – 1) %free-lunch eligibledistrict,(t – 1)) This measure is the difference between the percent of the student population who are free-lunch eligible at a teacher’s school and the average percent of the student population who are free-lunc h eligible at the dist rict level. It is designed to capture the socio-economic status (SES) of the student popul ation at a teacher’s in itial school relative to the SES of the district as a whole. Give n that teachers desire to sort away from low SES schools, this variable should ha ve a positive impact on transfers. To measure administrative quality ( Adm ), I utilize the variable: Poor Administration = a dummy variable, receives a valu e of 1 is a teacher reports that administrative support is poor within the past year ( t – 1 ) This variable captures a teacher’s per ception of how supportive/encouraging the administration at their school is towards its st aff. The predicted sign of this variable should be positive, since in schools where admi nistrative quality is perceived to be poor, teacher mobility should be higher.8 Some additional measures of the working conditions in my regressions specification include: 8 I have also measured poor administrative quality us ing the average teacher response for each school. My results do not change with this different measure.

PAGE 85

74 Student Threat = a dummy variable, equals 1 if a teacher reports being threatened by a student within the past year ( t – 1 ) ( Thr ). Again because receiving a student threat in dicates poor working conditions, the sign of the coefficient should be positive to express teachers’ preferences to sort away from undesirable schools.9 Bonus Pay = a dummy variable, equals 1 if a teacher reports receiving bonus pay, separate from salary and extracurri cular pay within the past year ( BP ). Bonus pay should give teachers an incentive to stay at their existing school, and should therefore be negatively correlated with transfer rates. I also use the self-reporte d data to construct a meas ure of job satisfaction ( Job ) that reflects a teacher’s general attit ude toward their workplace: Job Dissatisfaction = a dummy variable, equals 1 if a teac her reports being strongly dissatisfied with being a teacher at their school in year ( t – 1 ) ( Job ) This variable is intended to capture not ju st the quality of the students, but also a teacher’s perception about the quality of thei r co-workers, their administrative superiors, school resource availability, and general i ssues surrounding their employment. Since greater general job dissatisfaction should lead to greater mobility, I expect the sign of this variable to be positive. While it is true that salary differences between schools seem to affect teacher transfer decisions, this analysis is limited to those teachers whose movement is within their district. Given that sala ries schedules are generally se t at the district level and do 9 I have also measured student threat using the aver age teacher response for each school. My results do not change with this different measure.

PAGE 86

75 not vary across schools in the same district, base salary differences between schools are negligible10. As a result, I do not include base salary as a variable in my analysis.11 I also control for whether a teacher belongs to a union in this model for two main reasons. The first is that union membersh ip may provide existing teachers with preferential treatment in filling a job vacancy. Another reason is that districts may find unionized teachers more difficult to fire. As a result, problem teachers may be shuffled around more often. These two explanations of how unions affect tran sfer rates are very different, however in both scenarios union membership should positively affect the probability of transferring. Finally I add state fixed-effects into th e regression model to control for basic institutional and cultural differences across states, and a dummy = 1 for missing data in the % Free-Lunch ( L) variable.12 It would be ideal to use sc hool district fixed-effects, however, doing this results in a dram atic reduction of my sample size.13 Once this is done, the regression takes the following form: Transfer(t) = + 1Hours(t – 1) + 2L(t – 1) + 3Thr(t – 1) + 4BP(t – 1) + 5PT(t – 1) + 6Exp(t – 1) + 7Adm(t – 1) + 8Mar(t – 1) + 9Job(t – 1) + 10Age + 11Union + 12LDummy + 13StateFE where PT is a dummy for job type (1 if part-time, 0 otherwise), Exp is the total years of full-time experience, Age is teacher age, Union is a dummy for union membership (1 is 10 In the SASS, over 98% of school districts reported having a set salary schedule for teachers in their district. This means that virtually every teacher in the sample who moved within district did not face salary differences between the schools in their district 11 Although other types of compensation such as incentive pay are an important component to one’s decision to transfer, those types of compensati on could be endogenously determined and therefore inappropriate to include. 12 The Free-Lunch variable is the only variable in the regression where observations were missing. Roughly 157 (or 9%) of observations were missing this data. 13 Never the less I show on Table 3-7 that the inclusion of district fixed-effects does not alter the results.

PAGE 87

76 unionized, 0 otherwise) and Mar is a dummy for marriage status (1 is married, 0 otherwise). 3.3 Data The data for this model comes from the 1999 – 2000 Schools and Staffing Survey (SASS) and the 2000 – 2001 Teacher Followup Survey (TFS) produced by the National Center for Education Statistics (NCES).14 The data are comprised of individual (teacher) level survey information that can be matched across surveys, as well as relevant survey information for the relevant schools, principals and school districts. In the original 19992000 SASS survey there were over 52,000 respo ndents, but the TFS consists of only a subset of those orig inal respondents. In aggregate the TFS includes 6,758 teachers, however not all of the teachers in the survey are included in this analysis.15 Since this analysis focuses on public school teachers who move within district relativ e to those who do not change schools, I eliminate those respondents who either leave th eir district, or leave teaching altogether. Later in the paper I will use other types of movement (int erdistrict and interstate movements) to show that PDA hours do not influence other mobility decisions, particularly those decisions wher e networks should not matter. 14 The datasets utilized in this paper contain iden tifiable and sensitive information. While there are versions of these data that are available to the general public, the datasets used in this paper are only available at the discretion of the NCES and with their proper approval. 15 Those respondents in the TFS who le ft the teaching profession after th e 1999-2000 school year (2,374 respondents) are given a version of the TFS that is different from the survey given to those who did not leave teaching.

PAGE 88

77 In addition, I eliminate those teachers who initially teach at private, Indian, or charter schools, teachers who reporte d their move was mostly involuntary16, and those teachers who were not considered full or part time regular teachers in 1999 (e.g., specialists, substitutes, student teachers and itinerant teachers). Once these criteria are imposed on the data set, 1,688 observations remain. Of the 1,688 observations in the dataset, 349 (or approximately 20.7%) of those were teachers who transferred within their district. 3.4 Empirical Evidence Because the dependent variable is a binary choice variable with a successful transfer equal to 1, probit estimation is appr opriate. Column 1 of Table 3-3 reports a preliminary regression of PDA Hours onto th e dependent variable, with the specified covariates also included.17 At first glance it seems th at networking does not have a significant impact on the likelihood of transfer. Networks th ough are a story of collegial kinship, and like any friendship, benefits may not be realized until the parties involved are familiar and comfortable enough with one another to begin a relationship. Thus networks may be difficult to establish initiall y. Further, networks may require a level of maintenance. If that is the case, then t eachers who engage in PDA frequently may have an increasing advantage over those who do not. This may imply that there are increasing returns to hours spent in PDA in terms of networking. A simple test of this assumption is to include a squared te rm of PDA Hours. 16 It is also worth noting that an y involuntary teacher movements who ma y be unavoidably included in the sample (thus creating noisy measures of intradistrict m ovement), are likely to bias any positive estimates of the effect of networking towards zero. 17 All regression estimates are clustered at the district level to control for possible standard error correlation at the district level. In addition, all regression estimates include robu st standard errors, to account for possible heteroskedasticity.

PAGE 89

78 In Column 2 of Table 3-3, the addition of a squared measure of PDA Hours in the regression seems to suggest that there may be some value to networking. Specifically, the coefficient on the squared term is posit ive and significant at the 0.025 level, which may indicate that while ha ving only a few hours networki ng may have no impact on transferring, positive returns to networking may exist for large values of PDA Hours. To further test whether this is th e case, I use a spline to allow the effect of an additional hour of PDA to be greater at high levels of PD A hours then at low levels of PDA hours. The spline estimates a continuous, piecewise-li near relationship between PDA Hours and transfer probability, and is defined as follows: PDA Hours (H) : from [0 – H*] hours: = H if H < H* = H* if H H* PDA Hours (H) : from [H* Hmax] hours: = 0 if H H* = (H – H*) if H > H* Table 3-4 reports the coeffici ents of several regressions using the spline. I test various cutoffs of H*, where H* is chosen at decile intervals al ong the distribution of PDA Hours. Table 3-4 illustrates that the 2nd segment of the spline is positive and significant for most decile cutoffs18, but the best fit for the regression is where the cutoff is at the 80th percentile. This table confirms that additional PDA Hours, particularly for those teachers in the top 20% of the overall distribution of PDA Hours, seem to have a positive and significant impact on a teacher’s tr ansfer decision. Interpretation of the coefficients suggests that a one standard deviation increase in the total number if PDA 18 For every cutoff chosen at or above the 20th percentile of the distribution, the 2nd segment of the spline estimation is positive and significant.

PAGE 90

79 Hours (approximately 46.6) can lead to a 0.020 to 0.078 increase in the probability of transferring within oneÂ’s own school district. Other covariates also seem to have a si gnificant impact of the transfer decision. The evidence seems to indicate that teacher s with less experience and those who work only part-time are significantly more likely to move between schools. Intuitively this seems valid since teachers with li ttle experience have had arguably less time to sort into a desirable school and should therefore be more likely to transfer than teachers who are further along the career pat h. In addition, part-time empl oyees stand to secure higher wages if they search for full-time employment, and have lower time-cost to job searching. Besides the natural sorting taking place for younger/less experienced teachers, teachers who work in low SES schools a nd who face undesirable working conditions clearly have a greater incen tive to sort into better sc hools, especially given that additional/incentive pay is rarely available for teachers to work in less desirable schools. The evidence in Tables 3-3 and 3-4 seem to support this claim. Teachers at schools where student SES as a whole is lower than the district averag e, schools where their overall job satisfaction is generally low, and teachers that feel administrative quality is generally poor, seem to have a higher propensity to sort. Lastly, Tables 3-3 and 3-4 show that ther e is also some advantage to a teacher being part of a teacher union in terms of increased mobility. This could be because teachers who are part of a union receive priority in filling a position vacancy. Another interpretation of this union effect could also be that unionized teachers are difficult to fire, and therefore get shuffled around more of ten. Overall the results shown in Tables 3-

PAGE 91

80 3 and 3-4 are consistent with previous findings in the literature, and not at all surprising given the economics going on. These regressi on specifications provi des some valuable insight into the teacher transfer decision, and also tells us that teacher networks are not an integral component to teacher mobility. 3.5 Robustness Checks The evidence in the previous section sugge sts that there is a significant relationship between teacher networks and transfer behavi or. However it is important to test the robustness of these results. If the time a te acher spends in PDA is simply correlated with some other (possibly unobservable) characte ristic(s), then the explanation provided by networking may be inaccurate. If it is true th at teachers who spend a great deal of time doing professional development activities are s imply more likely to transfer, then the results are spurious in a networking framework. Before I test the robustness of the results in the previous section, I look to see if those teachers who engage in the most pr ofessional development look substantially different from the average teacher. I compar e the characteristics of teachers in every decile of the distribution of PD A hours to the average characteristics of the entire sample of full and part-time teachers (2,280 observations ). The results of these comparisons are reported in Figure 3-1. E ach graph on Figure 3-1 shows the mean of the given characteristic for each interval (0% – 10% 10% 20%, etc) of the PDA distribution, centered about the mean for the entire sample. I use the standard errors at each interval to construct a 95% confidence interval on which the mean lies, and test whether the sample average falls in the range of th e confidence interval. There are only a few characteristics wher e teachers who engage in the most PDA look significantly different from the sample average. The graphs of Age and Full-Time

PAGE 92

81 Experience illustrate that teachers in the top end of the distribution of PDA Hours are significantly older and more experienced. However, previous ev idence suggests that more experienced teachers, who are also generally older, are also less likely to transfer schools. Therefore, if older and more expe rienced teachers are ta king the most hours of professional development, then we should expe ct PDA Hours to be negatively correlated with the probability of transferring, which does not seem to be the case. The graphs in Figure 3-1 may also suggest th at those teachers at the very bottom of the distribution of PDA Hours may be disproportionately white, male part-time teachers, with respect to the entire dist ribution of hours in PDA. Howeve r it is not clear that this presents a problem of selection bias in who d ecides to take PDA as it relates to transfer rates. Currently there is no evidence to sugge st that qualities of teacher race or gender are effective at predicting/explaining teacher movement. In addition, Table 3-4 showed that part-time teachers are more likely to m ove than full-time teachers. If part-time teachers are more likely to spend only a small amount of time in PDA, this should also downwardly bias the estimated effect of PD A Hours on transfers. These results are promising, however they are not conclusive. Teachers may still be sorting into the distribution of PDA Hours based on some characteristic that is not observable. To test whether teachers ar e sorting into high or lo w levels of professional development activity, I begin by looking at other types of teacher moveme nt. If it is true that teachers with high levels of PDA participation are simply more likely to move, then one might expect these same teachers are also mo re likely to move in ways other than just intradistrict transfer. However, since mo st PDAs are activities limited to oneÂ’s own school district, then networks should be la rgely ineffective for movement outside the

PAGE 93

82 scope of the immediate district. So if PDA hours are a strong predicto r of other types of movement, then that may be an indication th at the network relations hip is unjustified. Tables 3-5 and 3-6 report similar regressions similar to those reported in Table 3-4. Table 3-5 shows the effect of a change in PDA Hours on the probability of movement to another school dist rict within one’s state. Ta ble 3-6 shows the effect of a change in PDA Hours on the probability of movement to another school district in another state. In both cases, the effect of increased PDA Hours is not significantly different from zero. For Tables 3-5 and 3-6, neither spline segment is significant, which suggests that the amount of time a teacher sp ends in PDAs has no significant impact on how likely they are to move out-o f-district (within the state), or even out-of-state. This helps to quell some suspicion that teachers with high levels of PDA are simply more mobile. It may also be possible that the rela tionship between PDA and mobility is a function of district characteristic s or behavior. For example, if districts with “naturally” high levels of teacher mobility also demand an unusually high number of PDA hours for their faculty (arguably this may be the case with some urban school districts), then the positive effect of networking is simply a reflection of high levels of mobility in a subset of districts. To account for this possibility, I re peat the regressions in Table 3-4 with district fixed effects. Because the use of district-fixed effects limits the sample to districts where both movement and nonmoveme nt occur, the number of observations drops to only 398. Naturally this limits the power of the analysis, and is the main problem with using district fixed effects thr oughout. Table 3-7 reports the results of these regressions. According to the data, high levels of PDA have a positive effect on the

PAGE 94

83 transfer decision and low levels of PDA have a negative effect on th e transfer decision. The results in Table 3-7 confir m that district specific beha vior does not fully explain the relationship hours in PDA and mobility. 3.6 Conclusion The evidence presented in this paper clea rly seems to indicate that time spent in professional development activities leads to networks that enhance teacher mobility within their own district. These PDAs ar guably provide a forum in which teachers can build networks to gather im portant information about sc hool characteristics, work environments and job availability from which th ey can sort in a less costly, more efficient way. It also seems plausible that teachers can use their networks with favored colleagues to reveal their aptitude for a job vacancy, which may allow teachers access to transfer options that might have previous ly been closed to them. Results show that a one standard deviat ion increase in the amount of networking leads to a 0.020 to 0.078 increase in the proba bility of transferri ng within oneÂ’s own school district. If teachers use these networks to more e ffectively sort away from less desirable schools, such as low-income or failing schools, then networks arising from PDAs may contribute to the disparities in teach er quality across school types. It is also possible that if networks make it more difficult for low achieving schools to retain qualified teachers, then student achievement at these schools may also be diminished by networks. Oddly enough, this unintended c onsequence of staff development policies would be subversive to the goal of PDAs, wh ich is to improve education through greater teacher quality.

PAGE 95

84 Table 3-1: Professional Developm ent Activities Summary Statistics Number of Teachers Who Reported Yes No Percent Who Participated University Course in your Main 785 1495 34.4 % Teaching Field University Course NOT in your Main 614 1666 26.9 % Teaching Field Observational Visits to Other Schools 766 1514 33.6 % Individual or Collaborative Reseach on 1032 1248 45.3 % a Topic of Interest Regularly Scheduled Collaboration with 1561 719 68.5 % other Teachers on Issues of Instruction Mentoring or Peer Observation 1014 1266 44.4 % Participating in a Network of Teachers 559 1721 24.5 % (e.g. Internet Organization) Attending Workshops, Conferences 2141 139 93.9 % or Training Presenter at a Workshop, Conference 452 1828 19.8 % or Training

PAGE 96

85 Table 3-2: Summary Statistics for Various Teacher Groups No Move Move Intra-District InterDistrict Inter-State Number of Teachers 1400 (61.4 %) 370 (16.2 %) 484 (21.2 %) 96 (3.9 %) Mean Std. Dev. Mean Mean Mean PDA Hours 55.728 46.662 59.000 49.636 48.220 % Free-Lunch Eligible 0.002 0.225 0.045 -0.014 -0.021 Bonus Pay 0.137 0.344 0.113 0.101 0.146 Student Threat 0.095 0.294 0.127 0.130 0.083 Job Dissatisfaction 0.021 0.144 0.048 0.066 0.072 Full-Time Experience 10.915 9.856 7.959 5.134 4.583 Part-time employment 0.040 0.197 0.081 0.041 0.010 Poor Administration 0.069 0.254 0.102 0.101 0.104 Union Member 0.783 0.411 0.778 0.652 0.667 Age 40.484 11.340 37.956 33.634 32.030 Gender (Male) 0.282 0.450 0.235 0.299 0.292 Race (White) 0.902 0.297 0.862 0. 902 0.927 Marital Status 0.695 0.460 0.678 0.665 0.667

PAGE 97

86 Table 3-3: Probit Regression of Intra-District Movement Variables 1 2 0.0010626 0.0006944 PDA Hours (1.27) (0.62) — 0.00003 (PDA Hours)2 — (2.30)** 0.301017 0.3069657 % Free-Lunch Eligible (1.84)* (1.87)* 0.1669206 0.1629131 Bonus Pay (1.40) (1.37) 0.1009573 0.1002069 Student Threat (0.87) (0.86) 0.4224374 0.4353552 Job Dissatisfaction (1.85)* (1.92)* 0.0143332 0.0131213 Full-Time Experience (2.34)** (2.13)** 0.5452476 0.5394231 Part-time employment (3.43)** (3.41)** 0.3542717 0.3600298 Poor Administration (2.55)** (2.59)** 0.2372197 0.2386774 Union Member (2.36)** (2.37)** -0.005178 0.0055949 Age (1.01) (1.09) 0.0237581 0.0261624 Marital Status (0.29) (0.32) Dummy for Missing 0.1770793 0.176266 % Free Lunch Eligible (1.20) (1.20) 0.2702219 0.2997908 Constant (1.37) (1.50) Pseudo R-squared 0.0955 0.0984 Number of Obs. 1627 1627 Robust SE, clustered Y Y State FE Y Y denotes significance at the 0.05 level ** denotes significance at the 0.025 level

PAGE 98

87 Table 3-4: Probit Regression of Intra-District Movement Regression of PDA Hours with H* Chosen at Decile Intervals Along the Distribution of PDA Hours H* = 9 H* = 17 H* = 22.5 H* = 29.5 H* = 40.5 H* = 51.5 H* = 68 H* = 87.5 H* = 122.5 Variables (10%) (20%) (30%) (40%) (50%) (60%) (70%) (80%) (90%) PDA Hours 0 H* -0.0262 -0.0170 -0.0117 -0.0074 -0.0047 -0.0033 -0.0022 -0.0018 -0.0006 (1.12) (1.68) (1.63) (1.43) (1.34) (1.27) (1.12) (1.15) (0.55) PDA Hours > H* 0.0013 0.0017 0.0018 0.0020 0.0022 0.0026 0.0031 0.0043 0.0061 (1.52) (1.85)* (1.94)* (1.94)* (2.00)** (2.08)** (2.16)** (2.45)** (2.51)** % Free-Lunch Eligible 0.3041 0.3066 0.3077 0.3074 0.3067 0.3086 0.3117 0.3081 0.3010 (1.86)* (1.87)* (1.88)* (1.88)* (1.88)* (1.89)* (1.91)* (1.88)* (1.83)* Job Dissatisfaction 0.4214 0.4254 0.4282 0.4285 0.4306 0.4338 0.4377 0.4385 0.4288 (1.86)* (1.87)* (1.88)* (1.88)* (1.89)* (1.90)* (1.92)* (1.93)* (1.89)* Full-Time Experience -0.0141 -0.0136 -0.0135 -0.0134 -0.0132 -0.0130 -0.0130 -0.0130 -0.0134 (2.29)** (2.22)** (2.19)** (2.17)** (2.13)** (2.10)** (2.10)** (2.10)** (2.18)** Part-time employment 0.5383 0.5362 0.5351 0.5373 0.5417 0.5458 0.5479 0.5477 0.5381 (3.37)** (3.36)** (3.36)** (3.38)** (3.41)** (3.44)** (3.45)** (3.46)** (3.41)** Poor Administration 0.3575 0.3616 0.3636 0.3641 0.3638 0.3621 0.3595 0.3596 0.3593 (2.57)** (2.59)** (2.61)** (2.61)** (2.61)** (2.60)** (2.58)** (2.58)** (2.59)** Union Member 0.2344 0.2406 0.2419 0.2411 0.2415 0.2416 0.2399 0.2387 0.2373 (2.32)** (2.38)** (2.40)** (2.39)** (2.39)** (2.39)** (2.37)** (2.37)** (2.36)** Student Threat 0.1011 0.1016 0.1022 0.1025 0.1016 0.1000 0.0984 0.0987 0.0991 (0.87) (0.87) (0.87) (0.88) (0.87) (0.86) (0.85) (0.85) (0.85) Bonus Pay -0.1674 -0.1639 -0.1639 -0.1657 -0.1658 -0.1650 -0.1635 -0.1633 -0.1658 (1.41) (1.38) (1.38) (1.39) (1.40) (1.39) (1.38) (1.38) (1.40) Age -0.0053 -0.0054 -0.0054 -0.0055 -0.0055 -0.0056 -0.0057 -0.0058 -0.0056 (1.02) (1.05) (1.06) (1.07) (1.08) (1.09) (1.10) (1.12) (1.09) Dummy for Missing 0.1789 0.1779 0.1798 0.1796 0.1790 0.1777 0.1772 0.1807 0.1776 % Free Lunch Eligible (1.22) (1.21) (1.22) (1.22) (1.22) (1.21) (1.21) (1.23) (1.21) Marital Status 0.0232 0.0224 0.0232 0.0245 0.0256 0.0265 0.0266 0.0277 0.0264 (0.28) (0.27) (0.28) (0.30) (0.31) (0.32) (0.33) (0.34) (0.32) Constant -0.9070 0.5313 0.4935 0.4260 -0.9811 0.3674 0.3596 -0.9753 0.3402 (1.29) (2.15)** (2.09)** (1.93)* (1.45) (1.78)* (1.75)* (1.44) (1.69) Pseudo RSquared 0.0993 0.1004 0.1004 0.1002 0.1002 0.1003 0.1005 0.1012 0.1009 Number of Obs. 1627 1627 1627 1627 1627 1627 1627 1627 1627 State FE Y Y Y Y Y Y Y Y Y Robust SE, clustered Y Y Y Y Y Y Y Y Y denoted significance at the 0.05 level ** denoted significance at the 0.025 level

PAGE 99

88 A) Percent of teachers receiving poor administrative support B) Percent of teachers receiving bonus pay C) Percent of teachers reporti ng overall job dissatisfaction Figure 3-1. Comparisons of Teacher Charact eristics Along the Distribution of PDA.

PAGE 100

89 D) Percent of teachers with a master's degree E) Percent of free-lunch eligible relative to the district average F) Percent of teachers who belong to a union Figure 3-1. Continued.

PAGE 101

90 G) Percent of teachers who are employed part-time H) Percent of teachers who are married I) Average teacher age Figure 3-1. Continued.

PAGE 102

91 J) Average years of full-time experience K) Percent of teachers who are white L) Percent of teachers who are male Figure 3-1. Continued.

PAGE 103

92 M) Percent of teachers who reporte d being threatened by a student Figure 3-1. Continued.

PAGE 104

93 Table 3-5: Within State Probit Regr ession of Interdistrict Movement Regression of PDA Hours with H* Chosen at Decile Intervals Along the Distribution of PDA Hours H* = 9 H* = 17 H* = 22.5 H* = 29.5 H* = 40.5 H* = 51.5 H* = 68 H* = 87.5 H* = 122.5 Variables (10%) (20%) (30%) (40%) (50%) (60%) (70%) (80%) (90%) PDA Hours 0 H* -0.0035 -0.0053 -0.0059 -0.0063 -0.0045 -0.0033 -0.0026 -0.0020 -0.0012 (0.16) (0.54) (0.86) (1.29) (1.34) (1.31) (1.36) (1.27) (1.03) PDA Hours > H* -0.0002 -0.0000 0.00016 0.00051 0.00079 0.00100 0.00151 0.00211 0.00348 (0.24) (0.05) (0.16) (0.51) (0.70) (0.81) (1.03) (1.17) (1.25) % Free-Lunch Eligible -0.1968 -0.1979 -0.1992 -0.2014 -0.2036 -0.2032 -0.2021 -0.2028 -0.2019 (1.11) (1.12) (1.12) (1.14) (1.15) (1.15) (1.14) (1.14) (1.13) Job Dissatifaction 0.76287 0.76015 0.76046 0.76279 0.76613 0.76885 0.77090 0.77125 0.76829 (3.67)** (3.65)** (3.65)** (3.65)** (3.65)** (3.66)** (3.67)** (3.67)** (3.68)** Full-Time Experience -0.0331 -0.0331 -0.0330 -0.0328 -0.0327 -0.0326 -0.0326 -0.0326 -0.0326 (4.51)** (4.51)** (4.50)** (4.48)** (4.44)** (4.42)** (4.42)** (4.40)** (4.41)** Part-time employment 0.00475 0.00152 -0.0011 -0.0045 -0.0010 0.00166 0.00252 0.00442 0.00656 (0.02) (0.01) (0.01) (0.02) (0.01) (0.01) (0.01) (0.02) (0.03) Poor Administration 0.03717 0.03778 0.03775 0.03936 0.03753 0.03605 0.03455 0.03386 0.03687 (0.24) (0.24) (0.24) (0.25) (0.24) (0.23) (0.22) (0.21) (0.23) Union Member -0.0888 -0.0900 -0.0909 -0.0910 -0.0904 -0.0887 -0.0877 -0.0874 -0.0863 (0.91) (0.92) (0.93) (0.93) (0.92) (0.91) (0.90) (0.89) (0.88) Student Threat 0.12178 0.12207 0.12313 0.12641 0.12717 0.12681 0.12620 0.12655 0.12332 (0.96) (0.97) (0.97) (1.00) (1.00) (1.00) (0.99) (1.00) (0.97) Bonus Pay -0.2414 -0.2381 -0.2352 -0.2327 -0.2327 -0.2332 -0.2324 -0.2331 -0.2390 (1.79)* (1.77)* (1.74)* (1.72)* (1.72)* (1.73)* (1.72)* (1.73)* (1.78)* Age -0.0157 -0.0156 -0.0156 -0.0155 -0.0156 -0.0157 -0.0157 -0.0158 -0.0159 (3.03)** (3.02)** (3.01)** (3.00)** (3.02)** (3.03)** (3.03)** (3.04)** (3.06)** Dummy for Missing -0.3381 -0.3357 -0.3342 -0.3336 -0.3331 -0.3369 -0.3376 -0.3342 -0.3324 % Free Lunch Eligible (1.80)* (1.79)* (1.78)* (1.78)* (1.78)* (1.79)* (1.79)* (1.78)* (1.77)* Marital Status -0.0054 -0.0049 -0.0043 -0.0019 -0.0013 -0.0014 -0.0006 0.00009 -0.0000 (0.06) (0.06) (0.05) (0.02) (0.02) (0.02) (0.01) (0.00) (0.00) Constant -0.0104 0.03170 0.32786 0.35249 0.33986 0.32819 0.32307 0.31912 0.28890 (0.02) (0.05) (0.69) (0.75) (0.72) (0.70) (0.68) (0.67) (0.6) Pseudo RSquared 0.1686 0.1689 0.1694 0.1695 0.1695 0.1696 0.1695 0.1694 0.1689 Number of Obs. 1635 1635 1635 1635 1635 1635 1635 1635 1635 State FE Y Y Y Y Y Y Y Y Y Robust SE, clustered Y Y Y Y Y Y Y Y Y denoted significance at the 0.05 level ** denoted significance at the 0.025 level

PAGE 105

94 Table 3-6: Out Of State Probit Regr ession of Interdistrict Movement Regression of PDA Hours with H* Chosen at Decile Intervals Along the Distribution of PDA Hours H* = 9 H* = 17 H* = 22.5 H* = 29.5 H* = 40.5 H* = 51.5 H* = 68 H* = 87.5 H* = 122.5 Variables (10%) (20%) (30%) (40%) (50%) (60%) (70%) (80%) (90%) PDA Hours 0 H* 0.02337 0.01003 -0.0003 -0.0014 0.00059 0.00215 0.00216 0.00292 0.00281 (0.65) (0.62) (0.03) (0.18) (0.10) (0.48) (0.63) (1.03) (1.35) PDA Hours > H* 0.00101 0.00089 0.00142 0.00168 0.00148 0.00088 0.00053 -0.0014 -0.0079 (0.65) (0.54) (0.85) (0.97) (0.78) (0.42) (0.21) (0.44) (1.27) % Free-Lunch Eligible -0.6608 -0.6655 -0.6616 -0.6636 -0.6633 -0.6593 -0.6593 -0.6549 -0.6604 (2.51)** (2.54)** (2.52)** (2.52)** (2.51)** (2.50)** (2.51)** (2.50)** (2.53)** Job Dissatifaction 0.41903 0.41354 0.41948 0.42437 0.41961 0.41483 0.41369 0.40598 0.40404 (1.05) (1.03) (1.05) (1.06) (1.05) (1.03) (1.03) (1.01) (1.00) Full-Time Experience -0.0440 -0.0440 -0.0434 -0.0433 -0.0435 -0.0437 -0.0438 -0.0440 -0.0443 (2.79)** (2.77)** (2.75)** (2.74)** (2.75)** (2.76)** (2.77)** (2.81)** (2.82)** Part-time employment -0.8839 -0.8876 -0.8871 -0.8903 -0.8874 -0.8842 -0.8842 -0.8852 -0.8912 (1.63) (1.64) (1.63) (1.64) (1.64) (1.63) (1.63) (1.62) (1.63) Poor Administration 0.27761 0.27740 0.26798 0.26595 0.26815 0.27250 0.27307 0.27725 0.27726 (1.05) (1.05) (1.01) (1.00) (1.01) (1.03) (1.03) (1.05) (1.05) Union Member -0.0159 -0.0188 -0.0172 -0.0164 -0.0170 -0.0186 -0.0191 -0.0213 -0.0249 (0.11) (0.13) (0.12) (0.11) (0.11) (0.12) (0.13) (0.14) (0.16) Student Threat 0.05111 0.05212 0.05034 0.05093 0.05066 0.05101 0.05180 0.05261 0.06140 (0.23) (0.24) (0.23) (0.23) (0.23) (0.23) (0.24) (0.24) (0.28) Bonus Pay 0.20266 0.20067 0.20190 0.20185 0.20131 0.20238 0.20180 0.20068 0.21043 (1.09) (1.08) (1.09) (1.09) (1.09) (1.10) (1.09) (1.08) (1.14) Age -0.0394 -0.0396 -0.0394 -0.0394 -0.0394 -0.0394 -0.0394 -0.0392 -0.0388 (3.67)** (3.69)** (3.67)** (3.67)** (3.67)** (3.66)** (3.65)** (3.64)** (3.59)** Dummy for Missing 0.03832 0.04128 0.05240 0.05589 0.05176 0.04787 0.04829 0.04654 0.03925 % Free Lunch Eligible (0.11) (0.12) (0.15) (0.16) (0.15) (0.14) (0.14) (0.14) (0.11) Marital Status 0.06546 0.06408 0.06081 0.06151 0.06151 0.06037 0.05910 0.05554 0.04640 (0.46) (0.45) (0.43) (0.44) (0.44) (0.43) (0.42) (0.39) (0.33) Constant 0.76726 0.07371 0.96390 0.98918 0.95216 0.91634 0.34887 0.87351 0.86222 (1.02) (0.10) (1.36) (1.42) (1.39) (1.33) (0.41) (1.28) (1.27) Pseudo RSquared 0.2489 0.2488 0.2484 0.2485 0.2484 0.2484 0.2485 0.2492 0.2508 Number of Obs. 1108 1108 1108 1108 1108 1108 1108 1108 1108 State FE Y Y Y Y Y Y Y Y Y Robust SE, clustered Y Y Y Y Y Y Y Y Y denoted significance at the 0.05 level ** denoted significance at the 0.025 level

PAGE 106

95 Table 3-7: Probit Regression of Intradistric t Movement with District Fixed Effects Regression of PDA Hours with H* Chosen at Decile Intervals Along the Distribution of PDA Hours H* = 9 H* = 17 H* = 22.5 H* = 29.5 H* = 40.5 H* = 51.5 H* = 68 H* = 87.5 H* = 122.5 Variables (10%) (20%) (30%) (40%) (50%) (60%) (70%) (80%) (90%) PDA Hours 0 H* -0.2037 -0.0707 -0.0443 -0.0282 -0.0147 -0.0099 -0.0065 -0.0047 -0.0035 (2.12)** (2.20)** (2.03)** (1.82)* (1.45) (1.31) (1.22) (1.25) (1.26) PDA Hours > H* 0.00242 0.00303 0.00330 0.00357 0.00371 0.00416 0.00509 0.00678 0.01231 (1.13) (1.38) (1.49) (1.52) (1.44) (1.48) (1.59) (1.79)* (2.34)** % Free-Lunch Eligible 0.57160 0.50149 0.46890 0.45768 0.44608 0.45569 0.45798 0.44090 0.45394 (1.10) (0.95) (0.89) (0.87) (0.85) (0.87) (0.88) (0.85) (0.87) Job Dissatifaction 1.24178 1.22087 1.22145 1.21289 1.21081 1.20811 1.20478 1.17851 1.13832 (2.09)** (2.02)** (1.98)** (1.96)** (1.96)** (1.96)** (1.96)** (1.93)* (1.88)* Full-Time Experience -0.0119 -0.0098 -0.0099 -0.0094 -0.0086 -0.0089 -0.0093 -0.0099 -0.0098 (0.68) (0.55) (0.56) (0.53) (0.48) (0.51) (0.53) (0.56) (0.56) Part-time employment 0.51540 0.69751 0.72376 0.73810 0.73554 0.73359 0.72474 0.71323 0.71233 (1.09) (1.51) (1.59) (1.65) (1.67) (1.67) (1.65) (1.62) (1.62) Poor Administration 0.77191 0.73451 0.73095 0.72802 0.73031 0.72283 0.71226 0.70846 0.67968 (1.79)* (1.68) (1.66) (1.65) (1.66) (1.64) (1.61) (1.60) (1.55) Union Member 0.01757 0.08987 0.09814 0.09679 0.09406 0.09087 0.08021 0.08181 0.09168 (0.06) (0.31) (0.33) (0.33) (0.32) (0.31) (0.27) (0.28) (0.32) Student Threat 0.41600 0.42832 0.42130 0.41446 0.41373 0.40975 0.40086 0.39627 0.39301 (1.30) (1.32) (1.29) (1.29) (1.30) (1.30) (1.27) (1.26) (1.26) Bonus Pay -0.4831 -0.4094 -0.3959 -0.3940 -0.3839 -0.3700 -0.3651 -0.3550 -0.3483 (1.26) (1.08) (1.03) (1.02) (0.99) (0.96) (0.95) (0.93) (0.90) Age -0.0154 -0.0161 -0.0156 -0.0154 -0.0155 -0.0154 -0.0153 -0.0152 -0.0156 (0.96) (1.01) (0.98) (0.97) (0.98) (0.98) (0.98) (0.98) (1.00) Dummy for Missing -0.0759 -0.0769 -0.0674 -0.0824 -0.1056 -0.1187 -0.1165 -0.0904 -0.0660 % Free Lunch Eligible (0.20) (0.19) (0.17) (0.21) (0.27) (0.30) (0.30) (0.23) (0.17) Marital Status -0.0479 -0.0328 -0.0319 -0.0328 -0.0396 -0.0419 -0.0485 -0.0521 -0.0487 (0.21) (0.15) (0.14) (0.15) (0.18) (0.19) (0.22) (0.23) (0.22) Constant 2.09461 1.55653 1.29478 1.16611 1.02917 0.98271 0.96313 0.99048 1.00050 (2.62)** (2.25)** (2.00)** (1.82)* (1.62) (1.57) (1.54) (1.57) (1.62) Pseudo RSquared 0.1760 0.1710 0.1693 0.1675 0.1640 0.1632 0.1631 0.1646 0.1683 Number of Obs. 398 398 398 398 398 398 398 398 398 State FE Y Y Y Y Y Y Y Y Y District FE Y Y Y Y Y Y Y Y Y Robust SE, clustered Y Y Y Y Y Y Y Y Y denoted significance at the 0.05 level ** denoted significance at the 0.025 level

PAGE 107

96 APPENDIX A SUMMARY STATISTICS AND SUPPLEMENTARY REGRESSION ANALYSIS Table A-1: Strike Summary Statistics Total Number of Strike and Strike Treatments in the 22 Year Sample: Total Strikes: 102 (64)** Total Strike Treatments: 2,351 Total Strikes and Strike Treatments in Each Community: Urban: Total Strikes: 74 (43)** Total Strike Treatments: 1,858 (1616)** Suburban Total Strikes: 31 (17)** Total Strike Treatments: 97 (67)** Rural Total Strikes: 56 (32)** Total Strike Treatments: 398 (275)** Longest Strike Length: 25 Days Mean Strike Length: 6.7 Days (Median is 5 Days) Distribution of Strike Days: Total Strike Treatments by Strike Year: 1 Day 321 1980 106 2 Days 211 1981 24 3 Days 202 1982 0 4 Days 197 1983 74 5 Days 188 1984 2 6 Days 177 1985 567 7 Days 179 1986 27 8 Days 170 1987 87 9 Days 58 1988 0 10 Days 49 1989 0 11 Days 45 1990 222 12 Days 44 1991 1016 13 Days 44 1992 9 14 Days 39 1993 4 15 Days 39 1994 90 16 Days 39 1995 75 17 Days 38 1996 0 18 Days 38 1997 0 19 Days 34 1998 18 20 Days 34 1999 1 21 Days 30 2000 0 22 Days 3 2001 11 23 Days 3 24 Days 3 25 Days 3 **The numbers in parenthesis represent the value in Type IV specifications

PAGE 108

97 Table A-2: Regressions of Full Data Set excluding Zip C odes with less than 500 students Variables ( ) ( ) ( ) ( V) (V) Strike Median Income Welfare Urban Poor Parental Educ. Juvenile Maleness Student Employment Alpha 0.187** 0.177** 0.200** 0.246** 0.212** (3.88)** (3.64)** (3.94)** (4.77)** (4.27)** -0.00003** -0.00003** -0.00003** -0.00003** (103.76)** (85.14)** (61.69)** (54.36)** 1.120** 1.149** 1.250** 1.572** (18.28)** (14.20)** (11.22)** (11.09)** 0.110** 0.089** 0.071** 0.046** (22.30)** (14.42)** (8.38)** (5.23)** 0.503** 0.523** 0.570** 0.802** (14.02)** (11.83)** (9.34)** (9.39)** 0.266** 0.195** 0.0829 0.265** (5.66)** (3.33)** (1.01) (2.66)** 0.224** 0.228** 0.225** 0.192** (11.67)** (9.61)** (6.88)** (5.07)** 0.922** 0.941** 0.927** 0.889** 0.705** (153.67)** (117.63)** (92.70)** (85.48)** (76.58)** Number of obs. Time Fixed Effects Zip Fixed Effects Pseudo R-Squared 879,441 582,543 301,056 249,745 249,745 Y Y Y Y Y N N N N Y 0.0397 0.0396 0.040 0.0308 0.0556

PAGE 109

APPENDIX B COMPLETE RESULTS OF SPECIFIED REGRESSIONS

PAGE 110

99 Table B-1: Reported on Table 3-4 Regression of PDA Hours with H* chosen at a cutoff on the Distri bution of PDA Hours Variables 10% 20% 30% 40% 50% 60% 70% 80% 90% PDA Hours 0 H* -0.0262 -0.0170 -0.0117 -0.0074 -0.0047 -0.0033 -0.0022 -0.0018 -0.0006 (1.12) (1.68) (1.63) (1.43) (1.34) (1.27) (1.12) (1.15) (0.55) PDA Hours > H* 0.0013 0.0017 0.0018 0.0020 0.0022 0.0026 0.0031 0.0043 0.0061 (1.52) (1.85)* (1.94)* (1.94)* (2.00)** (2.08)** (2.16)** (2.45)** (2.51)** % Free-Lunch Eligible 0.3041 0.3066 0.3077 0.3074 0.3067 0.3086 0.3117 0.3081 0.3010 (1.86)* (1.87)* (1.88)* (1.88)* (1.88)* (1.89)* (1.91)* (1.88)* (1.83)* Job Dissatifaction 0.4214 0.4254 0.4282 0.4285 0.4306 0.4338 0.4377 0.4385 0.4288 (1.86)* (1.87)* (1.88)* (1.88)* (1.89)* (1.90)* (1.92)* (1.93)* (1.89)* Full-Time Experience -0.0141 -0.0136 -0.0135 -0.0134 -0.0132 -0.0130 -0.0130 -0.0130 -0.0134 (2.29)** (2.22)** (2.19)** (2.17)** (2.13)** (2.10)** (2.10)** (2.10)** (2.18)** Part-time employment 0.5383 0.5362 0.5351 0.5373 0.5417 0.5458 0.5479 0.5477 0.5381 (3.37)** (3.36)** (3.36)** (3.38)** (3.41)** (3.44)** (3.45)** (3.46)** (3.41)** Poor Administration 0.3575 0.3616 0.3636 0.3641 0.3638 0.3621 0.3595 0.3596 0.3593 (2.57)** (2.59)** (2.61)** (2.61)** (2.61)** (2.60)** (2.58)** (2.58)** (2.59)** Union Member 0.2344 0.2406 0.2419 0.2411 0.2415 0.2416 0.2399 0.2387 0.2373 (2.32)** (2.38)** (2.40)** (2.39)** (2.39)** (2.39)** (2.37)** (2.37)** (2.36)** Race (White) -0.0906 -0.0924 -0.0931 -0.0950 -0. 0970 -0.0987 -0.10 00 -0.1015 -0.0953 (0.76) (0.77) (0.78) (0.80) (0.81) (0.83) (0.84) (0.85) (0.80) Marital Status 0.0232 0.0224 0.0232 0.0245 0.0256 0.0265 0.0266 0.0277 0.0264 (0.28) (0.27) (0.28) (0.30) (0.31) (0.32) (0.33) (0.34) (0.32) Dummy for Missing 0.1789 0.1779 0.1798 0.1796 0.1790 0.1777 0.1772 0.1807 0.1776 % Free Lunch Eligible (1.22) (1.21) (1.22) (1.22) (1.22) (1.21) (1.21) (1.23) (1.21) Bonus Pay -0.1674 -0.1639 -0.1639 -0.1657 -0.1658 -0.1650 -0.1635 -0.1633 -0.1658 (1.41) (1.38) (1.38) (1.39) (1.40) (1.39) (1.38) (1.38) (1.40) Student Threat 0.1011 0.1016 0.1022 0.1025 0.1016 0.1000 0.0984 0.0987 0.0991 (0.87) (0.87) (0.87) (0.88) (0.87) (0.86) (0.85) (0.85) (0.85) Age -0.0053 -0.0054 -0.0054 -0.0055 -0.0055 -0.0056 -0.0057 -0.0058 -0.0056 (1.02) (1.05) (1.06) (1.07) (1.08) (1.09) (1.10) (1.12) (1.09) Gender (Male) -0.1157 -0.1180 -0.1186 -0.1188 -0.1188 -0.1188 -0.1188 -0.1176 -0.1132 (1.35) (1.37) (1.38) (1.38) (1.38) (1.38) (1.38) (1.37) (1.32) Master's Degree 0.1211 0.1211 0.1212 0.1204 0.1202 0.1210 0.1215 0.1210 0.1180 (1.43) (1.43) (1.43) (1.42) (1.42) (1.43) (1.43) (1.43) (1.39) Constant -0.9070 0.5313 0.4935 0.4260 -0.9811 0.3674 0.3596 -0.9753 0.3402 (1.29) (2.15)** (2.09)** (1.93)* (1.45) (1.78)* (1.75)* (1.44) (1.69) R-Squared 0.0993 0.1004 0.1004 0.1002 0.1002 0.1003 0.1005 0.1012 0.1009 Number of Obs. 1627 1627 1627 1627 1627 1627 1627 1627 1627 State FE Y Y Y Y Y Y Y Y Y Robust SE, clustered Y Y Y Y Y Y Y Y Y denoted significance at the 0.05 level, ** denoted significance at the 0.025 level

PAGE 111

100 Table B-2: Reported on Table 3-5 Regression of PDA Hours with H* chosen at a cutoff on the Distri bution of PDA Hours Variables 10% 20% 30% 40% 50% 60% 70% 80% 90% PDA Hours 0 H* -0.0035 -0.0053 -0.0059 -0.0063 -0. 0045 -0.0033 -0.0026 -0.0020 -0.0012 (0.16) (0.54) (0.86) (1.29) (1. 34) (1.31) (1.36) (1.27) (1.03) PDA Hours > H* -0.0002 -0.0000 0.00016 0.00051 0.00079 0.00100 0.00151 0.00211 0.00348 (0.24) (0.05) (0.16) (0.51) (0. 70) (0.81) (1.03) (1.17) (1.25) % Free-Lunch Eligible -0.1968 -0.1979 -0.1992 -0.2014 -0. 2036 -0.2032 -0.2021 -0.2028 -0.2019 (1.11) (1.12) (1.12) (1.14) (1. 15) (1.15) (1.14) (1.14) (1.13) Job Dissatifaction 0.76287 0.76015 0.76046 0.76279 0. 76613 0.76885 0.77090 0.77125 0.76829 (3.67)** (3.65)** (3.65)** (3.65)** (3. 65)** (3.66)** (3.67)** (3.67)** (3.68)** Full-Time Experience -0.0331 -0.0331 -0.0330 -0.0328 -0. 0327 -0.0326 -0.0326 -0.0326 -0.0326 (4.51)** (4.51)** (4.50)** (4.48)** (4. 44)** (4.42)** (4.42)** (4.40)** (4.41)** Part-time employment 0.00475 0.00152 -0.0011 -0.0045 -0 .0010 0.00166 0.00252 0.00442 0.00656 (0.02) (0.01) (0.01) (0.02) (0. 01) (0.01) (0.01) (0.02) (0.03) Poor Administration 0.03717 0.03778 0.03775 0.03936 0. 03753 0.03605 0.03455 0.03386 0.03687 (0.24) (0.24) (0.24) (0.25) (0. 24) (0.23) (0.22) (0.21) (0.23) Union Member -0.0888 -0.0900 -0.0909 -0 .0910 -0.0904 -0.0887 -0.0877 -0.0874 -0.0863 (0.91) (0.92) (0.93) (0.93) (0. 92) (0.91) (0.90) (0.89) (0.88) Race (White) -0.1287 -0.1324 -0.1355 -0 .1382 -0.1378 -0.1367 -0.1374 -0.1366 -0.1335 (1.00) (1.03) (1.05) (1.07) (1. 07) (1.06) (1.06) (1.06) (1.03) Marital Status -0.0054 -0.0049 -0.0043 -0.0019 -0.0013 -0.0014 -0.0006 0.00009 -0.0000 (0.06) (0.06) (0.05) (0.02) (0. 02) (0.02) (0.01) (0.00) (0.00) Dummy for Missing -0.3381 -0.3357 -0.3342 -0.3336 -0. 3331 -0.3369 -0.3376 -0.3342 -0.3324 % Free Lunch Eligible (1.80)* (1.79)* (1.78)* (1.78)* (1. 78)* (1.79)* (1.79)* (1.78)* (1.77)* Bonus Pay -0.2414 -0.2381 -0.2352 -0. 2327 -0.2327 -0.2332 -0.2324 -0.2331 -0.2390 (1.79)* (1.77)* (1.74)* (1.72)* (1. 72)* (1.73)* (1.72)* (1.73)* (1.78)* Student Threat 0.12178 0.12207 0. 12313 0.12641 0.12717 0.12681 0.12620 0.12655 0.12332 (0.96) (0.97) (0.97) (1.00) (1. 00) (1.00) (0.99) (1.00) (0.97) Age -0.0157 -0.0156 -0.0156 -0.0155 -0 .0156 -0.0157 -0.0157 -0.0158 -0.0159 (3.03)** (3.02)** (3.01)** (3.00)** (3. 02)** (3.03)** (3.03)** (3.04)** (3.06)** Gender (Male) 0.0217 0.0199 0.0181 0.0167 0.0176 0.0194 0.0212 0.0227 0.0245 (0.26) (0.23) (0.21) (0.20) (0. 21) (0.23) (0.25) (0.27) (0.29) Master's Degree -0.0153 -0.0162 -0.0173 -0 .0156 -0.0156 -0.0129 -0.0115 -0.0126 -0.0139 (0.17) (0.18) (0.19) (0.17) (0. 17) (0.14) (0.13) (0.14) (0.15) Constant -0.0104 0.03170 0.32786 0.35249 0.33986 0.32819 0.32307 0.31912 0.28890 (0.02) (0.05) (0.69) (0.75) (0 .72) (0.70) (0.68) (0.67) (0.6) R-squared 0.1686 0.1689 0.1694 0.1695 0.1695 0.1696 0.1695 0.1694 0.1689 Number of Obs. 1635 1635 1635 1635 1635 1635 1635 1635 1635 State FE Y Y Y Y Y Y Y Y Y Robust SE, clustered Y Y Y Y Y Y Y Y Y denoted significance at the 0.05 level ** denoted significance at the 0.025 level

PAGE 112

101 Table B-3: Reported on Table 3-6 Regression of PDA Hours with H* chosen at a cutoff on the Distri bution of PDA Hours Variables 10% 20% 30% 40% 50% 60% 70% 80% 90% PDA Hours 0 H* 0.02337 0.01003 -0.0003 -0.0014 0.00059 0.00215 0.00216 0.00292 0.00281 (0.65) (0.62) (0.03) (0.18) (0. 10) (0.48) (0.63) (1.03) (1.35) PDA Hours > H* 0.00101 0.00089 0.00142 0.00168 0.00148 0.00088 0.00053 -0.0014 -0.0079 (0.65) (0.54) (0.85) (0.97) (0. 78) (0.42) (0.21) (0.44) (1.27) % Free-Lunch Eligible -0.6608 -0.6655 -0.6616 -0.6636 -0. 6633 -0.6593 -0.6593 -0.6549 -0.6604 (2.51)** (2.54)** (2.52)** (2.52)** (2. 51)** (2.50)** (2.51)** (2.50)** (2.53)** Job Dissatifaction 0.41903 0.41354 0.41948 0.42437 0. 41961 0.41483 0.41369 0.40598 0.40404 (1.05) (1.03) (1.05) (1.06) (1. 05) (1.03) (1.03) (1.01) (1.00) Full-Time Experience -0.0440 -0.0440 -0.0434 -0.0433 -0. 0435 -0.0437 -0.0438 -0.0440 -0.0443 (2.79)** (2.77)** (2.75)** (2.74)** (2. 75)** (2.76)** (2.77)** (2.81)** (2.82)** Part-time employment -0.8839 -0.8876 -0.8871 -0.8903 -0. 8874 -0.8842 -0.8842 -0.8852 -0.8912 (1.63) (1.64) (1.63) (1.64) (1. 64) (1.63) (1.63) (1.62) (1.63) Poor Administration 0.27761 0.27740 0.26798 0.26595 0. 26815 0.27250 0.27307 0.27725 0.27726 (1.05) (1.05) (1.01) (1.00) (1. 01) (1.03) (1.03) (1.05) (1.05) Union Member -0.0159 -0.0188 -0.0172 -0 .0164 -0.0170 -0.0186 -0.0191 -0.0213 -0.0249 (0.11) (0.13) (0.12) (0.11) (0. 11) (0.12) (0.13) (0.14) (0.16) Race (White) 0.23557 0.2448 0.24363 0.24274 0.24431 0.24747 0.24979 0.25548 0.24977 (0.89) (0.92) (0.92) (0.91) (0. 92) (0.93) (0.94) (0.95) (0.93) Marital Status 0.06546 0.06408 0. 06081 0.06151 0.06151 0.06037 0.05910 0.05554 0.04640 (0.46) (0.45) (0.43) (0.44) (0. 44) (0.43) (0.42) (0.39) (0.33) Dummy for Missing 0.03832 0.04128 0.05240 0.05589 0. 05176 0.04787 0.04829 0.04654 0.03925 % Free Lunch Eligible (0.11) (0.12) (0.15) (0.16) (0. 15) (0.14) (0.14) (0.14) (0.11) Bonus Pay 0.20266 0.20067 0.20190 0. 20185 0.20131 0.20238 0.20180 0.20068 0.21043 (1.09) (1.08) (1.09) (1.09) (1. 09) (1.10) (1.09) (1.08) (1.14) Student Threat 0.05111 0.05212 0. 05034 0.05093 0.05066 0.05101 0.05180 0.05261 0.06140 (0.23) (0.24) (0.23) (0.23) (0. 23) (0.23) (0.24) (0.24) (0.28) Age -0.0394 -0.0396 -0.0394 -0.0394 -0 .0394 -0.0394 -0.0394 -0.0392 -0.0388 (3.67)** (3.69)** (3.67)** (3.67)** (3. 67)** (3.66)** (3.65)** (3.64)** (3.59)** Gender (Male) -0.1214 -0.1187 -0.1257 -0 .1286 -0.1267 -0.1219 -0.1209 -0.1176 -0.1264 (0.78) (0.76) (0.80) (0.82) (0. 81) (0.78) (0.77) (0.75) (0.81) Master's Degree 0.53769 0.53978 0. 53363 0.53165 0.5316 0.53155 0.53111 0.53137 0.54125 (3.21) (3.23) (3.19) (3.19) (3. 19) (3.19) (3.18) (3.18) (3.23) Constant 0.76726 0.07371 0.96390 0. 98918 0.95216 0.91634 0.34887 0.87351 0.86222 (1.02) (0.10) (1.36) (1.42) (1. 39) (1.33) (0.41) (1.28) (1.27) R-squared 0.2489 0.2488 0.2484 0.2485 0.2484 0.2484 0.2485 0.2492 0.2508 Number of Obs. 1108 1108 1108 1108 1108 1108 1108 1108 1108 State FE Y Y Y Y Y Y Y Y Y Robust SE, clustered Y Y Y Y Y Y Y Y Y denoted significance at the 0.05 level, ** denoted significance at the 0.025 level

PAGE 113

102 Table B-4: Reported on Table 3-7 Regression of PDA Hours with H* chosen at a cutoff on the Distri bution of PDA Hours Variables 10% 20% 30% 40% 50% 60% 70% 80% 90% PDA Hours 0 H* -0.2037 -0.0707 -0.0443 -0.0282 -0. 0147 -0.0099 -0.0065 -0.0047 -0.0035 (2.12)** (2.20)** (2.03)** (1.82)* (1. 45) (1.31) (1.22) (1.25) (1.26) PDA Hours > H* 0.00242 0.00303 0.00330 0.00357 0. 00371 0.00416 0.00509 0.00678 0.01231 (1.13) (1.38) (1.49) (1.52) (1. 44) (1.48) (1.59) (1.79)* (2.34)** % Free-Lunch Eligible 0.57160 0.50149 0.46890 0.45768 0. 44608 0.45569 0.45798 0.44090 0.45394 (1.10) (0.95) (0.89) (0.87) (0. 85) (0.87) (0.88) (0.85) (0.87) Job Dissatifaction 1.24178 1.22087 1.22145 1.21289 1. 21081 1.20811 1.20478 1.17851 1.13832 (2.09)** (2.02)** (1.98)** (1.96)** (1. 96)** (1.96)** (1.96)** (1.93)* (1.88)* Full-Time Experience -0.0119 -0.0098 -0.0099 -0.0094 -0. 0086 -0.0089 -0.0093 -0.0099 -0.0098 (0.68) (0.55) (0.56) (0.53) (0. 48) (0.51) (0.53) (0.56) (0.56) Part-time employment 0.51540 0.69751 0.72376 0.73810 0. 73554 0.73359 0.72474 0.71323 0.71233 (1.09) (1.51) (1.59) (1.65) (1. 67) (1.67) (1.65) (1.62) (1.62) Poor Administration 0.77191 0.73451 0.73095 0.72802 0. 73031 0.72283 0.71226 0.70846 0.67968 (1.79)* (1.68) (1.66) (1.65) (1. 66) (1.64) (1.61) (1.60) (1.55) Union Member 0.01757 0.08987 0.09814 0.09679 0. 09406 0.09087 0.08021 0.08181 0.09168 (0.06) (0.31) (0.33) (0.33) (0. 32) (0.31) (0.27) (0.28) (0.32) Race (White) -0.0808 -0.083 -0.0838 -0 .0862 -0.0885 -0.0908 -0.0924 -0.0937 -0.0864 (0.65) (0.67) (0.68) (0.69) (0. 71) (0.73) (0.74) (0.75) (0.69) Marital Status -0.0479 -0.0328 -0.0319 -0 .0328 -0.0396 -0.0419 -0.0485 -0.0521 -0.0487 (0.21) (0.15) (0.14) (0.15) (0. 18) (0.19) (0.22) (0.23) (0.22) Dummy for Missing -0.0759 -0.0769 -0.0674 -0.0824 -0. 1056 -0.1187 -0.1165 -0.0904 -0.0660 % Free Lunch Eligible (0.20) (0.19) (0.17) (0.21) (0. 27) (0.30) (0.30) (0.23) (0.17) Bonus Pay -0.4831 -0.4094 -0.3959 -0. 3940 -0.3839 -0.3700 -0.3651 -0.3550 -0.3483 (1.26) (1.08) (1.03) (1.02) (0. 99) (0.96) (0.95) (0.93) (0.90) Student Threat 0.41600 0.42832 0. 42130 0.41446 0.41373 0.40975 0.40086 0.39627 0.39301 (1.30) (1.32) (1.29) (1.29) (1. 30) (1.30) (1.27) (1.26) (1.26) Age -0.0154 -0.0161 -0.0156 -0.0154 -0 .0155 -0.0154 -0.0153 -0.0152 -0.0156 (0.96) (1.01) (0.98) (0.97) (0. 98) (0.98) (0.98) (0.98) (1.00) Gender (Male) -0.107 -0.1103 -0.1111 -0 .1113 -0.1114 -0.1117 -0.1118 -0.1099 -0.1043 (1.30) (1.33) (1.34) (1.34) (1. 35) (1.35) (1.36) (1.33) (1.26) Master's Degree 0.1495 0.1497 0.15 0.149 0. 1485 0.1495 0.1502 0.1497 0.1463 (1.76)* (1.75)* (1.75)* (1.74)* (1. 73)* (1.74)* (1.75)* (1.74)* (1.71)* Constant 2.09461 1.55653 1.29478 1. 16611 1.02917 0.98271 0.96313 0.99048 1.00050 (2.62)** (2.25)** (2.00)** (1.82)* (1. 62) (1.57) (1.54) (1.57) (1.62) R-squared 0.1760 0.1710 0.1693 0.1675 0.1640 0.1632 0.1631 0.1646 0.1683 Number of Obs. 398 398 398 398 398 398 398 398 398 State FE Y Y Y Y Y Y Y Y Y Robust SE, clustered Y Y Y Y Y Y Y Y Y denoted significance at the 0.05 level, ** denoted significance at the 0.025 level

PAGE 114

103 LIST OF REFERENCES Allan, Emilie A., and Steffensmeier, Darrell J., “Youth, Underemployment, and Property Crime: Differential Effects of Job Availa bility and Job Quality on Juvenile and Young Adult Arrest Rates” American Sociological Review Vol. 54, No.1 (February 1989), pp.107-123 Angrist, Joshua D. and Guryan, Jonathan “D oes Teacher Testing Raise Teacher Quality? Evidence from State Certification Requirements” NBER Working Paper No 9545 March 2003. Ballou, Dale and Podgursky, Michael “Teach ers’ Attitudes Towards Merit Pay: Examining Conventional Wisdom” Industrial and Labor Relations Review Vol. 47, No. 1, (October 1993) p. 50-61 Becker, Gary S. and Tomes, Nigel, “Human Capital and the Rise and Fall of Families” Journal of Labor Economics Vol. 4, No. 3, Part 2: The Family and the Distribution of Economic Rewards. (July 1986), pp. S1-S39. Becker, Gary S. and Tomes, Nigel, “Child Endowments and the Quantity and Quality of Children” (in Part II: Labor Supply and the Family) The Journal of Political Economy Vol. 84, No. 4, Part 2: Essays in Labor Economics in Honor of H. Gregg Lewis. (August 1976), pp. S143-S162. Brewer, Dominic J. “Career Paths and Quits Decisions: Evidence From Teaching” Journal of Labor Economics Vol. 14, No. 2 (April 1996) p. 313-339 Cameron, Colin A., and Trivedi, Pravin K., “Econometric Models Based on Count Data: Comparisons and Applications of Some Estimators and Tests” Journal of Applied Econometrics Vol. 1, No. 1. (January 1986), pp. 29-53. Carson, R.T. and Grogger, Jefferey T. “Models for Truncated Counts” Journal of Applied Econometrics Vol. 6, No. 3. (July September 1991), pp. 225238. Corman, Hope, and Mocan, Naci H., “Carro ts, Sticks and Broken Windows”, National Bureau of Economic Research: Working Paper 9061 (July 2002). Cuellar, Alison Evans, Markowitz, Sara and Libby, Anne M., "The Relationships Between Mental Health and Substance Abuse Treatment and Juvenile Crime", National Bureau of Economic Research Working Paper No. W9952 (September 2003).

PAGE 115

104 Eberts, Randall, Hollenbeck, Kevin and Stone Joe “Teacher Performance Incentives and Student Outcomes” The Journal of Human Resources, Vol. 37, No. 4 (Autumn 2002), p. 913-927 Ehrlich, Isaac, “On the Useful ness of Controlling Individual s: An Economic Analysis of Rehabilitation, Incapacitation and Deterrence” The American Economic Review Vol. 71, No. 3. (June 1981), pp. 307-322. Ehrlich, Isaac and Gibbons, Joel C., “On the Measurement of the Deterrent Effect of Capital Punishment and the Theory of Deterrence” The Journal of Legal Studies Vol. 6, No. 1. (January 1977), pp. 35-50. Feld, Barry C., “Juvenile and Criminal Just ice Systems' Responses to Youth Violence” Crime and Justice Vol. 24, Youth Violence. (1998), pp. 189-261. Fleisher, Belton M., “The Effect of Unemployment on Juvenile Delinquency” The Journal of Political Economy Vol. 71, No. 6 (December 1963), pp.543-555 Freeman, Richard B., “Why Do So Many Y oung American Men Commit Crimes and What Might I Do About It?” Journal of Economic Perspectives Vol 10:1; pp.25-42 (Winter 1996). Freeman, Richard B., “Crime and The Empl oyment of Disadvantaged Youths” in Urban Labor Markets and Job Opportunity by George Peterson and WayneVroman (Washington, D.C.: Urban Press Institute, 1992). Glaeser, Edward L. and Sacerdote, Bruce, “Why is There More Crime in Cities?” The Journal of Political Economy Vol. 107, No. 6, Part 2: Symposium on the Economic Analysis of Social Behavior in Honor of Gary S. Becker. (December 1999), pp. S225-S258. Glaeser, Edward L. and Sacerdote, Bruce, “Crime and Social Interactions” The Quarterly Journal of Economics Vol. 111, No. 2. (May 1996), pp. 507-548. Greenberg, David and McCall, John “Teacher Mobility and Allocation” The Journal of Human Resources, Vol. 9, No. 4 (Autumn 1974), p. 480-502. Griliches, Zvi, Hall, Bronwyn H. and Haus man, Jerry, “Econometric Models for Count Data with an Application to the Patents-R & D Relationship” Econometrica Vol. 52, No. 4. (July 1984), pp. 909-938. Gritz, Mark R. and Theobald, Neil D. “The E ffects of School District Spending Priorities On Length of Stay in Teaching” The Journal of Human Resources, Vol. 31, No. 3 (Summer 1996), p. 477-512 Grogger, Jeffrey T., "Local Violence, E ducational Attainment, and Teacher Pay" National Bureau of Economic Research Working Paper No. 6003 (April 1997).

PAGE 116

105 Hanushek, Eric A., Kain, John F. and Rivkin, St even G. “Do Higher Salaries Buy Better Teachers?” NBER Working Paper No 7082, April 2002, JEL No. I2, J4 Hanushek, Eric A., Kain, John F. and Rivkin, Steven G. “Why Public Schools Lose Teachers” NBER Working Paper No 8599, November 2001, JEL No. I20, J45 Hanushek, Eric A., Kain, John F. and Rivkin, Steven G. “Teachers, Schools, and Academic Achievement” NBER Working Paper No 6691, August 1998, JEL No. I2, H4 Haveman, Robert and Wolfe, Barbara, “The Determinants of Children's Attainments: A Review of Methods and Findings” Journal of Economic Literature Vol. 33, No. 4. (December 1995), pp. 1829-1878. Hill, Anne M. and O’Neill, June, “Family Endowments and the Achievement of Young Children with Special Reference to the Underclass” The Journal of Human Resources Vol. 29, No. 4, Special Issue: Th e Family and Intergenerational Relations. (Autumn 1994), pp. 1064-1100. Hoxby, Caroline “Would School Choice Change The Teaching Profession?” Journal of Human Resources, Vol. 37, No. 4 (Autumn 2002), p. 846-891 Imazeki, Jennifer “Teacher Salaries and Teach er Attrition: How Much is Enough?” (May 2003) Imazeki, Jennifer “Teacher Attrition and Mobili ty in Urban Districts: Evidence From Wisconsin” In Fiscal Issues in Urban Schools; Research in Education: Fiscal Policy and Practice Volume 1, Jennifer King Rice and Christopher Roelke, eds. Jacob, Brian and Lefgren, Lars, “Are Idle Hands the Devil’s Workshop? Incapacitation, Concentration and Juvenile Crime” The American Economic Review Vol. 93, No. 5. (December 2003), pp.1560-1577. Lankford, Hamilton, Loeb, Susanna and Wyckoff, James “Teacher Sorting and the Plight of Urban Schools: A Descriptive Analysis” Educational Evaluation and Policy Analysis Vol. 24, No. 1 (Spring 2002) p. 37-62 Levitt, Steven, “Juvenile Crime and Punishment” The Journal of Political Economy Vol. 106, No. 6. (December 1998), pp. 1156-1185. Lochner, Lance, "Education, Work, and Cr ime: A Human Capital Approach" National Bureau of Economic Research Working Paper No. 10478 (May 2004). Lochner, Lance, "Individual Perceptions of the Criminal Justice System", National Bureau of Economic Research Working Paper No. 9474 (February 2003).

PAGE 117

106 Lochner, Lance and Moretti, Enrico, "The Effect of Education on Criminal Activity: Evidence from Prison Inmates, Arrests and Self-Reports" American Economic Review Vol. 94, No. 1. (March 2004). McDowall, David and Singer, Simon I., “C riminalizing Delinquency: The Deterrent Effects of the New York Juvenile Offender Law” Law & Society Review Vol. 22, No. 3. (1988), pp. 521-536. Mocan, Naci H., Rees, Daniel I., “Economic Conditions, Deterrence and Juvenile Crime: Evidence From Micro Data”, National Bureau of Economic Research: Working Paper No. 7405 (October 1999). Mocan, Naci H., Scafidi, Benjamin, and Tekin, Erdal, “Catholic Schools and Bad Behavior”, National Bureau of Ec onomic Research: Working Paper 9172 (September 2002). Murnane, Richard J. “Selection and Su rvival in the Teacher Labor Market” The Review of Economics and Statistics, Vol. 66, No. 3 (August 1984), p. 513-518 Murnane, Richard J. “Teacher Mobility Revisited” The Journal of Human Resources, Vol. 16, No. 1 (Winter 1981), p. 3-19 Murnane, Richard J. and Olsen, Randall J. “T he Effects of Salaries and Opportunity Costs on Length of Stay in Teaching: Evidence from North Carolina” The Journal of Human Resources, Vol. 25, No. 1 (Winter 1990), p. 106-124 Pfeiffer, Christian, “Juvenile Crime and Violence in Europe” Crime and Justice Vol. 23. (1998), pp. 255-328. Rees, Daniel “Grievance Procedur e Strength and Teacher Quits” Industrial and Labor Relations Review Vol. 45, No. 1, (October 1991), p. 31-43 Tatum, Becky L., “An Analysis of Factors Contributing to the Delinquency of the Black Youth” Journal of Black Studies Vol. 26, No. 3 (January 1996), pp. 356-368 U.S. Department of Education, National Center for Education Statistics (1997), America’s Teachers: Profile of a Profession, 1993-94 Washington, D.C.: U.S. Government Printing Office, NCES 97-460. Wilson, O.W., “How to Measure the Extent of Juvenile Delinquency” Journal of Criminal Law and Criminology (1931-1951) Vol. 41, No. 4. (November December 1950), pp. 435-438. Zabalza, A. “Internal Labour Mobi lity and the Teaching Profession” The Economic Journal Vol. 88, No. 350 (June 1978), p. 314-330

PAGE 118

107 BIOGRAPHICAL SKETCH Before beginning his doctoral studies, Jere my Luallen attended the University of Florida as an undergraduate where he receiv ed two bachelor degrees (one in economics and one in political science) in three years. As a graduate student Jeremy has been recognized numerous times for scholarly achie vement in his work. In 2004 he was the recipient of Edward Zabel Award, an award gi ven for excellence in dissertation research and publication potential. In addition he was also recognized for excellence in completed research, as well as potential for future res earch, in the field of Public Policy as a recipient of the Walter-Lanzillotti Award. Jeremy is a member of the American Education Finance Association, has been invited to present his research at several coll egial conferences, such as the annual meeting of the Southern Economic Association, a nd at the American Education Finance Association annual conference. In addition to his schooling Jeremy has worked as a consultant for the Naples Children and E ducation Foundation, and will begin working as a Senior Analyst for Abt A ssociates after graduation.