<%BANNER%>

Does High Cost-Sharing in Physician Care Reduce Health Care Utilization and Expenditures Differently for People with Sev...

Permanent Link: http://ufdc.ufl.edu/UFE0041643/00001

Material Information

Title: Does High Cost-Sharing in Physician Care Reduce Health Care Utilization and Expenditures Differently for People with Severe Disease from Those without?
Physical Description: 1 online resource (156 p.)
Language: english
Creator: Xin, Haichang
Publisher: University of Florida
Place of Publication: Gainesville, Fla.
Publication Date: 2010

Subjects

Subjects / Keywords: care, cost, differently, disease, expenditures, high, physician, reduce, severe, sharing, utilization
Health Services Research, Management, and Policy -- Dissertations, Academic -- UF
Genre: Health Services Research thesis, Ph.D.
bibliography   ( marcgt )
theses   ( marcgt )
government publication (state, provincial, terriorial, dependent)   ( marcgt )
born-digital   ( sobekcm )
Electronic Thesis or Dissertation

Notes

Abstract: This study examines whether high cost-sharing in physician care reduces health care utilization and expenditures differently for people with severe disease from those without. By providing empirical evidence, this study?s findings inform policies that help protect and prevent severely ill individuals from exacerbating their poor health status and financial burden, as well as save costs for society at the aggregate level. The study adopted a cross-sectional study design from the 2007 Medical Expenditure Panel Survey data. In STATA, weights and variance were adjusted to account for the complex multi-stage, unequal probability, and cluster sampling survey designs. The potential endogeneity problem between cost-sharing and health care utilization or expenditures was addressed by a valid instrumental variable. Negative binomial regressions and two-part models were employed to analyze the utilization and expenditure data, using the bootstrap technique to incorporate split results in two-part models. In response to increased physician care cost-sharing, the study revealed a differential impact in the probability of physician care and total care expenditure models and the emergency room care utilization model, and opposite differential impact in the overall physician care and total care expenditure models. The severely ill were significantly associated with poor income. Thus, although the severely ill, in response to high cost-sharing, had a desire to reduce essential and necessary medical care to a lesser extent than the general health population, they actually reduced more, because high cost-sharing policies, augmented by their financial difficulties, greatly distorted their desires and voluntary behaviors. In response to high cost-sharing pressure, severely ill individuals appear to have experienced both substantial physician care reduction and emergency room care increase, and they were in worse clinical conditions. Therefore, current high cost-sharing policies should be replaced with low cost-sharing policies for individuals with severe illnesses to reflect their situations. This study highlights the necessity and importance of value-based insurance design in terms of differentiation of and specification for its target population. Furthermore, this study contributes to the current debate on health care reform. Specialized plans for subpopulations instead of a universal plan may want to be considered to complement existing public and private designs.
General Note: In the series University of Florida Digital Collections.
General Note: Includes vita.
Bibliography: Includes bibliographical references.
Source of Description: Description based on online resource; title from PDF title page.
Source of Description: This bibliographic record is available under the Creative Commons CC0 public domain dedication. The University of Florida Libraries, as creator of this bibliographic record, has waived all rights to it worldwide under copyright law, including all related and neighboring rights, to the extent allowed by law.
Statement of Responsibility: by Haichang Xin.
Thesis: Thesis (Ph.D.)--University of Florida, 2010.
Local: Adviser: Harman, Jeffrey S.

Record Information

Source Institution: UFRGP
Rights Management: Applicable rights reserved.
Classification: lcc - LD1780 2010
System ID: UFE0041643:00001

Permanent Link: http://ufdc.ufl.edu/UFE0041643/00001

Material Information

Title: Does High Cost-Sharing in Physician Care Reduce Health Care Utilization and Expenditures Differently for People with Severe Disease from Those without?
Physical Description: 1 online resource (156 p.)
Language: english
Creator: Xin, Haichang
Publisher: University of Florida
Place of Publication: Gainesville, Fla.
Publication Date: 2010

Subjects

Subjects / Keywords: care, cost, differently, disease, expenditures, high, physician, reduce, severe, sharing, utilization
Health Services Research, Management, and Policy -- Dissertations, Academic -- UF
Genre: Health Services Research thesis, Ph.D.
bibliography   ( marcgt )
theses   ( marcgt )
government publication (state, provincial, terriorial, dependent)   ( marcgt )
born-digital   ( sobekcm )
Electronic Thesis or Dissertation

Notes

Abstract: This study examines whether high cost-sharing in physician care reduces health care utilization and expenditures differently for people with severe disease from those without. By providing empirical evidence, this study?s findings inform policies that help protect and prevent severely ill individuals from exacerbating their poor health status and financial burden, as well as save costs for society at the aggregate level. The study adopted a cross-sectional study design from the 2007 Medical Expenditure Panel Survey data. In STATA, weights and variance were adjusted to account for the complex multi-stage, unequal probability, and cluster sampling survey designs. The potential endogeneity problem between cost-sharing and health care utilization or expenditures was addressed by a valid instrumental variable. Negative binomial regressions and two-part models were employed to analyze the utilization and expenditure data, using the bootstrap technique to incorporate split results in two-part models. In response to increased physician care cost-sharing, the study revealed a differential impact in the probability of physician care and total care expenditure models and the emergency room care utilization model, and opposite differential impact in the overall physician care and total care expenditure models. The severely ill were significantly associated with poor income. Thus, although the severely ill, in response to high cost-sharing, had a desire to reduce essential and necessary medical care to a lesser extent than the general health population, they actually reduced more, because high cost-sharing policies, augmented by their financial difficulties, greatly distorted their desires and voluntary behaviors. In response to high cost-sharing pressure, severely ill individuals appear to have experienced both substantial physician care reduction and emergency room care increase, and they were in worse clinical conditions. Therefore, current high cost-sharing policies should be replaced with low cost-sharing policies for individuals with severe illnesses to reflect their situations. This study highlights the necessity and importance of value-based insurance design in terms of differentiation of and specification for its target population. Furthermore, this study contributes to the current debate on health care reform. Specialized plans for subpopulations instead of a universal plan may want to be considered to complement existing public and private designs.
General Note: In the series University of Florida Digital Collections.
General Note: Includes vita.
Bibliography: Includes bibliographical references.
Source of Description: Description based on online resource; title from PDF title page.
Source of Description: This bibliographic record is available under the Creative Commons CC0 public domain dedication. The University of Florida Libraries, as creator of this bibliographic record, has waived all rights to it worldwide under copyright law, including all related and neighboring rights, to the extent allowed by law.
Statement of Responsibility: by Haichang Xin.
Thesis: Thesis (Ph.D.)--University of Florida, 2010.
Local: Adviser: Harman, Jeffrey S.

Record Information

Source Institution: UFRGP
Rights Management: Applicable rights reserved.
Classification: lcc - LD1780 2010
System ID: UFE0041643:00001


This item has the following downloads:


Full Text





DOES HIGH COST-SHARING IN PHYSICIAN CARE REDUCE HEALTH CARE
UTILIZATION AND EXPENDITURES DIFFERENTLY FOR PEOPLE WITH SEVERE
DISEASE FROM THOSE WITHOUT?



















By

HAICHANG XIN


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

2010

































2010 Haichang Xin



































To my parents









ACKNOWLEDGMENTS

I feel fortunate to have had the opportunity to study in the PhD program in Health Services

Research, where I was granted help and support from numerous people. I would like to express

my gratitude to my supervisory committee chair, Dr. Jeffrey S. Harman, for his invaluable advice

and support. His guidance has led me to set up the framework and structure for this study and

resolve technical challenges throughout my research process. I give my sincere thanks to Dr. R.

Paul Duncan. During a difficult time, his encouragement boosted my confidence and helped me

through that critical moment. I also would like to offer my sincere thanks to Dr. Niccie L.

McKay, who has been a constant source of invaluable advice. Her serious and careful style and

attention to detail have influenced me, and I have applied these to my studies. As a committee

member, Dr. Jing Cheng helped me refine my methodology and provided psychological support

for me. I would like to thank her for her efforts. I am also indebted to Drs. Ning Li, Xiaohui Xu,

and Chunrong Ai. All have provided useful additional support for this study.

My sincere thanks go to Dr. Frederick Rohde who explained technical problems in detail

to me and helped me through some analytic difficulties. During my study, Nancy Hamilton and

Dustin Heinen also provided valuable editing help to me. I have improved my writing skills after

reading their comments.

It would be a long list of others I should acknowledge. During the process they indirectly

supported me academically, materially, or psychologically.

I would like to thank my parents and my brother. Their eternal love for me is my

inexhaustible resource, and the strength of my drive for study and to establish my future career is

greatly attributed to them.









TABLE OF CONTENTS

page

A C K N O W L E D G M E N T S ..............................................................................................................4

L IS T O F T A B L E S ................................................................................................. ..................... 8

LIST OF FIGURES ............................................. .. .......... ...........................9

L IST O F A B B R E V IA T IO N S ............................................. ..................................................... 11

A B S T R A C T .......................................................................................................... ..................... 12

CHAPTER

1 INTRODUCTION ............................................ .. .......... ................................... 14

O v erview .............................................................................................. ......... 14
Stu dy O bjectiv es .......................................................................... ............. .................. 16

2 BACKGROUND AND SIGNIFICANCE.........................................................................17

B a c k g ro u n d ............................................................................................................................. 1 7
The Function of Health Insurance ........................................................18
M o ral H az ard ................................................................................................................. .. 19
C o st-S h a rin g .................................................................................................................. .. 2 0
P h y sician C are ............................................................................................................... .. 2 2
P rice E lasticity of D em and ...................................................................... ................ 23
S ig n ifican ce ................................................. ........ .................................................. ........ .. 2 4
Im pact of Cost-sharing on H health Status.................................................... ................ 25
Cost-sharing and Out-of-Pocket Cost Burden............................................................26
High Cost-sharing and Total Expenditures ................................................................27
Literature Review ........... .. .. ........... ... ...... ................. 30
Price Elasticity of Demand for Medical Care ............... ... ................30
Impact of Cost-sharing for Different Types of Medical Care ....................................32
A ll service type cost-sharing ...................................... ...................... ................ 32
P physician care cost-sharing ....................................... ....................... ................ 33
P rev mention cost-sharing ........................................... ......................... .................. 34
Inpatient care cost-sharing ........................................ ....................... ................ 34
Pharm acy cost-sharing .............. ...... ............ ............................................... 35
Em ergency room visit cost-sharing..................................................... ................ 36
Effect of Cost-sharing in Physician Care by Health Status........................................37
Absence of differential impact ........................................................37
Presence of differential im pact............................................................ ................ 39
R elated fi n d in g s ..................................................... .. ........................................... 4 1
Gaps and Limitations of Prior Studies....................................................................42









3 CONCEPTUAL FRAMEWORK ...................................................................................45

G rossm an's M odel .................................................................................... ....................... 45
A ndersen's B ehavioral M odel .. ...................................................................... ................ 46
D em and C urv e ........................................................................... ............. .................. 4 8
P ric e E la stic ity ........................................................................................................................4 8
Hypotheses ............................................................................... 53

4 M E T H O D S .............................................................................................................................5 8

D ata D description ................................................................ ...... ..................... 58
Measures and Operationalization ........................................................................................60
M measures of O utcom e V ariables ....................... ......................................... ...............60
Measures of Explanatory Variables..............................................................................62
Severity level ............................................................................................. 62
C o st-sh arin g lev el .....................................................................................................6 3
Measures of control variables................................................................................65
Study D design ............................................................................... ..... ..................... 67
S statistic a l A n a ly sis ..................................................................................................................6 8

5 R E S U L T S ..................................................................................................... ..................... 7 8

Overview ............................................................................. 78
D description of the Sam ple ................................................................................................ 78
V ariab le O p eration alization ....................................................................................................79
Missing Cost-sharing Values Imputation ............................... ...................................... 79
Test for the Multicollinearity between Disease Severity and Priority Conditions..........80
Test for Potential Endogeneity and IV Validity ..................................................................80
T h e H au sm an T e st ...........................................................................................................8 0
Test for IV Relevance and Exogeneity.........................................................................81
Use IV to Get Predicted Cost-sharing Values .................. ...................................81
The Multivariate Analysis Results .......................................................................................82
The Health Care Utilization Results.............................................................. ...............82
Physician care utilization ......................................................................................82
Primary care physician utilization.........................................................................83
E R c are u tiliz atio n ....................................................................................................8 4
Inpatient care utilization ....................................... ......................... ................ 86
The H health C are Expenditure R results ........................................................... ...............86
Physician care expenditure ............................... ....................... ................ 86
E R v isit ex p en ditu re ............................................................................ ...............89
Inpatient care expenditu re ....................... ............................................ ...............9 1
T total care expenditure ............................................... ............... ................ 94

6 D IS C U S S IO N ....................................................................................................................... 1 3 5

Sum m ary of B asic F indings................................................... ........................................... 135
Physician C are Price E lasticity ....................................................................... ................. 137



6









O their Sensitivity A nalyses ................................................. ............................................ 138

7 CONCLUSIONS .................................. .. ........... ..................................... 144

Sum m ary and Interpretation ....................................................................... ..... ............... 144
Negative Consequence of High Cost-sharing Policies ......... ................................... 144
P o licy Im p lic atio n s ...............................................................................................................14 5
L im station s .............................................................................. ............. .................. 14 8

L IST O F R E FE R E N C E S ....................................................... ................................................ 150

B IO G R A P H IC A L SK E T C H .................................................... ............................................. 156










































7









LIST OF TABLES


Table page

4-1 Sum m ary of outcome e m measures ......................................... ........................ ................ 76

4-2 E explanatory variables........................................................................................... 77

5-1 D description of the study sam ple (N = 13,020).............................................. ................ 98

5-2 Study sam ple characteristics (N = 13,020) ................................................... ................ 99

5-3 Association between disease severity and priority conditions (P <0.01).......................101

5-4 The first stage of IV 2SLS results with cost-sharing as dependent variable ................ 102

5-5 The association between IV and the error terms in the structural equation of health
care utilization and expenditure .................. .......................................................... 103

5-6 Cost-sharing value description by insurance types....... ... ...................................... 104

5-7 Negative binomial regression estimates for physician visits................ ...................105

5-8 Negative binomial regression estimates for primary care physician visits................... 107

5-9 Negative binomial regression estimates for ER visits .......................... ...................109

5-10 Negative binomial regression estimates for hospital admissions ................................. 111

5-11 Logit regression predicting probability of having any physician care expenditures ....... 113

5-12 Log transformed OLS regression estimates for physician care expenditures................115

5-13 Logit regression predicting probability of having any ER visits expenditure ...............117

5-14 Log transformed OLS regression estimates for ER care expenditures..........................119

5-15 Logit regression predicting probability of having any inpatient care expenditure ..........121

5-16 Log transformed OLS regression estimates for inpatient care expenditures .................123

5-17 Logit regression predicting probability of having any medical care expenditure .........125

5-18 Log transformed OLS regression estimates for total medical care expenditures ..........127

6-1 Differential impact summ ary by service types ....... .......... ....................................... 140

6-2 Significant differential impact summary by service types.................... ...................140









LIST OF FIGURES


Figure page

2-1 M oral hazard and social w welfare ........................................ ........................ ................ 44

2-2 Income transfer helps distinguish between efficient and inefficient moral hazard ...........44

3-1 Demand curve for severely ill and reference group...................................... ................ 55

3-2 Demand curve for severely ill and reference group with switched X and Y axis .............55

3-3 Demand curve for severely ill and reference group with switched X and Y axis in
M D v isits ....................................................................................................... ....... .. 56

3-4 Demand curve for severely ill and reference group with switched X and Y axis in ER
o r in p atien t c are ................................................................................................................ .. 5 6

3-5 Demand curve for severely ill and reference group with switched X and Y axis in
to tal c are ........................................................................................................... ........ .. 5 7

5-1 P-P plot for log transformed OLS regression on physician care expenditures .............129

5-2 Q-Q plot for log transformed OLS regression on physician care expenditures.............129

5-3 Residual-fitted plot for log transformed OLS regression on physician care
ex p e n d itu re s .................................................................................................................. ... 1 3 0

5-4 P-P plot for log transformed OLS regression on ER care expenditures ........................130

5-5 Q-Q plot for log transformed OLS regression on ER care expenditures .......................131

5-6 Residual-fitted plot for log transformed OLS regression on ER care expenditures ........ 131

5-7 P-P plot for log transformed OLS regression on inpatient care expenditures ..............132

5-8 Q-Q plot for log transformed OLS regression on inpatient care expenditures ..............132

5-9 Residual-fitted plot for log transformed OLS regression on inpatient care
ex p e n d itu re s .................................................................................................................. ... 1 3 3

5-10 P-P plot for log transformed OLS regression on total medical care expenditures.........133

5-11 Q-Q plot for log transformed OLS regression on total medical care expenditures .........134

5-12 Residual-fitted plot for log transformed OLS regression on total medical care
ex p e n d itu re s ..................................................................................................................... 1 3 4









6-1 The differential impact by health status in probability of having any physician care
ex p e n d itu re s .................................................................................................................. ... 14 1

6-2 The differential impact by health status in ER care visits .................... ...................141

6-3 The differential impact by health status in probability of having any total medical
care ex p en ditu res ............................................................................................................. 14 2

6-4 The differential impact by health status in integrated physician care expenditures ........142

6-5 The differential impact by health status in integrated total medical care expenditures... 143









LIST OF ABBREVIATIONS


MEPS-HC

ADL

IADL

CDHP

AMI

OLS

HIE

FFS

HMO

MCBS

MPC

DRG

NHIS

PSU

AHRQ

SBD

PCS

MCS

SF-36

BMI

SCHIP

IV

2SLS

H-L

AIC


Medical expenditure panel survey household component

Activity of daily living

Instrumental activity of daily living

Consumer directed health plan

Acute myocardial infarction

Ordinary least squares

Health insurance experiment

Fee for service

Health maintenance organization

Medicare Current Beneficiary Survey

Medical provider component

Diagnosis-related group

National Health Interview Survey

Primary sampling unit

Agency for Healthcare Research and Quality

Separately billing doctor

Physical component score

Mental component score

Short-form 36-item

Body mass index

State Children's Health Insurance Program

Instrumental variable

Two stage least squares

Hosmer-Lemeshow test

Akaike's Information Criterion









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

DOES HIGH COST-SHARING IN PHYSICIAN CARE REDUCE HEALTH CARE
UTILIZATION AND EXPENDITURES DIFFERENTLY FOR PEOPLE WITH SEVERE
DISEASE FROM THOSE WITHOUT?


By

Haichang Xin

August 2010

Chair: Jeffrey S. Harman
Major: Health Services Research

This study examines whether high cost-sharing in physician care reduces health care

utilization and expenditures differently for people with severe disease from those without. By

providing empirical evidence, this study's findings inform policies that help protect and prevent

severely ill individuals from exacerbating their poor health status and financial burden, as well as

save costs for society at the aggregate level.

The study adopted a cross-sectional study design from the 2007 Medical Expenditure

Panel Survey data. In STATA, weights and variance were adjusted to account for the complex

multi-stage, unequal probability, and cluster sampling survey designs. The potential endogeneity

problem between cost-sharing and health care utilization or expenditures was addressed by a

valid instrumental variable. Negative binomial regressions and two-part models were employed

to analyze the utilization and expenditure data, using the bootstrap technique to incorporate split

results in two-part models.

In response to increased physician care cost-sharing, the study revealed a differential

impact in the probability of physician care and total care expenditure models and the emergency









room care utilization model, and opposite differential impact in the overall physician care and

total care expenditure models.

The severely ill were significantly associated with poor income. Thus, although the

severely ill, in response to high cost-sharing, had a desire to reduce essential and necessary

medical care to a lesser extent than the general health population, they actually reduced more,

because high cost-sharing policies, augmented by their financial difficulties, greatly distorted

their desires and voluntary behaviors.

In response to high cost-sharing pressure, severely ill individuals appear to have

experienced both substantial physician care reduction and emergency room care increase, and

they were in worse clinical conditions.

Therefore, current high cost-sharing policies should be replaced with low cost-sharing

policies for individuals with severe illnesses to reflect their situations.

This study highlights the necessity and importance of value-based insurance design in

terms of differentiation of and specification for its target population. Furthermore, this study

contributes to the current debate on health care reform. Specialized plans for subpopulations

instead of a universal plan may want to be considered to complement existing public and private

designs.









CHAPTER 1
INTRODUCTION

Overview

People with severe illness carry a heavy disease burden; beside their overall poor health

status and functional status, either physically or mentally, they have a heavy out-of-pocket

financial burden (Kaiser Family Foundation, 2007). Their disease burdens are directly reflected

in high levels of health care utilization and spending (Conwell et al., 2005). In order to cut down

on outsized medical expenditures, current insurance policies rely heavily on high cost-sharing or

other price-related approaches. In this way, cost-sharing aims to reduce social welfare waste by

insured individuals due to moral hazard. Unfortunately, while cost-sharing can reduce utilization

of unnecessary services, current high cost-sharing policies might produce a new problem by

reducing utilization of effective and necessary care. Thus, a legitimate question arises: Is high

cost-sharing necessary to reduce health care utilization and expenditure for those with severe

health conditions?

This study focuses on cost-sharing in physician care, among which primary care plays a

central role in a health care delivery system. It is the first contact a patient makes with the health

care system, and is ideally the means by which health service delivery is coordinated in a

comprehensive and continuous manner to meet an individual's health care needs (Shi et al.,

2005). Beside the frontline role of primary care, as a whole physician care is an upstream service

that maintains people's health status and prevents individuals from avoidable downstream

emergency room (ER) visits, hospitalizations, and worse health outcomes (McCauley et al.,

1998, Chandra et al., 2007).

As a quantitative measure, price elasticity of demand can also assist us to evaluate people's

perception about whether high cost-sharing in physician care is appropriate for severely ill









patients. A related concept is cross-price elasticity of demand. Clarification of the relationship

between service types for severely ill patients can help us better estimate cost saving at a societal

level given the stringent budget for medical care, which also contributes to our discussion of

whether current high cost-sharing policies increase net expenditures.

High cost-sharing in physician care may result in worse health outcomes and heavier

financial burdens for individuals, and higher aggregate cost because it reduces needed health care

utilization and leads to a greater need for acute and inpatient care for those with severe health

conditions. In order to examine this research question, the study needs to compare the utilization

and expenditure reduction magnitude due to high cost-sharing for the severely ill group and for a

less severely ill group. This study focuses on the impact of cost-sharing on utilization of

physician care. Because physician care is likely responsive to price, decreased utilization of this

type of care due to cost-sharing might lead to subsequent ER visits or hospitalizations by

individuals with more severe and advanced illnesses. This study examines the effect of cost-

sharing for physician services on ER and inpatient care as well, specifically, if it has an offset

effect, which can help us better understand the potential differential impact on physician care and

overall health care utilization.

The research question is examined using a cross-sectional study design with a nationally

representative sample using the 2007 medical expenditure panel survey, household component

(MEPS-HC), data for analysis. Subjects are included who meet certain age and insurance

coverage criteria. Captured by the interaction of dummy variables severity and cost-sharing, the

differential impacts are examined in both the upstream physician care and the downstream ER

and inpatient care, in both utilization and expenditure measures and their total expenditures.

Weights and variance are adjusted to account for the complex multi-stage, unequal probability,









and cluster sampling survey designs in STATA. The potential endogeneity problem between

cost-sharing and health care utilization or expenditure is tested and confirmed. An instrumental

variable is then introduced to address the endogeneity. In multivariate analysis, negative

binomial regressions and two-part models are employed to analyze the utilization and

expenditure data. In addition, the study estimates the price elasticity and cross-price elasticity of

demand of physician care. Sensitivity analysis is also conducted to test the findings' robustness.

Study Objectives

The main purpose of this study is to examine the potential differential impact of high cost-

sharing in physician care on health care utilization and expenditure for those with and without

severe health conditions in the United States.

The specific aims of the study are

Aim 1: To examine the differential impact of high cost-sharing in physician care on

utilization and expenditures between severely ill people and healthier people.

Aim 2: To examine the cross-price elasticity of physician care on ER and inpatient care

utilization and expenditures between severely ill people and healthier people.

Aim 3: To compare the differential impact of cost-sharing in physician care on

expenditures for all services between severely ill people and healthier people.









CHAPTER 2
BACKGROUND AND SIGNIFICANCE

Background

Based on 2002 national health care spending data, a statistical brief reports that 5% of the

population accounted for 49% (almost half) of total health care expenses, while the bottom half

of the population only spent 3% (Conwell et al., 2005). That top 5% spent, on average, more than

17 times as much per person as those in the healthier half (Stanton, 2006). Estimated national

health care spending in 2002 was $1.6 trillion (Cowan et al., 2004), which means the top 5%

incurred about $800 billion. A close look at the statistical brief reveals they had much worse

health status. This 5% of the population were 11 times (45% vs. 4%) more likely to be in fair or

poor physical health as people in the healthier half of the distribution and 7 times (21% vs. 3%)

as likely to be in fair or poor mental health as compared to people in the lower half (Conwell et

al., 2005). Meanwhile, in terms of limitations in activities of daily living (ADL), or instrumental

activities of daily living (IADL), the top 5% of the group was 26 times (26% vs. 1%) more likely

to need help than those in the healthier half of the population (Conwell et al., 2005). This

indicates that this 5% of the population, although a small group, have severe diseases and carry

catastrophic disease burdens.

Meanwhile, Stanton's report indicates that a small number of medical conditions

accounted for most of the growth in total health care spending between 1987 and 2000, with the

top five medical conditions (heart disease, pulmonary disorders, mental disorders, cancer, and

trauma) accounting for 31%, and the 15 most expensive health conditions accounting for 44% of

total health care expenses (Stanton, 2006). These expensive health conditions can serve as a

rough measure for disease severity, although there are some discrepancies, that further warn that

severely ill individuals require much more attention and protection.









Obviously, this sub-set of the population has severe physical and mental problems. Poor

health and functional status can push them in the direction of life-threatening conditions. Their

huge medical expenditures were paid either by patients themselves or a third party. Due to

coverage limitation by insurance policies, catastrophic disease can easily impose heavy out-of-

pocket financial burdens on them; it is especially disastrous for those who are uninsured.

According to the Commonwealth Fund 2007 Biennial Health Insurance Surveys, adults who

were uninsured or underinsured for any time during a year had more than twice the rate of

medical bill problems and debt than those who were insured all year long (61% vs. 26%; Doty, et

al., 2008). Even for the insured, insurance policies vary widely in terms of plan type, service

type, and enrollee characteristics. Many people may still face wide coverage gaps and have to

pay out-of-pocket for certain services or pay much of the cost for services already covered by an

existing insurance plan (Doty, et al., 2008). Catastrophic disease can thus threaten severely ill

people to the point of impoverishment. In fact, the above report (Conwell et al., 2005) revealed

that persons in the top 5% of the expenditure distribution were 6.8 times (34% vs. 5%) more

likely to have out-of-pocket medical expenses exceeding 10% of the family income, and 6 times

(18% vs. 3%) more likely to have out-of-pocket medical expenses exceeding 20% of the family

income than those in the healthier half of the population.

In order to curb medical expenditures, a common insurance practice is to impose high cost-

sharing or use other price-related approaches (Remler et al., 2009). Before discussing the cost-

sharing role, this study will briefly introduce the health insurance function.

The Function of Health Insurance

Health insurance is basically modeled after traditional business insurance, and shares some

of its functions. By spending a fixed amount on premiums, people can be protected against

uncertain higher costs of medical care for undesirable health events. Thus health insurance









improves the affordability of health care to the individual (Remler et al., 2009). Moreover, health

insurance provides financial protection, preserving people's resources or income in case of

catastrophic disease and preventing them from becoming impoverished (Remler et al., 2009). In

addition, health insurance has two unique functions compared to business insurance: insulation

and social justice. It insulates people's medical care decisions from most fiscal consequences and

also protects their health. Relative to individuals without insurance, health insurance affords

people with equitable rights of access to health care benefits (Daniels, 2001). Numerous

empirical studies have consistently shown that health insurance makes a difference in whether

and when people get necessary medical care, where they get their care, and ultimately, the level

of health people maintain. Health insurance increases utilization and improves health outcomes

(Kaiser Family Foundation, 2008; Freeman, 2008). However, health insurance is also

accompanied by moral hazard.

Moral Hazard

Moral hazard means that people consume more health services when insured than they

would if uninsured. Pauly (1968) provided the original analysis of the welfare effect due to moral

hazard. Traditionally, it is believed that moral hazard wastes social resources. This argument is

illustrated in Figure 2-1. If an individual is uninsured, no one would pay a price higher than

market cost, so in the demand curve the marginal cost equals market price. The average cost of

medical care, at intersection price point A, corresponds to demanded quantity of medical service

(Qu). At this point, the benefit is equal to social expenditure, E = P*Qu and there is no social

welfare loss. On top of the Qu amount, suppose an individual is fully insured and thus faces an

out-of-pocket expense of zero, he may consume more services until reaching the maximum

amount Qi, where the demand curve intersects the zero price line. The insured individual actually









enjoys an extra benefit of area AQuQi, but society has to pay, as shown in the corresponding area

of AQuQiB, so there is a social welfare loss of area AQiB.

However, Nyman's revised theory about moral hazard argues for its beneficial component

(Nyman, 2004). The basis for his argument is that an insurer transferring income to an ill person

would allow us to distinguish between efficient and inefficient moral hazard. If the patient takes

advantage of the income transfer and spends it all on medical care due to moral hazard, the

income transfer would serve to relieve the individual's financial constraint, so his demand curve

would shift outward and parallel to the original demand curve (Figure 2-2). Then initial point A

will move to a new point, Ai, and point Qi to Qil, which indicates that both efficient and

inefficient moral hazard will increase. Although the inefficient moral hazard is still harmful, the

increased efficient moral hazard is desirable and beneficial, which is the area of AQuQu1A1.

Nyman's theory suggests insurance policies should be designed to match the appropriate level of

cost-sharing, effectively protecting the individual's efficient moral hazard while cutting down the

waste brought on by inefficiency.

Cost-Sharing

In order to curb escalating health care expenditures and reduce the resource waste

produced by the moral hazard embedded in health insurance, a cost-sharing mechanism is

devised to cut down the use of unnecessary service and bring the health care consumed closer to

what it would be without insurance, thus helping to save cost. Unfortunately, current insurance

company policy relies heavily on the high cost-sharing strategy, and excessively deviates from

what it was originally intended to do. Therefore, while resolving the moral hazard problem, the

current high cost-sharing policy may produce a new problem: cutting effective and necessary

care for severely ill individuals.









Basically there are three widely used forms of cost-sharing: copayments, coinsurance, and

deductibles. For patients with copayments, they pay a fixed fee amount for each medical service

sought or product purchased. Similarly, coinsurance requires patients to pay a fixed percentage

of the cost for each care episode. The third, deductibles, refer to the amount one must pay out-of-

pocket before insurance coverage begins. In addition, other commonly used variants include caps

and an out-of-pocket maximum.

While cost-sharing is designed to reduce the unnecessary care induced by moral hazard to

cut down health care expenditure and save cost, it has been shown to be a "blunt instrument,"

reducing medical care of both low value and high value, both cost effective and ineffective, and

both needed and unneeded (Manning, 1987; Haren, 2009). There may be many reasons for the

reduction of effective care. An important one is that cost-sharing tends to push the consumers to

look for cheaper yet similarly effective medical care or products, which assumes that they have

complete information to compare across providers or have trade-offs between price and

effectiveness of medical care. Unfortunately, the health care market is dominated by asymmetric

information. Ordinary consumers do not have the necessary tools, education, knowledge, time, or

money to acquire full information about provider performance, care efficacy, and the

consequences of forgoing care. Furthermore, in sophisticated cost-sharing designs, like consumer

directed health plans (CDHPs), uncertainty will contribute to their "bounded rationality" (Simon,

1957) and thus irrational decisions (Haren, 2009). This explanation is based on the assumption

that people have the ability to pay for medical care. Meanwhile, it may also be possible that

patients sometimes have to forgo or delay their care, even needed care, due to financial

constraints. Even if they are aware of health consequences, enrollees' health care decisions are

greatly influenced by financial considerations (Rukavina, 2009), and they have to trade off









between two difficult options; incurring higher cost burdens and getting treatment or choosing

not to consume health care service with potentially worse health outcomes.

Increased cost-sharing serves to undermine insurance functions by weakening an enrollee's

affordability and insulation. Moreover, as mentioned above, cost-sharing also threatens the

financial protection function of health insurance, resulting in some people's inability to pay for

needed care. Willingness to pay gives way to ability to pay, so access disparity violates the social

equality goal that everyone has equal access to health care. As Remler et al. (2009) pointed out,

by its very nature, cost-sharing partially undoes insurance, and can undermine exactly what

people with insurance hope to protect. As a result, beside the immediate consequence of

reducing needed or unneeded care, cost-sharing has further and greater impacts on health care

utilization, expenditures, and health outcomes.

Physician Care

It seems that there is no appropriate cost-sharing level that applies universally across

medical service types. Cost-sharing can be implemented in many service types, including

preventive care, physician care, emergency room visits, hospitalizations, prescription drugs, and

so on. In order to evaluate the impact of the current high cost-sharing level, this study will focus

on cost-sharing in physician care. Rather than the term "outpatient care," "physician care" is a

stricter and more specified one, covering primary care that reflects preventive care and specialist

care including clinical diagnostic and therapeutic services by MDs that do not require an

overnight stay in a health care institution, but can be provided in a physician's office or hospital

outpatient setting. Primary care is the conceptual foundation for physician services and plays a

central role in a health care delivery system (Shi et al., 2005). It is the first contact a patient

makes with the health care system so that health service delivery is coordinated in a

comprehensive and continuous manner to meet an individual's health care needs (Shi et al.,









2005). Beside the frontline role of primary care, physician care is an upstream service that

ensures an individual's access to and use of the health care system. Necessary physician care will

prevent individuals from avoidable downstream ER visits, hospitalizations, and worse health

outcomes. Thus, it plays a key role in maintaining health status. Although important, physician

care may be susceptible to cost-sharing and thus forgone because of out-of-pocket costs,

potentially leading to subsequent avoidable ER visits or hospitalizations. The reason for the

relatively high price-responsiveness of this type of care may be that people might not have the

professional knowledge to judge the importance of physician care.

The potential differential impact for the severely ill will highlight the significance of

physician care. Forfeiture of this frontline and upstream service for the sick may worsen their

health conditions and widen the gap of health status among the population.

Price Elasticity of Demand

One role of price elasticity of demand is that it provides a quantitative guide to the effect of

cost-sharing on health care utilization. Thus, it further assists us to evaluate people's perceptions

about whether cost-sharing in physician care affects severely ill patients differently from people

without illness or who have less severe illnesses. It measures responsiveness of quantity

demanded to changes in price. Ed= AQ% / AP%. Specifically, demand is:

Elastic if Ed > 1

Unit Elastic if Ed = 1

Inelastic if Ed < 1

Here, price refers to cost-sharing level. In this study, price elasticity of demand is

calculated from a comparison of change in the price and demand quantity. It can be influenced

by many factors, like income and preference, and used in many areas, like disease types, service

types, and subpopulation types.









A related concept is cross-price elasticity of demand, which measures the responsiveness

of the demand of one good to a change in the price of another good. Clarification of the

relationship between physician care and its downstream ER visits or hospitalizations for severely

ill patients will help us better estimate cost saving or net expenditures at a societal level given the

stringent budget for medical care, which also contributes to whether current high cost-sharing

policies for physician care are appropriate for the severely ill. Price elasticity always has a

negative sign, the focus is on elasticity magnitude. While in terms of cross-price elasticity, the

focus will be transferred to its sign. This study will implicitly examine whether physician care

and ER or inpatient care are complements or substitutes. Positive cross-price elasticity indicates

the existence of an offset effect, and physician care and ER or inpatient care are substitutes. ER

or hospitalization care is much more expensive than office-based or outpatient doctor visits and,

at a societal level, this will become a great waste of social medical resources, which highlights

the importance of physician care. Physician care, if forgone, will incurred higher prices either

clinical or financial or both. A RAND study (Manning et al., 1987) indicates that physician care

and hospitalizations were complements, but the result was not statistically significant, and the

relationship may refer to the overall study sample, but not the sicker subgroup.

Physician care is the upstream service that ensures an individual's access to and use of the

health care system. Thus, lower utilization of necessary physician care by cost-sharing pressure

could worsen a patient's health status and lead to avoidable and expensive ER visits and

hospitalizations that result in greater overall health care costs.

Significance

Having briefly introduced the background information above, the study will focus on the

significance of the research question: If high cost-sharing in physician care is necessary to

reduce health service utilization and expenditure for people with severe disease, as these process









indicators-people's behaviors-can indicate, there is the potential for severe consequences on

the following aspects.

Impact of Cost-sharing on Health Status

If this study confirms the differential impact by health status, it will reveal the

phenomenon that high cost-sharing will have a unique impact on severely ill individuals, who

tend to resist a reduction of their current health care level-one of the important health

determinants that directly influences health outcomes, since forfeiture or delay of needed care

will worsen their existing condition. By setting up financial barriers, high cost-sharing for this

vulnerable group will deter or deprive them from getting access to needed primary care and

specialists' services that are essential to maintain their health.

This is indirectly evidenced by the RAND findings (Manning et al., 1987). For low income

individuals with health conditions, reduced care due to increased cost-sharing adversely affected

their health. In the RAND study, cost-sharing was implemented for all service types, but not

limited to physician care only. In addition, the result referred to people who were both sick and

poor, not just sick. Other indirect evidence comes from a study examining the elderly with

chronic conditions, which can be taken as a proxy measure for severe conditions (Chandra et al.,

2007). They found that increases in physician care and prescription drug copayments had little

effect on hospital use for an average elderly person, but for chronically ill elderly patients there

was a significant offsetting rise in hospital admissions as physician and drug use fell. This is also

indirect evidence that increased cost-sharing level in physician care is harmful to the health

status for the severely ill, as subsequent downstream inpatient care is used as a proxy measure

for health outcomes, although the result cannot be attributed exclusively to copayment increase

in physician care.









Cost-sharing and Out-of-Pocket Cost Burden

Severely ill individuals usually have a higher out-of-pocket financial burden than less

severely ill individuals. Conwell et al. (2005) indicated that people in the top 5% of the

expenditure distribution were 6.8 times (34% vs. 5%) more likely to have out-of-pocket medical

expenses exceeding 10% of the family income and 6 times (18% vs. 3%) more likely to have

out-of-pocket medical expenses exceeding 20% of the family income than those in the healthier

half of the population. In 2003, 12% of all non-elderly adults had out-of-pocket costs exceeding

5% of family income, while 19% of those with chronic conditions faced this same level of out-

of-pocket costs (Kaiser Family Foundation, 2007). Roughly 14% of all U.S. families reported

problems paying their medical bills, but among families with a member in fair or poor health, the

proportion increased to 25% (Kaiser Family Foundation, 2007).

Increased cost-sharing in physician care or other medical service will increase individuals'

financial burden, and put them in a difficult situation. It is reported that among insured adults

whose health plans limited the total amount they could spend, 43% incurred medical bill

problems and unpaid debt, compared with 27% of adults who did not have total-dollar limits

(Doty et al., 2008). Facing this dilemma, patients would spend as much of their resources as

possible to maintain health. Rukavina (2009) reports that 24% of Americans buying health

insurance still incurred debt for medical bills, which indicated that they had exhausted available

resources to maintain necessary medical care when insurance was based on cost-sharing.

Furthermore, the financial burden from cost-sharing will also influence individuals' health

status by intensifying their stress and distress, exacerbating their existing poor health status. For

example, mood plays a key role in the disease aggravating process in acute myocardial infarction

(AMI) (Bax et al., 2008, Pearte et al., 2006). Although many patients can ultimately overcome









these difficulties and get necessary care, their physical and psychological condition may still be

indirectly harmed.

High Cost-sharing and Total Expenditures

At a societal level, the severely ill incur tremendous medical expenditures. High cost-

sharing is designed to reduce unnecessary service utilization to save costs, but, for severely ill

patients it may not achieve this desired goal. Ethically, even if it could, this should not come at

the expense of worsening their health outcomes and increasing their financial burdens, which

will undermine the public health goal to maintain a high level of population health.

Empirically, some studies suggested that the use of cost-sharing may result in higher

overall costs for the low income population on Medicaid (Helms et al., 1978; Tamblyn et al.,

2001). The Helms and colleagues' study reported that after implementation of a $1.00 copay for

physician care, ambulatory care utilization declined by 8%, but subsequent inpatient care use

increased by 17%, which resulted in 3% to 8% higher total Medicaid costs. Low income

individuals often suffer from poor health, since their financial constraints usually deter them

from getting necessary care to maintain their health. Among the Medicaid population, 18%

reported fair and poor health status (Julie et al., 2003).

Elderly individuals also often serve as a proxy for individuals in poor health. For the

Medigap beneficiaries, a study indicated that cost-sharing policies would lead to a smaller cost

saving by at least 2% to 7% for sick people compared to healthy people (Remler et al., 2003).

Similarly, a recent study suggested that, for elderly people, reduced drug use and physician visits

due to cost-sharing were offset by increased subsequent health care utilization, especially

hospitalizations (Chandra et al., 2007), which is contrary to the RAND findings (Manning et al.,

1987) of no offset rising hospitalizations. The subsequent medical expenditure may offset or

even exceed the saved cost, especially for Medicare, although the offset effect in response to









cost-sharing needs further exploration because the cost-sharing policy was instituted on

physician care and pharmaceuticals at the same time. The proportion who reported fair and poor

health in Medicare was even higher than in Medicaid, as much as 28% in 2008 (Kaiser Family

Foundation, 2008). Medicare is publicly funded, so policy makers and taxpayers are concerned

about whether Medicare has best used limited public funding to achieve maximum utility. It is

reported that, for the Medicare insurance trust fund, costs will exceed income from 2008 to 2010,

which will be exhausted by 2019 (Eppig, Medicare Current Beneficiary Survey, ca.2008),

although there are conflicting data and opinions. No matter what the year may be, the goal of

cost containment has become imperative.

Some other studies examining the pharmaceutical cost-sharing policy also provide indirect

evidence. One study suggested that for elderly people who use inhaled medications cost-sharing

may increase net expenditures (Dormuth et al., 2009). By assembling the data of pharmacy and

medical claims from 1997 to 2002 from 88 health plans and 25 employers, Goldman et al. (2006)

simulated a prescription copayment design by health status. Specifically, relative to the universal

$10 copayment policy, high and medium risk people were exempt from cost-sharing, while the

low risk group would have either $10 or $22 copayment. The study estimated that this new

design would reduce the number of hospitalizations by between 80,000 to 90,000 annually and

the number of ER visits by 30,000 to 35,000, resulting in net aggregate savings of more than $1

billion. This study provides indirect evidence that current cost-sharing is inefficient. If a high

cost-sharing policy ends up losing money, whether in Medicare, Medicaid, or private plans, it is

a considerable waste of social and medical resources at the aggregate level. These results from

the low income and elderly population indirectly imply that the cost saving performance of the

high cost-sharing policy for the severely ill is unsuccessful. If so, for this vulnerable group the









widely adopted current insurance policy design might be problematic, not only clinically, but

also financially. Under these circumstances, an evaluation of an insurance policy design for the

target population has practical policy implication. If the differential impact in expenditure is

confirmed, the results will provide persuasive evidence to inform insurance plans that policies

denying coverage for physician care or primary care will end up reducing profits, which was

labeled penny wise, pound foolish, and to motivate them to adopt a smarter benefit structure that

covers physician care. This will both benefit an enrollee's health status and save the insurers

money.

In summary, high cost-sharing for severely ill people potentially may not only pose a threat

to an individual's health status and increase the out-of-pocket financial burden, but also waste

limited medical resources.

The present study will provide empirical evidence to evaluate the potential differential

effect of high cost-sharing for physician care between the severely ill and healthier population.

By observing severely ill people's behavior, specifically their health care utilization change in

response to high cost-sharing pressures, this study will allow researchers to evaluate the effect of

high cost-sharing policies and whether they are necessary to cut down health care utilization and

expenditures for the severely ill.

A high cost-sharing policy naturally will directly reduce health care utilization, regardless

of subpopulations. If the differential impact of cost-sharing is confirmed, that means severely ill

people, in response, may reduce less health care utilization and expenditure, especially the

efficient moral hazard in utilization that is crucial to maintain their health. This phenomenon

may further reveal the potential underlying truth that efficient and inefficient moral hazard share

differs by severity. Severely ill people usually have less inefficient moral hazard and more









efficient moral hazard than healthy people. Therefore, these vulnerable people should be treated

differently. Specifically, a low cost-sharing policy should be designed to reflect and match their

situation. This study will contribute to inform the necessity and importance of insurance policy

design in terms of differentiation and specification for its target population, so as to best protect

and prevent that population from exacerbating its poor health status and financial burden, and to

save costs for the society at the aggregate level.

Literature Review

Price Elasticity of Demand for Medical Care

Price elasticity of demand measures the responsiveness of quantity demanded to changes

in price. This quantitative measure helps us better understand the actual effect of cost-sharing.

Numerous studies have examined price elasticity of demand across service types and so far their

findings are inconsistent. Using state variations in coinsurance rates from the American Hospital

Association's hospital survey and National Center of Health Statistics data with time series

regression, Feldstein (1971) estimated a price elasticity of -0.49 for hospital bed days (Cutler et

al., 1999). Based on the study conducted at Stanford University with a quasi-experiment design,

Phelps and Newhouse (1972) calculated an elasticity of physician visits of -0.14 with ordinary

least squares (OLS) estimation and -0.118 with the Tobit approach (Rice et al., 1994; Cutler et

al., 1999). Using the Tobit estimate of 1960 Survey of Consumer Expenditure, a cross sectional

study, Rosett and Huang (1973) estimated a price elasticity of -0.35 to -1.5 for hospitalizations

and physician service (Cutler et al., 1999). The RAND health insurance experiment (HIE)

reported an overall estimated elasticity of medical service spending of -0.2 (Manning et al.,

1987), which is supported by a summary of more than 20 studies, indicating that the total price

elasticity of demand for medical services was approximately -0.2 (Cutler et al., 1999). Goldman

et al. (2007) summarized hundreds of papers and found that for the prescription drug service,









elasticity ranges from -0.2 to -0.6. This is consistent with a study by Landsman et al. (2005),

indicating that elasticity of demand for drugs was generally low for asymptomatic conditions,

ranging from -0.16 to -0.10, and moderate for symptomatic conditions, ranging from -0.60 to

-0.24. The Landsman et al. paper also cited the results of several previous studies in the 1980s

and early 1990s, indicating that demand for prescription drugs was highly inelastic, with values

ranging between -0.33 and -0.10 for small absolute changes in price (Smith et al., 1993).

Goldman et al.(2006) another paper revealed that specialty drug use was largely price

insensitive, with price elasticity ranging from -0.01 to -0.21. Most of these studies agreed that

price elasticity of medical services and drugs varied within the range of 0 to -0.6. Contrary to

these results, however, a recent study by Chandra et al. (2007) found that physician office visits

and prescription drug utilization were very price sensitive. For pharmaceuticals the price

elasticity varied between -0.20 and -1.4, a range that crossed the boundary of unit elasticity. For

office visits, the estimated price elasticity was between -1.38 and -1.90, again completely elastic.

These findings and those of Rosett and Huang (1973) are surprising since they pose a significant

challenge to the traditional belief that the demand for medical care is inelastic.

In summary, the price elasticity for each service type is as follows: from -0.2 to -1.4 for

prescription drugs; from -0.14 to -1.9 for physician care; from -0.35 to -0.49 for hospitalizations;

and -0.2 for overall medical services. The reason for the mixed results may be many because

each service type covers a wide range of disease severity and therapeutic categories. It may also

suggest that efforts to evaluate price elasticity for each individual service type may be

incomplete, since they can be substitutes or complements. A holistic evaluation for all service

types may improve our understanding.









Moreover, for physician care, the range of price elasticity from -0.14 to -1.9 could be due

to differences in the health status of the samples used to create these estimates. The wide range

indicates that if price elasticity is near -0.14, physician care has little room to be cut down for the

severely ill, and is uniquely valuable and essential to maintain health. Thus, physician care

cannot be replaced or substituted; there would be a differential impact of high cost-sharing by

disease severity. On the other hand, if price elasticity is approximately -1.9, then physician care

is elastic and sensitive to price, and cannot be uniquely valuable because it can be substituted. In

this case, even for severely ill people, high cost-sharing would likely cut back on their needed

care, which would result in greater demand for more expensive acute care services downstream.

Namely, there could be no differential impact between groups in physician care itself, but

possible differential impacts downstream, in ER or inpatient care. If so, the whole picture still

indicates the potential differential impact of overall service amount between the severely ill and

the reference group.

Given the quantitative measure of cost-sharing effect by service types, this study will

discuss findings from empirical literature reports on cost-sharing effects. Rice et al. (1994) and

Remler et al. (2009) systematically reviewed the empirical evidence of cost-sharing on

utilization, expenditure, and health outcomes. Their findings are summarized below, as are a few

recent publications.

Impact of Cost-sharing for Different Types of Medical Care

All service type cost-sharing

The RAND-HIE (Manning et al., 1987) is so far the only social experiment to determine

the effect of patient cost-sharing on the utilization and cost of medical services and on patients'

health status for all kinds of service types. The insurance design incorporated varying levels of

cost-sharing and the out-of-pocket maximums, as well as a deductible plan. All the plans covered









a comprehensive set of services that included outpatient services, inpatient services, prescription

drugs, and preventive services for twelve fee for service (FFS) and two health maintenance

organization (HMO) plans. The RAND-HIE study focused on the non-elderly-a population

younger than 62 years old. This study produced a series of results, among them: increasing

coinsurance levels reduced medical use and expenditure. Specifically, individuals with a 25%

and 95% coinsurance policy had medical costs 23% and 46% lower than those receiving free

care, respectively. By contrast, there were no differences in utilization by groups with differing

out-of-pocket maximums. Deductibles also appeared to reduce service usage. Cost-sharing

reduced both ineffective and effective care, and increased cost-sharing did not affect health status

for people of average health, but did adversely affect that for the sick poor (Manning et al., 1987;

Rice et al., 1994).

Physician care cost-sharing

One quasi-experimental study examined the health care use and cost for faculty and staff

and their dependents at Stanford University from 1966 to 1968 (Scitovsky et al., 1972), during

which a 25% coinsurance requirement was instituted on physician inpatient services and all

outpatient services, including, in 1967, that for ancillary services. The cost-sharing requirements

for hospitalizations did not change during the same time frame. Results revealed that the

utilization and cost of physician services fell considerably, by about 25%; a follow up study

indicated that this effect remained stable over four years. Another early quasi-experimental study

in 1977 found a similar effect, revealing an average $7.50 copayment for physician visits to be

associated with a dramatic drop in utilization for both inpatient and outpatient services among

numbers of the United Mine Workers (Scheffler, 1984). However, the author also discussed

some potential threats to the internal validity of the findings. Most other later studies also found

that greater physician visit cost-sharing was associated with fewer office visits (Chandra et al.









2007, Cherkin et al. 1989, Roddy et al. 1986). However, one study that followed enrollees for

three years is an exception to the findings, reporting an increase in utilization in the third year

(Feldman et al., 2007). Regarding reduced types of office visits, these studies also produced

mixed results. Cherkin et al. (1989) found that for those with cardiac disease, implementation of

office visits copayment selectively decreased physical exams and primary care visits, but did not

influence immunizations, cancer screenings, or specialist visits. Roddy et al. (1986) found that

the reductions in physician visits were substantial for both prevention and acute self-limiting

conditions (ones that would clear up on their own). Similarly, Hibbard et al. (2008) found

reductions for both high and low effective care (defined by the State of Oregon's Prioritized List

of Health Services). Examining the impact duration of cost-sharing, Hibbard et al. found that the

reduction in office visits among CDHP enrollees lasted for two years. Roddy et al., on the other

hand, found that the first year reduction in office visits returned to the baseline rate beginning in

the second year.

Prevention cost-sharing

Studies consistently suggested that cost-sharing for preventive care was associated with

less preventive services across a wide range, including pap smears, preventive counseling,

clinical breast exams, and self-monitoring of blood glucose for diabetics (Karter et al., 2003;

Solanki et al., 1999). So far, no direct evidence has shown whether or not reductions in

preventive care due to cost-sharing affect health.

Inpatient care cost-sharing

Two studies (Feldman et al., 2007; Parente et al., 2004) examined CDHPs with a

deductible of $1,500 for individual policies, and they found that, compared to other plans,

hospitalizations actually increased for CDHPs, although other forms of care fell. There may be a

number of reasons for these findings. One, for example, is that the increased deductible reduced









CDHP members' outpatient care, they got sicker and in turn used more inpatient care.

Alternatively, specialist care preceding hospitalizations would have consumed a substantial

portion of the deductible, allowing patients to easily exceed the threshold and receive excessive

care without copays.

Pharmacy cost-sharing

As expenditure grows dramatically in pharmaceuticals, numerous studies examined newly

emerged cost-sharing devices in this area. Pharmacy cost-sharing schedules work through a

mechanism of different rates for different drug types. The lowest rate is for generic drugs, the

middle rate for preferred brand-name drugs, and the highest rate for non-preferred brand names.

In general, sizable changes in incentive-based formulary cost-sharing affect people's behavior,

including those with chronic illness. Goldman et al. (2007) and Gibson et al. (2005)

comprehensively reviewed pharmaceutical cost-sharing publications from 1985-2006 and 1974-

2005, respectively, and found that for every 10% increase in pharmaceutical copayment or

coinsurance, there was a decrease of 2% to 6% in drug spending, depending on the class of drugs

and condition of the patient.

For chronically ill subgroups, some studies show that drug cost-sharing increases

utilization in at least one of the following service types: office visits, hospitalizations, or

emergency care. Also greater use of inpatient and emergency medical services are associated

with higher copayments or cost-sharing for prescription drugs or benefit caps. Similarly, lower

cost-sharing was associated with considerable increases in drug use that, in turn, might be

associated with significant reductions in ER and hospital usage (Chernew et al., 2008; Goldman

et al., 2006). When the study population was not limited to those with certain chronic illnesses,

increased drug copayments were not associated with more outpatient visits, hospitalizations, or

ER visits (Fairman et al.,2003; Motheral and Fairman, 2001; Johnson et al.,1997; Smith and









Kirking, 1992), as the reason for this difference might be that the healthier group makes up the

majority of the study population and dilutes the effect of their severely ill counterparts.

In terms of direct health outcomes, Zeber et al. (2007) found that increased medication

copayment for schizophrenic veterans reduced their refills of psychiatric drugs, resulting in a

modest increase in inpatient admissions and may have hurt the veterans' health and society at

large. Another study using a before-and-after design found that increased pharmaceutical cost-

sharing in Quebec reduced the use of essential drugs, which in turn increased ER visits, but not

heart attack mortality (Tamblyn et al., 2001). The study also reported reduced use of nonessential

medications without apparent adverse effect.

Emergency room visit cost-sharing

Studies have consistently found that emergency room visit cost-sharing reduces ER

utilization (Hsu et al., 2006; Selby et al., 1996; Wharam et al., 2007). With respect to the type

and value of ER visits reduced by cost-sharing, the studies consistently found that visits defined

by high value-"high severity," "time sensitive," or "always an emergency"-were not

significantly reduced by ER cost-sharing; conversely, large reductions were observed in ER

visits considered "low severity" or "often not an emergency." Furthermore, no increases were

observed in hospitalizations, intensive care unit admissions, or mortality rates. There may be

some threats to the internal validity of these studies because they tracked individuals for just one

year and because adverse outcomes are uncommon occurrences.

For other service type cost-sharing, studies showed that a $20 copayment for outpatient

mental health services also reduced the likelihood of receiving outpatient mental health care

(Simon et al., 1996).

In summary, cost-sharing is a blunt instrument. In its immediate impact, it can reduce both

unnecessary and necessary care and thus undermine health insurance functions. Cost-sharing can









further reduce health care utilization and expenditures for most service types, including

physician care, preventive care, ER visits and pharmaceuticals, and for most subpopulations,

including both the healthy and the sick, most of the time. Moreover, cost-sharing may influence

harmful health outcomes.

Effect of Cost-sharing in Physician Care by Health Status

Only a few analyses have studied this exact topic for physician care cost-sharing, although

quite a number of papers touch on this topic of cost-sharing for other services, such as pharmacy.

Results are mixed.

Absence of differential impact

The RAND-HIE study (Manning et al., 1987) provides an important and representative

result. As described in that paper, an important goal of HIE is to study how the response to cost-

sharing varied across subgroups. RAND results explicitly indicated that higher patient

copayments reduced medical utilization for a variety of subpopulations (Gruber et al., 2006).

There was no differential response in expenditure levels to health insurance coverage between

the healthy and the sick (Manning et al., 1987). This indicates that both groups had similar high

or low expenditure changes. The study explained that the reason for this striking fact was due to

the upper limit feature; namely that the sicker group was less responsive to the increased

coinsurance and individuals were more likely to exceed their upper limits on out-of-pocket

expenditures and receive some free care. However, a closer look at the direction of the healthy

group indicated that, lacking a differential response in expenditure level for the two groups, the

healthy group should follow the same pattern and also be less responsive to increased cost-

sharing and to exceed the upper limit. This may not be reasonable. As mentioned before and

consistent with RAND results, reduced care for people of average health due to increased

coinsurance does not adversely affect their health. Thus, they would be more responsive to









increased coinsurance and would reduce their care use. For this group, cost dominates health

status considerations and becomes the priority utility concern unless such individuals value the

free care much more than their out-of-pocket expenditures, which they pay first. More

reasonable, however, is the opposite, tentative explanation in that paper: given no interaction

between plan and health status, the sickly exhibit more discretion at the margin and are similar to

the healthy group by being more responsive to increased cost-sharing. If so, that means the

disease condition for the sick group was not considered severe enough, allowing group members

to forgo some discretionary care. This group, then, does not represent the severely ill

subpopulation, which was not our main interest.

Alternatively, the sickly group may have offset care utilization with more ER visits or

inpatient care, in the short run reducing needed care in a fashion similarly to that of the reference

group. Gruber (2006) stated that there were no offset effects in the RAND result, but he did not

specify whether this finding refers to the subgroup of average health, the sick, or as a whole.

Moreover, the recent study by Chandra et al. (2007) mentioned above found significant offset

utilization for chronically ill patients, indirectly suggesting a significant insurance and health

status interaction. For these people, forgoing needed care even temporarily means they would

have to make it up later on. Also the RAND results suggest no "short run bias," implying that the

sick group had only a mild or at most a moderate condition, rather than "severe," allowing these

individuals to forgo some discretionary services to the extent exhibited by the healthy group.

Still another explanation for the lack of differential impact is that service types in this study

where expenditures are examined consist of a combination of both physician care and

hospitalization. Their different price elasticities may cancel the potential differential impact. As

mentioned above, physician care has high price elasticity either within or beyond the unit









elasticity boundary. In a 2007 study (Chandra et al.), the estimated price elasticity for office

visits is between -1.38 and -1.90, so that the sick group can be similarly price sensitive as the

healthy group. The potential pronounced differential impact in inpatient care, due to its narrow

range from -0.35 to -0.49 in price elasticity, may then be mitigated by physician care. Moreover,

the interaction effect is examined by expenditure instead of utilization, which may be further

confounded by price in each service type and make this result even more complicated.

In summary, when used for the purpose of this paper, the RAND study has some

limitations. The potential differential impact is only examined by expenditure, but not by

utilization; cost-sharing is implemented for all service types, and not limited only to physician

care. Furthermore, the RAND study population is the non-elderly, and thus not representative of

the general population. In addition to the RAND study, a number of studies in the literature on

pharmacy cost-sharing also indirectly found no differential impact by health status (Goldman et

al., 2006; Fairman et al., 2003; Motheral et al., 1999; Doshi et al., 2009).

Presence of differential impact

Link et al. (1980) found that Medigap policies increased physician visits by different

amounts according to chronic conditions: 42% for those without chronic conditions, only 5% for

those with at least one chronic condition. Here, health status or disease severity is indicated by

chronic conditions. A similar study by Cartwrite et al. (1992) found that Medigap policies

increased medical expenditure by different amounts according to health status: 25% for poor

health, 35% for fair health, 45% for good health, and 95% for excellent health. Both studies

indicated the healthier beneficiaries had a greater utilization response than less healthy ones,

their care usage likely reflecting a greater share of unnecessary need and thus a greater

sensitivity to cost-sharing change. Given the same change in cost-sharing levels, their utilization

will increase or decrease proportionately more than that of the severely ill group. Using 1995









Medicare Current Beneficiary Survey (MCBS) data, one recent study (Remler et al., 2003) also

examined the Medicare beneficiaries with and without supplemental insurance either privately

purchased or from an employer-sponsored retiree plan. Investigators found that the severely ill

group, measured by worse self-reported health status or functional health was less sensitive to

cost-sharing for hospital care but not for physician care. The above three studies share certain

limitations; results can only be applied to the elderly population, cost-sharing or supplemental

insurance is not focused on physician care, and investigators did not directly compare the effect

of cost-sharing by health status or with interactions. Moreover, these studies use Medigap

employer-sponsored insurance and individually purchased policies relative to the group without

any supplemental plans to measure different cost-sharing levels, which is a convenient but rough

measure. If the common practice of 20% of cost-sharing level applies to both Medicare and

private insurance plans, a rough estimate of cost-sharing level for the group lacking supplemental

plan (Medicare only) is 20%, and that for the group with a supplemental plan is 20%*20% = 4%.

Thus, these studies end up comparing high and low cost-sharing level with two relatively fixed

values-20% and 4% without any range. These two values are only a special case and lack

generalizability for high and low cost-sharing levels. It would be better to have a range or

variation for each group.

A counter result of differential impact by health status was found by McCall et al. (1991).

This study found that having a Medigap policy increased inpatient hospital use by 31%, part B

services by 42%, and total charges by 36% for people in poor or fair health, but had little effect

on use of inpatient hospital and physician services for people in good or excellent health. It is

natural that possession of a Medicare supplemental policy would relieve a person's financial

anxieties and sick people tended to increase their health care utilization. If this increased amount









for health care use and cost is needed, one expects that the counterpart would be larger for people

in good health due to their potential greater share of inefficient moral hazard. However, the

opposite results were found, suggesting there was no moral hazard effect for the healthy

associated with a more generous insurance policy.

Related findings

Furthermore, some studies provide indirect evidence of a differential impact. The study by

Chandra et al. (2007) reported that increases in physician care and prescription drug copayments

had little effect on hospital use for an average elderly person; for chronically ill elderly patients,

however, there was a significant offsetting rise in hospital admissions as physician and drug use

declined. The temporary health care reduction was later made up to some extent, which

indirectly suggests the differential health care utilization reduction between severely ill and

healthier groups. A related but opposite result from the RAND study (Manning et al., 1987)

indicated that physician care and inpatient care are complements rather than substitutes. This

finding is not statistically significant, however, and the relationship may refer to the overall study

sample but not the sicker subgroup.

Cherkin et al. (1989) found that for those with cardiac disease, copayment for office visits

selectively decreased physical exams and primary care visits, but did not influence

immunizations, cancer screenings, or specialist visits. Here, cost-sharing and disease severity

would compete against each other to affect health care use. The result depended on comparison

of degree of disease severity and cost-sharing; namely, utility priority. Individuals will choose

the less serious consequence to minimize their adverse utility. The results convey an implicit

message that physical exams and primary care visits-the relative discretionary services-can be

cut down, while cancer screenings or specialty care-services needed for cardiac patient

survival-cannot.









The RAND-HIE result indicated that people with higher coinsurance rates were less likely

to seek any inpatient or outpatient care. However, once care was sought, the amount received and

its cost did not vary by coinsurance (Rice et al., 1994). This may imply that the services actually

sought are not discretionary and would not allow some delay. Under this condition, the cost-

sharing level would not affect care use and cost for the sick, unlike the general effect of cost-

sharing for average healthy people. Thus, there might be a differential impact on care utilization

by health status. In addition, a number of studies on pharmacy cost-sharing also indirectly found

differential impacts by health status (Gilman et al., 2008; Harris et al., 1990; Landsman et al.,

2005; Goldman et al., 2004, Stuart et al., 1999).

Gaps and Limitations of Prior Studies

In summary, the literature results are mixed on this topic. Some found no differential

impact, while others found that there was, within which results both positive and negative were

also reported. Similarly conflicting results were found for the offset effect. For the mixed results,

the reasons may be due to different study samples, plan benefit structure, service types, different

methodologies, and time frame. For this study's purpose some gaps exist in prior studies that

limit both their internal and external validity. First, cost-sharing is not focused on physician care,

it is implemented either in combination with some other service types or in general, covering

most service types. It is, therefore, difficult to attribute the differential impact to physician care

cost-sharing. Second, these results can only be applied to some subpopulations, like the elderly

or non-elderly, and are therefore not representative of the whole population. Third, most of the

prior literature suffers from methodological drawbacks. The studies examine the relationship

within severe and healthy groups respectively, without directly comparing them or using an

interaction term. It may be possible that the main effect of cost-sharing is significant in one

group while not in another group when, by putting the severely ill and the reference group









together, the difference of the main effect disappears. Fourth, most of the prior studies use a

rough measure for cost-sharing level. They suffer from lack of generalizability because they

actually evaluate a certain insurance type's effect rather than comparing cost-sharing level

differences. Similarly, severity level is measured by self-reported health status in five categories,

and those in poor and fair health are grouped as severely ill and the rest becomes the reference

group, thus using crude measures that lack precision. This current study will fill in these gaps,

focusing only on physician care cost-sharing using a nationally representative sample and more

precise measures, plus adopting more rigorous methods to improve the understanding of the

research question.










Price of medical care


Average cost of medical care






Medical care quantity demanded


Figure 2-1. Moral hazard and social welfare








Price of medical care


Average cost of medical care


Medical care quantity demanded


Q1 Ql


Figure 2-2. Income transfer helps distinguish between efficient and inefficient moral hazard


A









CHAPTER 3
CONCEPTUAL FRAMEWORK

The main purpose of this study is to examine the potential differential impact of high cost-

sharing in physician care on the change in the amount of health care utilization and expenditures

between severely ill people and healthier people.

Grossman's Model

Grossman's model of health capital and demand for medical care can guide our analysis

(Grossman, 1972). It suggests that health can be viewed as a durable capital stock that produces

an output of healthy time. It is assumed that individuals inherit an initial stock of health that

depreciates with age and can be increased by investment. As the depreciation rate rises with age,

it is not unlikely that old people who are unhealthy will make larger gross investments than will

younger, healthier people. This indicates that unhealthy people have a greater depreciation rate

than the healthy, and that they would be likely to invest even more on health. Thus, unhealthy

people value medical care more than the healthy. Therefore, when considering the dilemma

between high cost-sharing and health status, they prefer health and are less sensitive to cost

pressure.

Grossman's model measures the magnitude of elasticity F by evaluating the relationship

between "need" or illness, measured by the rate of health depreciation and utilization of medical

services. The smaller the value of F, the greater the explanatory power of "need" relative to that

of the other variables in the demand curve for medical care.

s = 1lnHi / O1n i

clnHi refers to the depreciation rate of the stock of health, which describes the need for

health care or disease severity; aln yi refers to the demand for health care utilization. Because 0










and health care utilization. As F is equal to the negative of the ratio, so the bigger the ratio's

absolute value, the smaller the F. Holding everything else constant, demand for health care

utilization can be determined by depreciation rate or severity magnitude. In the case of the

severely ill people, the higher depreciation rate relative to price (i.e., high cost-sharing), the

greater the explanatory power of disease severity in affecting demands for health care utilization.

Thus, the smaller the F, the less responsive severely ill people become to cost-sharing pressure.

In the case of healthy people, the lower depreciation rate relative to high cost-sharing, the less

important disease severity is in influencing health care utilization. Therefore, s indicates the

differential impact of cost-sharing on health care utilization by disease severity, which moderates

the main effect of cost-sharing.

Andersen's Behavioral Model

Interestingly, Andersen's behavioral model of health service utilization makes a similar

argument (Andersen, 1995). It suggests that peoples' use of health service is a function of their

predisposition to use service, factors that enable or impede use, and their need for care. More

enabling resources may help realize potential access to actual utilization by ways of need. As an

enabling factor, cost-sharing reduction would effectively relieve the financial constraint and

provide individuals access to necessary health care. Need is defined as "illness level, which is the

most immediate cause of health service use," especially evaluated need is closely related to

utilization. It represents disease severity and the degree of price elasticity holding everything else

constant.

Andersen's behavioral model indicates that enabling and need factors would have a

differential ability to explain use, depending on which type of service was examined. Hospital

services received in response to more serious problems and conditions would be primarily

explained by need and demographic characteristics, while dental services-considered more









discretionary-would more likely be explained by social structure, beliefs, and enabling factors.

Here, service type is used to represent disease severity level, with discretionary service

representing less severe conditions, under which health care use is mainly influenced by enabling

factors rather than need. For serious conditions, however, use is mainly determined by need

factors rather than by enabling factors. Disease severity (representing the need factor) and cost-

sharing level (representing the enabling factor) -two opposite forces-may compete against

each other to influence utilization; the result depends on their strength comparison, the utility

priority. Facing the dilemma of a higher cost burden and a potentially worse health outcome,

individuals have to evaluate two adverse consequences and choose the less serious one to

minimize adverse utility. For less-severely ill individuals, given increased cost-sharing, the need

for utilization may become discretionary and thus less important. Utility priority will give way to

the enabling factor-cost-sharing pressure-the dominant predictor for utilization. In the case of

a severe condition, individuals are less responsive to high cost-sharing because they have more

urgent need, the most immediate and imminent cause of health service use despite constrained

enabling factors. Therefore, they tend to maintain their current adequate health care and reduce

less necessary utilization, since any delay or forfeiture would cause more serious health

consequences and higher future medical expenditure, so the need dominates the enabling factor

in influencing use. Now the main explanatory power transfers to need, the more severe

condition, with less variation in use explained by an initial enabling factor.

Obviously, the effect of the enabling factor on utilization is contingent on the need factor,

which bridges and moderates the impact of the enabling factor. To put it quantitatively, the

relationship can be represented by an interaction term. For both situations, this study focuses on

the relationship between the enabling factor and use, but treats need as a moderating factor.









Enabling and need factors are the two main effects. As cost-sharing increases, health care use

will decrease, but this relationship may differ depending on the need factor, or the severity level.

Specifically, this relationship is unique to the severely ill, who will decrease utilization less than

the healthy group, since for them the need factor dominates the enabling one. Here, disease

severity can be used to represent need as a main effect: the more severe one's disease, the

stronger need he or she has for health care use.

Demand Curve

In economic theory, the demand curve also provides insight into this relationship. In Figure

3-1, the average health group's demand curve is used as the reference, while the counterpart for

the severely ill tends to be less elastic. In this study, as price is the independent variable and

demand quantity the dependent variable, the X and Y axis is switched. Figure 3-2 reflects Figure

3-1, with price referring to cost-sharing, and demand quantity referring to health care utilization.

It suggests that as cost-sharing increases, an individual's health care consumption decreases.

However, utilization among the severely ill group decreases less than for the average health

group, which marks the differential impact.

Price Elasticity

Quantitatively, the definition of price elasticity of demand further serves as a conceptual

framework to provide mathematical precision for the potential differential impact. Price elasticity

of demand measures responsiveness of the quantity demanded to changes in price. It can be

influenced by many factors, such as income and preference. Thus, holding everything else

constant, price elasticity of demand serves as an indicator reflecting disease severity. In the

extreme case of a life-threatening condition, price elasticity | Ed | would approach 0. For each

severely ill and reference group AQ% = AP% Ed, utilization change percentage AQ% is also

a function of cost-sharing price change percentage AP% and price elasticity of demand Ed. For









the purpose of clarification, this study now denotes AP% as APe%, c referring to cost-sharing.

Thus, given the same cost-sharing change APe% between severely ill and reference groups, the

differential impact in utilization change AQ% depends solely on the price elasticity discrepancy.

AEd= Edl-Ed2 (1)

AQi% AQ2% = APc%* Edl APc%* Ed2= APe%* (EdI Ed2) (2)

The wider the gap between these groups in disease severity level or price elasticity, the

larger the difference between the two groups in percentage of utilization change.

Similarly, health care expenditure is a function of price and demand quantity with

TE = P* Q, with P being the unit price for certain service type. Here, this study denotes P as PMD,

which refers to doctor visit unit price. PMD is different from above APc%, which refers to the

percent of change in cost-sharing price. Between the severely ill and the reference group, PMD is

the same, since they are compared for the same type of service-physician care. For each group,

the expenditure difference between high and low cost-sharing level is

ATE = TEh TEl = PMD *Qh PMD *Q = PMD *(Qh Qi) = PMD *AQ. (3)

ATE% = (TEh TEi) / TEl = PMD *(Qh Q) / (PMD *Q) = (Qh Q) / Q1 AQ% =

APc%* Ed (4)

TE1 and TEh refer to health care expenditure for low and high cost-sharing levels within

each group. Qi and Qh refer to health care use amount for low and high cost-sharing levels within

each group.

Thus, the differential impact in expenditure change ATE% also depends solely on the

price elasticity discrepancy.

ATEi% ATE2% = APc%* Edl APc%* Ed2= APc%* (Edi Ed2) (5)

1 and 2 refer to the severely ill and the reference groups, respectively.









It should be noted that expenditure for each service type consists of two components: out-

of-pocket spending and third-party payment. Cost-sharing policy is intended to influence demand

side by adjusting individual consumer behavior, so the potential differential impact should be

directly demonstrated in out-of-pocket expenditure. Since coinsurance is fixed for any service

within a plan, out-of-pocket expenditure is proportional to total expenditure, so the potential

differential impact can also be indirectly demonstrated in total expenditures for a service as well.

In summary, the economic theory of price elasticity of demand indicates that the

differential impact between the severely ill and the reference group may be demonstrated in both

health care utilization and expenditures.

Initially, price change AP, refers to the same group of people experiencing cost-sharing

change over time, but this study can treat AP, as two groups' cost-sharing differences in a cross

sectional study.

The demand curve slope goes downward, which means the direction of health care demand

is always negative. The magnitude of the negative sign of price elasticity | Ed | determines the

absolute reduction effect of cost-sharing on care utilization, whether for the elastic or inelastic

group, severely ill or reference group. Therefore, as cost-sharing increases, health care utilization

will always decrease. It is no surprise that many papers report the cost-sharing effect held even

for chronically ill people. Only in extreme cases-a life-threatening condition when| Ed I

approaches 0-increased cost-sharing will not reduce utilization and expenditures. What really

matters is the reduction magnitude for severely ill people, whether the medical utilization and

expenditure reduction is similar to or substantially less than those for the low health risk group.

Cost-sharing may not sizably reduce health service utilization and expenditure for people with

severe diseases since the services they seek are mainly of a life sustaining and life saving nature,









which is necessary and exhibits less elastic demand. The essence of this study is whether there

exists an interaction effect between disease severity and cost-sharing.

Beside health status heterogeneity, service type may also influence the differential impact.

Specifically, emergency room care and inpatient care are relatively insensitive and resistant to

cost-sharing, and thus more likely to demonstrate the potential differential impact due to cost-

sharing in ER and inpatient care. Physician care, however, is likely to be sensitive to cost-

sharing. Therefore, the potential differential impact may be attenuated for the severely ill group

and the healthier group. More generally, service types with higher price elasticity may offset the

differential impact by health status. For physician care, price elasticity estimates range from

-0.14 to -1.9 as summarized in the literature review section. The wide range indicates that if price

elasticity is near the lower limit, -0.14, similar to those for ER and inpatient care, there could be

a differential impact of high cost-sharing by disease severity in its own service. On the other

hand, if price elasticity is near the upper limit, -1.9, there would not be a differential impact

between groups in physician care itself. Both situations can be examined as follows:

First, there is a direct differential impact on physician care. That means severely ill patients

can successfully maintain their necessary care in MD visits that are adequate to maintain their

health status and meet their requirements, so they need little downstream ER or inpatient care to

make up for potential deficient care. As a result, there will not be an offset effect, and the overall

health care utilization and expenditure for all service types reflect the direct physician care

differential impact.

Second, there is no direct differential impact on physician care due to cost-sharing

pressure. It is possible that the severely ill are sensitive to high cost-sharing pressure, so they

reduce physician care to a similarly large extent as the general health population. However, their









suppressed physician care utilization will be released in downstream ER or inpatient services

since the sick's severe conditions do not allow them to forgo needed physician care, while the

healthy need little ER or inpatient care because their larger health stock allows them to reduce

more unnecessary physician care. Thus, severely ill individuals have increased ER or inpatient

care utilization more than the general health population, so there still may be an indirect

differential impact in ER or inpatient care (Figures 3-3 and 3-4).

The indirect differential impact on ER or inpatient care expenditure can be demonstrated as

follows:

For each severely ill and healthier group, net expenditure change is

PMD* AQMD1 + PERinpatient* AQER, inpatient (6)

PMD* AQMD2 + PERinpatient* AQER, inpatient2 (7)

The expenditure change PMD*AQMD is negative, indicating it is savings, while the

expenditure change PER,inpatient* AQER, inpatient is positive, indicating it is actual spending. The

differential expenditure change is (6) (7), as there is no differential impact in doctor

expenditure, so the first item is canceled:

(6) (7) = PER,inpatient* AQER, inpatient PER,inpatient* AQER, inpatient2 = P ER,inpatient*

(AQER, inpatient AQER, inpatient2) (8)

As there is little expected offset effect for the healthier group, AQER, inpatient2 Z 0, so the

ultimate differential net expenditure change is PER,inpatient* AQER, inpatient, which is positive.

"Differential" refers to the comparison between the severely ill and healthier groups; while

"net" refers to overall expenditure for all service types; "expenditure change" is due to

expenditure comparison from low to high cost-sharing.









The sick individual's significant amount of care reduction in physician care will be offset

to a certain extent by ER or inpatient care, while there is little offset effect for the general health

population. As a whole, for all service types, those severely ills' total care reduction amount is

still less than the healthy, and there still may be total differential impact (Figure 3-5). It is less

likely that there are both differential impacts in direct physician care and indirect offset effect in

ER or inpatient care.

Although the net health care utilization amount cannot be derived across different service

types, expenditure has the same unit that allows an estimation of the potential differential impact

in total expenditures. For the overall expenditure in the second situation, it is expected that the

increased ER and inpatient care expenditure is less than the reduced physician care expenditure

for the severely ill group. As a whole, the high cost-sharing group has less overall mean

expenditure than the low cost-sharing group, and the severely ill group has higher overall mean

expenditure than the general health population. Finally, the severely ill group reduces its

expenditure amount less than the general health population.

Hypotheses

Based on these analyses, two alternative sets of hypotheses are proposed.

Set 1:

1. Cost-sharing level directly affects physician care utilization change differently by health

status; specifically, high cost-sharing reduces physician care utilization less for the severely ill

group than for the healthier group.

2. Cost-sharing level directly affects physician care expenditure change differently by

health status, specifically high cost-sharing reduces physician care expenditures less for the

severely ill group than the healthier group.









If Set 1 hypotheses are confirmed, then the research question has been answered. In the

event there is no direct differential impact in physician care, this study will proceed to the next

set of hypotheses.

Set 2

3. There is an indirect differential impact of cost-sharing in physician care on ER and

inpatient care utilization change by health status; specifically, high cost-sharing in physician care

increases ER and inpatient care utilization for severely ill people more than healthier people.

4. There is an indirect differential impact of cost-sharing in physician care on ER and

inpatient care expenditure change by health status; specifically, high cost-sharing in physician

care increased ER and inpatient care expenditures for severely ill people more than healthier

people.

5. Cost-sharing level in physician care affects expenditure change in overall service type

differently by health status; specifically, high cost-sharing reduces total care expenditures less

for the severely ill group than for the healthier group.

In summary, the expected results would be either hypothesis 1 and 2, or 3, 4 and 5. Either

hypothesis set is expected to occur, but not at the same time.









Price of medical care

t Reference group


Severely ill group


I __ _Demand quantity for medical care


Figure 3-1. Demand curve for severely ill and reference group


Utilization of medical care


Severely ill group





Reference group

-- Price of medical care


Low cost-sharing High cost-sharing


Figure 3-2. Demand curve for severely ill and reference group with switched X and Y axis










MD utilization and expenditure

A


> \ Severely ill group


Reference group
MD price of medical care

Low cost-sharing High cost-sharing

Figure 3-3. Demand curve for severely ill and reference group with switched X and Y axis in
MD visits





ER or inpatient care utilization and expenditure


Severely ill group





Reference group


I--- MD price of medical care

Low cost-sharing High cost-sharing

Figure 3-4. Demand curve for severely ill and reference group with switched X and Y axis in
ER or inpatient care









Utilization of total medical care


Severely ill group





Reference group


-- Price of medical care

Low cost-sharing High cost-sharing

Figure 3-5. Demand curve for severely ill and reference group with switched X and Y axis in
total care









CHAPTER 4
METHODS

Data Description

This section focuses on the design, content, and procedure of the MEPS data, and the

rationale for using the MEPS for this study.

Medical expenditure panel survey household component (MEPS-HC) data from 2007 was

used for analysis in this study. The MEPS-HC is a national representative survey of the non-

institutionalized civilian population of the United States and is designed to produce national and

regional estimates of the health care use, expenditures, sources of payment, and insurance

coverage of the U.S. population. The MEPS includes surveys of medical providers component

(MPC), furnishing information on providers to supplement the data provided by household

respondents. The MEPS design permits both individually based and family-level estimates. In

this study, the unit of analysis was each individual.

The MEPS-HC has an overlapping panel design. Each year a new MEPS-HC panel is

established. Information is collected from each household to cover a two-year period, so the data

can be used to track changes and trends over time to estimate health care utilization, expenditure,

and insurance coverage. The MEPS-HC sample is drawn from about a one-quarter subsample

from the previous year's National Health Interview Survey (NHIS), which follows a stratified

multi-stage area probability design. The complete NHIS sample consists of 358 primary

sampling units (PSUs), which are counties or groups of contiguous counties. The sample PSUs

are stratified by geographic area, metropolitan status, and socio-demographic measures. In

MEPS-HC, Hispanics and African-Americans have been over-sampled each year since 2004.

From 2002 on, the MEPS sample design over-sampled Asians and persons predicted to have

incomes less than 200% of the poverty level (Agency for Healthcare Research and Quality,









AHRQ, 2007). Over-sampling holds the advantage of improving the precision of estimates for

specific subgroups.

Each new MEPS annual sample, referred to as a panel, cover a series of five rounds of in-

person interviews over a 30-month field period to yield annual circle data for two full calendar

years. The computer-assisted personal interview is the principal data collection mode; each

interview takes on average 90 minutes to conduct. Individuals' data can be linked across rounds

to build longitudinal data for up to two years of survey participation.

During each calendar year since 1997, data are collected simultaneously for two MEPS

panels. One panel is in its first year of interviews, including rounds 1, 2, and 3, while the prior

year's panel is in its second year of data collection, including rounds 3, 4, and 5; and round 3 for

each MEPS panel overlaps two calendar years. These two panels comprise the annual estimates.

The MEPS-HC contains detailed data that meets this study purpose, including

demographic characteristics, family structure, household income, health and functional status,

health insurance coverage, access to care, health care use, and expenditures. Meanwhile,

information from medical provider component (MPC) is used to supplement and validate the

information from MEPS-HC about diagnosis, charges, payments, and specific services provided.

As a result, the final data include information on diagnoses, procedures, inpatient stays classified

by diagnosis-related group (DRG), prescriptions (medication names, strengths, and quantities

dispensed), charges, and payments. The data allow us to take advantage of above-detailed

information relevant to this study to make nation-wide inferences.

Like many other national survey designs, the complex multi-stage, unequal probability,

and cluster sampling methods require adjustment for sample weight to reflect the unequal

probability of selection, household non-response (MEPS Round 1), attrition of persons









(subsequent rounds), post-stratification (census population estimates), and trimming of extreme

weights. A sample weight is assigned to each sample person. It is a measure of the number of

people in the population represented by that sample person in MEPS (AHRQ 2007; Cohen et al.,

1999). There are two groups of weighting variables: person-level weights in the annual data file

and longitudinal weights in panel files. The longitudinal weight variables should be used when

the sample includes persons participating in both years of one panel. This study will use person-

level weights in the annual data file. Variance also needs to be adjusted to account for the

differential weighting and the correlation among sample persons within a cluster. Typically,

individuals within a cluster are more similar to one another than those in other clusters; as a

result, the error term in health care utilization or expenditure for individuals within a cluster is

correlated. Failure to consider this correlation at cluster level will result in variance estimates

that are too small. This, in turn, will tend to overestimate the significance of the estimates.

Measures and Operationalization

Measures of Outcome Variables

Two groups of outcome measures will be examined: utilization and expenditures of care.

In MEPS, both utilization amount and expenditure are specified by service type: physician care,

ER, and inpatient care. This also permits examination of the effect of cost-sharing in physician

care on its own service and ER or inpatient care. This study used both specific expenditures by

service type and the overall expenditure. The total health care expenditure is the total expense for

these service types, rather than existing total expenditure in the dataset with service types, such

as medication, being included. Expenditures in the MEPS are defined as the sum of payments

made, including those made out-of-pocket and by third-party payers (Zuvekas et al., 2002). This

study used expenditures instead of charges when measuring health care costs, since charges are

the fees billed to patients and insurance companies by providers, while expenditures are actual









direct payments to providers and health care organizations due to the negotiated discounts

between payers and providers. The expenditure variables were constructed from the original

event files by summing the total expenditures for each event by person. Expenditures inccurred

in the following types of settings-inpatient, outpatient, and emergency room-consist of two

expense variables per setting: facility expenses and "separately billing doctor" (SBD) expenses.

Hospital facility expenses include all expenses for direct hospital care, including room and

board, diagnostic and laboratory work, X-rays, and similar charges, as well as any physician

services included in the hospital charge. SBD expenses typically cover services provided to

patients in hospital settings by providers like radiologists, anesthesiologists, and pathologists,

whose charges are often not included in hospital bills. This study will use the sum of facility and

SBD expenditure for each of above services. Beside physician care, this study also examines its

primary care component, which can help us better understand the potential differential impact in

physician care.

The medical utilization measures are the quantity of each type of care in 2007. Those

variables are constructed from MEPS Events Files by counting the number of events within the

year for each type of care. The MEPS asks respondents the primary reason for the visit and

prompts respondents with a list of possible reasons. The primary care visit number results from

counting the numbers for primary-care related reasons from both office visits and hospital

outpatient visits, which include general and family practice, internal medicine, and geriatric

services for the elderly. Similarly, the measurement of physician care in this study is the number

of MD visits identified by office and hospital outpatient visits. Physician visits is a commonly

used and validated measure in health care use and its relationship to health insurance coverage.

In the MEPS survey, the question was asked: Did the patient see the doctor or was it a phone









call? Only in the former case is the visit coded as a physician visit. The measure of inpatient care

is the number of hospitalizations. The variable comes from the Hospital Inpatient Stay file.

Emergency care utilization is measured by the number of ER visits. The MEPS contains an

event file that has information on each ER visit.

Measures of Explanatory Variables

Severity level

Some studies use presence of chronic conditions, self-reported health status by five

categories, or therapeutic classes to measure disease severity; however, those suffer from a lack

of precision or universality. Instead, Ware et al. (1992) adopted a physical component score

(PCS) and a mental component score (MCS) from a 36-item short-form (SF-36) in the Medical

Outcomes Study to reflect health related quality of life in medical conditions and illness severity,

and this practice was followed by other studies, such as the one by Harman et al. (2010). A short-

form questionnaire SF-12 was derived from the SF-36 (Ware et al., 1992), and both

questionnaires are widely used and well-validated generic instruments for measuring health

status (Lowe et al., 2004). SF-12 includes the following eight concepts: physical functioning,

role functioning-physical, bodily pain, general health, vitality, social functioning, role

functioning-emotional, and mental health, summarizing these 12 items into an overall physical

and mental function score. This study focuses on physical health status, but not mental health

status; MCS is not used, since mentally ill patients may make irrational judgments and decisions

on physician care usage in response to high cost-sharing. PCS is constructed with a continuum

range from 0 to 100, a higher score indicating better health status to precisely reflect an

individual's health status and co-morbidity severity. As a measure for health status, PCS

provides us with several advantages. First, judgment of illness severity based on co-morbidity

type and number makes it difficult to compare people's disease severity. PCS goes beyond this









limitation by providing this study with a universal measure. Second, PCS as a continuous index

allows us to examine health status and disease or condition severity in a precise manner. It

overcomes the difficulty of determining the categories for some conditions when the response

scale is simply dichotomous or encompasses several categories-the essential and discretionary

drug category as a proxy measure for disease severity, or the five categories of poor, fair, good,

very good, and excellent for self-reported health status. These measures may suffer from

measurement error problem. With the continuous measure of PCS, disease severity or health

status can be classified into precise categories. Like the World Health Organization classification

for body mass index (BMI), different PCSs can be classified into different illness severity

categories. According to the study by Harman et al.(2010), when using the 100 score range, top

tertile and bottom tertile scores in actual data distribution are divided into severity groups, with

any score in the lowest tertile categorized as severely ill, and any score in the highest tertile

serving as the reference group. This study will follow suit. Note that PCS is available in MEPS.

Cost-sharing level

Before measuring cost-sharing, certain concerns about observation selection need to be

dealt with. An individual may move out of a plan and become uninsured for some time or may

transfer to another plan. If insurance coverage status is defined by less than one full year

coverage, this measure of insurance coverage is subject to measurement error. Therefore, this

study only selected those who have retained insurance coverage for a full year, which best serves

the research question. In MEPS, there are two kinds of insurance duration status variables:

insurance coverage in a year and insurance coverage by month, defined as having insurance

coverage any time in a year or month. This study used the month indicator and respondents must

have insurance in all twelve calendar months to be included in the study.









Similar to severity level, there is not a direct cost-sharing level measure in MEPS,

however, it can be calculated by dividing out-of-pocket expenditures by total expenditures for

each event for each person; both data are available. In this way, cost-sharing can be quantified

precisely in a continuous percent range. Unlike copayment measured by a fixed amount, this

coinsurance measure can be compared across plan types and service types, thus will be used in

this study. On the other hand, it should be noted that the cost-sharing level cannot be derived

from those whose out-of-pocket and total expenditures for 2007 are zero. Since there may be a

non-trivial number of such observations, listwise deletion of these missing values may cause

selection bias problem, leaving the remaining observations unrepresentative for the whole

population. A safer approach to address this problem is to assign these missing data with the

mean of the existing values by each insurance type.

With complete cost-sharing data, this study can then move on to determine cutoff points in

each plan to classify people into high or low cost-sharing groups. As is common practice,

Medicare and many private insurance plans adopt a 20% cost-sharing level. One study adopts

20% to differentiate low or high levels of cost-sharing in medications, arguing that it was

consistent with a common co-insurance level and that it would be important to know if an effect

was present at a lower limit of high cost-sharing (Ungar et al., 2008). This study followed suit by

using 20% to classify individuals into high and low cost-sharing groups, but made slight

modifications: zero cost-sharing individuals were not singled out as a group separate from the

low and high ones since it is not likely that people with a free health care plan would behave

markedly differently from those in the low cost-sharing group.

This 20% criteria, however, does not apply to Medicaid because medical care places a

relatively heavier financial burden on the poor. The Medicaid program was designed to serves a









low-income population that lacks substantial resources and often has significant health care

needs. It prohibited or limited premiums and cost-sharing to nominal levels (Kaiser Family

Foundation, 2003). One study that surveyed cost-sharing policy in Medicaid and State Children's

Health Insurance Program (SCHIP) outlined this standard by these facts: the median co-payment

for non-preventive office visits was $5 for all family income levels, and ranged from $5 to $25

for emergency room use, and from $5 to $18 for hospitalizations (Selden et al., 2009). In this

study a relative cut off point-the mean cost-sharing level in this study sample-was used to

differentiate individuals into high and low cost-sharing groups.

Measures of control variables

The control variables, a comprehensive range of factors likely or known to affect health

care utilization available in the data, include a person's age, gender, race, geographic area,

rural/urban location, marital status, education level, income, other health conditions, and

gatekeeper plan status. Since different age groups may not affect health care use in similar ways,

age is grouped into three categories-young, middle, and elderly in the ranges 18-34 years, 35-

64 years, and older than 65 years, respectively. This study excludes children because the study is

focused on people who are most likely to make health care decisions for themselves. Gender, of

course, is treated as a dichotomous variable: female and male. Race and ethnicity were classified

into four groups: non-Hispanic white, non-Hispanic African-American, Hispanic, and other.

Regions of residence were classified as Northeast, South, Midwest, and West. Rural versus urban

residence is determined by whether or not a respondent lives in a metropolitan statistical area.

Geography and location may represent an individual's socioeconomic condition, and this affects

health care use. Marital status is classified as married and non-married, and marriage, too, may

influence health care, since it represents a type of social support. Education level may affect

people's awareness of their health and value of health care, and it is represented by years of









education achieved. Different education levels may not affect health care use in similar ways, so

the education variable is categorized into three groups: 0-8 years, 9-12 years, and 13-17 years.

Similarly, family income or poverty status is categorized into five groups: poor, near poor, low

income, middle income, and high income. This income variable is categorized by percentage of

the federal poverty level, which is consistent with the 2007 poverty statistics developed by the

Current Population Survey.

In addition to disease severity as a main effect, other co-morbidity conditions are also

controlled, since our severity measure PCS comes from SF-12, self-reported perceived overall

health status. Andersen's model indicates that both perceived need and evaluated need influence

health care utilization, so the remaining evaluated need has to be controlled as well. Therefore,

this study includes other co-morbidities: hypertension, diabetes, asthma, arthritis, stroke, brain

injury, and cancer, which are dummy variables.

Gatekeeper status also affects people's health care since the primary care physician

controls and decides, depending on their medical condition, whether or not to refer patients to

specialists. Dummy variables will be used to indicate if individuals have primary care

physicians.

When the offset effect is examined for ER or inpatient care, this study should also control

the cost-sharing level for ER visits or hospitalizations since utilization or expenditure is

influenced directly by cost-sharing level in their own service types, and not simply by cost-

sharing for physician care. Here, cost-sharing level for ER visits or hospitalizations is a

continuous variable, calculated directly from the proportion of out-of-pocket expenditure out of

total event expenditure.









Based upon the above considerations, this study may organize the following variables. The

dependent variable is health care utilization and expenditure, including physician care, primary

care, emergency care, and inpatient care. The independent variables are cost-sharing level and

severity level. These are the two main effects, and their interaction is our major interest in

examining cost-sharing's effect on health care utilization or expenditure between the severely ill

and the control groups. The covariates include age, gender, race, geographic area, rural/urban

location, marital status, education level, income, and gatekeeper plan status. In the ER and

inpatient care equation, cost-sharing levels of their own services also need to be controlled.

Study Design

This study used a cross-section design from MEPS-HC data in 2007, the latest available

data. Since a universal cost-sharing policy change does not occur at a time point across different

insurance types, a cost-sharing policy change may occur at different times for different plans at

different levels, or there may be no change at all. This makes it difficult to take advantage of a

longitudinal design of panel data to track utilization and expenditure changes over time in

response to insurance changes. Instead, this study selects a single year to examine the research

questions. Subject selection criteria should conform to research question requirements and study

design; thus, only observations that meet the criteria below were included in the analysis.

1) Subjects aged 18 or older were selected, since children do not make health care

utilization decisions.

2) Those with the same insurance status for a full year were selected, since partial year

insurance coverage may produce measurement error.

3) As a representative sample of public and private plans, individuals enrolled in Medicare,

Medicaid, and private plans were selected.









As it is assumed that individuals will behave similarly at the same cost-sharing level

regardless of plan type, this study does not need to specify the potential differential impact

within Medicare, Medicaid, and private plans. However, cut off points to divide enrollees into

high and low cost-sharing groups may differ by plan types. Specifically, cost-sharing level above

or equal to 20% and below 20% is selected as the high and low groups for Medicare and private

insurance plans. In Medicaid, cost-sharing levels above or equal to and below the mean cost-

sharing value are divided into high and low groups. Meanwhile, the study specifies service types

as physician care, emergency care, and inpatient care. Since MD visits are the main form of

physician care, this study uses these two terms interchangeably.

Statistical Analysis

Descriptive analysis was conducted first in order to describe study population

characteristics. Next, multivariate analysis was employed.

As mentioned above, a rigorous design should account for the potential endogeneity

between health care utilization or expenditure and cost-sharing level, as it is obvious that

increasing cost-sharing level reduces health care utilization and expenditures, and those with

high health care utilization and expenditures will seek to enroll accordingly in low cost-sharing

plans. In this case, individuals with different cost-sharing levels are likely to differ in ways that

are related to their utilization of care, which, however, are not readily observable by researchers

(i.e., unobserved heterogeneity). As a result, differences in people's utilization could reflect the

combination of a causal effect of cost-sharing level and the effect of unmeasured characteristics

that are correlated with cost-sharing and use of care. The existence of omitted variables could

threaten internal validity. If the omitted variables affect the cost-sharing level and utilization in

the same direction, the predicted cost-sharing effect will be overestimated. The Hausman test

(Hausman, 1978) can be employed to test for endogeneity by regressing cost-sharing on all









exogenous variables, then adding the residual as a new variable into the initial structural

equation. The significant coefficient for the residuals indicates the existence of endogeneity. The

Hausman test compares the OLS and two-stage least squares (2SLS) estimates and determine

whether these differences are statistically significant. Both OLS and 2SLS are consistent if the

cost-sharing variable is exogenous. If OLS and 2SLS differ significantly, then the suspectible

regressor must be endogenous (Wooldridge, 2005) since now the 2SLS estimate is consistent,

while the OLS estimate is inconsistent. In this case, as a solution, instrumental variables (IVs)

should be introduced that only affect cost-sharing level, but not utilization and expenditure. An

individual's out-of-pocket premium within each plan type can be an eligible IV. As an

individual's out-of-pocket premium increases, some enrollees tend to drop out of plans. In order

to maintain enrollees, insurers will reduce cost-sharing level to make their plan competitive in

the market. Therefore out-of-pocket premiums influence cost-sharing levels, but once their levels

are determined there is no way to influence actual utilization and expenditures. The IV approach

was used to account for this endogenous problem because it is unlikely to identify and measure

all the unobserved heterogeneity.

With the aid of IVs, the predicted cost-sharing level for each person was obtained,

replacing the actual ones for the main effect and interaction term. Specifically, two-stage least

squares techniques were employed. The analysis used the first stage by regressing the

endogenous independent variable on IVs and other exogenous variables, and getting the

predicted cost-sharing value to replace the actual value.

This study also needs to test the goodness and effectiveness of IVs. A good instrumental

variable must have two properties: first, it must be correlated with the endogenous regressor;

second, it should not be correlated with the error term of structural equation (Wooldridge, 2005).









In the first stage, an IV is usually valid so long as the F statistic, as a rule of thumb, is

greater than 10. Meanwhile, this study directly regressed the error term of utilization or

expenditure on the IV, with an IV being deemed valid if its coefficient was not significant,

indicating that it was not related to dependent variables.

In the final analysis, analytic model selection should account for characteristics of count

data and expenditure data. Measured by number of visits, utilization is count data. Count data are

non-negative values, with the possibility of a substantial number of zeros. As a random variable,

small value counts may have higher probability than large value counts. In addition, the variance

may not be equal to its mean. A negative binomial model may best reflect these characteristics

for the count data than any other model.

Expenditure is a continuous variable; accordingly, a generalized linear model could be

employed. Specifically, the two-part model is suitable for analysis, using logit in the first part to

estimate the probability of any expenditure. For the second part, conditional on any expenditure,

non-linear regression could be employed to account for the skewed distribution. This usually

means using log transformation with residual diagnosis to check the model fit. Specifically the

statistical analysis or model set up for each hypothesis is as follows:

1. Cost-sharing level directly affects physician care utilization change differently by health

status; specifically, high cost-sharing reduces physician care utilization less for the severely ill

group than for the healthier group.

A negative binomial regression was used.

In E(primary care visits) = Po0 + 31*cost-sharing + 02*severity + 03 cost-sharing*severity +

04*covariates + P.









In E(MD visits) = 3o + 11*cost-sharing + 02*severity + 03 cost-sharing* severity +

04*covariates + P.

E is the expected value. Based on theory prediction and variables coding, it is expected that

P1 will be negative, indicating that the high cost-sharing group has lower mean utilization value;

02 will be positive, indicating that the severely ill group has higher mean utilization value; and 13

will be positive, indicating that the severely ill group reduces their utilization amount less than

the reference group.

Note the interaction term is a reduced form that only represents the simplest case where

both cost-sharing and severity are dichotomous variables. In the final analysis where cost-sharing

and severity, alone or in combination, can be multiple groups, this reduced form of interaction

may represent a series of dummy interactions.

2. Cost-sharing level directly affects physician care expenditure change differently by

health status; specifically, high cost-sharing reduces physician care expenditure less for the

severely ill group than for the healthier group.

A two-part model was used.

In the first part, logit was used to estimate the probability of any expenditure.

In odds (MD expenditure) = 3o + 31*cost-sharing + 02*severity + 03 cost-sharing*severity

+ 04*covariates + P.

Odds are the ratio of probability of having any physician care expenditures relative to that

of not having any expenditure.

Odds = P / (1-P)

For the second part:









In MD expenditure = Po + 31*cost-sharing + 02*severity + 03 cost-sharing*severity +

04*covariates + F.

The coefficients should follow the same pattern as that in hypothesis 1. It is expected that

31 will be negative, indicating that the high cost-sharing group has lower mean expenditure

value; 02 will be positive, indicating that the severely ill group has higher mean expenditure

value; and 03 will be positive, indicating that the severely ill group reduces their expenditure

amount less than the reference group.

3. There is an indirect differential impact of cost-sharing in physician care on ER and

inpatient care utilization change by health status; specifically, high cost-sharing in physician care

leads to increased ER and inpatient care utilization for severely ill people, but little change for

the reference group.

A negative binomial regression was used.

In E( ER / inpatient care visits) = Po + 31*cost-sharing + 02*severity + 03 cost-

sharing*severity + 04*covariates + F.

It is expected that P3 will be positive, indicating that the high cost-sharing group has

overall higher mean increased offset ER or inpatient care utilization; 32 will be positive,

indicating that the severely ill group has higher mean increased offset ER or inpatient care

utilization; and f3 will be positive, indicating that the severely ill group increased ER or inpatient

care utilization more than the reference group; the latter can be around zero.

4. There is an indirect differential impact of cost-sharing in physician care on ER and

inpatient care expenditure change by health status; specifically, high cost-sharing in physician

care leads to increased ER and inpatient care expenditure for severely ill people, but little change

for the reference group.









A two-part model was used.

In the first part, logit was used to estimate the probability of any expenditure.

In odds (ER and inpatient care expenditure) = Po + P1*cost-sharing + 02*severity + 03 cost-

sharing*severity + 04*covariates + P.

For the second part:

In ER / inpatient care expenditure = Po + P1*cost-sharing + 02*severity + 03 cost-

sharing*severity + 04*covariates + s

It is expected that 31 will be positive, indicating that the high cost-sharing group has

overall higher mean increased offset ER or inpatient care expenditure; 32 will be positive,

indicating that the severely ill group has higher mean increased offset ER or inpatient care

expenditure; and 33 will be positive, indicating that the severely ill group increased ER or

inpatient care expenditure more than the reference group; the latter can be around zero.

5. Cost-sharing level in physician care affects expenditure change in overall service type

differently by health status; specifically, high cost-sharing reduces physician care expenditure

less for the severely ill group than for the healthier group.

A two-part model was used.

In the first part, logit was used to estimate the probability of any expenditure.

In odds (total care expenditure) = Po + 31*cost-sharing + 02*severity + 03 cost-

sharing*severity + 04*covariates + P.

For the second part:

In total expenditure = Po + 31*cost-sharing + 02*severity + 03 cost-sharing*severity +

04*covariates + P.









It is expected that P1 will be negative, indicating that the high cost-sharing group has lower

mean expenditure value; 12 will be positive, indicating that the severely ill group has higher

mean expenditure value; and 13 will be positive, indicating that the severely ill group reduces

their expenditure amount less than the reference group.

For any main effect and interaction, the predicted expenditure amount was re-transformed

to obtain estimates on the original scale, since it has been transformed to satisfy OLS

assumptions. As heteroskedasticity will be a potential issue, subgroup-specific smearing factors

were used to obtain unbiased estimates to quantify the magnitude of the interaction effect if the

differential impact was significant (Manning, 1998). Meanwhile, the bootstrap technique was

employed to get standard error and significance for the differential impacts.

This study also calculated price elasticity and cross-price elasticity in physician care for the

severely ill and the reference group.

Sensitivity analysis was also conducted to test finding robustness for the following model

specifications: First, if there was a sizable proportion of individuals whose cost-sharing was

exactly 20% for Medicare and private plans or the mean value for Medicaid, the results may be

influenced when they are categorized into high or low cost-sharing groups. This study examined

both situations. Second, this study may also consider different cut-off points for cost-sharing,

since there could be some concern about whether it is really a high or low level if cost-sharing is

a little higher or lower than 20% for Medicare and private plans or the mean value for Medicaid.

This study can also test 25% and 15% as cut-off points for Medicare and private plans, and 5%

away from either side of the mean value for Medicaid. This way can produce three categories in

cost-sharing: low, middle, and high cost-sharing groups. Third, this study also included the

middle tertile subjects of severity level and created another severity dummy variable to examine









the differential impact for this moderately ill group within the whole picture. Fourth, alternative

ways of categorizing PCS were employed to examine the results' robustness, such as using both

sides of two standard deviations from the PCS mean as cut off points.









Table 4-1. Summary of outcome measures
Variable
Utilization
Number of primary care visits
Number of MD visits
Number of ER visits
Number of hospitalizations


Expenditure
Physician care expenditures
Emergency care expenditures
Inpatient care expenditures
Total expenditures


Continuous
Continuous
Continuous
Continuous


Type

Count
Count
Count
Count










Table 4-2. Explanatory variables
Variable Type
Cost-sharing level Categorical


Disease severity level


Interaction

Age (yrs)


Categorical


Categorical

Categorical


Gender


Categorical


Categorical


Race


Geographic area




Rural/urban location


Marital status


Education level (yrs)


Income


Categorical


Categorical


Categorical


Categorical


Categorical


Co-morbidities


Gatekeeper status


Categorical


Categorical


0 = Top tertile PCS
1 = Middle tertile PCS
2 = Bottom tertile PCS

Cost-sharing*Disease severity


18-34
35-64
65 and older

Female
Male


Other
Non-Hispanic white
Non-Hispanic African-American
Hispanic

Midwest
Northeast
South
West

Rural
Urban

Not married
Married


0-8
9-12
13-17

Poor
Near poor
Low income
Middle income
High income

No
Yes

No gatekeeper
Gatekeeper


Category
0 = Low
1 = High









CHAPTER 5
RESULTS

Overview

The results of the study are presented in the following sections. First, the demographic and

socioeconomic characteristics of selected sample are described. Second, before any analysis,

variable operationalization is reported. Specifically, the missing cost-sharing values were

imputed by insurance type, and the variable selection issue was addressed due to the concern

about multicollinearity. Third, the potential endogenous problem between cost-sharing and

health care utilization or expenditure was tested and confirmed, then an instrumental variable

was introduced to address the endogeneity. Fourth, the results from the multivariate analysis of

the differential impact are presented. This section is organized by outcome measures in the

sequence of health care utilization and expenditures by service types. Specifically, physician

visits, primary care visits, ER visits, hospitalization admissions, physician visit expenditures, ER

visit expenditures, hospitalization expenditures, and total expenditures. The "svy" commands of

Stata 10.0 (StataCorp, 2007) were employed for all the statistical inference tests.

Description of the Sample

Table 5-1 presents the number and composition of individuals eligible for the study. The

final study sample consisted of 13,020 individuals who met the inclusion criteria (Table 5-1).

Altogether there were 60,595 observations, since each individual may have had several instances

of health care utilization with different service types or doctor specialties that could not be

combined into a single value when the service type variable was kept in the data. For example, if

an individual's physician visits in 2007 were hematology service, primary care, and cardiology

service, these service types could not be combined into a composite value, thus total observations

may be more than the number of individuals.









The age of the study sample ranged between 18 and 85 years, with a mean age of 50.19

years and a standard deviation of 17.75 years. The study sample consisted of 3,482 (26.7%)

Medicare beneficiaries, 1,630 (12.5%) Medicaid beneficiaries, and 9,747 (74.9%) private plan

enrollees. There were some overlaps, since some individuals may have a full year of insurance

coverage with Medicare and Medicaid dual eligibility, plus Medicare supplemental plans. The

study sample had a mean PCS score of 48.19 ranging from 5.68 to 70.98, with a standard

deviation of 11.35.

In Table 5-2, the characteristics of the study sample are described. The majority of this

study sample are middle-aged individuals (54.7%), female (55.4%), non-Hispanic White

(61.7%), living the South (35.4%), and are urban (83.0%), married individuals (60.1%), with

college education (50.1%), and high income (41.1%).

Variable Operationalization

Missing Cost-sharing Values Imputation

In this study cost-sharing values in physician care were calculated by dividing out-of-

pocket expenditures by total expenditures for each event by each person. Because of how cost-

sharing was calculated, some missing cost-sharing values would be created in STATA for

individuals because of their zero medical expenditure in the denominator. Eliminating these

missing values may cause potential sample selection bias if the remaining observations are

unrepresentative of the population. To avoid this problem, these missing values were replaced

with mean values within corresponding insurance types. Insurance types can be separated into

seven subgroups in this study: Medicare, Medicaid, private plans, Medicare-Medicaid, Medicare-

private plans, Medicaid-private plans, and Medicare-Medicaid-private plans. Since 9,906

individuals-the majority of this study sample-have cost-sharing values, this imputation

approach should not cause substantial bias. In this way all the missing values were imputed.









Test for the Multicollinearity between Disease Severity and Priority Conditions

The disease severity dummy variables and priority condition variables may be correlated,

because the former are self-perceived summary health status directly based on each individual's

health conditions. The covariance matrix of disease severity and priority conditions indicated all

of them are consistently highly correlated (p <0.01, Table 5-3). Andersen's model states that

both perceived and evaluated needs influence an individual's health care utilization. Technically,

Andersen's model applies to statistical models that contain either perceived or evaluated needs,

but not both. These two types of needs can be substitutes when available data only contains

measures of one or the other type. For this study's purposes, since disease severity is the main

variable of interest and represents one's overall health status, those priority condition variables

were dropped from the model. Keeping both group variables would hurt the model by inflating

the variance and underestimating the significance of coefficients.

Test for Potential Endogeneity and IV Validity

The Hausman Test

The Hausman test can be employed to test for endogeneity by regressing cost-sharing on

all exogenous variables, then adding the residual as a new variable into the initial structural

equation. The significant coefficient for the residuals indicates the existence of endogeneity. In

this case, as a solution, an instrumental or identifying variable would be introduced that affects

cost-sharing level, but not utilization and expenditures. In the first step of the Hausman test, the

physician care cost-sharing variable was used as a dependent variable and regressed on all

exogenous variables, then its error term was obtained. In the second step, when the error term of

the cost-sharing in physician care was entered into the initial structural equation, this error term

was highly significantly associated with physician care utilization and expenditure variables (p

<0.01 for both tests).









Test for IV Relevance and Exogeneity

As planned, an instrumental variable was then introduced to address endogeneity. The F

statistic was examined in the first stage of 2SLS. Meanwhile, this study examined the association

between the error term of utilization or expenditure and IV. The results are reported in Tables 5-

4 and 5-5. In Table 5-4, the F statistic in the first stage of 2SLS was much higher than 10 (p

<0.01), indicating the identifying variable premium was relevant to the endogenous regressor of

the cost-sharing variable. Next, the association between the identifying variable and the error

terms in the structural equation of health care utilization and expenditure was tested, with p=

0.23 andp = 0.27, respectively, indicating the identifying variable was not significantly

associated with these two dependent variables. Therefore, the premium variable met the

requirement of a good IV.

Use IV to Get Predicted Cost-sharing Values

Using the first-stage regression of 2SLS, predicted cost-sharing values can be obtained,

and the initial cost-sharing values can be replaced by these predicted ones. Based on this

continuous cost-sharing variable, a dichotomous cost-sharing variable was then created with 20%

as a cutoff point in Medicare and private plans, and the mean cost-sharing value of 2.9% as a

cutoff point for Medicaid. This dummy cost-sharing variable examines whether high cost-sharing

relative to low cost-sharing in physician care reduces health care utilization and expenditures

differently by health status. Next, two interaction terms were further generated as products of the

dummy cost-sharing variable and two disease severity variables-severe and moderate

conditions. Captured by the interaction term of severe condition and cost-sharing dummy

variables, the differential impact can be examined.

In the final analysis, the interaction term was dropped in both physician care expenditure

and total expenditure models-the findings for two important outcome dimensions-due to









"perfect success prediction." "Perfect success prediction" means that dependent variable did not

vary by the predictor-interaction term. In order to get complete findings, this study used 19% as a

proxy for the 20% cutoff point to divide individuals into high and low cost-sharing groups,

which reclassified 2,416 observations into high cost-sharing group, 3.98% of the total

observations. The results remained similar, but with a gain of complete results. The description

of cost-sharing variable is then reported in Table 5-6. Note that in this table Medicaid insurance

refers to both Medicaid coverage alone and Medicaid in combination with Medicare or private

plans, as do Medicare and private plans. When a 19% cutoff point in private and Medicare plans

is determined first, this may include some Medicaid enrollees due to their dual eligibility. Thus,

the above mean cost-sharing value of 2.9% for Medicaid refers to the remaining beneficiaries

with Medicaid coverage alone, so this mean value may differ slightly from the 2.7% mean value

reported in Table 5-6.

The Multivariate Analysis Results

The Health Care Utilization Results

Physician care utilization

The hypothesis is that high cost-sharing reduces physician care utilization less for the

severely ill group than for the healthier group. The regression is expressed as

log(g) = 3o + 31*cost-sharing + 02*severity + 03 cost-sharing*severity + 04*covariates + F.

Coefficient 0 could be interpreted as the difference between the log of expected counts [,

where 0 = log( [x+1) log( [x) = log( tx+1 / [tx), meaning change in a dependent variable given

each unit change or group difference in the independent variable x from x to x+1. Since the

counts of an event occur in a given period by default, the count can then be considered an

incidence rate, thus the coefficient 3 can be interpreted as the log incidence rate ratio. Intuitively,









the incidence rate ratio is the exponentiated coefficient P. In this study the coefficient 0 could be

interpreted as the log incidence rate ratio of seeing a physician in 2007 between groups.

The differential impact in physician visits was analyzed in a negative binomial regression.

The results are reported in Table 5-7. Holding covariates constant, the high cost-sharing group

had an insignificant 15% less incidence rate of seeing a physician for individuals relative to the

low cost-sharing group (p = 0.09). The severely ill had significantly higher incidence rate (156%)

of physician care visits relative to healthier individuals (p <0.01). The incidence rate difference

in physician care visits between high and low cost-sharing groups was 8% different between the

severely ill and the healthier group, which was not significant (p = 0.67).

In this model significant results were also found for other covariates. Compared with the

reference group, the significant higher physician care visits were found for the White individuals

(p = 0.02) living in urban areas (p <0.01) with low income (p = 0.03), with moderate conditions

(p <0.01) and primary care physicians (p = 0.04).

In summary, the high cost-sharing group was not different in physician care visits from the

low cost-sharing group, the severely ill had more physician care visits than the healthier group,

and the physician care visit difference between high and low cost-sharing groups was not

different between the severely ill and the healthier group.

Primary care physician utilization

The hypothesis is that high cost-sharing reduces primary care utilization less for the

severely ill group than for the healthier group.

The differential impact in primary care physician visits was analyzed in a negative

binomial regression. The results are reported in Table 5-8. Holding covariates constant, the high

cost-sharing group had an insignificant 6% higher incidence rate of seeing a primary care

physician relative to the low cost-sharing group (p = 0.16). The severely ill had a significant 9%









lower incidence rate of primary care physician visits relative to healthier individuals (p = 0.05).

The incidence rate difference in primary care physician visits between high and low cost-sharing

groups was 16% different between the severely ill and the healthier, which was not significant (p

=0.09).

In this model, significant results were also found for other covariates. Compared with the

reference group, the significant higher physician care visits were found for the middle-aged (p

<0.01) and older (p <0.01) African-American individuals (p = 0.05) living in the West (p = 0.02)

and rural areas (p <0.01), unmarried (p <0.01), and with primary care physicians (p <0.01).

Meanwhile, the significant lower physician care visits were found for the high income

individuals (p = 0.02) and those with college education (p <0.01).

In summary, the high cost-sharing group was not different in primary care physician visits

from the low cost-sharing group. The severely ill had fewer primary care physician visits than

the healthier group since most of the time they may go to see specialty physicians based on their

diagnosis and condition. The primary care physician visit difference between high and low cost-

sharing groups was not different between the severely ill and the healthier. Although

insignificant, the positive sign of cost-sharing and interaction term indicated these anomalous

results went against an individual's rational behavior, which will be further explored in the later

part of this study.

ER care utilization

As stated in the hypotheses, the direct differential impact in MD visits was not significant,

which may suggest that the severely ill perhaps use more downstream services to meet their

needs but suppressed health care need due to high cost-sharing pressure for MD visit. Thus, this

study also examined and reported the results in the downstream services. The hypothesis was









that high cost-sharing in physician care leads to more increased ER care utilization for severely

ill people than for the healthier group.

The differential impact in ER visits was analyzed in a negative binomial regression. The

results are reported in Table 5-9. Holding covariates constant, the high cost-sharing group had an

insignificant 4% lower incidence rate of ER visits relative to the low cost-sharing group (p =

0.80). The severely ill had an insignificant 19% higher incidence rate of ER visits relative to

healthier individuals (p = 0.08). The incidence rate difference in ER visits between high and low

cost-sharing groups was 122% significantly different between the severely ill and the healthier (p

= 0.02). The positive coefficient and over 100% incidence rate ratio means the severely ill

reduced ER visits less than the healthier.

In this model, significant results were also found for other covariates. Compared with the

reference group, significantly higher ER visits were found for unmarried individuals (p = 0.01),

while cost-sharing in ER was significantly associated with fewer ER visits (p <0.01).

In summary, the two main effects of cost-sharing and severity variables did not affect ER

care utilization. Individuals with high physician care cost-sharing had similar ER visits to those

with low physician care cost-sharing, the severely ill had similar ER visits to the healthier. The

ER visit difference between high and low cost-sharing groups was less for the severely ill than

the healthier.

These results were consistent with those for physician care visits, but did not demonstrate

an expected offset effect. At this point, it is hard to judge whether these results were reasonable

or not. It is better to examine the expenditure results in ER visit as well and see if there is a

systematic pattern. If those results are inconsistent with this one, further analysis would be

needed.









Inpatient care utilization

The hypothesis was that high cost-sharing in physician care leads to more increased

inpatient care utilization for severely ill people than for the healthier group.

The differential impact in hospitalizations was analyzed in a negative binomial regression.

The results are reported in Table 5-10. Holding covariates constant, the high cost-sharing group

had an insignificant 12% higher incidence rate of hospital admissions relative to the low cost-

sharing group (p = 0.31). The severely ill had a significant 40% higher incidence rate of hospital

admissions relative to healthier individuals (p <0.01). The incidence rate difference in hospital

admissions between high and low cost-sharing groups was 9% different between the severely ill

and the healthier, which was not significant (p = 0.68).

In this model, significant results were also found for other covariates. Compared with the

reference group, the significant higher inpatient care visits were found for the middle-aged (p =

0.01) male (p = 0.03) African-American individuals (p = 0.03) who were in the near poor group

(p = 0.01), while cost-sharing in hospitalizations was significantly associated with lower

inpatient care (p = 0.03).

In summary, the high physician care cost-sharing group was not different in hospital

admissions from the low physician care cost-sharing group, the severely ill had more hospital

admissions than the healthier, the hospital admission difference between high and low physician

care cost-sharing groups was not different between the severely ill and the healthier.

The Health Care Expenditure Results

Physician care expenditure

The hypothesis was that high cost-sharing reduces physician care expenditures less for the

severely ill group than for the healthier group. The differential impact in physician visit

expenditures was analyzed in a two-part model. In part one, a logit regression was used to predict









the probability of any expenditure reduction difference between the severely ill and the reference

group. In part two, a log transformed regression was employed to indicate the difference in

magnitude in expenditure reduction given those who had any medical expenditures. The results

of the regressions are presented in Tables 5-11 and 5-12.

Part I. High cost-sharing policy holders insignificantly had 0.91 times the odds of having

any physician care expenditures relative to the low cost-sharing group (p = 0.46), the severely ill

had 13.80 times the odds of having any physician care expenditures relative to healthier

individuals (p <0.01), and their probability of reducing any physician care expenditures was also

significantly less than the reference group, with 10.81 times the odds (p = 0.02).

In this model, significant results were also found for other covariates. Compared with the

reference group, the significant higher odds of having any physician care expenditures were

found for the middle-aged (p <0.01) and older (p <0.01), White (p <0.01), female individuals (p

<0.01), who had high school (p <0.01) and college education (p <0.01), moderate conditions (p

<0.01), and primary care physicians (p <0.01). However, the near poor and low income

individuals both had significantly lower odds of having any physician care expenditures than the

poor (p = 0.02 and 0.00, respectively).

Since commands for goodness of fit assessments, such as a likelihood ratio test, Hosmer-

Lemeshow (H-L) test, R2, or Akaike's Information Criterion (AIC), cannot be obtained with
"svy" analysis in Stata, this study directly evaluated the logit model fit with the area under a

ROC curve as a proxy approach. An area under a ROC curve greater than 0.7 is a predictor of

good model fit according to common practice. The area under a ROC curve for the logit

regression was 0.8450, indicating the logit model fit was fairly good.









In summary, individuals with high cost-sharing had similar odds of having any physician

care expenditures than those with low cost-sharing, the severely ill had a higher probability for

having any physician visit expenditures than the healthier, and their odds of having lower

physician care expenditures were significantly less than for the healthier.

Part II. High cost-sharing policy holders had significantly lower physician care

expenditures compared to the low cost-sharing group (p = 0.01). The severely ill had

significantly higher physician care expenditures than the healthier (p <0.01), but expenditure

difference between high and low cost-sharing groups was not different between the severely ill

and the healthier (p = 0.70).

In this model, significant results were also found for other covariates. Compared with the

reference group, the significant higher physician care expenditures were found for the middle-

aged (p = 0.01) and older (p = 0.01), White (p = 0.04) individuals, who were married (p <0.01),

living in urban areas (p <0.01) with moderate health conditions (p <0.01), and near poor income

(p = 0.02), middle income (p <0.01), and high income (p <0.01). However, those living in the

South had significant lower physician care expenditures than those in the Midwest (p = 0.01).

The log-transformation model of physician care expenditures for those who had some

expenditures yielded an error term with a skewness of -0.05 and kurtosis of 3.16. Figure 5-1 is

the standardized normal probability (P-P) plot ("pnorm" command). It is sensitive to deviation

from normality in the middle range of data. Figure 5-2 is the quantiles of residuals against

quantiles of normal distribution plot ("qnorm" command). It is sensitive to non-normality near

the tails. These figures show that the residuals were close to a normal distribution. The

Pregibon's Link test indicated good linearity with an insignificant quadratic term of model fit (p

= 0.95). Model goodness-of-fit statistics, such as the Hosmer-Lemeshow test and AIC, cannot be









obtained after a "svy" regression command in Stata. The residual-prediction plot (Figure 5-3)

shows that the model fit was well. Glesjer's test detected heteroskedasticity for the overall model

(p <0.01), but no heteroskedasticity for the two main effects of cost-sharing and severity

variables and their interaction variable (p = 0.76, 0.17, and 0.46, respectively), which are the

study's main interests. The marginal magnitude would be calculated if the differential impact

was significant in the log transformed regression. Smearing estimators for each severity group

would then be calculated after OLS regression retransformation to correct for heteroskedasticity,

since the "robust" option cannot be used with a "svy" command in Stata.

In summary, individuals with high cost-sharing had less physician care expenditures than

those with low cost-sharing, the severely ill had higher physician care expenditures than the

healthier, but expenditure differences between high and low cost-sharing groups were not

different between the severely ill and the healthier.

The two parts were inconsistent in the signs of the interaction variables; specifically, the

differential impact was significant in part I, but not in part II. The difference in physician care

expenditures between high and low physician care cost-sharing groups was less for the severely

ill than the healthier group for the probability of having any expenditure, but not in their actual

expenditures.

ER visit expenditure

The hypothesis was that high cost-sharing in physician care leads to more increased ER

care expenditures for severely ill people than for healthier people. The differential impact in ER

visit expenditures was analyzed using a two-part model. In part one, a logit regression was used

to predict the probability of any expenditure reduction difference between the severely ill and the

reference group. In part two, a log transformed regression was employed to indicate the









difference magnitude in expenditure reduction given those who had any medical expenditures.

The results of the regressions are presented in Tables 5-13 and 5-14.

Part I. High cost-sharing policy holders had insignificant 3.41 times the odds of having any

ER care expenditures relative to the low cost-sharing group (p = 0.16), the severely ill had

insignificant 0.47 times the odds of having any ER care expenditures relative to healthier

individuals (p = 0.23), and their probability of increasing any ER expenditures was also

insignificantly higher than the reference group, with 1.99 times more odds (p = 0.60).

In this model, significant results were also found for other covariates. Compared with the

reference group, significant higher odds of having any ER care expenditures were found for

middle-income individuals (p = 0.04). However, individuals with moderate conditions had

significant 0.08 times lower odds for having any ER care expenditures (p = 0.04).

The area under the ROC curve for the logit regression was 0.7726, indicating logit model

fit was fairly good.

In summary, individuals with high physician care cost-sharing did not have a different

probability of having any ER expenditures than those with low physician care cost-sharing, the

severely ill did not have a different probability of having any ER expenditures than the healthier,

and the probability of having any ER expenditure difference between high and low physician

care cost-sharing groups was not different between the severely ill and the healthier.

Part II. High cost-sharing policy holders had insignificantly higher ER care expenditures

relative to the low cost-sharing group (p = 0.48), the severely ill had insignificantly higher ER

care expenditures relative to healthier individuals (p = 0.41), and the ER expenditure difference

between high and low physician care cost-sharing groups was not different between the severely

ill and the healthier (p = 0.64).









In this model, significant results were also found for other covariates. Compared with the

reference group, the significantly higher ER care expenditures were found for middle-aged

individuals (p = 0.04), while cost-sharing in ER was significantly associated with lower ER care

expenditures (p <0.01).

The log-transformation model of physician care expenditures for those who had actual

expenditures yielded an error term with a skewness of -0.21 and kurtosis of 3.35. Figure 5-4 is

the standardized normal probability (P-P) plot ("pnorm" command). It is sensitive to deviation

from normality in the middle range of data. Figure 5-5 shows the quantiles of residuals against

the quantiles of normal distribution plot ("qnorm" command). It is sensitive to non-normality

near the tails. These figures show that the residuals were close to a normal distribution. The

residual-prediction plot (Figure 5-6) shows that the model fit was well. Model goodness-of-fit

statistics, such as the Hosmer-Lemeshow test and AIC, cannot be obtained after a "svy"

regression command in Stata. The Pregibon's Link test indicated good linearity with an

insignificant quadratic term of model fit (p = 0.96). Glesjer's test detected no heteroskedasticity

for the overall model (p = 0.54).

In summary, individuals with high physician care cost-sharing did not have different ER

expenditures than those with low physician care cost-sharing, the severely ill did not have

different ER expenditures than the healthier, and ER expenditure differences between high and

low physician care cost-sharing groups was not different between the severely ill and the

healthier.

Neither of the two parts demonstrated a significant differential impact.

Inpatient care expenditure

The hypothesis was that high cost-sharing in physician care leads to more increased

inpatient care expenditures for severely ill people than for the healthier group. The differential









impact in inpatient care expenditures was analyzed in a two-part model. In part one, a logit

regression was used to predict the probability of any expenditure reduction difference between

the severely ill and the reference group. In part two, a log transformed regression was employed

to indicate the difference magnitude in expenditure reduction given those who had any medical

expenditure. The results of the regressions are presented in Tables 5-15 and 5-16.

Part I. High cost-sharing policy holders had 0.64 times the odds of having any inpatient

care expenditures relative to the low cost-sharing group (p = 0.05), the severely ill had 3.37 times

the odds of having any inpatient care expenditures relative to healthier individuals (p <0.01), and

the probability of having any inpatient care expenditure difference between high and low

physician care cost-sharing groups was not different between the severely ill and the healthier,

with 1.00 times the odds (p = 1.00).

In this model, significant results were also found for other covariates. Compared with the

reference group, significantly higher odds in inpatient care expenditures were found for

individuals with moderate conditions (p = 0.01). However, the middle-aged (p <0.01), Hispanic

(p = 0.04) individuals with middle incomes (p <0.01) and high income (p <0.01) had

significantly lower odds of having any inpatient care expenditures.

The area under the ROC curve for the logit regression was 0.6816, near 0.7, indicating

logit model fit was basically satisfactory.

In summary, individuals with high physician care cost-sharing had lower odds of having

any inpatient care expenditures, the severely ill had higher odds of having any inpatient care

expenditures, and the probability of having any inpatient care expenditure difference between

high and low physician care cost-sharing groups was not different between the severely ill and

the healthier.









Part II. High cost-sharing policy holders had insignificantly higher inpatient care

expenditures relative to the low cost-sharing group (p = 0.43), the severely ill had significantly

higher inpatient care expenditures relative to healthier individuals (p <0.01), and their

expenditure increase was not significantly different from the reference group (p = 0.45).

In this model, significant results were also found for other covariates. Compared with the

reference group, the significantly higher inpatient care expenditures were found for the middle-

aged (p = 0.01) and older (p = 0.05), married (p <0.01) individuals, who were White (p = 0.01)

or African-American (p <0.01). However, those living in the South had significantly lower

inpatient care expenditures than in the Midwest (p = 0.01). Cost-sharing in inpatient care was

significantly associated with lower inpatient care expenditures (p = 0.01).

The log-transformation model of physician care expenditures for those who had actual

expenditures yielded an error term with a skewness of -0.86 and kurtosis of 6.90. Figure 5-7 is

the standardized normal probability (P-P) plot ("pnorm" command). It is sensitive to deviation

from normality in the middle range of data. Figure 5-8 is the quantiles of residuals against

quantiles of normal distribution plot ("qnorm" command). It is sensitive to non-normality near

the tails. Although a kurtosis of 6.9 is not near normal, these figures show that the residuals were

basically close to a normal distribution. The residual-prediction plot (Figure 5-9) showed that the

model fit was well. The Pregibon's Link test indicated good linearity with an insignificant

quadratic term of model fit (p = 0.93). Glesjer's test detected no heteroskedasticity for the

overall model (p = 0.07), and no heteroskedasticity for the two main effects of cost-sharing and

severity variables and their interaction variable (p = 0.85, 0.69, and 0.41, respectively).

In summary, individuals with high cost-sharing in physician care did not have different

inpatient care expenditures from those with low physician care cost-sharing, the severely ill had









higher inpatient care expenditures than the healthier, and inpatient care expenditure difference

between high and low physician care cost-sharing groups was not different between the severely

ill and the healthier.

Similar to the consistent results of insignificant differential impact on ER care

expenditures, neither of the two parts in inpatient care expenditures demonstrated a significant

differential impact. However, in a utilization model the differential impact was significant for ER

visits, but insignificant for hospital admissions. This may suggest that ER and inpatient care are

complements that together serve as the downstream services for physician care. If the ER service

can meet the unmet medical care needs for physician care for the severely ill, then

hospitalizations would be redundant. It may also be possible that the offset effect was evenly

distributed between ER and inpatient care, so these two options attenuate the significant

differential impact of either of them.

Total care expenditure

The hypothesis was that high cost-sharing in physician care reduces total expenditures for

the severely ill group less than for the healthier group. The differential impact in total care

expenditures was analyzed in a two-part model. Total care refers to the sum of physician care,

ER visits and inpatient care. In part one, a logit regression was used to predict the probability of

any expenditure reduction difference between the severely ill and the reference group, and in part

two, a log transformed regression was employed to indicate the difference magnitude in

expenditure reduction given those who had any medical expenditures. The results of the

regressions are presented in Tables 5-17 and 5-18.

Part I. High cost-sharing policy holders had similar odds of having any medical care

expenditures relative to the low cost-sharing group (odds ratio = 0.95, p = 0.65), the severely ill

had 15.05 times the odds of having any medical care expenditures relative to healthier









individuals (p <0.01), and their odds of having lower total care expenditures was significantly

less than the healthier, with 8.85 times the odds (p = 0.04).

In this model, significant results were also found for other covariates. Compared with the

reference group, the significantly higher odds of having any medical care expenditures were

found for the middle-aged (p <0.01), older (p <0.01), White (p <0.01), female individuals (p

<0.01), who had high school (p <0.01) or college education (p <0.01), moderate conditions (p

<0.01), and primary care physicians (p <0.01) living in an urban area (p = 0.04). However, the

near poor (p = 0.03) and low income (p = 0.03) individuals and individuals who live in the West

(p = 0.02) had significantly lower odds of having any medical care expenditures than the

reference groups.

The area under the ROC curve for the logit regression was 0.8466, indicating logit model

fit was fairly good.

In summary, individuals with high physician care cost-sharing did not have a different

probability of having any total medical care expenditures relative to low cost-sharing group, the

severely ill had higher odds of having any total medical care expenditures relative to healthier

individuals, and their odds of having lower total medical expenditures associated with high

physician care cost-sharing was less than the healthier individuals. These results were essentially

consistent with the hypotheses.

Part II. High cost-sharing policy holders had significantly lower total medical care

expenditures relative to the low cost-sharing group (p = 0.03), the severely ill had significantly

higher total medical care expenditures relative to healthier individuals (p <0.01), but total

medical care expenditure difference between high and low physician care cost-sharing groups

was not different between the severely ill and the healthier (p = 0.55).









In this model, significant results were also found for other covariates. Compared with the

reference group, significantly higher total medical care expenditures were found for married

individuals (p <0.01), those living in an urban area (p <0.01), individuals with moderate

conditions (p <0.01), and high income (p <0.01). However, those living in the South and West

had significantly lower total medical care expenditures than those in the Midwest (p = 0.01),

while cost-sharing in ER visits was significantly associated with lower total medical care

expenditures (p <0.01).

The log-transformation model of physician care expenditures for those who had actual

expenditures yielded an error term with a skewness of -0.11 and kurtosis of 2.75. Figure 5-10 is

the standardized normal probability (P-P) plot ("pnorm" command). It is sensitive to deviation

from normality in the middle range of data. Figure 5-11 is the quantiles of residuals against

quantiles of normal distribution plot ("qnorm" command). It is sensitive to non-normality near

the tails. These figures showed that the residuals were close to a normal distribution. The

residual-prediction plot (Figure 5-12) showed that the model fit was well. The Pregibon's Link

test indicated good linearity with an insignificant quadratic term of model fit (p = 0.96).

Glesjer's test detected no heteroskedasticity for the overall model (p = 0.07), and no

heteroskedasticity for the two main effects of cost-sharing and severity variables and their

interaction variable (p = 1.00, 0.68, and 0.10, respectively).

In summary, individuals with high cost-sharing in physician care had less total medical

care expenditures relative to the low cost-sharing group, the severely ill had higher total medical

care expenditures relative to healthier individuals, but the expenditure difference between high

and low physician care cost-sharing groups was not different between the severely ill and the

healthier.









The differential impact was significant in part I, but not in part II. The probability of

having any total medical expenditures, rather than actual total medical expenditures in the

difference between high and low physician care cost-sharing groups was less for the severely ill

individuals than the healthier individuals.

Since the differential impact in expenditure models in physician care and total medical care

was split, this study further examined an overall effect that incorporated these two parts using the

bootstrap technique. Samples were replicated 1,000 times. The results indicated that, for

physician care, the differential impact was -$3,052.36, with a 95% confidence interval ranging

from -$5,180.53 to -$924.20. Similarly, for total medical care, the differential impact was

-$12,853.23, with a 95% confidence interval ranging from -$17,582.86 to -$8,123.60. Both

results were significant, meaning severely ill individuals actually reduced expenditure in

physician care and total medical care more than the general health population; the sick

population was even more sensitive to high cost-sharing pressure than the general health

population. The results further indicated a significant physician care reduction within the

severely ill group (p <0.01 for both utilization and expenditures).









Table 5-1. Description of the study sample (N= 13,020)
Variable Mean SD Minimum Maximum
Age 50.19478 17.75128 18 85
PCS 48.19457 11.35009 5.68 70.98









Table 5-2. Study sample characteristics (N = 13,020)
Variable Number %
Age (yrs)
18-34 2,832 21.75
35-64 7,122 54.70
>65 3,066 23.55
Gender
Male 5,809 44.62
Female 7,211 55.38
Race
Non-Hispanic white 8,035 61.71
Non-Hispanic African-American 2,021 15.52
Hispanic 1,987 15.26
Other 977 7.51
Geographic area
Northeast 2,203 16.92
Midwest 3,006 23.09
South 4,606 35.38
West 3,205 24.62
Rural/urban location
Urban 10,804 82.98
Rural 2,216 17.02
Marital status
Married 7,827 60.12
Not married 5,193 39.88
Education level (yrs)
0-8 1,047 8.04
9-12 5,453 41.88
13-17 6,520 50.08
Income
Poor 1,503 11.54
Near poor 551 4.23
Low income 1,648 12.66
Middle income 3,973 30.51
High income 5,345 41.05
Co-morbidities
Hypertension 4,661 35.89
No hypertension 8,326 64.11
Coronary heart disease 744 5.73
No coronary heart disease 12,238 97.47
Angina 407 3.13
No angina 12,578 96.87
Heart attack 536 4.13
No heart attack 12,457 95.87









Table 5-2. Continued
Variable Number %
Other heart disease 1,188 9.15
No other heart disease 11,799 90.85
Stroke 552 4.25
No stroke 12,445 95.75
Diabetes 1,543 11.87
No diabetes 11,452 88.13
Arthritis 3,565 27.48
No arthritis 9,406 75.52
Asthma 1,284 9.89
No asthma 11,705 90.11
Gatekeeper status
Gatekeeper 4,854 37.28
No gatekeeper 8,166 62.72
Cost-sharing level
Low 6,939 53.29
High 6,081 46.71
Disease severity level
Top tertile PCS 4,369 33.56
Middle tertile PCS 4,316 33.15
Bottom tertile PCS 4,335 33.29









Table 5-3. Association between disease severity and priority conditions (p <0.01)
Pearson Chi-square df = 1 Severe Moderate
Hypertension 3.8e+03 189.64
Coronary heart disease 2.4e+03 26.81
Angina 2.3e+03 88.05
Heart attack 1.8e+03 28.02
Other heart disease 2.3e+03 10.43
Stroke 2.2e+03 70.55
Diabetes 3.5e+03 16.03
Arthritis 7.5e+03 50.55
Asthma 1.2e+03 55.66









Table 5-4. The first stage of IV 2SLS results with cost-sharing as dependent variable
Variable Coef. S.E. p value
Age (yrs)
18-34 Reference
35-64 0.01 0.00 0.00
>65 -0.09 0.00 0.00
Gender
Female Reference
Male -0.00 0.00 0.00
Race
Other Reference
Non-Hispanic white 0.02 0.00 0.00
Non-Hispanic
African-American -0.01 0.00 0.00
Hispanic -0.01 0.00 0.00
Geographic area
Midwest Reference
Northeast 0.01 0.00 0.00
South 0.03 0.00 0.00
West 0.03 0.00 0.00
Rural/urban location
Rural Reference
Urban -0.01 0.00 0.00
Marital status
Not married Reference
Married -0.03 0.00 0.00
Education level (yrs)
0-8 Reference
9-12 0.01 0.00 0.00
13-17 0.01 0.00 0.00
Income
Poor Reference
Near poor -0.00 0.00 0.01
Low income -0.00 0.00 0.57
Middle income 0.01 0.00 0.00
High income 0.02 0.00 0.00
Gatekeeper status
No gatekeeper Reference
Gatekeeper 0.00 0.00 0.65
Disease severity level
Top tertile PCS Reference
Middle tertile PCS -0.03 0.00 0.00
Bottom tertile PCS -0.07 0.00 0.00
Premium 2.45e-06 0.00 0.00









Table 5-5. The association between IV and the error terms in the structural equation of health
care utilization and expenditure
Variable Coef. S.E. p value
Physician care
utilization error -0.00 0.00 0.23
Physician care
expenditure error -0.11 0.10 0.27









Table 5-6. Cost-sharing value description by insurance types
Insurance type Low level (%) High level (%) Mean SD
Medicare 91.67 8.33 0.086 0.111
Medicaid 90.44 9.56 0.027 0.088
Private plans 56.59 43.41 0.174 0.064
Total sample 67.00 33.00 0.137 0.107









Table 5-7. Negative binomial regression estimates for physician visits
Variable Coef. S.E. IRR S.E. p value
Cost-sharing
Low Reference
High -0.16 0.09 0.85 0.08 0.09
Disease severity level
Top tertile PCS Reference
Middle tertile PCS 0.51 0.09 1.66 0.14 0.00
Bottom tertile PCS 0.94 0.08 2.56 0.21 0.00
Cost-sharing*
Bottom tertile PCS 0.08 0.19 1.08 0.20 0.67
Middle tertile PCS -0.03 0.12 0.97 0.12 0.82
Age (yrs)
18-34 Reference
35-64 0.06 0.08 1.06 0.08 0.48
>65 0.15 0.09 1.16 0.11 0.10
Gender
Female Reference
Male -0.07 0.05 0.93 0.05 0.16
Race
Other Reference
Non-Hispanic white 0.22 0.10 1.25 0.12 0.02
Non-Hispanic
African-American 0.17 0.13 1.19 0.16 0.20
Hispanic 0.24 0.13 1.28 0.17 0.07
Geographic area
Midwest Reference
Northeast 0.09 0.08 1.09 0.08 0.25
South -0.05 0.07 0.95 0.06 0.45
West -0.09 0.07 0.91 0.06 0.18
Rural/urban location
Rural Reference
Urban 0.30 0.06 1.36 0.08 0.00
Marital status
Not married Reference
Married -0.01 0.05 0.99 0.05 0.87
Education level (yrs)
0-8 Reference
9-12 -0.23 0.13 0.80 0.10 0.08
13-17 -0.02 0.12 0.98 0.12 0.88
Income
Poor Reference
Near poor 0.21 0.20 1.23 0.24 0.30









Table 5-7. Continued
Variable Coef. S.E. IRR S.E. p value
Low income -0.20 0.09 0.82 0.08 0.03
Middle income -0.03 0.10 0.97 0.09 0.72
High income 0.09 0.09 1.10 0.10 0.30
Gatekeeper status
No gatekeeper Reference
Gatekeeper 0.11 0.05 1.12 0.06 0.04









Table 5-8. Negative binomial regression estimates for primary care physician visits
Variable Coef. S.E. IRR S.E. p value
Cost-sharing
Low Reference
High 0.07 0.05 1.07 0.05 0.16
Disease severity level
Top tertile PCS Reference
Middle tertile PCS -0.03 0.05 0.97 0.04 0.50
Bottom tertile PCS -0.10 0.05 0.91 0.04 0.05
Cost-sharing*
Bottom tertile PCS 0.15 0.09 1.16 0.10 0.09
Middle tertile PCS -0.00 0.06 1.00 0.06 0.96
Age (yrs)
18-34 Reference
35-64 0.33 0.04 1.39 0.06 0.00
>65 0.33 0.04 1.39 0.06 0.00
Gender
Female Reference
Male 0.04 0.02 1.04 0.03 0.10
Race
Other Reference
Non-Hispanic white -0.01 0.05 0.99 0.05 0.79
Non-Hispanic
African-American 0.12 0.06 1.12 0.07 0.05
Hispanic 0.06 0.06 1.06 0.06 0.33
Geographic area
Midwest Reference
Northeast -0.08 0.04 0.92 0.04 0.05
South -0.04 0.04 0.97 0.04 0.36
West 0.09 0.04 1.09 0.04 0.02
Rural/urban location
Rural Reference
Urban -0.11 0.03 0.89 0.03 0.00
Marital status
Not married Reference
Married -0.08 0.03 0.92 0.02 0.00
Education level (yrs)
0-8 Reference
9-12 -0.05 0.05 0.95 0.05 0.33
13-17 -0.20 0.05 0.82 0.04 0.00
Income
Poor Reference
Near poor -0.07 0.09 0.93 0.08 0.44









Table 5-8. Continued
Variable Coef. S.E. IRR S.E. p value
Low income -0.00 0.05 1.00 0.05 0.96
Middle income -0.02 0.04 0.98 0.04 0.72
High income -0.10 0.04 0.90 0.04 0.02
Gatekeeper status
No gatekeeper Reference
Gatekeeper 0.10 0.03 1.11 0.03 0.00









Table 5-9. Negative binomial regression estimates for ER visits
Variable Coef. S.E. IRR S.E. p value
Cost-sharing
Low Reference
High -0.04 0.17 0.96 0.17 0.80
Disease severity level
Top tertile PCS Reference
Middle tertile PCS 0.12 0.11 1.13 0.12 0.28
Bottom tertile PCS 0.17 0.10 1.19 0.12 0.08
Cost-sharing*
Bottom tertile PCS 0.80 0.32 2.22 0.72 0.01
Middle tertile PCS -0.13 0.18 0.87 0.16 0.45
ER cost-sharing -0.66 0.14 0.51 0.07 0.00
Age (yrs)
18-34 Reference
35-64 0.16 0.13 1.17 0.15 0.23
>65 -0.10 0.13 0.91 0.11 0.44
Gender
Female Reference
Male 0.01 0.07 1.01 0.07 0.87
Race
Other Reference
Non-Hispanic white 0.08 0.10 1.09 0.11 0.40
Non-Hispanic
African-American 0.13 0.16 1.14 0.19 0.43
Hispanic -0.03 0.11 0.97 0.11 0.78
Geographic area
Midwest Reference
Northeast 0.06 0.14 1.06 0.15 0.66
South -0.06 0.12 0.94 0.11 0.60
West -0.14 0.11 0.87 0.10 0.22
Rural/urban location
Rural Reference
Urban -0.02 0.07 0.98 0.07 0.82
Marital status
Not married Reference
Married -0.25 0.09 0.78 0.07 0.01
Education level (yrs)
0-8 Reference
9-12 0.02 0.12 1.02 0.12 0.87
13-17 0.02 0.12 1.02 0.12 0.88
Income
Poor Reference
Near poor -0.18 0.11 0.84 0.09 0.11









Table 5-9. Continued
Variable Coef. S.E. IRR S.E. p value
Low income -0.06 0.11 0.94 0.10 0.57
Middle income 0.04 0.14 1.04 0.15 0.78
High income 0.09 0.11 1.09 0.13 0.45
Gatekeeper status
No gatekeeper Reference
Gatekeeper -0.03 0.09 0.97 0.09 0.76









Table 5-10. Negative binomial regression estimates for hospital admissions
Variable Coef. S.E. IRR S.E. p value
Cost-sharing
Low Reference
High 0.11 0.11 1.12 0.12 0.31
Disease severity level
Top tertile PCS Reference
Middle tertile PCS 0.19 0.10 1.21 0.12 0.06
Bottom tertile PCS 0.34 0.07 1.40 0.10 0.00
Cost-sharing*
Bottom tertile PCS 0.09 0.21 1.09 0.23 0.68
Middle tertile PCS -0.29 0.13 0.75 0.10 0.03
Inpatient care cost-
sharing -0.56 0.26 0.57 0.15 0.03
Age (yrs)
18-34 Reference
35-64 0.18 0.07 1.20 0.09 0.01
>65 0.11 0.09 1.12 0.10 0.22
Gender
Female Reference
Male 0.16 0.07 1.17 0.09 0.03
Race
Other Reference
Non-Hispanic white 0.16 0.09 1.17 0.10 0.07
Non-Hispanic
African-American 0.30 0.14 1.35 0.19 0.03
Hispanic 0.17 0.10 1.19 0.12 0.09
Geographic area
Midwest Reference
Northeast 0.01 0.10 1.01 0.11 0.91
South -0.01 0.08 0.99 0.08 0.90
West -0.07 0.08 0.93 0.08 0.43
Rural/urban location
Rural Reference
Urban 0.00 0.09 1.00 0.09 0.99
Marital status
Not married Reference
Married -0.01 0.06 0.99 0.06 0.86
Education level (yrs)
0-8 Reference
9-12 0.14 0.09 1.16 0.10 0.11
13-17 0.16 0.09 1.17 0.10 0.08









Table 5-10. Continued
Variable


Coef.


IRR


p value


Income
Poor Reference
Near poor 0.42 0.15 1.52 0.23 0.01
Low income 0.06 0.09 1.06 0.10 0.53
Middle income -0.02 0.10 0.98 0.10 0.86
High income 0.02 0.08 1.02 0.08 0.81
Gatekeeper status
No gatekeeper Reference
Gatekeeper -0.03 0.07 0.97 0.06 0.67









Table 5-11. Logit regression predicting probability of having any physician care expenditures
Variable Coef. S.E. OR S.E. p value
Cost-sharing
Low Reference
High -0.09 0.12 0.91 0.11 0.46
Disease severity level
Top tertile PCS Reference
Middle tertile PCS 1.02 0.13 2.76 0.35 0.00
Bottom tertile PCS 2.62 0.18 13.80 2.54 0.00
Cost-sharing*
Bottom tertile PCS 2.38 1.04 10.81 11.27 0.02
Middle tertile PCS 0.12 0.17 1.13 0.19 0.49
Age (yrs)
18-34 Reference
35-64 0.55 0.08 1.73 0.14 0.00
>65 2.02 0.14 7.51 1.09 0.00
Gender
Female Reference
Male -1.18 0.07 0.31 0.02 0.00
Race
Other Reference
Non-Hispanic white 0.57 0.15 1.77 0.27 0.00
Non-Hispanic
African-American 0.08 0.17 1.08 0.19 0.66
Hispanic 0.23 0.17 1.26 0.21 0.17
Geographic area
Midwest Reference
Northeast 0.09 0.13 1.10 0.14 0.46
South -0.07 0.10 0.93 0.10 0.51
West -0.21 0.11 0.81 0.09 0.05
Rural/urban location
Rural Reference
Urban 0.17 0.10 1.19 0.12 0.09
Marital status
Not married Reference
Married 0.03 0.07 1.03 0.07 0.65
Education level (yrs)
0-8 Reference
9-12 0.59 0.12 1.81 0.22 0.00
13-17 1.06 0.13 2.88 0.38 0.00









Table 5-11. Continued
Variable


Coef.


p value


Income
Poor Reference
Near poor -0.50 0.22 0.60 0.13 0.03
Low income -0.48 0.15 0.62 0.09 0.00
Middle income -0.22 0.15 0.81 0.12 0.16
High income -0.07 0.16 0.94 0.15 0.68
Gatekeeper status
No gatekeeper Reference
Gatekeeper 0.61 0.08 1.85 0.15 0.00









Table 5-12. Log transformed OLS regression estimates for physician care expenditures
Variable Coef. S.E. p value
Cost-sharing
Low Reference
High -0.28 0.10 0.01
Disease severity level
Top tertile PCS Reference
Middle tertile PCS 0.46 0.09 0.00
Bottom tertile PCS 1.08 0.09 0.00
Cost-sharing*
Bottom tertile PCS -0.08 0.19 0.70
Middle tertile PCS 0.07 0.11 0.49
Age (yrs)
18-34 Reference
35-64 0.20 0.07 0.01
>65 0.21 0.08 0.01
Gender
Female Reference
Male -0.08 0.05 0.11
Race
Other Reference
Non-Hispanic white 0.18 0.09 0.04
Non-Hispanic
African-American 0.10 0.12 0.41
Hispanic 0.03 0.12 0.81
Geographic area
Midwest Reference
Northeast -0.01 0.07 0.89
South -0.17 0.07 0.01
West -0.13 0.08 0.12
Rural/urban location
Rural Reference
Urban 0.18 0.06 0.00
Marital status
Not married Reference
Married 0.15 0.05 0.00
Education level (yrs)
0-8 Reference
9-12 0.01 0.12 0.96
13-17 0.23 0.12 0.05









Table 5-12. Continued
Variable


Coef.


p value


Income
Poor Reference
Near poor 0.42 0.17 0.02
Low income 0.09 0.09 0.31
Middle income 0.32 0.09 0.00
High income 0.46 0.09 0.00
Gatekeeper status
No gatekeeper Reference
Gatekeeper 0.01 0.04 0.87









Table 5-13. Logit regression predicting probability of having any ER visits expenditure
Variable Coef. S.E. OR S.E. p value
Cost-sharing
Low Reference
High 1.23 0.88 3.41 3.00 0.16
Disease severity level
Top tertile PCS Reference
Middle tertile PCS 0.45 0.72 1.56 1.13 0.54
Bottom tertile PCS -0.75 0.63 0.47 0.30 0.23
Cost-sharing*
Bottom tertile PCS 0.69 1.31 1.99 2.59 0.60
Middle tertile PCS -2.57 1.26 0.08 0.10 0.04
ER cost-sharing 3.27 2.45 26.28 64.49 0.18
Age (yrs)
18-34 Reference
35-64 0.87 0.60 2.38 1.43 0.15
>65 0.27 0.63 1.30 0.82 0.67
Gender
Female Reference
Male 0.16 0.51 1.17 0.60 0.75
Race
Other Reference
Non-Hispanic white 1.19 0.66 3.28 2.15 0.07
Non-Hispanic
African-American 0.60 0.73 1.82 1.33 0.41
Hispanic 1.14 0.72 3.13 2.25 0.11
Geographic area
Midwest Reference
Northeast -0.87 0.64 0.42 0.27 0.17
South -0.38 0.49 0.68 0.33 0.43
West -0.54 0.75 0.58 0.44 0.47
Rural/urban location
Rural Reference
Urban 0.54 0.43 1.71 0.74 0.21
Marital status
Not married Reference
Married -0.10 0.59 0.90 0.53 0.86
Education level (yrs)
0-8 Reference
9-12 0.25 0.60 1.28 0.77 0.68
13-17 -0.30 0.61 0.74 0.46 0.63









Table 5-13. Continued
Variable


Coef.


p value


Income
Poor Reference
Near poor -0.87 0.74 0.42 0.31 0.24
Low income 0.60 0.63 1.83 1.15 0.34
Middle income 1.44 0.68 4.23 2.89 0.04
High income 0.19 0.66 1.20 0.79 0.78
Gatekeeper status
No gatekeeper Reference
Gatekeeper 0.46 0.40 1.58 0.62 0.25









Table 5-14. Log transformed OLS regression estimates for ER care expenditures
Variable Coef. S.E. p value
Cost-sharing
Low Reference
High 0.16 0.22 0.48
Disease severity level
Top tertile PCS Reference
Middle tertile PCS 0.15 0.18 0.42
Bottom tertile PCS 0.16 0.19 0.41
Cost-sharing*
Bottom tertile PCS 0.26 0.54 0.64
Middle tertile PCS -0.06 0.25 0.81
ER cost-sharing -1.38 0.31 0.00
Age (yrs)
18-34 Reference
35-64 0.33 0.16 0.04
>65 -0.16 0.18 0.37
Gender
Female Reference
Male 0.11 0.10 0.28
Race
Other Reference
Non-Hispanic white -0.12 0.22 0.59
Non-Hispanic
African-American 0.05 0.24 0.85
Hispanic -0.26 0.28 0.36
Geographic area
Midwest Reference
Northeast -0.09 0.13 0.50
South -0.24 0.13 0.06
West -0.28 0.16 0.08
Rural/urban location
Rural Reference
Urban 0.03 0.12 0.79
Marital status
Not married Reference
Married 0.20 0.11 0.07
Education level (yrs)
0-8 Reference
9-12 0.15 0.20 0.45
13-17 0.21 0.21 0.32









Table 5-14. Continued
Variable


Coef.


p value


Income
Poor Reference
Near poor -0.04 0.22 0.84
Low income 0.24 0.17 0.16
Middle income 0.06 0.18 0.75
High income 0.35 0.18 0.06
Gatekeeper status
No gatekeeper Reference
Gatekeeper -0.04 0.12 0.77









Table 5-15. Logit regression predicting probability of having any inpatient care expenditure
Variable Coef. S.E. OR S.E. p value
Cost-sharing
Low Reference
High -0.44 0.22 0.64 0.14 0.05
Disease severity level
Top tertile PCS Reference
Middle tertile PCS 0.58 0.22 1.78 0.39 0.01
Bottom tertile PCS 1.21 0.19 3.37 0.65 0.00
Cost-sharing*
Bottom tertile PCS 0.00 0.31 1.00 0.31 1.00
Middle tertile PCS 0.31 0.29 1.37 0.40 0.28
Inpatient care cost-
sharing 0.07 1.85 1.07 1.99 0.97
Age (yrs)
18-34 Reference
35-64 -0.51 0.15 0.60 0.09 0.00
>65 -0.30 0.16 0.74 0.12 0.07
Gender
Female Reference
Male -0.10 0.10 0.91 0.10 0.36
Race
Other Reference
Non-Hispanic white -0.28 0.19 0.76 0.14 0.14
Non-Hispanic
African-American -0.40 0.23 0.67 0.16 0.09
Hispanic -0.50 0.24 0.60 0.14 0.04
Geographic area
Midwest Reference
Northeast -0.14 0.15 0.87 0.13 0.35
South 0.14 0.13 1.15 0.15 0.30
West -0.14 0.14 0.87 0.13 0.34
Rural/urban location
Rural Reference
Urban 0.22 0.13 1.25 0.17 0.10
Marital status
Not married Reference
Married 0.16 0.10 1.17 0.12 0.14
Education level (yrs)
0-8 Reference
9-12 -0.07 0.18 0.93 0.17 0.70
13-17 0.06 0.19 1.06 0.20 0.76









Table 5-15. Continued
Variable


Coef.


p value


Income
Poor Reference
Near poor 0.01 0.25 1.01 0.25 0.97
Low income -0.14 0.17 0.87 0.15 0.39
Middle income -0.48 0.16 0.62 0.10 0.00
High income -0.48 0.17 0.62 0.10 0.00
Gatekeeper status
No gatekeeper Reference
Gatekeeper 0.09 0.10 1.09 0.11 0.38









Table 5-16. Log transformed OLS regression estimates for inpatient care expenditures
Variable Coef. S.E. p value
Cost-sharing
Low Reference
High 0.15 0.19 0.43
Disease severity level
Top tertile PCS Reference
Middle tertile PCS 0.18 0.19 0.35
Bottom tertile PCS 0.52 0.16 0.00
Cost-sharing*
Bottom tertile PCS -0.23 0.30 0.45
Middle tertile PCS -0.15 0.23 0.52
Inpatient care cost-sharing -1.46 0.58 0.01
Age (yrs)
18-34 Reference
35-64 0.28 0.10 0.01
>65 0.24 0.12 0.05
Gender
Female Reference
Male 0.18 0.11 0.11
Race
Other Reference
Non-Hispanic white 0.35 0.14 0.01
Non-Hispanic
African-American 0.65 0.18 0.00
Hispanic 0.24 0.21 0.26
Geographic area
Midwest Reference
Northeast -0.16 0.12 0.18
South -0.24 0.11 0.04
West -0.08 0.13 0.57
Rural/urban location
Rural Reference
Urban -0.06 0.10 0.56
Marital status
Not married Reference
Married 0.34 0.09 0.00
Education level (yrs)
0-8 Reference
9-12 0.01 0.16 0.94
13-17 0.05 0.17 0.76









Table 5-16. Continued
Variable


Coef.


p value


Income
Poor Reference
Near poor 0.35 0.30 0.25
Low income 0.03 0.14 0.83
Middle income 0.06 0.14 0.65
High income 0.06 0.14 0.66
Gatekeeper status
No gatekeeper Reference
Gatekeeper -0.01 0.10 0.95









Table 5-17. Logit regression predicting probability of having any medical care expenditure
Variable Coef. S.E. OR S.E. p value
Cost-sharing
Low Reference
High -0.06 0.13 0.95 0.12 0.65
Disease severity level
Top tertile PCS Reference
Middle tertile PCS 1.07 0.12 2.90 0.36 0.00
Bottom tertile PCS 2.71 0.20 15.05 3.01 0.00
Cost-sharing*
Bottom tertile PCS 2.18 1.05 8.85 9.27 0.04
Middle tertile PCS 0.10 0.17 1.10 0.18 0.57
ER cost-sharing -0.18 0.33 0.83 0.27 0.58
Inpatient care
cost-sharing -0.75 0.53 0.47 0.25 0.16
Age (yrs)
18-34 Reference
35-64 0.50 0.08 1.65 0.14 0.00
>65 1.97 0.15 7.17 1.09 0.00
Gender
Female Reference
Male -1.19 0.07 0.30 0.02 0.00
Race
Other Reference
Non-Hispanic white 0.54 0.15 1.72 0.26 0.00
Non-Hispanic
African-American 0.09 0.17 1.09 0.19 0.61
Hispanic 0.21 0.17 1.24 0.21 0.21
Geographic area
Midwest Reference
Northeast 0.09 0.12 1.10 0.13 0.46
South -0.08 0.10 0.93 0.09 0.44
West -0.25 0.10 0.78 0.08 0.02
Rural/urban location
Rural Reference
Urban 0.20 0.10 1.23 0.12 0.04
Marital status
Not married Reference
Married 0.02 0.07 1.02 0.07 0.74
Education level (yrs)
0-8 Reference
9-12 0.56 0.13 1.75 0.22 0.00
13-17 1.03 0.14 2.81 0.39 0.00









Table 5-17. Continued
Variable Coef. S.E. OR S.E. p value
Income
Poor Reference
Near poor -0.52 0.23 0.60 0.14 0.03
Low income -0.58 0.16 0.56 0.09 0.00
Middle income -0.30 0.16 0.74 0.12 0.07
High income -0.17 0.17 0.84 0.14 0.31
Gatekeeper status
No gatekeeper Reference
Gatekeeper 0.60 0.08 1.82 0.15 0.00









Table 5-18. Log transformed OLS regression estimates for total medical care expenditures
Variable Coef. S.E. p value
Cost-sharing
Low Reference
High -0.25 0.11 0.03
Disease severity level
Top tertile PCS Reference
Middle tertile PCS 0.58 0.10 0.00
Bottom tertile PCS 1.35 0.10 0.00
Cost-sharing*
Bottom tertile PCS -0.12 0.21 0.55
Middle tertile PCS 0.08 0.13 0.52
ER cost-sharing -1.13 0.39 0.00
Inpatient care cost-sharing -0.88 1.01 0.39
Age (yrs)
18-34 Reference
35-64 0.04 0.07 0.52
>65 0.03 0.09 0.76
Gender
Female Reference
Male -0.09 0.05 0.10
Race
Other Reference
Non-Hispanic white 0.17 0.10 0.09
Non-Hispanic
African-American 0.14 0.13 0.30
Hispanic -0.00 0.13 0.99
Geographic area
Midwest Reference
Northeast -0.05 0.08 0.52
South -0.14 0.07 0.04
West -0.16 0.08 0.04
Rural/urban location
Rural Reference
Urban 0.19 0.07 0.00
Marital status
Not married Reference
Married 0.19 0.05 0.00
Education level (yrs)
0-8 Reference
9-12 -0.02 0.12 0.83
13-17 0.21 0.11 0.06









Table 5-18. Continued
Variable


Coef.


p value


Income
Poor Reference
Near poor 0.33 0.18 0.08
Low income 0.04 0.09 0.63
Middle income 0.12 0.10 0.22
High income 0.24 0.09 0.01
Gatekeeper status
No gatekeeper Reference
Gatekeeper 0.01 0.05 0.88






































0.00 0.25 0.50 0.75 1.00
Empirical P[i] = i/(N+1)



Figure 5-1. P-P plot for log transformed OLS regression on physician care expenditures


1.0-






C -






0






o


0
Inverse Normal


Figure 5-2. Q-Q plot for log transformed OLS regression on physician care expenditures














5.40609 +


Linear prediction


8.85922


Figure 5-3. Residual-fitted plot for log transformed OLS regression on physician care

expenditures


0.25 0.50 0.75
Empirical P[i] = i/(N+1)


1.00


Figure 5-4. P-P plot for log transformed OLS regression on ER care expenditures


*
*************** ** ** **
**** **************************** **** ** ******
*** ********************************** ****** *****
** **** *** ******** **** **** *** ******** **** *** ******** **** *
**** *** ******** **** **** *** ******** **** *** ******** **** ***
* *** **** *** ******** **** **** *** ******** **** *** ******** **** *** *
***** **** *** ******** **** **** *** ******** **** *** ******** **** ***
* ***** **** *** ******** **** **** *** ******** **** *** ******** **** ***
* **** **** *** ******** **** **** *** ******** **** *** ******** **** *
**** *** ******** **** **** *** ******** **** *** ******** **** ***
*** **** **** *** ******** **** *** **** ******** *** **** ** *
k ** ************************** ******* **** **
*** *** ** *** *** ***** **** **
** ** ** *
***
*


-6.89628 +


5.95516


0.00


































0 -
















LO -
I


Inverse Normal


Figure 5-5. Q-Q plot for log transformed OLS regression on ER care expenditures


3.89166 +


** ** *
** **** *
**** ****** ** ********** ****
** ** ******** ********************* *


***
*** *
*
* *** *
***
*


***** ********************************* *
*** *************** **** **** *** **** *** **
* **************************************** *
**************************************** *
* *************************************** *
**** *********** *** **** **** *** **** ******** **
*** ** *********** **** **** *** **** ******** *
** ** ****** ********** **** **
** ** ******** ******* *
**** *
** ** *
*


-4.92886 +
4.55936
4.55936


Linear prediction


Figure 5-6. Residual-fitted plot for log transformed OLS regression on ER care expenditures



















131


** *


7.21069


































0.00 0.25 0.50 0.75 1.00
Empirical P[i] = i/(N+l)

Figure 5-7. P-P plot for log transformed OLS regression on inpatient care expenditures


0-


* m


-4 -2 0 2 4
Inverse Normal

Figure 5-8. Q-Q plot for log transformed OLS regression on inpatient care expenditures














3.67583 +




I *


-7.22596 +

7.20208


Linear prediction


Figure 5-9. Residual-fitted plot for log transformed OLS regression on inpatient care

expenditures


0.00


0.25 0.50 0.75
Empirical P[i] = i/(N+1)


Figure 5-10. P-P plot for log transformed OLS regression on total medical care expenditures


** *** ** *
***** ************ ****
**************************** *****
**** ********************************* *
* *** ***************************** ***
**** ****************************** *
*** ************************** *
** ** **** ******** *** **
** **** *** *** **
* ****** *
**
*
** *
*


10.3167














O






LC

















0
O









O


-10 -5 0 5
Inverse Normal


Figure 5-11. Q-Q plot for log transformed OLS regression on total medical care expenditures



5.75448 +
R *
I *
** *
| *********** ** *** *
R ** ********************************** ** **** *
e ****************************************** **
S I *** ************************************************
i I ** ****** ********************************************* *
d I ** ** *********************************************** *
U | ************************************************ *
a I ***********************************************
1 I ************************************************
S ******************************************** ***
**************************************** *** *
S** ************************************* **
** *** ** **** ***** ********* *
** **** ** ** **
S* *
-7.32696 + *
+----------------------------------------------------------------+
5.65631 Linear prediction 9.48733


Figure 5-12. Residual-fitted plot for log transformed OLS regression on total medical care

expenditures














134









CHAPTER 6
DISCUSSION

Summary of Basic Findings

The overall results have laid a framework to address the research question. This study

summarizes systematic patterns within each service type across health care utilization and

expenditure models.

In physician care, the differential impact was not significant in the utilization model and

the actual expenditure model, but significant in the probability expenditure model indicating less

expenditure reduction (Figure 6-1). In ER care, this differential impact was significant in the

utilization model (Figure 6-2), but was not significant in either part of the expenditure model.

The inpatient care results demonstrated consistent patterns; no differential impact was found in

either the utilization or the expenditure model. As a component of physician care, the primary

care physician visit model also produced an insignificant differential impact. Last, in total

medical care, this differential impact was not significant in the actual expenditure model, but

significant in the probability expenditure model (Figure 6-3).

In summary, the differential impact was less pronounced since it was only demonstrated in

the ER utilization model and the expenditure probability model in physician care and total

medical care (Table 6-1). The integrated medical care expenditure models revealed opposite

differential impacts, indicating severely ill individuals actually reduced expenditures in physician

care (Figure 6-4) and total medical care (Figure 6-5) more, thus were even more sensitive to high

cost-sharing pressure than the general health population. All the significant differential impacts

can be summarized by service types as follows (Table 6-2): ER visits, the probability of having

any physician care and total care expenditures, integrated physician care and total care









expenditures. As ER utilization result is an intermediate outcome for total care results, so this

study will focus on discussing physician care and total care results.

The negative significant differential impacts in integrated physician care and total care

expenditure models were opposite to expectations. As discussed in this study, severe diseases

and a high cost-sharing financial burden compete against each other and present the severely ill

with a dilemma. Under high cost-sharing pressure, either choice-retaining needed physician

care or saving money-will come by sacrificing the other. Thus, it is possible that this opposite

differential impact may be due to a poverty constraint or fear of excessive debt burden in the

severely ill group, and this reasoning was confirmed by the study results. Poverty was measured

by family income categories and Medicaid enrollment status. Chi square test results consistently

indicated a significantly higher proportion of poor individuals within the severely ill group than

the healthier group (p <0.01 and 0.01, respectively).

The results further indicated a significant physician care reduction within the severely ill.

Since the severely ill group contained a substantial proportion of poor individuals, their

significant care reduction in both utilization and expenditures in response to high cost-sharing

clearly indicated that high cost-sharing policies were harmful; it pushed them to forgo substantial

physician care. A natural concern would arise as to whether this substantial physician care

reduction was needed or essential to maintain their health, or whether this loss would have

detrimental consequences. The answer would be demonstrated in their downstream ER service

utilization or direct health outcome change. The results indicated within the severely ill group,

relative to the low MD cost-sharing group, the high MD cost-sharing group had significantly

higher ER utilization and expenditures, and significantly lower PCS. Their sizable offset effects









in downstream ER service and worse health status, in conjunction with their substantial care

reduction, suggested that the severely ill may have forgone essential physician care.

Thus, high physician care cost-sharing policies could deprive the severely ill of essential

physician care, thus worsening their health conditions and pushing them into the undesired

downstream ER service, a more expensive service for worse clinical conditions, and thus may

hurt both their health and finances.

Recall that in the physician care probability expenditure model, the differential impact was

positively significant, and in the physician care integrated expenditure model, the differential

impact was negatively significant. Considered together, this means that severely ill individuals

had a strong desire to reduce physician care expenditures less than the general health population.

However, they actually reduced more because high cost-sharing policies, reinforced by their

financial difficulties, resulted in more elastic demand. Thus, this finding, in combination with

offset ER utilization and worse health status, demonstrated that high cost-sharing policies could

greatly hurt and penalize severely ill individuals.

Physician Care Price Elasticity

Calculation of price elasticity of physician care also contributes to the understanding of

the research questions. Price elasticity is calculated with the midpoint approach.

E = AQ% / AP% = {(Qi-Qo) / [(QI+Qo) / 2]} / {(Pi- Po) / [(Pi+Po) / 2]} = [(Q1-Qo)*(Pi+Po)] /

[(Qi+Qo)*(Pi-Po)]

The price elasticity for the severely ill was -0.139 in physician care utilization and -0.399

in physician care expenditures. The cross-price elasticity of physician care for the severely ill

was 0.381 in ER care utilization, and 0.183 in ER care expenditures.

These results are comparable to those found in the literature, with a range of price

elasticity from -0.14 to -1.9 for physician care. The severely ill had the upper limit price









elasticity of -0.14, which shows that their physician care had little room to be cut back, and was

uniquely valuable and essential to maintain their health. Meanwhile, these positive cross-price

elasticity estimates further indicate that physician care and ER are substitutes rather than

complements. ER care can act as a backup service when physician care is unavailable to the

severely ill due to high cost-sharing, but ER care, a more expensive service for worse clinical

conditions, is by no means a better choice for the severely ill.

Other Sensitivity Analyses

This study also examined other sensitivity analyses to test finding robustness. First, if there

are a number of individuals whose cost-sharing is exactly 20%, or 19% for Medicare and private

plans, the results may be influenced when they are categorized into high or low cost-sharing

groups. The results indicated that there were no observations whose cost-sharing was exactly

20% or 19% for Medicare or private plans.

Second, this study plan also considered different cut-off points for cost-sharing, since there

could be some doubt about whether it is really a high or low level if cost-sharing is a little higher

or lower than 20% for Medicare and private plans or the mean value for Medicaid. This study

considered 25% and 15% as cut-off points for Medicare and private plans, and 5% away from

either side of the mean value for Medicaid. However, the proxy treatment of 19% for 20% due to

the constraint of "perfect success prediction" precluded the model specification of cut off point

extension. Nevertheless, this study did adopt the same cut off point of 19% across insurance plan

types, including Medicaid, to examine the result robustness. Basically, the universal cut off point

did not change the results, since only 597 observations out of 60,595 transferred from high to

low cost-sharing groups. As a result, all the results kept the same significance and signs except

primary care physician visit and total care expenditure part I results, where significance changed

slightly from marginally significant to significant, or vice versa. Thus, the results are robust.









Third, this study also planned and had included the middle tertile subjects in the severity

level as a dummy variable in the model, and its interaction with the cost-sharing variable

produced a less pronounced differential impact than the interaction by the severely ill and cost-

sharing.

Fourth, alternative ways of categorizing PCS was considered to examine the results'

robustness, such as using both sides of two standard deviations from the PCS mean as cut off

points. This approach, however, did not apply to the skewed distribution of PCS, which yielded

only a few observations in the severely ill group.










Table 6-1. Differential impact summary by service types
Service type Coefficient S.E. p value
Physician care visits 0.08 0.19 0.67
Physician care expenditure part I 2.38 1.04 0.02
Physician care expenditure part II -0.08 0.19 0.70
Primary care visits 0.15 0.09 0.09
ER visits 0.08 0.32 0.01
ER expenditure part I 0.69 1.31 0.60
ER expenditure part II 0.26 0.54 0.64
Hospital admissions 0.09 0.21 0.68
Inpatient care expenditure part I 0.00 0.31 1.00
Inpatient care part II -0.23 0.30 0.45
Total care expenditure part I 2.18 1.05 0.04
Total care expenditure part II -0.12 0.21 0.55


Table 6-2. Significant differential impact summary by service types
Service type Differential impact sign
Probability of physician care expenditures +
Integrated physician care expenditures
Emergency room care visits +
Probability of total care expenditures +
Integrated total care expenditures









Probability of having any physician care expenditures


Severely ill group





Reference group


- Price of medical care


Low cost-sharing


High cost-sharing


Figure 6-1. The differential impact by health status in probability of having any physician care
expenditures




Utilization of ER care


Severely ill group





Reference group


* Price of medical care


Low cost-sharing


High cost-sharing


Figure 6-2. The differential impact by health status in ER care visits








Probability of having any total medical care expenditures


Severely ill group




Reference group
- Price of medical care


Low cost-sharing


High cost-sharing


Figure 6-3. The differential impact by health status in probability of having any total medical
care expenditures





Integrated physician care expenditures






Severely ill group


Reference group


Price of medical care


Low cost-sharing


High cost-sharing


Figure 6-4. The differential impact by health status in integrated physician care expenditures








Integrated total medical care expenditures


Severely ill group

Reference group
Price of medical care

Low cost-sharing High cost-sharing

Figure 6-5. The differential impact by health status in integrated total medical care expenditures









CHAPTER 7
CONCLUSIONS

Summary and Interpretation

In reality, severely ill individuals most usually have low incomes. As was identified in this

study, the severely ill group had a significantly higher proportion of poor patients than the

healthier group. Thus, the sick and poor group's financial difficulties made them sensitive and

vulnerable to high cost-sharing policies that constrained them from maintaining adequate

necessary health care, and their utilization reduction magnitude was likely an involuntary

behavior rather than an active choice. Thus, in response to high cost-sharing pressure, severely ill

individuals could experience substantial physician care reduction, ER care increase, and worse

clinical conditions. Their price elasticity of demand for physician care was small, indicating that,

in severely ill individuals' perception, physician care was essential to maintain their health.

However, high cost-sharing policies actually thwarted their need for more frequent physician

care.

Negative Consequence of High Cost-sharing Policies

Severe diseases and a high cost-sharing financial burden compete against each other for

severely ill and poor individuals and present the sick and poor group with a dilemma. Either

choice-maintaining health or lowering cost-will come at the price of the other. Under this

condition, the existence of positive differential impact in the expenditure probability models for

severely ill individuals indicated that they had a strong desire and tendency to reduce less needed

medical care. However, the negative differential impact in the integrated expenditure models

indicated that severely ill individuals actually reduced more needed medical care. As a whole,

severely ill individuals demonstrated that, although they wished to reduce less needed medical

care because it was essential to maintain their health, they actually reduced it more than the









general population because high cost-sharing policies, in conjunction with their financial

difficulties, distorted their desires and hindered their voluntary behaviors; in short, they were

more sensitive and vulnerable to high cost-sharing pressure.

The positively significant differential impact in the probability expenditure model and the

negatively significant differential impact in the integrated expenditure model indicated an even

more harmful effect of high cost-sharing policies for this especially vulnerable subpopulation.

High cost-sharing policies acted as a counter force to undo the severely ills' health maintenance

efforts by imposing high financial pressure, eroding their bottom line of affordability, and then

forcing them to delay or forgo needed physician care. In fact, high cost-sharing policies for

physician care had already pushed the severely ill away from needed physician care and into the

less desirable downstream ER service, a more expensive service for worse clinical conditions,

and thus may hurt them both clinically and financially. Moreover, high cost-sharing policies also

may have exacerbated these vulnerable people's current severe conditions.

In summary, opposite differential impacts in expenditure models in combination with

offset ER utilization and worse health status results consistently demonstrated that high cost-

sharing policies could greatly distort severely ill individuals' willingness to pay for needed care.

Policy Implications

A high cost-sharing policy will naturally directly reduce health care utilization, regardless

of subpopulations. In theory, severely ill individuals should be price inelastic because care is

necessary to maintain their health. However, high cost-sharing policies acted as a counter force

to undo the severely ills' health maintenance efforts by imposing a financial barrier, and these

policies were especially harmful to sick and poor individuals. In reality, severely ill individuals

are usually associated with low incomes. Because of this, this sick and poor group was even

more vulnerable to high cost-sharing policies than the general health population. The opposite









differential impact for the sick and poor group highlighted their financial difficulties or even risk,

which distorted their willingness to maintain their health by reducing needed physician care to a

larger extent than healthier individuals. Thus, current high cost-sharing policies should be

replaced with low cost ones to reflect and match severely ill individuals' situation.

Recall that the existence of a positive differential impact in the expenditure probability

models for severely ill individuals indicated that they had a strong tendency and desire to reduce

needed medical care less than the general health population in response to high cost-sharing

pressure, which is a "pure" result, not being contaminated by their practical financial difficulties.

This phenomenon may reveal the potential underlying truth that the efficient and inefficient

moral hazard fractions differ by disease severity. Severely ill people usually have less inefficient

moral hazard and more efficient moral hazard than healthy people. They have a smaller fraction

of discretionary care and a larger fraction of needed care than the general health population, so

they "waste" less medical care. Cost-sharing was originally introduced in insurance policies to

control inefficient health care utilization by increasing an individual's awareness of cost. Based

on the moral hazard level difference between a subpopulation's health status, cost-sharing's

levels should be designed differently and uniquely for subpopulations. Specifically, since

severely ill individuals have a smaller fraction of discretionary care, they should be treated

differently with a low cost-sharing policy that should cut down smaller inefficient health care

utilization.

As the results indicated, high cost-sharing policies for physician care cut back needed care

for the severely ill to a greater extent, which highlights these people's financial difficulties and

vulnerability to cost pressure. Consequently, high cost-sharing policies for physician care could

hurt and penalize the severely ill both clinically and financially. A low cost-sharing policy would









remove the financial threat from the severely ill population's decision-making so as to provide

them with access to needed care.

Considering these facts and results together, low cost-sharing is not only necessary in

theory, but also feasible and beneficial in practice for severely ill individuals. It protects and

prevents them from exacerbating their poor health status and financial burden, and saves costs

for the society at the aggregate level.

In addition to copayment, coinsurance, and deductible, cost-sharing can also take the form

of an out-of-pocket maximum. An appropriate out-of-pocket maximum level can effectively

protect severely ill individuals' health status if it is within the low cost-sharing categories.

RAND-HIE found no differential impact in health care expenditures. One possible explanation

discussed in a paper by Manning et al. (1987) may be due to the low out-of-pocket maximum

level, so that not only the sick, but also healthy people could easily exceed this boundary and

enjoy free inpatient care. That level may be appropriate for sick individuals, but questionable for

healthy people. In the 1970s when the RAND-HIE was conducted, the out-of-pocket maximum

level was only $1,050. The findings of this present study call for a low out-of-pocket maximum

level for the severely ill, and today this level should be different from several decades ago and

adjusted based on the current situation. More importantly, this level should not be high, so that it

can provide the severely ill with adequate access to medical care to maintain their health and

minimize their financial burden.

This study highlighted the necessity and importance of value-based insurance design in

terms of differentiation and specification of its target population, so as to best protect and prevent

them from exacerbating their poor health status and financial burden, and to save cost for the

society at the aggregate level. Furthermore, this study will contribute to the current debate on









health care reform, especially specialized plans for subpopulations instead of a universal plan,

may want to be considered to complement the existing public and private designs.

Limitations

There are some limitations in the data, measurements, and operationalization.

MEPS data only represents non-institutionalized individuals, which may limit this study's

external validity. In addition, self-reporting system in MEPS data may cause measurement error

concern. Moreover, the data do not contain information from providers if induced demand exists.

There may be concerns that health care utilization is also influenced by supply-side dynamics,

when a provider's revenue decreases substantially due to reduced utilization by cost-sharing

from the demand side. In order to compensate and maintain their revenue level, providers will

increase utilization by inducing demand so that the healthier group may not reduce medical use

substantially. By doing so, they may demonstrate a similarly lesser amount of care reduction as

the severely ill group. In this situation, no differential impact will be observed. Fortunately, the

results indicated a significant physician care reduction within each group, the severely ill (p

<0.01) and the reference group (p <0.01) for both utilization and expenditures, thus eliminating

the supply side information concern.

In addition, the PCS measure focuses more on health status and functional status that

indirectly reflect an individual's health condition, so it is an imperfect measure of severity of

illness. This study examined service types, including physician care, ER visits, and inpatient

care. However, medication is another service type that might influence the relationships among

these service types and the results. Inclusion of medications will be one of the future research

directions.

Despite these limitations, this study significantly adds to the literature by focusing on cost-

sharing only in physician care, using a nationally representative sample and more precise









measures, and adopting more rigorous methods to improve the understanding of the research

question.

By observing severely ill people's behavior, specifically their health care utilization change

in response to high cost-sharing pressure, this study allows researchers to evaluate the effect of

high cost-sharing policies, and whether they are necessary to cut down health care utilization and

expenditures for the severely ill. Severely ill individuals demonstrated a strong desire to reduce

necessary medical care to a lesser extent than the general health population, however, high cost-

sharing policies went against their wishes, and pushed the severely ill away from needed

physician care and into the less desired downstream ER service, and may hurt them both

clinically and financially.

Therefore, these vulnerable people should be treated differently. Specifically, a low cost-

sharing policy should be designed to reflect and match their situation. This study contributes to

inform the necessity and importance of insurance policy design in terms of differentiation and

specification for its target population, so as to best protect and prevent them from exacerbating

their poor health status and financial burden, and to save costs for society at the aggregate level.

Beside adding service types, future research plans to explore the differential impact of high

cost-sharing in physician care on health care utilization and expenditures by age (elderly and

non-elderly populations) and income (poor and non-poor populations), which corresponds to

Medicare (relative to non-Medicare) and Medicaid (relative to non-Medicaid) plans. The purpose

and significance of this series of inquiries is to determine whether insurance policy design,

specifically the cost-sharing level, should be unique to Medicare and Medicaid or

subpopulations, which will contribute to the insurance policy debate on universal and specialty

health plans.









LIST OF REFERENCES


Agency for Healthcare Research and Quality. Medical Expenditure Panel Survey: Household
Component Weighting and Estimation Issues. Agency for Healthcare Research and Quality.
2007.

Andersen, R. "Revisiting the Behavioral Model and Access to Medical Care: Does It Matter?"
Journal of Health and Social Behavior. 1995. 36: 1-10

Bax, L., A. Algra, W.P Mali, M. Edlinger, J.J. Beutler, and Y. van der Graaf. "Renal function as
a risk indicator for cardiovascular events in 3216 patients with manifest arterial disease."
A heit,,,le,,ii. 2008. 200(1): 184.

Cartwright, W. S., T.-W. Hu, and L.-F. Huang. "Impact of varying medigap insurance coverages
on the use of medical services of the elderly." Applied Economics. 1992. 24: 529-39.

Chandra, Amitabh, Jonathan Gruber, and Robin McKnight. "Patient Cost-/hii ing.
Hospitalization Offsets, and the Design of Optimal Health Insurance for the Elderly. Working
Paper 12972. 2007. http://www.nber.org/papers/wl2972. Accessed 09/04/2009.

Cherkin, D.C., L. Grothaus, and H. Wagner. "The effect of office visit copayments on utilization
in a health maintenance organization." Medical Care. 1989. 27(7): 669-78.

Chernew. M.E., M.R. Shah, A. Wegh, S.N. Rosenberg, I.A. Juster, et al. "Impact of decreasing
copayments on medication adherence within a disease management environment." Health
Affairs. 2008. 27: 103-12.

Cohen, S.B., R. DiGaetano, and H. Goksel. "Methodology Report #5: Estimation Procedures in
the 1996 Medical Expenditure Panel Survey Household Component. Agency for Healthcare
Research and Quality. 1999.

Conwell, Leslie J., and Joel W. Cohen. "Characteristics of Persons with High Medical
Expenditures in the U.S. Civilian Noninstitutionalized Population." 2002. Statistical Brief 73.
2005.

Cowan, Cathy, Aaron Catlin, Cynthia Smith, and Arthur Sensenig. "National health
expenditures, 2002. Health Care Financing Review. 2004. 25: 4.

Cutler, David M., and Richard J. Zeckhauser. "The Anatomy of Health Insurance. "Working
Paper 7176. 1999. http://www.nber.org/papers/w7176.pdf. Accessed 09/06/2009.

Daniels, Norman. "Justice, health, and healthcare." American Journal of Bit'ethi/, 2001. 1(2): 2-
16

Dormuth, Colin R., Peter Neumann, Malcolm Maclure, Robert J. Glynn, and Sebastian
Schneeweiss. "Effects of prescription coinsurance and income-based deductibles on net health
plan spending for older users of inhaled medications." Medical Care. 2009. 47: 508-516.










Doty, Michelle M., Sara R. Collins, Sheila D. Rustgi, and Jennifer L. Kriss. Seeing Red: The
Growing Burden of Medical Bills and Debt Faced by U.S. Families. Commonwealth Fund pub.
1164, vol. 42.

Eppig, Franklin. "Medicare current beneficiary survey: Centers for Medicare & Medicaid
Services. http://aging.senate.gov/award/cms3.pdf. Accessed 09/04/2009.

Fairman, Kathleen A., Brenda R. Motheral, and Rochelle R. Henderson. "Retrospective, long-
term follow-up study of the effect of a three-tier prescription drug copayment system on
pharmaceutical and other medical utilization and costs." Clinical Therapeutics. 2003. 25(12):
3147-61.

Feldman, R., S.T. Parente, and J.B. Christianson "Consumer-directed health plans: New
evidence on spending and utilization." Inquiry. 2007. 44: 26-40.

Follmann, Dean, Michael P. Fay, and Michael Proschan. "Chop-lump tests for vaccine trials."
Biometrics. 2009. 65: 885-893.

Freeman, J.D., S. Kadiyala, J.F. Bell, and D.P. Martin "The causal effect of health insurance on
utilization and outcomes in adults: A systematic review of US studies." Medical Care. 2008.
46(10): 1023-32.

Gibson, T.B., R.J. Ozminkowski, and R.Z. Goetzel. "The effects of prescription drug cost
sharing: A review of the evidence." American Journal of Managed Care. 2005. 11:730-40.

Goldman, Dana P., Geoffrey F. Joyce, and Pinar Karaca-Mandic. "Varying pharmacy benefits
with clinical status: The case of cholesterol-lowering therapy." American Journal of Managed
Care. 2006. 12: 21-28.

Goldman, Dana P., Geoffrey F. Joyce, Grant Lawless, William H.Crown, and Vincent Willey.
"Benefit design and specialty drug use." Health Affairs. 2006. 25(5): 1319-1331.

Goldman, Dana P., Geoffrey F. Joyce, Jose J. Escarce, Jennifer E. Pace, Matthew D. Solomon,
Marianne Laouri, Pamela B. Landsman, and Steven M. Teutsch. "Pharmacy benefits and the use
of drugs by the chronically ill." Journal of the American Medical Association. 2004. 291: 2344-
2350.

Goldman, Dana P., Geoffrey F. Joyce; and Yuhui Zheng. "Prescription drug cost sharing:
Associations with medication and medical utilization and spending and health." The Journal of
the American Medical Association. 2007. 298(1): 61-69.

Grossman, Michael. "On the concept of health capital and the demand for health.: The Journal of
Political Economy. 1972. 80(2): 223-55.









Gruber, Jonathan. The Role of Consumer Copayments for Health Care: Lessons from the RAND
Health Insurance Experiment and Beyond. The Kaiser Family Foundation. 2006

Haren, Melinda C., Kirk McConnell, and Arthur F. Shinn. "Increased patient cost-sharing, weak
US economy, and poor health habits: Implications for employers and insurers." American Health
& Drug Benefits. 2009. 2(3): 134-41.

Harman, J.S., S.H. Scholle, R.E. Mardon, A.S. Bierman, S.C. Haffer, J.H. Ng, S. Shih, and L.G.
Pawlson. "Association of health plan's HEDIS performance with outcomes of enrollees with
diabetes". Medical Care. 2010. 48(3): 217-223.

Hausman, J.A. "Specification Tests in Econometrics", Econometrica, 1978.46 (6):1251-71.

Helms, Jay, Joseph Newhouse, and Charles Phelps. "Copayments and demand for medical care:
The California Medicaid experience." Bell Journal of Economics. 1978. 9: 192-208.

Hibbard, J, J. Greene, and M. Tusler. "Does enrollment in a CDHP stimulate cost-effective
utilization?" Medical Care Research and Review. 2008. 65(4): 437-49.

Hildebrand, D.K., J.D. Laing, and H. Rosenthal. 1977. Prediction Analysis of Cross
Classifications. New York: John Wiley and Sons, Inc.

Hsu, J., M. Price, R. Brand, G.T. Ray, B. Fireman, et al. "Cost-sharing for emergency care and
unfavorable clinical events: Findings from the safety and financial ramifications of ED
copayments study." Health Services Research. 2006. 41: 1801-20.

Hudman, Julie, and Molly O'Malley. Health Insurance Premiums and Cost-h/,i ing Findings
from the Research on Low-Income Populations. Kaiser Family Foundation. 2003.

Johnson, Richard E., Michael J., Goodman, Mark C. Hombrook, and Michael B. Eldredge. "The
impact of increasing patient prescription drug cost sharing on therapeutic classes of drugs
received and on the health status of elderly HMO members." Health Services Research. 1997.
32(1): 103-122.

Kaiser Family Foundation. Health Coverage for Low-Income Americans: An Evidence-Based
Approach to Public Policy, 1. Kaiser Family Foundation. 2007.

Kaiser Family Foundation. Medicare at a Glance. Kaiser Family Foundation. 2008, 11.

Kaiser Family Foundation. The Uninsured: A Primer. Key Facts about Americans n init
Health Insurance, 10. Kaiser Family Foundation. 2008.

Karter, A.J, M.R. Stevens, W.H. Herman, S. Ettner, D.G. Marrero, et al. "Out-of-pocket costs
and diabetes preventive services: The Translating Research into Action for Diabetes (TRIAD)
study." Diabetes Care. 2003. 26: 2294-99.









Landsman, P.B., W. Yu, X. Liu, S.M. Teutsch, and M.L. Berger. "Impact of 3-tier pharmacy
penefit design and increased consumer cost-sharing on drug utilization." American Journal of
Managed Care. 2005. 11(10): 621-28.

Link, C.R., S.H. Long, and R.F. Settle. "Cost sharing, supplementary insurance, and health
services utilization among the Medicare elderly." Health Care Financing Review. 1980. 2(2): 25-
31.

Lowe, Bernd, Levke Willand, Wolfgang Eich, Stephan Zipfel, Anthony D. Ho, Wolfgang
Herzog, and Christoph Fiehn. "Psychiatric Comorbidity and Work Disability in Patients with
Inflammatory Rheumatic Diseases." Psychosomatic Medicine. 2004. 66: 395-402.

Manning, W.G. "The logged dependent variable, heteroscedasticity, and the retransformation
problem." Journal of Health Economics. 1998. 17(3): 283-295.

Manning, Willard G., Joseph P. Newhouse; Naihua Duan; Emmett B. Keeler; and Arleen
Leibowitz. "Health insurance and the demand for medical care: Evidence from a randomized
experiment." The American Economic Review. 1987. 77(3): 251-277.

McCall, N., T. Rice, J. Boismier, et al. "Private health insurance and medical mare utilization:
Evidence from the Medicare population." Inquiry. 1991. 28(3): 276-87.

McCauley, J., M.W. Jenckes, C. Saeger-Procter, M.J. McGuire, D.E. Ford, and A.M. Warwick.
"Frequent fliers: Managed care patients with multiple emergency room visits". Abstract,
Association for Health Services Research Annual Meeting. 1998. 39-40.

Motheral, Brenda R., and Rochelle Henderson. "The effect of a copay increase on
pharmaceutical utilization, expenditures, and treatment continuation." American Journal of
Managed Care. 1999. 5: 1383-1394.

Nyman, John A. "Is 'Moral Hazard' inefficient? The policy implications of a new theory."
Health Affairs. 2004. 23(5): 194-199.

Parente, S.T., R. Feldman, and J.B. Christianson. "Evaluation of the effect of a consumer-driven
health plan on medical care expenditures and utilization." Health Services Research. 2004. 39:
1189-210.

Pauly, Mark V. "The Economics of Moral Hazard: Comment." The American Economic Review.
1968. 58(3): 531-537.

Pearte, C.A., C.D. Furberg, E.S. O'Meara, et al. "Characteristics and baseline clinical predictors
of future fatal versus nonfatal coronary heart disease events in older adults: The Cardiovascular
Health Study:". Circulation. 2006. 113(18): 2177-85.

Remler, Dahlia K., and Jessica Greene. "Cost-Sharing: A Blunt Instrument." Annual Review of
Public Health. 2009. 30: 293-311.










Remler, Dahlia K., and Adam J. Atherly. "Health status and heterogeneity of cost-sharing
responsiveness: How do sick people respond to cost-sharing?" Health Economics. 2003. 12: 269-
280.

Rice, Thomas, and Kathleen R. Morrison. "Patient cost sharing for medical services: A review of
the literature and implications for health care reform." Medical Care Research and Review.
1994. 51: 235.

Roddy, P.C., J. Wallen, and S.M. Meyers. "Cost sharing and use of health services. The United
Mine Workers of America Health Plan." Medical Care. 1986. 24:873-76.

Rukavina, Mark. The Financial Burden of Health Care. Boston: The Access Project. 2009.

Scheffler, R.M. :United Mine Workers' Health Plan: An analysis of the cost-sharing program."
Medical Care. 1984. 22(3): 247-54.

Scitovsky, A.A., and N.M. Snyder. "Effect of coinsurance on use of physician services." Social
Security Bulletin 35. 1972. (6): 3-19.

Selby, J.V., B.H. Fireman, and B.E., Swain. "Effect of a copayment on use of the emergency
department in a health maintenance organization." New England Journal of Medicine. 1996. 334:
635-41.

Selden, Thomas M., Genevieve M. Kenney, Matthew S. Pantell, and Joel Ruhter. "Cost sharing
in Medicaid and CHIP: How does it affect out-of-pocket spending?" Health Affairs. 2009.:28(4):
w607-w619.

Shi, Leiyu, and Douglas A. Singh. 2005. Essentials of the US Health Care System. Sudbury,
MA: Jones and Barlett.

Simon, G.E., L. Grothaus, M.L., Durham, M. VonKorff, and C. Pabiniak C. "Impact of visit
copayments on outpatient mental health utilization by members of a health maintenance
organization." American Journal of Psychiatry. 1996. 153: 331-38.

Simon, Herbert. "A behavioral model of rational choice." In Models of Man, Social and
Rational. Mathematical Essays on Rational Human Behavior in a Social Setting. New York:
John Wiley. 1957.

Smith, D.G. "The effects of copayments and generic substitution on the use and costs of
prescription drugs." Inquiry. 1993. 30: 189-198.

Smith, D.G., and D.M. Kirking. "Impact of consumer fees on drug utilisation." Pharmaco-
Economics. 1992. 2(4): 335-42.









Solanki, G., and H.H. Schauffler. "Cost-sharing and the utilization of clinical preventive
services." American Journal of Prevention Medicine. 1999. 17: 127-33.

Stanton, M.W., and M.K. Rutherford. "The high concentration of U.S. health care expenditures."
Rockville, MD: Agency for Healthcare Research and Quality. 2005. Research in Action Issue 19.
AHRQ Pub. No. 06-0060.

StataCorp. Stata Statistical Software: Release 10. College Station, TX: StataCorp LP. 2007.

Stuart, Bruce, and Christopher Zacker. "Who bears the burden of Medicaid drug copayment
policies?" Health Affairs, 1999. 18(2): 201-212.

Tamblyn, Robyn, Rejean Laprise, James A. Hanley, Michael Abrahamowicz, Susan Scott,
Nancy Mayo, Jerry Hurley, Roland Grad, Eric Latimer, Robert Perreault, Peter McLeod, Allen
Huang, Pierre Larochelle, and Louise Mallet. "Adverse events associated with prescription drug
cost-sharing among poor and elderly persons." Journal of the American Medical Association.
2001. 285: 421-29.

Ungar, Wendy J., Anita Kozyrskyj, Michael Paterson, and Fida Ahmad. "Effect of cost-sharing
on use of asthma medication in children." Archives of Pediatrics & Adolescent Medicine. 2008.
162(2): 104-10.

Ware, J.E., and C.D. Sherbourne. "The MOS 36-item short-form health survey (SF-36). I.
Conceptual framework and item selection." Medical Care. 1992. 30(6): 473-83.

Wharam, J.F., B.E. Landon, A.A. Galbraith, K.P. Kleinman, S.B. Soumerai, and D. Ross-
Degnan. "Emergency department use and subsequent hospitalizations among members of a high-
deductible health plan." Journal of the American Medical Association. 2007. 297: 1093-102.

Wooldridge, J.M. Introductory Econometrics: A Modern Approach, Third Edition. Mason, OH:
Thomson South-Western. 2005.

Zeber, J.E., K.L. Grazier, M. Valenstein, F.C. Blow, and P.M. Lantz. "Effect of a medication
copayment increase in veterans with schizophrenia." American Journal of Managed Care. 2007.
13: 335-46.

Zuvekas, S.H., and J.W. Cohen. "A guide to comparing health care expenditures in the 1996
MEPS to the 1987 NMES." Inquiry. 2002. 39(1): 76-86.









BIOGRAPHICAL SKETCH

Haichang Xin was born in Beijing, China. His parents always encouraged him to pursue as

much education as possible, especially in medicine and the health professions to treat diseases

and save lives. His undergraduate study in the field of medicine and public health at Shanghai

Medical University laid a foundation for his future career. During this period, he realized that

each physician can treat only a limited number of patients in a certain area, while an effective

health policy would benefit thousands of people in a community and even a society. Bearing this

philosophy in mind, he later acquired a master's degree in health management and social

medicine from Capital University of Medical Sciences, Beijing. In 2006, he came to the United

States to pursue a PhD degree in health services research at the University of Florida. At UF, he

received systematic training in the area of health policy and system, health care utilization and

quality, health outcome, health economics, and health insurance. He focused on health services

research methods and applied them in this research area. He graduated in summer 2010.





PAGE 1

1 DOES HIGH COST SHARING IN PHYSICIAN CARE REDUCE HEALTH C ARE UTILIZATION AND EXPENDITURE S DIFFERENTLY FOR PEOP LE WITH SEVERE DISEASE FROM THOSE WITHOUT? By HAICHANG XIN 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 2010

PAGE 2

2 2010 H aichang X in

PAGE 3

3 To my parents

PAGE 4

4 ACKNOWLEDGMENTS I feel fortunate to have had the opportunity t o study in the PhD program in Health Services Research where I was granted help and support from numerous people. I would like to express my gratitude to my supervisory committee chair, Dr. Jeffrey S. Harman for his invaluable advice and support. His gui dance has led me to set up the framework and structure for this study and resolve technical challenges throughout my research process I give my sincere thanks to Dr. R. Paul Duncan During a difficult time, his encouragement boosted my confidence and help ed me through that critical moment. I also would like to offer my sincere thanks to Dr. Niccie L. McKay who has been a constant source of invaluable advice. Her serious and careful style and attention to detail ha ve influenced me, and I have applied these to my studies. As a committee member, Dr. Jing Cheng helped me refine my methodology and provided psychological support for me. I would like to thank her for her efforts I am also indebted to Dr s Ning Li, Xiaohui Xu, and Chunrong Ai. A ll have provided u seful additional support for this study. My sincere thanks go to Dr. Frederick Rohde who explain ed technical problems in detail to me and helped me through some analytic difficulties. During my study, Nancy Hamilton and Dustin Heinen also provided valuabl e editing help to me. I have improved my writing skills after read ing their comments. It would be a long list of others I should acknowledge During the process they indirectly support ed me academic ally material ly or psychological ly I would like to th ank my parents and my brother. Their eternal love for me is my inexhaustible resource, and the strength of my drive for study and to establish my future career is greatly attributed to them

PAGE 5

5 TABLE OF CONTENTS page ACKNOWLEDGMENTS ................................ ................................ ................................ ............... 4 LIST OF TABLES ................................ ................................ ................................ ........................... 8 LIST OF FIGURES ................................ ................................ ................................ ......................... 9 LIST OF ABBREVIATIONS ................................ ................................ ................................ ........ 11 ABSTRACT ................................ ................................ ................................ ................................ ... 12 C H A P T E R 1 INTRODUCTION ................................ ................................ ................................ .................. 14 Overview ................................ ................................ ................................ ................................ 14 Study Objectives ................................ ................................ ................................ ..................... 16 2 BACKGROUND AND SIGNIFICANCE ................................ ................................ .............. 17 Background ................................ ................................ ................................ ............................. 1 7 The Function of Health Insurance ................................ ................................ ................... 18 Moral Hazard ................................ ................................ ................................ ................... 19 Cost Sharing ................................ ................................ ................................ .................... 20 Physician Care ................................ ................................ ................................ ................. 22 Price Elasticity of Demand ................................ ................................ .............................. 23 Significance ................................ ................................ ................................ ............................ 24 Impact of Cost sharing on Health Status ................................ ................................ ......... 25 Cost sharing and Out of Pocket Cost Burden ................................ ................................ 26 High Cost sharing and Total Expenditures ................................ ................................ ..... 27 Literature Review ................................ ................................ ................................ ................... 30 Price Elasticity of Demand for Medical Care ................................ ................................ .. 30 Impact of Cost sharing for Different Types of Medical Care ................................ ......... 32 All service type cost sharing ................................ ................................ .................... 32 Physician care cost sharing ................................ ................................ ...................... 33 Prevention cost sharing ................................ ................................ ............................ 34 Inpatient care cost sharing ................................ ................................ ....................... 34 Pharmacy cost sharing ................................ ................................ ............................. 35 Emergency room visit cost shari ng ................................ ................................ .......... 36 Effect of Cost sharing in Physician Care by Health Status ................................ ............. 37 Absence of differential impact ................................ ................................ ................. 37 Presence of differential impact ................................ ................................ ................. 39 Related findings ................................ ................................ ................................ ........ 41 Gaps and Limitations of Prior Studies ................................ ................................ ............. 42

PAGE 6

6 3 CONCEPTUAL FRAMEWORK ................................ ................................ ........................... 45 ................................ ................................ ................................ .................. 45 ................................ ................................ ................................ 46 Demand Curve ................................ ................................ ................................ ........................ 48 Price Elasticity ................................ ................................ ................................ ........................ 48 Hypotheses ................................ ................................ ................................ .............................. 53 4 METHODS ................................ ................................ ................................ ............................. 58 Data Description ................................ ................................ ................................ ..................... 58 Measures and Op erationalization ................................ ................................ ........................... 60 Measures of Outcome Variables ................................ ................................ ..................... 60 Measures of Explanatory Variables ................................ ................................ ................. 62 Severity level ................................ ................................ ................................ ............ 62 Cost sharing level ................................ ................................ ................................ ..... 63 Measures of control variables ................................ ................................ ................... 65 Study Design ................................ ................................ ................................ ........................... 67 Statistical Analysis ................................ ................................ ................................ .................. 68 5 RESULTS ................................ ................................ ................................ ............................... 78 Overview ................................ ................................ ................................ ................................ 78 Description of the Sample ................................ ................................ ................................ ...... 78 Variable Operationalization ................................ ................................ ................................ .... 79 Missing Cost sharing Values Imputation ................................ ................................ ........ 79 Test for the Multicollinearity between Disease Severity and Priority Conditions .......... 80 Test for Potential Endogeneity and IV Validity ................................ ................................ ..... 80 The Hausman Test ................................ ................................ ................................ ........... 80 Test for IV Relevance and Exogeneity ................................ ................................ ............ 81 Use IV to Get Predicted Cost sharing Values ................................ ................................ 81 The Multi variate Analysis Results ................................ ................................ .......................... 82 The Health Care Utilization Results ................................ ................................ ................ 82 Physician care utilization ................................ ................................ ......................... 82 Primary care physician utilization ................................ ................................ ............ 83 ER care utilization ................................ ................................ ................................ .... 84 Inpatient care utilization ................................ ................................ ........................... 86 The Health Care Expenditure Results ................................ ................................ ............. 86 Physician care e xpenditure ................................ ................................ ....................... 86 ER visit expenditure ................................ ................................ ................................ 89 Inpatient care expenditure ................................ ................................ ........................ 91 Total care expenditure ................................ ................................ .............................. 94 6 DISCUSSION ................................ ................................ ................................ ....................... 135 Summary of Basic Findings ................................ ................................ ................................ .. 135 Physician Care Price Elasticity ................................ ................................ ............................. 137

PAGE 7

7 Other Sensitivity Analyses ................................ ................................ ................................ ... 138 7 CONCLUSIONS ................................ ................................ ................................ .................. 144 Summary and Interpretation ................................ ................................ ................................ 144 Negative Consequence of High Cost sharing Policies ................................ ......................... 144 Policy Implications ................................ ................................ ................................ ............... 145 Limitations ................................ ................................ ................................ ............................ 148 LIST OF REFERENCES ................................ ................................ ................................ ............. 150 BIOGRAPHICAL SKETCH ................................ ................................ ................................ ....... 156

PAGE 8

8 LIS T OF TABLES Table page 4 1 Summary of outcome measures ................................ ................................ ......................... 76 4 2 Explanatory variables ................................ ................................ ................................ ......... 77 5 1 Description of the study sample (N = 13,020) ................................ ................................ ... 98 5 2 Study sample characteristics (N = 13,020) ................................ ................................ ........ 99 5 3 Association between disease severity and priority conditions ( P < 0.01) ......................... 101 5 4 T he first stage of IV 2SLS results with cost sharing as dependent variable ................... 102 5 5 The association between IV and the error terms in the structural equation of health care utilization and expenditure ................................ ................................ ....................... 103 5 6 Cost sharing value descr iption by insurance types ................................ .......................... 104 5 7 Negative binomial regression estimates for physician visits ................................ ........... 105 5 8 Negative binomial regressi on estimates for primary care physician visits ...................... 107 5 9 Negative binomial regression estimates for ER visits ................................ ..................... 109 5 10 Negative bi nomial regression estimates for hospital admissions ................................ .... 111 5 11 Logit regression predicting probability of having any physician care expenditures ....... 113 5 12 Log transformed OLS regression estimates for physician care expenditures .................. 115 5 13 Logit regression predicting probability of having any ER visits expenditure ................. 117 5 14 Log transformed OLS regression estimates for ER care expenditures ............................ 119 5 15 Logit regression predicting probability of having any inpatient care expenditure .......... 121 5 16 Log transformed OLS regression estimates for inpatient care expenditures ................... 123 5 17 Logit regr ession predicting probability of having any medical care expenditure ........... 125 5 18 Log transformed OLS regression estimates for total medical care expenditures ............ 127 6 1 Differential impact summary by service types ................................ ................................ 140 6 2 Significant differential impact summary by service types ................................ ............... 140

PAGE 9

9 LIST OF FI GURES Figure page 2 1 Moral hazard and social welfare ................................ ................................ ........................ 44 2 2 Income transfer helps distinguish between efficient and i nefficient moral hazard ........... 44 3 1 Demand curve for severely ill and reference group ................................ ........................... 55 3 2 Demand curve for severely ill and refere nce group with switched X and Y axis ............. 55 3 3 Demand curve for severely ill and reference group with switched X and Y axis in MD visits ................................ ................................ ................................ ............................ 56 3 4 Demand curve for severely ill and reference group with switched X and Y axis in ER or inpatient care ................................ ................................ ................................ .................. 56 3 5 Demand curve for severely ill and reference group with switched X and Y axis in total care ................................ ................................ ................................ ............................. 57 5 1 P P plot for log transformed OLS regression on physician care expenditures ............... 129 5 2 Q Q plot for log transformed OLS regression on physician care expenditures ............... 129 5 3 Residual fitted plot for log transformed OLS regression on physician care expenditures ................................ ................................ ................................ ..................... 130 5 4 P P plot for log transformed OLS regression on ER care expenditures .......................... 130 5 5 Q Q plot for log transformed OLS regression on ER care expenditures ......................... 131 5 6 Residual fitted plot for log transformed OLS regression on ER care expenditures ........ 131 5 7 P P plot for log transformed OLS regression on inpatient care expenditures ................ 132 5 8 Q Q plot for log transformed OLS regression on inpatient care expenditures ................ 132 5 9 Residual fitted plot for log transformed OLS regression on inpatient care expenditures ................................ ................................ ................................ ..................... 13 3 5 10 P P plot for log transformed OLS regression on total medical care expenditures ........... 133 5 11 Q Q plot for log transformed OLS regression on total medical care expenditures ......... 134 5 12 Residual fitted plot for log transformed OLS regression on total medical care expenditures ................................ ................................ ................................ ..................... 134

PAGE 10

10 6 1 The differential impact by health status in probability of having any physician care expenditures ................................ ................................ ................................ ..................... 141 6 2 The differential impact by health status in ER care visits ................................ ............... 141 6 3 The differential impact by health status in probability of having any total medical care expenditures ................................ ................................ ................................ ............. 142 6 4 The differential impact by health status in integrated physician care expenditures ........ 142 6 5 The differential impact by heal th status in integrated total medical care expenditures ... 143

PAGE 11

11 LIST OF ABBREVIATIONS MEPS HC Medical expenditure panel survey household component ADL A ctivity of daily living IADL I nstrumental activity of daily living CDHP C onsumer directed health plan AMI A cute myocardial infarction OLS Ordinary least squares HIE H ealth insurance experiment FFS F ee for service HMO H ealth maintenance organization MCBS Medicare Current Beneficiary Survey MPC M edical provider component DRG Diagnosis related group NHIS National Health Interview Survey PSU P rimary sampling unit AHRQ Agency for Healthcare Research and Quality SBD S eparately billing doctor PCS P hysical component score MCS M ental component score SF 36 S hort form 36 i tem BMI B ody mass index SCHIP State Children's Health Insurance Program IV I nstrumental variable 2SLS T wo stage least square s H L Hosmer Lemeshow test AIC

PAGE 12

12 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 DOES HIGH COST SHARING IN PHYSICIAN CARE RE DUCE HEALTH CARE UTILIZATION AND EXPE NDITURE S DIFFERENTLY FOR PEOP LE WITH SEVERE DISEAS E FROM THOSE WITHOUT? By H aichang X in August 2010 Chair: Jeffrey S. Harman Major: Health Services Research This study examines whether high cost sharing in physician care reduces health care utilization and expenditures differently for people with sev ere disease from those without By severely ill individuals from exacerbating their poor health status and financial burden, as well as save cost s for society at the aggregate level. The study adopted a cross sectional study design from the 2007 Medical Expenditure Panel Survey data. In STATA w eights and variance were adjusted to account for the complex multi stage, unequal probability, and cluster sampling s urvey designs. T he potential endogen eity problem between cost sharing and health care utilization or expenditures was addressed by a valid instrumental variable. Negative binomial regressions and two part models were employed to analyze the utilization and expenditure data using the bootstrap technique to incorporate split results in two part models. In response to increased physician care cost sharing the study revealed a differential impact in the probability of physician care and total care expenditur e models and the emergency

PAGE 13

13 room care utilization model, and opposite differential impact in the overall physician care and total care expenditure models. The severely ill were significantly associated with poor income. T hus, although the severely ill, in r esponse to high cost sharing, had a desire to reduce essential and necessary medical care to a lesser extent than the general health population, they actually reduced more because high cost sharing policies, augmented by their financial difficult ies, grea tly distort ed their desires and voluntary behaviors. In response to high cost sharing pressure, severely ill individuals appear to have experienced both substantial physician care reduction and emergency room care increase, and they were in worse clinica l conditions. Therefore, current high cost sharing policies should be replaced with low cost sharing polic ies for individuals with severe illnesses to reflect their situations This study highlight s the necessity and importance of value based insurance d esign in terms of differentiation of and specific ation for its target population Furthermore, this study contribute s to the current debate on health care reform S pecialized plans for subpopulations instead of a universal plan may want to be considered to complement existing public and private designs

PAGE 14

14 CHAPTER 1 INTRODUCTION Overview People with severe illness carry a heavy disease burden; beside their overall poor health status and functional status, either physically or mentally, they have a heavy out of pocket financial burden ( Kaiser Family Foundation 2007) Their disease burdens are directly reflected in high levels of health care utilization and spending ( Conwell et al. 2005 ). In order to cut down on outsized medical expenditures, current insuranc e policies rely heavily on high cost sharing or other price related approaches. In this way, cost sharing aims to reduce social welfare waste by insured individuals due to moral hazard Unfortunately, while cost sharing can reduce utilization of unnecessar y services current high cost sharing policies might produce a new problem by reducing utilization of effective and necessary care Thus, a legitimate question arises: I s high cost sharing necessary to reduce health care utilization and expenditure for tho se with severe health conditions? This study focus es on cost sharing in physician care, among which primary care plays a central role in a health care delivery system. It is the first contact a patient makes with the health care system, and is ideally the means by which health service delivery is coordinated in a et al. 2005). Beside the frontline ro le of primary care, as a whole physician care is an upstream service that m emergency room (ER) visits, hospitalization s, and worse health outcomes ( McCauley et al. 1998, Chandra et al. 2007). As a quantitative measure, price elasticity of demand can also assist us to evaluate perception about whether high cost sharing in physician care is appropriate for severely ill

PAGE 15

15 patients. A related concept is cross price elasticity of demand Clarification of the relationship between service types f or severely ill patients can help us better estimate cost saving at a societal level given the stringent budget for medical care, which also contributes to our discussion of whether current high cost sharing policies increase net expenditures. H igh cost s haring in physician care may result in worse health outcomes and heavier financial burdens for individuals and higher aggregate cost because it reduces needed health care utilization and leads to a greater need for acute and inpatient care for those with severe health conditions In order to examine this research question, the study needs to compar e the utilization and expenditure reduction magnitude due to high cost sharing for the severely ill group and for a less severely ill group This study focuses o n the impact of cost sharing on utilization of physician care. Because physician care is likely responsive to price, decreased utilization of this type of care due to cost sharing might lead to subsequent ER visits or hospitalization s by individuals with m ore severe and advanced illnesses. This study examine s the effect of cost sharing for physician services on ER and inpatient care as well s pecifically if it has an offset effect, which can help us better understand the potential differential impact on ph ysician care and overall health care utilization. The research question is examined using a cross sectional study design with a nationally representative sample using the 2007 m edical expenditure panel survey household component (MEPS HC) data for analys is. Subjects are included who meet certain age and insurance coverage criteria. C aptured by the interaction of dummy variables severity and cost sharing the differential impact s are examined in both the upstream physician care and the downstream ER and in patient care, in both utilization and expenditure measures and their total expenditures. Weights and variance are adjusted to account for the complex multi stage, unequal probability,

PAGE 16

16 and cluster sampling survey designs in STATA. The potential endogen eity problem between cost sharing and health care utilization or expenditure is tested and confirmed An instrumental variable is then introduced to address the endogeneity. In multivariate analysis, negative binomial regressions and two part models are employe d to analyze the utilization and expenditure data. In addition, the study estimate s the price elasticity and cross price elasticity of demand of physician care. Sensitivity analysis is Study Objectives The m ain purpose of this study is to examine the potential differential impact of high cost sharing in physician care on health care utilization and expenditure for those with and without severe health conditions in the United States. The specific aims of the s tudy are Aim 1: To examine the differential impact of high cost sharing in physician care on utilization and expenditures between severely ill people and healthier people. Aim 2: To examine the cross price elasticity of physician care on ER and inpatient care utilization and expenditure s between severely ill people and healthier people. Aim 3: To compare the differential impact of cost sharing in physician care on expenditure s for all services between severely ill people and healthier people.

PAGE 17

17 CHAPTER 2 BACKGROUND AND SIGNI FICANCE Background Based on 2002 national health care spending data, a statistical brief reports that 5% of the population account ed for 49 % (almost half) of total health care expenses, while the bottom half of the population only spent 3% ( Conwell et al. 2005 ). Th at top 5 % spent, on average, more than 17 times as much per person as those in the healthier half (Stanton, 2006). E stimated national health care spending in 2002 was $1.6 trillion ( Cowan et al. 2004 ), which means the top 5 % incurred about $800 billion. A close look at the statistical brief reveals the y had much worse health status. This 5% of the population were 11 times (45 % vs. 4 % ) more likely to be in fair or poor physical health as people in the healthier half of the dist ribution and 7 times (21 % vs. 3 % ) as likely to be in fair or poor mental health as compared to people in the lower half ( Conwell et al. 2005 ). Meanwhile, in terms of limitation s in activit ies of daily living (ADL), or instrumental activit ies of daily livi ng (IADL), the top 5 % of the group wa s 26 times ( 26 % vs. 1 % ) more likely to need help than those in the healthier half of the population ( Conwell et al. 2005 ) This indicates that this 5% of the population, although a small group, have severe disease s and carry catastrophic disease burden s accounted for most of the growth in total health care spending between 1987 and 2000, with the top five medical conditions (heart disease, pulmonary disorders, mental disorders, cancer, and trauma) accounting for 31 % and the 15 most expensive health conditions accounting for 44 % of total health care expenses (Stanton, 2006). The se expensive health conditions can serve as a rough measure for disease severity although there are some discrepanc ies that further warn that severely ill individuals require much more attention and protection.

PAGE 18

18 Obviously, this sub set of the population has severe physical and mental problems. Poor health and functi onal status can push them in the direction of life threatening conditions. The ir huge medical expenditure s w ere paid either by patients themselves or a third party. Due to coverage limitation by insurance policies, catastrophic disease can easily impose he avy out of pocket financial burdens on them; it is especially disastrous for those who are uninsured. According to the Commonwealth Fund 2007 Biennial Health Insurance Surveys, adults who were uninsured or underinsured for any time during a year had more t han twice the rate of medical bill problems and debt than those who were insured all year long (61 % vs. 26 % ; Doty, et al. 2008). Even for the insured, insurance policies vary widely in terms of plan type, service type, and enrollee characteristics. Many p eople may still face wide coverage gaps and have to pay out of pocket for certain services or pay much of the cost for services already covered by an existing insurance plan ( Doty, et al. 2008 ). Catastrophic disease can thus threaten severely ill people t o the point of impoverishment. In fact, the above report ( Conwell et al. 2005 ) revealed that persons in the top 5 % of the expenditure distribution were 6.8 times ( 34 % vs. 5 % ) more likely to have out of pocket medical expenses exceeding 10 % of the family i ncome, and 6 times ( 18 % vs. 3 % ) more likely to have out of pocket medical expenses exceeding 20 % of the family income than those in the healthier half of the population. In order to curb medical expenditures, a common insurance practice is to impose high cost sharing or use other price related approaches (Remler et al. 2009). Before discussing the cost sharing role this study will briefly introduce the health insurance function. The Function of Health Insurance Health insurance is basically modeled afte r traditional business insurance, and share s some of its functions. By spending a fixed amount on premiums, people can be protected against uncertain higher costs of medical care for undesirable health events T hus health insurance

PAGE 19

19 improve s the affordabili ty of health care to the individual (Remler et al. 2009). Moreover, health insurance provides financial protection, preserv resources or income in case of catastrophic disease and preventing them from becoming impoverished (Remler et al. 2009 ). In addition, health insurance has two unique functions compared to business insurance: insulation s from most fiscal consequence s and also protects their health. Relative to individuals with out insurance, health insurance affords people with equ itable rights of access to health care benefits (Daniels, 2001). Numerous empirical studies have consistently shown that health insurance makes a difference in whether and when people get necessary med ical care, where they get their care, and ultimately, the level of health people maint ain. Health insurance increases utilization and improves health outcomes (Kaiser Family Foundation, 2008; Freeman, 2008). However, health insurance is also accompanied by moral hazard. Moral H azard Moral hazard means that people consume more health services when insured than they would if uninsured. Pauly (1968) provided the original analysis of the welfare effect due to moral hazard. Traditionally, it is believed that mor al hazard wastes social resources. This argument is illustrated in Figure 2 1. If an individual is uninsured, no one would pay a price higher than market cost, so in the demand curve the marginal cost equals market price. The average cost of medical care, at intersection price point A, corresponds to demanded quantity of medical service (Q u ). At this point, the benefit is equal to social expenditure, E = P*Q u and there is no social welfare loss. On top of the Q u amount, suppose an individual is fully insur ed and thus faces an out of pocket expense of zero, he may consume more services until reaching the maximum amount Q i where the demand curve intersects the zero price line. The insured individual actually

PAGE 20

20 enjoys an extra benefit of area AQ u Q i but society has to pay a s shown in the corresponding area of AQ u Q i B, so there is a social welfare loss of area AQ i B. (Nyman, 2004). The basis for his argument is that an insurer transferring income to an ill person would allow us to distinguish between efficient and inefficient moral hazard. If the patient take s advantage of the income transfer and spend s it all on medical care due to moral hazard, the income transfer would serve would shift outward and parallel to the original demand curve ( Figure 2 2) Then initial point A will move to a new point A 1 and point Q i to Q i1 which indicates that both efficient an d inefficient moral hazard will increase. Although the inefficient moral hazard is still harmful, the increased efficient moral hazard is desirable and beneficial, which is the area of AQ u Q u1 A 1 s insurance polic ies should be designed to match the appropriate level of cost sharing waste brought on by inefficiency Cost S haring In order to curb escalating health care expenditure s and reduce the resou rce waste produced by the moral hazard embedded in health insurance, a cost sharing mechanism is devised to cut down the use of unnecessary service and bring the health care consumed closer to what it would be without insurance, thus helping to save cost. Unfortunately, current insurance company policy relies heavily on the high cost sharing strategy, and excessively deviates from what it was originally intended to do. Therefore, while resolving the moral hazard problem, the current high cost sharing policy m ay produce a new problem: cutting effective and necessary care for severely ill individuals.

PAGE 21

21 Basically there are three widely used forms of cost sharing: copayments, coinsurance, and deductibles. For patients with copayments they pay a fixed fee amount for each medical service sought or product purchased. Similarly, coinsurance requires patients to pay a fixed percentage of the cost for each care episode. The third, deductibles refer to the amount one must pay out of pocket before insurance coverage be gins In addition, other commonly used variant s include caps and an out of pocket maximum. While cost sharing is designed to reduce the unnecessary care induced by moral hazard to blunt instrument reducing medical care of both low value and high value, both cost effective and ineffective, and both needed and unneeded (Manning, 1987; Haren, 2009). There may be many reasons for the reduction of effective care. An important one is t hat cost sharing tends to push the consumers to look for cheaper yet similar ly effective medical care or products, which assumes that they have complete information to compare across providers or have trade off s between price and effectiveness of medical c are. Unfortunately, the health care market is dominated by asymmetric information. Ordinary consumers do not have the necessary tools, education, knowledge, time or money to acquire full information about provider performance, care efficacy, and the conse quence s of forgoing care Furthermore, in sophisticated cost sharing design s like consumer directed health plans (CDHPs), uncertainty will contribute 1957) and thus irrational decisions (Haren, 2009). This explanatio n is based on the assumption that people have the ability to pay for medical care. Meanwhile, it may also be possible that patients sometimes have to forgo or delay their care, even needed care due to financial constraints. Even if they are aware of healt greatly influenced by financial considerations (Rukavina, 2009), and they have to trade off

PAGE 22

22 between two difficult options; incurring higher cost burden s and getting treatment or choosing not to consume h ealth care service with potential ly worse health outcome s Increased cost sharing serves to undermine insurance functions by weakening an affordability and insulation. Moreover, as mentioned above, cost sharing also threatens the financial prot ection function of health insurance, resulting in ina bility to pay for needed care. Willingness to pay gives way to ability to pay, so access disparity violates the social equality goal that everyone has equal access to health care. As Remler et al. (2009) pointed out, b y its very nature, cost sharing partially undoes insurance, and can undermine exactly what people with insurance hope to protect. As a result, beside the immediate consequence o f reducing needed or unneeded care, cost sharing has further and great er impact s on health care utilization, expenditures and health outcomes. Physician Care It seems that there is no appropriate cost sharing level that applies universally across medical service types. Cost sharing can be implemented i n many service types, including preventive care, physician care, emergency room visit s hospitalization s prescription drug s and so on. In order to evaluate the impact of the current high cost sharing level this study will focus on cost sharing in physic stricter and more specified one, cover ing primary care that reflects preventive care and specialist care including clinical diagnostic and therapeutic services by MDs that do not requi re an overnight stay in a health care institution but or hospital outpatient setting Primary care is the conceptual foundation for physician services and plays a central role in a health care delivery system (Shi e t al. 2005). It is the first contact a patient mak es with the health care system so that health service delivery is coordinated in a et al.

PAGE 23

23 2005). Beside the frontline rol e of primary care, physician care is an upstream service that prevent individuals from avoidable downstream ER visits, hospitalization s, and worse health out comes T hus it plays a key role in maintain ing health status. Although important, physician care may be susceptible to cost sharing and thus forgone because of out of pocket costs, potentially lead ing to subsequent avoidable ER visits or hospitalizations. The reason for the relat ively high price responsiveness of this type of care m ay be that people m ight not have the professional knowledge to judge the importance of physician care. The potential differential impact for the severely ill will highlight the significance of physician care. Forfeiture of this frontline and upstream service for the sick may worsen their health conditions and widen the gap of health status among the population. P rice E lasticity of D emand One role of price elasticity of demand is that it provides a quantitative guid e to the effect of cost sharing on health care utilization. Thus, it further assists us to evaluate pe o s about whether cost sharing in physician care affects severely ill patients differently from peop le without illness or who have less severe illnesses. It measures responsiveness of quantity demanded to changes in price. E d = Q% / P% S pecifically, demand is: Elastic if | E d | > 1 Unit Elastic if | E d | = 1 Inelas tic if | E d | < 1 Here, price refers to cost sharing level. In this study, price elasticity of demand is calculated from a comparison of change in the price and demand quantity. It can be influenced by many factors, like income and preference, and used in many areas, like disease types, service types, and subpopulation types.

PAGE 24

24 A related concept is cross price elasticity of demand, which measures the responsiveness of the demand of one good to a change in the price of another good. Clarification of the relationship between physician care and its downstream ER visits or hospitalization s for severely ill patients will help us better estimate cost saving or net expenditures at a societal level given the stringent budget for medical care, which also contributes to whether current high cost sharing policies for physician care are appropriate for the severely ill. P rice elasticity always has a negative sign, the focus is on elasticity magnitude. While in terms of cross price elasticity, the focus will be transferred to its sign. This study will implicitly examine whether physician care and ER o r inpatient care are complements or substitutes. Positive cross price elasticity indicates the existence of an offset effect, and physician care and ER or inpatient care are substitutes. ER or hospitalization care is much more expensive than office based o r outpatient doctor visit s and at a societal level this will become a great waste of social medical resources, which highlights the importance of physician care. Physician care, if forgone will incurred higher prices either clinical or financial or both A RAND study (Manning et al. 1987) indicates that physician care and hospitalizations we re complements, but the result wa s not statistically significant, and the relationship may refer to the overall study sample but not the sicker subgroup. Physicia n care is the upstream service that ensures an of the health care system Thus, lower utilization of necessary physician care by cost sharing pressure could worsen health status and lead to avoidable and expensive ER visits and hospitalization s that result in greater overall health care costs. Significance Having briefly introduced the background information above, the study will focus on the significance of the research question: I f high cost sharing in physician care is necessary to reduce health service utilization and expenditure for people with severe disea se, as these process

PAGE 25

25 indicators can indicate there is the potential for severe consequence s on the following aspects. Impact of Cost sha ring on Health Status If this study confirms the differential impact by health status, it will reveal the phenomenon that high cost sharing will have a unique impac t on severely ill individuals who tend to resist a reduction of their current health care level one of the important health determinants that directly influence s health outcome s since forfeit ure or delay of needed care will worsen their existing condition. By setting up financial barriers, high cost sharing for this vulnerable group will deter or deprive them from getting access to needed primary care and specialist service s that are essential to maintain their health. This is indirectly evidenced by the RAND findings (Manning et al. 1987) For low income individuals with health conditions, reduced care due to increased cost sharing adversely affected their health. In the RAND study, cost sharing was implemented for all service types, but not limited to physician care only. In addition, the result referred to people who we re both sick and po or, not just sick. Other indirect evidence comes from a study examining the elderly with chronic conditions, which can be taken as a proxy measure for severe conditions (Chandra et al. 2007). They found that increases in physician care and prescription dr ug copayments had little effect on hospital use for an average elderly person, but for chronically ill elderly patients there was a significant offsetting rise in hospital admissions as physician and drug use f ell This is also indirect evidence that incre ased cost sharing level in physician care is harmful to the health status for the severely ill, as subsequent downstream inpatient care is used as a proxy measure for health outcome s although the result cannot be attributed exclusively to copayme nt increa se in physician care.

PAGE 26

26 Cost sharing and Out of Pocket Cost Burden Severely ill individuals usually have a higher out of pocket financial burden than less severely ill individuals. Conwell et al. ( 2005 ) indicated that pe ople in the top 5 % of the expenditur e distribution were 6.8 times ( 34 % vs. 5 % ) more likely to have out of pocket medical expenses exceeding 10 % of the family income and 6 times ( 18 % vs. 3 % ) more likely to have out of pocket medical expenses exceeding 20 % of the family income than those in th e healthier half of the population. In 2003 12% of all non elderly adults had out of pocket costs exceeding 5% of family income, while 19% of those with chronic conditions faced this same level of out of pocket costs ( Kaiser Family Foundation 2007). Roug hly 14% of all U.S. families reported problems paying their medical bills, but among families with a member in fair or poor health, the proportion increased to 25% ( Kaiser Family Foundation 2007). Increased cost sharing in physician care or other medical service will increase financial burden, and put them in a difficult situation. It is reported that among insured adults whose health plans limited the total amount they could spend, 43 % incurred medical bill problems and unpaid debt compared with 27 % of adults who did not have total dollar limits (Doty et al. 2008). Facing this dilemma, patients would spend as much of their resources as possible to maintain health. Rukavina (2009) reports that 24% of Americans buying health insurance still in curred debt for medical bills, which indicated that they had exhausted available resources to maintain necessary medical care when insurance was based on cost sharing Furthermore, t he financial burden from cost sharing will also influence he alth status by intensifying their stress and distress, exacerbat ing their existing poor health status. For example, mood plays a key role in the disease aggravating process in acute myocardial infarction (AMI) ( Bax et al. 2008, Pearte et al. 2006 ). Alth ough many patients can ultimately overcome

PAGE 27

27 these difficulties and get necessary care, their physical and psychological condition may still be indirectly harmed. High Cost sharing and Total Expenditures At a societal level, the severely ill incur tremendo us medical expenditures. High cost sharing is designed to reduce unnecessary service u tilization to save costs, but for severely ill patients it may not achieve this desired goal. Ethically, even if it c ould this should not come at the expense of worseni ng their health outcome s and increasing their financial burden s which will undermine the public health goal to maintain a high level of population health. Empirically, some studies suggested that the use of cost sharing may result in higher overall costs for the low income population o n Medicaid (Helms et al. 1978; Tamblyn et al. 2001). The Helms and colleagues physician care, ambulatory care utilization declined by 8 % but subsequent inpati ent care use increased by 17 % which resulted in 3 % to 8 % higher total Medicaid costs. Low income individuals often suffer from poor health, since their financial constraint s usually deter them from getting necessary care to maintain their health. Among th e Medicaid population, 18% reported fair and poor health status (Julie et al. 2003). Elderly individuals also often serve as a proxy for individuals in poor health. For the Medigap beneficiaries, a study indicated that cost sharing policies would lead to a smaller cost saving by at least 2% to 7% for sick people compared to healthy people ( Remler et al. 2003). Similarly, a recent study suggested that for elderly people, reduced drug use and physician visits due to cost sharing were offset by increased su bsequent health care utilization, especially hospitalization s (Chandra et al. 2007), which is contrary to the RAND findings (Manning et al. 1987) of no offset rising hospitalization s The subsequent medical expenditure may offset or even exceed the saved cost, especially for Medicare, although the offset effect in response to

PAGE 28

28 cost sharing needs further exploration because the cost sharing policy was instituted on physician care and pharmac euticals at the same time. The proportion who reported fair and poo r health in Medicare was even higher than in Medicaid, as much as 28% in 2008 (Kaiser Family Foundation, 2008). Medicare is public ly fund ed so policy makers and taxpayers are concerned about whether Medicare has best used limited public funding to achieve maximum utility. It is reported that for the Medicare insurance trust fund costs will exceed income from 2008 to 2010 which will be exhausted by 2019 (Eppig, Medicare Current Beneficiary Survey ca.2008 ) although there are conflicting data and opinion s No matter what the year may be, t he goal of c ost containment has become imperative. Some other studies examining the pharmaceutical cost sharing policy also provide indirect evidence. One study suggested that for elderly people who use inhaled medicati ons cost sharing may increase net expenditures (Dormuth et al. 2009). By assembling the data of pharmacy and medical claims from 1997 to 2002 from 88 health plans and 25 employers, Goldman et al. (2006) simulated a prescription copayment design by health status. Specifically, relative to the universal $10 copayment policy, high and medium risk people were exempt from cost sharing while the low risk group would have either $10 or $22 copayment. The study estimated that this new design would reduce the numb er of hospitalizations by between 80,000 to 90,000 annually and the number of ER visits by 30,000 to 35,000, resulting in net aggregate savings of more than $1 billion. This study provides indirect evidence that current cost sharing is inefficient. If a hi gh cost sharing policy ends up losing money, whether in Medicare, Medicaid or private plans, it is a considerable waste of social and medical resources at the aggregate level. These results from the low income and elderly population indirectly imply that the cost saving performance of the high cost sharing policy for the severely ill is unsuccessful. If so, for this vulnerable group the

PAGE 29

29 widely adopted current insurance policy design might be problematic not only clinically but also financially. Under th e se circumstance s an evaluation of an insurance policy design for the target population has practical policy implication. If the differential impact in expenditure is confirmed, the results will provide persuasive evidence to inform insurance plans that po licies denying coverage for physician care or primary care will end up reduc ing profit s which was labeled penny wise, pound foolish, and to motivate them to adopt a smarter benefit structure that covers physician care. This will both benefit an health status and save the insurers money In summary, high cost sharing for severely ill people potentially may not only pose a threat increase the out of pocket financial burden, but also waste limited medical resou rces. The present study will provide empirical evidence to evaluate the potential differential effect of high cost sharing for physician care between the severely ill and healthier population. r health care utilization change in response to high cost sharing pressure s this study will allow researchers to evaluate the effect of high cost sharing policies and whether they are necessary to cut down health care utilization and expenditures for the severely ill. A high cost sharing policy naturally will directly reduce health care utilization, regardless of subpopulations. If the differential impact of cost sharing is confirmed, that means severely ill people, in response, may reduce less health car e utilization and expenditure, especially the efficient moral hazard in utilization that is crucial to maintain their health. This phenomenon may further reveal the potential underlying truth that efficient and inefficient moral hazard share differs by sev erity Severely ill people usually have less inefficient moral hazard and more

PAGE 30

30 efficient moral hazard than healthy people. Therefore, th ese vulnerable people should be treated differently Specifically a low cost sharing policy should be designed to refle ct and match their situation. This study will contribute to inform the necessity and importance of insurance policy design in terms of differentiation and specification for its target population, so as to best protect and prevent th at population from exace rbating its poor health status and financial burden, and to save cost s for the society at the aggregate level. Literature Review Price Elasticity of Demand for Medical Care Price elasticity of demand measures the responsiveness of quantity demanded to chan ges in price. This quantitative measure helps us better understand the actual effect of cost sharing Numerous studies have examined price elasticity of demand across service types and so far their findings are inconsistent Using state variation s in coins urance rates from the American Hospital regression, Feldstein (1971) estimated a price elasticity of 0.49 for hospital bed days (Cutler et al. 1999). Based on th e study conducted at Stanford University with a quasi experiment design, Phelps and Newhouse (1972) calculated an elasticity of physician visits of 0.14 with ordinary least square s (OLS) estimation and 0.118 with the Tobit approach (Rice et al. 1994; Cu tler et al. 1999). Using the Tobit estimate of 1960 Survey of Consumer Expenditure, a cross sectional study, Rosett and Huang (1973) estimated a price elasticity of 0.35 to 1.5 for hospitalizations and physician service (Cutler et al. 1999). The RAND h ealth insurance experiment (HIE) reported an overall estimated elasticity of medical service spending of 0.2 (Manning et al. 1987), which is supported by a summary of more than 20 studies, indicating that the total price elasticity of demand for medical services was approximately 0.2 (Cutler et al. 1999). Goldman et al. ( 2007) summarized hundreds of papers and found that for the prescription drug service

PAGE 31

31 elasticity ranges from 0.2 to 0.6 This is consistent with a study by Landsman et al. (2005), ind icating that elasticity of demand for drugs was generally low for asymptomatic conditions, ranging from 0.16 to The Landsman et al. paper also cited the results of several previo us studies in the 1980s and early 1990s, indicating that demand for prescription drugs w as highly inelastic, with values ranging between 0.33 and 0.10 for small absolute changes in price (Smith et al. 1993) Goldman et al. (2006) an other paper revealed t hat specialty drug use was largely price insensitive, with price elasticity ranging from 0.01 to 0.21. Most of these studies agreed that price elasticity of medical services and drugs varied within the range of 0 to 0.6 Contrary to these results, howev er, a recent study by Chandra et al. (2007) found that physician office visits and prescription drug utilization were very price sensitive. For pharmaceuticals the price elasticity varied between 0.20 and 1.4, a range that crossed the boundary of unit el asticity. For office visits, the estimated price elasticity was between 1.38 and 1.90, again completely elastic. These findings and those of Rosett and Huang (1973) are surprising since they pose a significant challenge to the traditional belief that the demand for medical care is inelastic In summary, the price elasticity for each service type is as follows: from 0.2 to 1.4 for pres cription drugs; from 0.14 to 1.9 for physician care; from 0.35 to 0.49 for hospitalizations ; and 0.2 for overall medical services. The reason for the mixed results may be many because each service type covers a wide range of disease severity and thera peutic categories I t may also suggest that effort s to evaluate price elasticity for each individual service type may be incomplete since they can be substitutes or complements A holistic evaluation for all service types may improve our understanding.

PAGE 32

32 M oreover, for physician care the rang e of price elasticity from 0.14 to 1.9 could be due to differences in the health status of the sam ple s used to create these estimates. The wide range indicates that if price elasticity is near 0.14, physician care ha s little room to be cut down for the severely ill, and is uniquely valuable and essential to maintain health. Thus, physician care cannot be replaced or substituted; there would be a differential impact of high cost sharing by disease severity. On the othe r hand, if price elasticity is approximately 1.9, then physician care is elastic and sensitive to price, and cannot be uniquely valuable because it can be substitute d In this case, even for severely ill people, high cost sharing would likely cut back on their needed care which would result in greater demand for more expensive acute care services downstream Namely, there c ould be no differential impact between groups in physician care itself, but possible differential impact s downstream, in ER or inpatie nt care. If so, the whole picture still indicates the potential differential impact of overall service amount between the severely ill and the reference group. Given the quantitative measure of cost sharing effect by service types, this study will discuss findings from empirical literature reports on cost sharing effects. Rice et al. (1994) and Remler et al. (2009) systematically reviewed the empirical evidence of cost sharing on utilization, expenditure and health outcome s T heir findings are summarize d below as are a few recent publications. Impact of Cost sharing for Different Types of Medical Care All service type cost sharing The RAND HIE (Manning et al. 1987) is so far the only social experiment to determine the effect of patient cost sharing on t health status for all kinds of service types. The insurance design incorporated varying levels of cost sharing and the out of pocket maximums, as well as a deductible plan. All the plans covered

PAGE 33

33 a comprehensive set of services that included outpatient services, inpatient services, prescription drugs and preventive services for twelve fee for service (FFS) and two health maintenance organization (HMO) plans. The RAND HIE study focuse d on the non e lderly a population younger than 62 years old. This study produced a series of results, among them: increasing coinsurance level s reduced medical use and expenditure S pecifically, individuals with a 25 % and 95 % coinsurance policy had medical cost s 23 % and 46% lower than those receiving free care respectively. By contrast, there were no differences in utilization by groups with differing out of pocket maximums. Deductibles also appeared to reduce service usage. Cost sharing reduced both ineffective and eff ective care, and increased cost sharing did not affect health status for people of average health, but did adversely affect that for the sick poor (Manning et al. 1987; Rice et al. 1994). Physician care cost sharing One quasi experimental study examined the health care use and cost for faculty and staff and their dependents at Stanford University from 1966 to 1968 (Scitovsky et al., 1972), during which a 25% coinsurance requirement was instituted on physician inpatient services and all outpatient service s, including, in 1967, that for ancillary services. The cost sharing requirements for hospitalizations did not change during the same time frame. Results revealed that the utilization and cost of physician services fell considerably by about 25 % ; a follow up study indicated that this effect remained stable over four years. Another early quasi experimental study in 1977 found a similar effect, revealing an average $7.5 0 copayment for physician visits to be associated with a dramatic drop in utilization for both inpatient and outpatient services among numbers of the United Mine Workers (Scheffler, 1984). However, the author also discussed some potential threats to the internal validity of the findings. Most other later studies also found that greater physicia n visit cost sharing was associated with fewer office visits (Chandra et al.

PAGE 34

34 2007, Cherkin et al. 1989, Roddy et al. 1986). However, one study that followed enrollees for three years is an exception to the findings, reporting an increase in utilization in the third year (Feldman et al. 2007). Regarding reduced types of office visits, these studies also produced mixed results. Cherkin et al. (1989) found that for those with cardiac disease, implementation of office visits copayment selectively decreased phy sical exams and primary care visits, but did not influence immunizations, cancer screenings, or specialist visits. Roddy et al. (1986) found that the reductions in physician visits were substantial for both prevention and acute self limiting conditions (on es that would clear up on their own). Similarly, Hibbard et al. (2008) found of Health Services). Examining the impact duration of cost sharing, Hibbard et a l. found that the reduction in office visits among CDHP enrollees lasted for two years. Roddy et al. on the other hand, found that the first year reduction in office visits returned to the baseline rate beginning in the second year. Prevention cost shari ng Studies consistently suggested that cost sharing for preventive care was associated with less preventive services across a wide range, including pap smears, preventive counseling, clinical breast exams, and self monitoring of blood glucose for diabetics (Karter et al. 2003; Solanki et al. 1999). S o far no direct evidence ha s shown whether or not reductions in preventive care due to cost sharing affect health. I npatient care cost sharing Two studies (Feldman et al. 2007; Parente et al. 2004) examine d CDHPs with a deductible of $1 500 for individual policies, and they found that compared to other plans, hospitalizations actually increased for CDHPs although other forms of care fell. There may be a number of reasons for th ese finding s. O ne, for examp le, is that the increased deductible reduced

PAGE 35

35 t hey got sicker and in turn used more inpatient care. Alternatively, specialist care preceding hospitalizations would have consumed a substantial portion of the deductible, allowin g patients to easily exceed the threshold and receive excessive care without copays Pharmacy cost sharing As expenditure grows dramatically in pharmaceutical s numerous studies examined newly emerged cost sharing devices in this area. Pharmacy cost shari ng schedules work through a mechanism of different rates for different drug types. The lowest rate is for generic drugs, the middle rate for preferred brand name drugs, and the highest rate for non preferred brand names. In general, sizable changes in ince ntive based formulary cost including those with chronic illness. Goldman et al. (2007) and Gibson et al. (2005) comprehensively reviewed pharmaceutical cost sharing publications from 1985 2006 and 1974 2005 respectively, and found that for every 10% increase in pharmaceutical copayment or coinsurance, there was a decrease of 2% to 6% in drug spending, depending on the class of drugs and condition of the patient. For chronically ill subgroups, some studies show that drug c ost sharing increases utilization in at least one of the following service types: office visits, hospitalizations, or emergency care. Also greater use of inpatient and emergency medical services are associated with higher copayments or cost sharing for pre scription drugs or benefit caps. Similarly, lower cost sharing was associated with considerable increases in drug use that, in turn, might be associated with significant reductions in ER and hospital usage (Chernew et al. 2008; Goldman et al. 2006). When the study population was not limited to those with certain chronic illnesses, increased drug copayments were not associated with more outpatient visits, hospitalizations, or ER visits (Fairman et al. ,2003; Motheral and Fairman, 2001; Johnson et al. ,1997; Smith and

PAGE 36

36 Kirking,1992), as the reason for this difference might be that the healthier group m akes up the majority of the study population and dilutes the effect of their severely ill counterpart s In terms of direct health outcome s Zeber et al. (2007) f ound that increased medication copayment for schizophrenic veterans reduced their refills of psychiatric drugs, resulting in a modest increase in inpatient admissions and may have hurt the large. Another study using a before and after design found that increased pharmaceutical cost sharing in Quebec reduced the use of essential drugs, which in turn increased ER visits but not heart attack mortality ( Tamblyn et al. 2001 ). The study also reported reduced use of nonessential m edications without apparent adverse effect. Emergency r oom v isit cost sharing Studies have consistently found that emergency room visit cost sharing reduces ER utilization (Hsu et al. 2006; Selby et al. 1996; Wharam et al. 2007). With respect to the typ e and value of ER visits reduced by cost sharing, the studies consistently found that visits defined by high value were not significantly reduced by ER cost sharing ; conversely, large reductions w ere observed in ER observed in hospitalizations, intensive care unit admissions, or mortality rates. There may be some threats to the internal validity of these st udies because they tracked individuals for just one year and because adverse outcomes are uncommon occurrences. For other service type cost sharing studies showed that a $20 copayment for outpatient mental health services also reduced the likelihood of receiving outpatient mental health care (Simon et al. 1996). In summary, cost sharing is a blunt instrument. In its immediate impact it can reduce both unnecessary and necessary care and thus undermine health insurance functions. Cost sharing can

PAGE 37

37 further reduce health care utilization and expenditure s for most service types, including physician care, preventive care, ER visits and pharmac euticals and for most subpopulations, including both the healthy and the sick most of the time. Moreover, cost sharin g may influence harm ful health outcome s Effect of Cost sharing in P hysician C are by H ealth S tatus Only a few analyses have stud ied th is exact topic fo r physician care cost sharing although quite a number of papers touch on this topic of cost sharing for other services such as pharmacy. Results are mixed. Absence of d ifferential i mpact The RAND HIE study (Manning et al. 1987) provides an important and representative result. As described in that paper, an important goal of HIE is to study how the respons e to cost sharing varied across subgroups. RAND results explicitly indicated that higher patient copayments reduced medical utilization for a variety of subpopulations (Gruber et al. 2006). There was no differential response in expenditure level s to healt h insurance coverage be tween the healthy and the sick (Manning et al. 1987). This indicates that both groups ha d similar high or low expenditure change s The study explained that the reason for this striking fact wa s due to the upper limit feature ; namely that the sicker group wa s less responsive to the increased coinsurance and individuals we re more likely to exceed their upper limit s on out of pocket expenditures and receive some free care. However, a closer look at the direction of the healthy group ind icate d that lacking a differential response in expenditure level for the two groups, the healthy group should follow the same pattern and also be less responsive to increased cost sharing and to exceed the upper limit. This may not be reasonable. As menti oned before and consistent with RAND results, reduced care for people of average health due to increased coinsurance does not adversely affect their health. Thus, they would be more responsive to

PAGE 38

38 increased coinsurance and would reduce their care use. For t his group, cost dominates health status consideration s and becomes the priority utility concern unless such individuals value the free care much more than their out of pocket expenditures, which they pay first. More reasonable, however, is the opposite te ntative explanation in that paper: given no interaction between plan and health status, the sickly exhibit more discretion at the margin and are similar to the healthy group by being more responsive to increased cost sharing If so, that means the disease condition for the sick group wa s not considered severe enough, allowing group members to forgo some discretionary care. This group, then, does not represent the severely ill subpopulation, which wa s not our main interest. Alternatively, the sickly group may have offset care utilization with more ER visits or inpatient care, in the short run reducing needed care in a fashion similarly to that of the reference group. Gruber (2006) stated that there were no offset effects in the RAND result, but he did not s pecify whether this finding refers to the subgroup of average health, the sick, or as a whole. Moreover, the recent study by Chandra et al. (2007) mentioned above found significant offset utilization for chronically ill patients, indirectly suggesting a si gnificant insurance and health status interaction. For these people, forgoing needed care even temporarily means they would have to make it up later on. Also the sick group ha d only a mild or at m ost a mode rate condition individuals to forgo some discretionary services to the extent exhibited by the healthy group. Still a nother explanation for the lack of differential impact is that service types in this study where expenditures are examined consist of a combination of both physician care and hospitalization. Their different price elasticit ies may cancel the potential differential impact. As mentioned above, physician care has high price elasticity either within or beyond the unit

PAGE 39

39 elasticity boundary. In a 2007 study ( Chandra et al.) the estimated price elasticity for office visits is between 1.38 and 1.90, so that the sick group can be similarly price sensitive as the healthy group. T he potential pronounced d ifferential impact in inpatient care, due to its narrow range from 0.35 to 0.49 in price elasticity may th en be mitigate d by physician care. Moreover, the interaction effect is examined by expenditure instead of utilization, which may be further confoun ded by price in each service type and make this result even more complicated. In summary, when used for the purpose of this paper, the RAND study has some limitations. The potential differential impact is only examined by expenditure, but not by utilizati on; cost sharing is implemented for all service types, and not limited only to physician care. Furthermore, the RAND study population is the non elderly, and thus not representative of the general population. In addition to the RAND study, a number of stud ies in the literature on pharmacy cost sharing also indirectly found no differential impact by health status (Goldman et al. 2006 ; Fairman et al. 2003 ; Motheral et al. 1999 ; Doshi et al. 2009 ) Pre sence of d ifferential i mpact Link et al. (1980) found that Medigap policies increased physician visits by different amounts according to chronic conditions: 42% for those without chronic conditions, only 5% for those with at least one chronic condition. Here, health status or disease severity is indicated by chronic conditions. A similar study by Cartwrite et al. (1992) found that Medigap policies increased medical expenditure by different amounts according to health status: 25% for poor health, 35% for fair health, 45% for good health, and 95% for excellent h ealth. Both studies indicated the healthier beneficiaries had a greater utilization response than less healthy ones, their care usage likely reflecting a greater share of unnecessary need and thus a greater sensitivity to cost sharing change. Given the sam e change in cost sharing levels, their utilization will increase or decrease proportionately more than that of the severely ill group. Using 1995

PAGE 40

40 Medicare Current Beneficiary Survey (MCBS) data, one recent study (Remler et al. 2003) also examined the Medi care beneficiaries with and without supplemental insurance either privately purchased or from an employer sponsored retiree plan. Investigators found that the severely ill group, measured by worse self reported health status or functional health was less s ensitive to cost sharing for hospital care but not for physician care. The above three studies share certain limitations; r esults can only be applied to the elderly population cost sharing or supplemental insurance is not focused on physician care, and in vestigators did not directly compare the effect of cost sharing by health status or with interaction s Moreover, these studies use Medigap employer sponsored insurance and individual ly purchased polic ies relative to the group without any supplemental plans to measure different cost sharing level s which is a convenient but rough measure. If the common practice of 20% of cost sharing level applies to both Medicare and private insurance plans, a rough estimate of cost sharing level for the group lacking suppl emental plan (Medicare only) is 20%, and that for the group with a supplemental plan is 20%*20% = 4%. Thus, these studies end up comparing high and low cost sharing level w ith two relatively fixed values 20% and 4% without any range. These two values are only a special case and lack generalizability for high and low cost sharing level s. I t would be better to have a range or variation for each group. A counter result of differential impact by health status was found by McCall et al. (1991). This study foun d that having a Medigap policy increased inpatient hospital use by 31%, part B services by 42%, and total charges by 36% for people in poor or fair health, but had little effect on use of inpatient hospital and physician services for people in good or exce llent health. It i s anxieties and sick people tend ed to increase their health care utilization. If this increased amount

PAGE 41

41 for health care use and cost is needed, on e expects that the counterpart would be larger for people in good health due to their potential greater share of inefficient moral hazard. However, the opposite results were found, suggesting there was no moral hazard effect for the healthy associated with a more generous insurance policy. Related f inding s Furthermore, some studies provide indirect evidence of a differential impact. The study by Chandra et al. (2007) reported that increases in physician care and prescription drug copayments had little effe ct on hospital use for an average elderly person; for chronically ill elderly patients, however, there was a significant offsetting rise in hospital admissions as physician and drug use declined. The temporary health care reduction was later made up to som e extent, which indirectly suggests the differential health care utilization reduction between severely ill and healthier group s A related but opposite result from the RAND study (Manning et al. 1987) indicated that physician care and inpatient care are complements rather than substitutes T his finding is not statistically significant, however, and the relationship may refer to the overall study sample but not the sicker subgroup Cherkin et al. (1989) found that for those with cardiac disease, copayment for office visits selectively decreased physical exams and primary care visits, but did not influence immunizations, cancer screenings, or specialist visits. Here, cost sharing and disease severity would compete against each other to affect health care us e. The result depend ed on comparison of degree of disease severity and cost sharing ; namely, utility priority. Individuals will choose the less serious consequence to minimize their adverse utility. The results convey an implicit message that physica l exam s and primary care visits the relative disc retionary services can be cut down, while cancer screenings or specialty care services need ed for cardiac patient survival cannot.

PAGE 42

42 The RAND HIE result indicated that people with higher coinsurance rates were less likely to seek any inpatient or outpatient care. However, once care was sought, the amount received and its cost did not vary by coinsurance (Rice et al. 1994). This may imply that the services actually sought are not discretionary and would not allow som e delay. Under this condition, the cost sharing level would not affect care use and cost for the sick, unlike the general effect of cost sharing for average health y people. Thus, there might be a differential impact on care utilization by health status. In addition, a number of studies on pharmacy cost sharing also indirectly found differential impact s by health status ( Gilman et al. 2008 ; Harris et al. 1990 ; Landsman et al. 2005 ; Goldman et al. 2004 Stuart et al., 1999 ) Gaps and Limitation s of Prior Studies In summary, the literature results are mixed on this topic. Some found no differential impact, while others found that there was, within which results both positive and negative were also reported. Similarly conflicting results were found for the offset effect. For the mixed results the reasons may be due to different study samples, plan benefit structure, service types, different methodologies limit both their interna l and external validity. First, cost sharing is not focused on physician care, it is implemented either in combination with some other service types or in general, covering most service types. It is, therefore, difficult to attribute the differential impac t to physician care cost sharing Second, these results can only be applied to some subpopulations, like the elderly or non elderly, and are therefore not representative of the whole population. Third, most of the prior literature suffers from methodologic al drawbacks The studies examine the relationship within severe and healthy group s respectively, without directly comparing them or using an interaction term. It may be possible that the main effect of cost sharing is significant in one group while not in another group when by putting the severely ill and the reference group

PAGE 43

43 together, the difference of the main effect disappears. Fourth, most of the prior studies use a rough measure for cost sharing level They suffer from lack of generalizability because they cost sharing level differences. Similarly, severity level is measured by self reported health status in five categories, and those in poor and fair health are grouped as severe ly ill and the rest becomes the reference group thus using crude measure s that lack precision. This current study will fill in these gaps, focusing only on physician care cost sharing using a nationally representative sample and more precise measure s pl us adopting more rigorous methods to improve the understanding of the research question.

PAGE 44

44 Figure 2 1. Moral hazard and social welfare Figure 2 2 Income transfer helps distinguish between eff icient and inefficient moral hazard Medical care quantity demanded Price of medical care Average cost of medical care A Q u Q u1 A 1 B Q i Q i1 Medical care quantity dema nded Price of medical care A Average cost of medical care Q u Q i B

PAGE 45

45 CHAPTER 3 CONCEPTUAL FRAMEWORK The main purpose of this study is to examine the potential differential impact of high cost sharing in physician care on the change in the amount of health care utilization and expenditure s between severely ill people and healthier people M odel o f health capital and demand for medical care can guide our analysis (Grossman, 1972). It suggests that health can be viewed as a durable capital stock that produces an o utput of healthy time. It is assumed that individuals inherit an initial stock of health that depreciates with age and can be increased by investment. As the depreciation rate rises with age, it is not unlikely that old people who are unhealthy will make l arger gross investments than will younger, health ier people. This indicates that unhealthy people have a greater depreciation rate than the healthy, and that they would be likely to invest even more on health. Thus, unhealthy people value medical care more than the healthy. Therefore, when considering the dilemma between high cost sharing and health status, they prefer health and are less sensitive to cost pressure. measures by evaluating the relationship betw een "need" or illness, measured by the rate of health depreciation and utilization of medical of the other variables in the demand curve for medical care. i i i refers to the depreciation rate of the stock of health which describes the need for i refers to the demand for health care utilization. Because 0 learly indicates the relationship between disease severity

PAGE 46

46 's utiliz ation can be determined by depreciation rate or severity magnitude. In the case of the severely ill people, the higher depreciation rate relative to price (i.e. high cost sharing ) the greater the explanatory power of disease severity in affecting demands for health care utilization. cost sharing pressure. In the case of healthy people, the lower depreciation rate relative to high cost sharing the less important disease severity is differential impact of cost sharing on health care utilization by disease severity, which moderates the main effect of cost sharing Anders e B ehavioral M odel Interestingly, Andersen behavioral model of health service utilization makes a similar argument ( Andersen predisposition to use service, factors that enable or impede use, and their need for care. More enabling resources may help realize potential access to actual utilization by ways of need. As an enabling factor, cost sharing reduction would effectively relieve the financial constraint and provide individuals access to necessary health care. Need is d most immediate cause of health service use utilization. It represents disease severity and the degree of price elasticity holding everything else constant Andersen behavioral model indicates that enabling and need factors would have a differential ability to explain use, depending on which type of service was examined. Hospital services received in response to more serious problems and conditions would be primarily explained by need and demographic characteristics, while dental services considered more

PAGE 47

47 discretionary would more likely be explained by social structure, beliefs, and enabling factors. Here, service type is used to represent disease severity level, with d iscretionary service representing less severe conditions, under which health care use is mainly influenced by enabling factors rather than need. For serious conditions, however, use is mainly determined by need factor s rather than by enabling factor s Dise ase severity (representing the need factor) and cost sharing level (representing the enabling factor) two opposite forces may compete against each other to influence utilization; the result depen ds on their strength comparison, the utility priority. Facin g the dilemma of a higher cost burden and a potentially worse health outcome individuals have to evaluate two adverse consequences and choose the less serious one to minimize adverse utility. For less severely ill individuals, given increased cost sharing the need for utilization may become discretionary and thus less important. Utility priority will give way to the enabling factor cost sharing pressure the dominant predictor for utilization. In the case of a severe condition, individuals are less respons ive to high cost sharing because they have more urgent need, the most immediate and imminent cause of health service use despite constrained enabling factors. Therefore, they tend to maintain their current adequate health care and reduce less necessary uti lization, since any delay or forfeiture would cause more serious health consequences and higher future medical expenditure, so the need dominates the enabling factor in influencing use. Now the main explanatory power transfers to need, the more severe cond ition, with less variation in use explained by an initial enabling factor. Obviously, the effect of the enabling factor on utilization is contingent on the need factor, which bridges and moderates the impact of the enabling factor. To put it quantitativel y, the relationship can be represented by an interaction term. For both situations this study focuses on the relationship between the enabling factor and use, but treats need as a moderati ng factor.

PAGE 48

48 Enabling and need factors are the two main effects. As c ost sharing increases, health care use will decrease, but this relationship may differ depending on the need factor, or the severity level. Specifically, this relationship is unique to the severely ill, who will decrease utilization less than the hea lthy g roup, since for them the need factor dominates the enabling one. Here, disease stronger need he or she has for health care use. Demand Cu rve In economic theory, the demand curve also provides insight into this relationship. In Figure 3 the severely ill tends to be less elastic. In this study, as price is the independent var iable and demand quantity the dependent variable, the X and Y axis is switched. Figure 3 2 reflects Figure 3 1, with price referring to cost sharing and demand quantity referring to health care utilization. It suggests that as cost sharing increases, an i However, utilization among the severely ill group decreases less than for the average health group, which marks the differential impact. P rice E lasticity Quantitatively, the definition of price elasticity of demand further serves as a conceptual framework to provide mathematical precision for the potential differential impact. Price elasticity of demand measures responsiveness of the quantity demanded to changes in price. It can be influenced by many factors, such as income and preference. Thus, holding everything else constant, price elasticity of demand serve s as an indicator reflecting disease severity. I n the extreme case of a life threatening condition, p rice elasticity | E d | would approach 0 For each se verely ill and reference group Q % = P % E d utilization change percentage Q % is also a function of cost sharing price change percentage P % and price elasticity of demand E d For

PAGE 49

49 the purpose of clarification, this study now denotes P % as P c %, c refe r ring to cost sharing Thus, given the same cost sharing change P c % between severely ill and reference groups, the differential impact in utilization change Q % depends solely on the price elasticity discrepancy. E d = E d1 E d2 (1) Q 1 % Q 2 % = P c % E d1 P c % E d2 = P c % (E d1 E d2 ) (2) The wider the gap between these groups in disease severity level or price elasticity, the large r the difference between the two groups in percentage of utilization change. Similarly, health care expenditure is a function of price and demand quantity with TE = P* Q, with P being the unit price for certain service type. Here, this study den otes P as P MD, which refers to doctor visit unit price. P MD is different from above P c %, which refers to the percent of change in cost sharing price. Between the severely ill and the reference group, P MD is the same, since they are compared for the same t ype of service physician care. F or each group, the expenditure difference between high and low cost sharing level is TE = TE h TE l = P MD *Q h P MD *Q l = P MD *(Q h Q l ) = P MD Q. (3) T E % = ( TE h TE l ) / TE l = P MD *(Q h Q l ) / ( P MD *Q l ) = (Q h Q l ) / Q l = Q % = P c % E d (4) TE l and TE h refer to health care expenditure for low and high cost sharing level s within each group. Q l and Q h refer to health care use amount for low and high cost sharing level s within each group. Thus, the differential impact in expenditure change TE % also depends solely on the price elasticity discrepancy TE 1 % TE 2 % = P c % E d1 P c % E d2 = P c % (E d1 E d2 ) ( 5 ) 1 and 2 refer to the severely ill and the reference groups respectively.

PAGE 50

50 It should be noted that expenditure for each service type consists of two components: out of pocket spending and third party payment. Cost shari ng policy is intended to influence demand side by adjusting individual consumer behavior, so the potential differential impact should be directly demonstrated in out of pocket expenditure. Since coinsurance is fixed for any service within a plan, out of po cket expenditure is proportional to total expenditure, so the potential differential impact can also be indirectly demonstrated in total expenditure s for a service as well. In summary, the economic theory of price elasticity of demand indicates that the d ifferential impact between the severely ill and the reference group may be demonstrate d in both health care utilization and expenditure s Initially, price change P c refers to the same group of people experiencing cost sharing change over time, but this s tudy can treat P c as two groups cost sharing difference s in a cross sectional study. The demand curve slope go es downward, which means the direction of health care demand is always negative. The magnitude of the negative sign of price elasticity | E d | determines the absolute reduction effect of cost sharing on care utilization, whether for the elastic or inelastic group, severely ill or reference group. Therefore, as cost sharing increases, health care utilization will always decrease. It is no surprise that many papers report the cost sharing effect held even for chronically ill people. O nly in extreme case s a life threatening condition when| E d | approaches 0 increased cost sharing will not reduce utilization and expenditure s What really matters is th e reduction magnitude for severely ill people, whether the medical utilization and expenditure reduction is similar to or substantially less than those for the low health risk group. Cost sharing may not sizably reduce health service utilization and expend iture for people with severe diseases since the services they seek are mainly of a life sustaining and life s aving nature,

PAGE 51

51 which is necessary and exhibits less elastic demand. The essence of this study is whether there exists an interaction effect between disease severity and cost sharing Beside health status heterogeneity, service type may also influence the differential impact. Specifically, emergency room care and inpatient care are relatively insensitive and resistant to cost sharing and thus more li kely to demonstrate the potential differential impact due to cost sharing in ER and inpatient care. Physician care, however, is likely to be sensitive to cost sharing Therefore, the potential differential impact may be attenuated for the severely ill grou p and the healthier group More generally, service types with higher price elasticity may offset the differential impact by health status. For physician care, price elasticity estimates range from 0.14 to 1.9 as summarized in the literature review secti on The wide range indicates that if price elasticity is near the lower limit 0.14, similar to those for ER and inpatient care, there could be a differential impact of high cost sharing by disease severity in its own service. On the other hand, if price elasticity is near the upper limit 1.9, there would not be a differential impact between groups in physician care itself. Both situations can be examined as follows: First, there is a direct differential impact on physician care. That means severely ill patients can successfully maintain their necessary care in MD visits that are adequate to maintain their health status and meet their requirements, so they need little downstream ER or inpatient care to make up for potential deficient care. As a result, t here will not be an offset effect, and the overall health care utilization and expenditure for all service types reflect the direct physician care differential impact. Second, there is no direct differential impact on physician care due to cost sharing pr essure It is possible that the severely ill are sensitive to high cost sharing pressure, so they reduce physician care to a similarly large exten t as the general health population. However, t he ir

PAGE 52

52 suppressed physician care utilization will be released in d ownstream ER or inpatient services since s do not allow them to forgo needed physician care w hile the healthy need little ER or inpatient care because their larger health stock allows them to reduce more unnecessary physician ca re Thus, severely ill individuals have increase d ER or inpatient care utilization more than the general health population, so there still may be an indirect differential impact in ER or inpatient care (Figure s 3 3 and 3 4). Th e indirect differential impa ct o n ER or inpatient care expenditure can be demonstrated as follows: For each severely ill and healthier group, net expenditure change is P MD Q MD1 + P ER,inpatient Q ER, inpatient1 ( 6 ) P MD Q MD2 + P ER,inpatient Q ER, inpatient2 ( 7 ) The expenditure change P MD Q MD is n egative, indicating it is sav ing s while the expenditure cha nge P ER,inpatient Q ER, inpatient is positive, indicating it is actual spending. The differential expenditure change is ( 6 ) ( 7 ), as there is no differential impact in doctor expenditure, so the first item is canceled: ( 6 ) ( 7 ) = P ER,inpatient Q ER, inpa tient1 P ER,inpatient Q ER, inpatient2 = P ER,inpatient ( Q ER, inpatient1 Q ER, inpatient2 ) ( 8 ) As there is little expected offset effect for the healthier group, Q ER, inpatient 2 0, so the ultimate differential net expenditure change is P ER,inpatient Q ER, inpatient1 which is positive. healthier group s ; while expenditure comparison from low to high cost sharing

PAGE 53

53 significant amount of care reduction in physician care will be offset to a certain extent by ER or inpatient care while there is little off set effect for the general health population. As a whole for all service types, those severely ill total care reduction amount is still less than the healthy, and there still may be total differential impact (Figure 3 5) It is less likely that there ar e both differential impact s in direct physician care and indirect offset effect in ER or inpatient care. Although the net health care utilization amount cannot be derived across different service types, expenditure has the same unit that allow s an estimat ion of the potential differential impact in total expenditure s For the overall expenditure in the second situation, it is expected that the increased ER and inpatient care expenditure is less than the reduced physician care expenditure for the severely il l group. As a whole, the high cost sharing group has less overall mean expenditure than the low cost sharing group, and the severely ill group has higher overall mean expenditure than the general health population Finally the severely ill group reduces i ts expenditure amount less than the general health population Hypotheses Based on these analyses, two alternative sets of hypothes e s are proposed. Set 1: 1. Cost sharing level directly affects physician care utilization change differently by health status ; specifically, high cost sharing reduces physician care utilization less for the severely ill group than for the healthier group. 2. Cost sharing level directly affects physician care expenditure change differently by health status, specifically high cos t sharing reduces physician care expenditure s less for the severely ill group than the healthier group.

PAGE 54

54 If S et 1 hypotheses are confirmed, then the research question has been answered. In the event there is no direct differential impact in physician care, this study will proceed to the next set of hypotheses. Set 2 3. There is an indirect differential impact of cost sharing in physician care on ER and inpatient care utilization change by health status; specifically, high cost sharing in physician care inc rease s ER and inpatient care utilization for severely ill people more than health ier people 4. There is an indirect differential impact of cost sharing in physician care on ER and inpatient care expenditure change by health status; specifically, high cost sharing in physician care increased ER and inpatient care expenditure s for severely ill people more than health ier people 5. Cost sharing level in physician care affects expenditure change in overall service type differently by health status; specifica lly, high cost sharing reduces total care expenditure s less for the severely ill group than for the healthier group. In summary, the expected results would be either hypothesis 1 and 2, or 3, 4 and 5. Either hypothesis set is expected to occur but not at the same time.

PAGE 55

55 Figure 3 1. Demand curve for severely ill and reference group Figure 3 2. Demand curve for severely ill and reference group with switched X and Y axis Demand quantity for medical care Severely ill group Reference group Price of medical care Price of medical care Low cost sharing Hig h cost sharing Utilization of medical care Severely ill group Reference group

PAGE 56

56 Figur e 3 3. Demand curve for severely ill and reference group with switched X and Y axis in MD visits Figure 3 4. Demand curve for severely ill and reference group with switched X and Y axis in ER or inpatient care MD p rice of medical care Low cost sharing High cost sharing Severely ill group Reference group MD u tilization and expenditure MD p rice of medical care Low cost sharing High cost sharing ER or inpatient care u tilization and ex penditure Severely ill group Reference group

PAGE 57

57 Figure 3 5 Demand curve for severely ill and reference group with switched X and Y axis in total care Price of medical care Low cost sharing High cost sharing Utilization of total medical care Severely ill group Reference group

PAGE 58

58 CHAPTER 4 METHODS Data Description This section focuses on the design, content and procedure of the MEPS data, and the rationale for us ing the MEPS for this study. Medical expenditure panel survey household component (MEPS HC) data from 2007 was used for analysis in this study. The MEPS HC is a national representative survey of the non institutionalized civilian population of the United S tates and is designed to produce national and regional estimates of the health care use, expenditures, sources of payment, and insurance coverage of the U.S. population. The MEPS includes surveys of medical providers component (MPC), furnishing information on providers to supplement the data provided by household respondents. The MEPS design permits both individually based and family level estimates. In this study, the unit of analysis was each individual. The MEPS HC has an overlapping panel design. Each y ear a new MEPS HC panel is established. Information is collected from each household to cover a two year period, so the data can be used to track changes and trends over time to estimate health care utilization, expenditure and insurance coverage. The ME PS HC sample is drawn from about a one quarter subsample from the multi stage area probability design. The complete NHIS sample consists of 358 primary sampling units (PSUs ), which are counties or groups of contiguous counties. The sample PSUs are stratified by geographic area, metropolitan status, and socio demographic measures. In MEPS HC, Hispanics and African Americans have been over s ampled each year since 2004. From 20 02 on, the MEPS sample design over sample d Asians and persons predicted to have incomes less than 200 % of the poverty level ( Agency for Healthcare Research and Quality

PAGE 59

59 AHRQ, 2007). Over sampling holds the advantage of improving the precision of estimates for specific subgroups. Each new MEPS annual sample, referred to as a panel, cover a series of five rounds of in person interviews over a 30 month field period to yield annual circle data for two full calendar years. The computer assisted personal intervi ew is the principal data collection mode; each to build longitudinal data for up to two years of survey participation. During each calendar year since 1997, dat a are collected simultaneously for two MEPS panels. One panel is in its first year of interviews, including rounds 1, 2, and 3, while the prior each MEPS p anel overlaps two calendar years. These two panels comprise the annual estimates. The MEPS HC contains detailed data that meets this study purpose, including demographic characteristics, family structure, household income, health and functional status, he alth insurance coverage, access to care, health care use, and expenditures. Meanwhile, information from medical provider component (MPC) is used to supplement and validate the information from MEPS HC about diagnosis, charges, payments, and specific servic es provided. As a result, the final data include information on diagnoses, procedures, inpatient stays classified by d iagnosis related group ( DRG ) prescriptions (medication names, strength s and quantit ies dispensed), charges, and payments. The data allow us to take advantage of above detailed information relevant to this study to make nation wide inference s Like many other national survey designs, the complex multi stage, unequal probability, and cluster sampling methods require adjustment for sample w eight to reflect the unequal probability of selection, household non response (MEPS Round 1), attrition of persons

PAGE 60

60 (subsequent r ounds), post stratification (census population estimates), and trimming of extreme weights. A sample weight is assigned to each sample person. It is a measure of the number of people in the population represented by that sample person in MEPS (AHRQ 2007; Cohen et al. 1999 ). There are two groups of weighting variables: person level weights in the annual data file and longitudinal w eights in panel files. The longitudinal weight variables should be used when the sample includes persons participating in both years of one panel. This study will use person level weights in the annual data file. Variance also needs to be adjusted to accou nt for the differential weighting and the correlation among sample persons within a cluster. Typically, individuals within a cluster are more similar to one another than those in other clusters; as a result, the error term in health care utilization or exp enditure for individuals within a cluster is correlated. Failur e to consider this correlation at cluster level will result in variance estimates that are too small This, in turn, will tend to overestimate the significance of the estimates. Measures and Op erationalization Measures of Outcome Variables Two groups of outcome measures will be examined: utilization and expenditures of care. In MEPS both utilization amount and expenditure are specified by service type: physician care, ER and inpatient care. T his also permits examination of the effect of cost sharing in physician care on its own service and ER or inpatient care. This study use d both specific expenditures by service type and the overall expenditure. The total health care expenditure is the total expense for these service types, rather than existing total expenditure in the dataset with service types such as medication being included. Expenditures in the MEPS are defined as the sum of payments made, including those made out of pocket and by thir d party payers (Zuvekas et al. 2002). This study used expenditures instead of charges when measuring health care costs, since charges are the fees billed to patients and insurance companies by providers, while expenditures are actual

PAGE 61

61 direct payments to pr oviders and health care organizations due to the negotiated discounts between payers and providers. The expenditure variables were constructed from the original event files by summing the total expenditures for each event by person. Expenditures in ccurred in the following type s of setting s inpatient, outpatient and emergency room consist of two Hospital facility expenses include all expenses for direct hospital care, including room and board, diagnostic and laboratory work, X rays, and similar charges, as well as any physician services included in the hospital charge. SBD expenses typically cover services provided to patients in hospital settings by providers li ke radiologists, anesthesiologists, and pathologists, whose charges are often not included in hospital bills. This study will use the sum of facility and SBD expenditure for each of above services. Beside physician care, this study also examine s its primar y care component, which can help us better understand the potential differential impact in physician care. The medical utilization measures are the quantity of each type of care in 2007 Those variables are constructed from MEPS Events Files by counting t he number of events within the year for each type of care. The MEPS asks respondents the primary reason for the visit and prompts respondents with a list of possible reasons. The primary care visit number results from counting the numbers for primary care related reasons from both office visits and hospital outpatient visits which include general and family practice, internal medicine, and geriatric services for the elderly. Similarly, t he measurement of physician care in this study is the number of MD vis its identified by office and hospital outpatient visits. Physician visits is a commonly used and validated measure in health care use and its relationship to health insurance coverage. In the MEPS survey, th e question was asked : Did the patient see the doc tor or was it a phone

PAGE 62

62 call ? Only in the former case is the visit coded as a physician visit. The measure of inpatient care is the number of hospitalizations. The variable comes from the Hospital Inpatient Stay file. Emergency care utilization is measured b y the number of ER visits. The MEPS contains an event file that has information on each ER visit. Measures of Explanatory Variables Severity l evel Some studies use presence of chronic conditions, self reported health status by five categories, or therape utic classes to measure disease severity; however, those suffer from a lack of precision or universality. Instead, Ware et al. (1992) adopted a physical component score (PCS) and a mental component score (MCS) from a 36 item short form (SF 36) in the Medic al Outcomes Study to reflect health related quality of life in medical conditions and illness severity, and this practice was followed by other studies such as the one by Harman et al. ( 2010 ). A short form questionnaire SF 12 wa s derived from the SF 36 (W are et al. 1992), and both questionnaires are widely used and well validated generic instruments for measuring health status (Lowe et al. 2004). SF 12 includes the following eight concepts: physical functioning, role functioning physical, bodily pain, ge neral health, vitality, social functioning, role functioning emotional, and mental health, summarizing these 12 items into an overall physical and mental function score. This study focus es on physical health status but not mental health status; M CS is not used since mentally ill patients may make irrational judgments and decisions on physician care usage i n response to high cost sharing PCS is constructed with a continuum range from 0 to 100, a higher score indicating better health status to precisely re flect an morbidity severity. As a measure for health status, PCS provides us with several advantages. First, judg ment of illness severity based on co morbidity type and number makes disease severity. PCS goes beyond this

PAGE 63

63 limitation by providing this study with a universal measure. Second, PCS as a continuous index allows us to examine health status and disease or condition severity in a precise manner. It overcomes the difficulty of determin ing the categories for some conditions when the response scale is simply dichotomous or encompasses several categories the essential and discretionary drug category as a proxy measure for disease severity or the five categories of poor, fair, good, very g ood, and excellent for self report ed health status. These measures may suffer from measurement error problem. With the continuous measure of PCS, disease severity or health status can be classified into precise categories. Like the World Health Organizatio n classification for body mass index (BMI), different PCSs can be classified into different illness severity categories. According to the study by Harman et al. (2010) when us ing the 100 score range, top tertile and bottom tertile scores in actual data dis tribution are divid ed into severity groups, with any score in the lowest tertile categorized as severely ill, and any score in the highest tertile serving as the reference group. This study will follow suit. Note that PCS is available in MEPS. Cost shari ng l evel Before measuring cost sharing certain concerns about observation selection need to be dealt with. An individual may move out of a plan and become uninsured for some time or may transfer to another plan. If insurance coverage status is defined by less than one full year cover age th is measure of insurance coverage is subject to measurement error. Therefore, this study only select ed those who have retained insurance coverage for a full year, which best serves the research question. In MEPS, there ar e two kinds of insurance duration status variables: insurance coverage in a year and insurance coverage by month, defined as having insurance coverage any time in a year or month. This study use d the month indicator and respondents must have insurance in a ll twelve calendar months to be included in the study.

PAGE 64

64 Similar to severity level, there is not a direct cost sharing level measure in MEPS however, it can be calculated by dividing out of pocket expenditures by total expenditures for each event for each person; both data are available. In this way, cost sharing can be quantified precisely in a continuous percent range. Unlike copayment measured by a fixed amount, this coinsurance measure can be compared across plan types and service types, thus will be us ed in this study. On the other hand, it should be noted that the cost sharing level cannot be derived from those whose out of pocket and total expenditures for 2007 are zero Since there may be a non trivial number of such observations, listwise deletion o f these missing values may cause selection bias problem, leaving the remaining observations unrepresentative for the whole population. A safer approach to address this problem is to assign the se missing data with the mean of the existing values by each ins urance type. With complete cost sharing data, this study can then move on to determine cutoff points in each plan to classify people into high or low cost sharing group s As is common practice, Medicare and many private insurance plans adopt a 20% cost s haring level. One study adopts 2 0% to differentiate low or high levels of cost sharing in medications, arguing that it was consistent with a common co insurance level and that it would be important to know if an effect was present at a lower limit of high cost sharing ( Ungar et al. 2008) This study follow ed suit by using 20% to classify individuals into high and low cost sharing group s, but mad e slight modifications: zero cost sharing individuals were not singl ed out as a group separate from the low and h igh ones since it is not likely that people with a free health care plan would behave markedly different ly from those in the low cost sharing group. This 20% criteria, howe ver, does not apply to Medicaid because medical care places a relatively heavier f inancial burden on the poor. The Medicaid program was designed to serves a

PAGE 65

65 low income population that lacks substantial resources and often has significant health care needs. It prohibited or limited premiums and cost sharing to nominal levels (Kaiser Fami ly Foundation, 2003). One study that surveyed cost sharing policy in Medicaid and State Children's Health Insurance Program (SCHIP) outlined this standard by these facts: the median co payment for non preventive office visits was $5 for all family income l evels, and ranged from $5 to $25 for emergency room use, and from $5 to $18 for hospitalizations (Selden et al. 2009). In this study a relative cut off point the mean cost sharing level in this study sample was use d to differentiate individuals into high and low cost sharing groups. Measures of control v ariables The control variables, a comprehensive range of factors likely or known to affect health care utilization available in the data, include a person rural/urban location, marital status, education level, income, other health conditions, and gatekeeper plan status. Since different age groups may not affect health care use in si milar way s age is grouped into three categories young, middle, and elderly in the range s 18 34 years, 35 64 years, and older than 65 years respectively This study excludes children because the study is focused on people who are most likely to mak e health care decisions for themselves. Gender, of course, is treated as a dichotomous variable : female and male. Race and ethnicity were classified into four groups: non Hispanic white, non Hispanic African American, Hispanic, and other. Regions of residence were classified as Northeast, South, Midwest, and West. Rural versus urban residence is det ermined by whether or not a respondent lives in a metropolitan statistical area. health care use. Marital status is classified as married and non married, and ma rriage, too, may influence health care since it represents a type of social support. Education level may affect of health care, and it is represented by years of

PAGE 66

66 education achieved Different education levels m ay not affect health care use in similar way s so the education variable is categorized into three groups: 0 8 years, 9 12 years, and 13 17 years. Similarly, family income or poverty status is categorized into five groups: poor, near poor, low income midd le income and high income. This income variable is categorized by percentage of the federal poverty level, which is consistent with the 2007 poverty statistics developed by the Current Population Survey In addition to disease severity as a main effect, other co morbidity conditions are also controlled since our severity measure PCS comes from SF 12, self reported perceived overall health status. Andersen health care utilization, so the remaining evaluated need has to be controlled as well. Therefore, this study includes other co morbidities: hypertension, diabetes, asthma, arthritis, stroke, brain injury, and cancer, which are dummy variables. health care since the primary care physician controls and decides depending on their medical condition whether or not to refer patients to specialists. Dummy variables will be used to indicate if individuals have primary care physicians. When the offset effect is examined for ER or inpatient care t his study should also control the cost sharing level for ER visits or hospitalization s since utilization or expenditure is influenced directly by cost sharing level in their own service types, and not simply by cost sharing for physician care. Here, cost sharing level for ER visits or hospitalization s is a continuous variable, calculated directly from the proportion of out of pocket expenditure out of total event expenditure.

PAGE 67

6 7 Based upon the above considerations, this study may organize the following variables The dependent variable is health care utilization and expenditure, including physician care, primary care, emergency care, and inpatient care. The independent variables are cost sharing level and severity l evel. These are the two main effects, and their interaction is our major interest in examining cost sharing and the control groups. The covariates include age, gender, race, geogra phic area, rural/urban location, marital status, education level, income, and gatekeeper plan status. In the ER and inpatient care equation, cost sharing levels of their own services also need to be controlled. Study Design This study use d a cross section design from MEPS HC data in 2007, the latest available data. Since a universal cost sharing policy change does not occur at a time point across different insurance types, a cost sharing policy change may occur at different times for different plans at diff erent levels, or there may be no change at all. This makes it difficult to take advantage of a longitudinal design of panel data to track utilization and expenditure changes over time in response to insurance changes. Instead, this study selects a single y ear to examine the research questions. Subject selection criteria should conform to research question requirements and study design; thus, only observations that meet the criteria below w ere included in the analysis. 1) S ubjects age d 18 or older were selecte d since children do not make health care utilization decisions 2) Those with the same insurance status for a full year were selected since partial year insurance coverage may produce measurement error 3) As a representative sample of public and private plans individuals enrolled in Medicare, Medicaid and private plans were selected

PAGE 68

68 As it is assumed that individuals will behave similarly at the same cost sharing level regardless of plan type, this study does not need to specify the potential differential im pact within Medicare, Medicaid, and private plans. However, cut off points to divide enrollees into high and low cost sharing groups may differ by plan types. Specifically, cost sharing level above or equal to 20% and below 20% is selected as the high and low groups for Medicare and private insurance plans. In Medicaid, cost sharing levels above or equal to and below the mean cost sharing value are divided into high and low groups Meanwhile, the study specifies service types as physician care, emergency ca re, and inpatient care. Since MD visits are the main form of physician care, this study use s these two terms interchangeably. Statistical Analysis Descriptive analysis was conducted first in order to describe study population characteristics. Next, multiva riate analysis was em ployed. As mentioned above, a rigorous design should account for the potential endogeneity between health care utilization or expenditure and cost sharing level, as it is obvious that increasing cost sharing level reduce s health care utilization and expenditure s and those with high health care utilization and expenditures will seek to enroll accordingly in low cost sharing plans. In this case, i ndividuals with different cost sharing levels are likely to differ in ways that are relat ed to their utilization of care, which, however, are not readily observable by researchers could reflect the combination of a causal effect of cost sharing level and the effec t of unmeasured characteristics that are correlated with cost sharing and use of care. The existence of omitted variables could threaten internal validity. If the omitted variables affect the cost sharing level and utilization in the same direction, the pr edicted cost sharing effect will be overestimated. The Hausman test (Hausman, 1978) can be employed to test for endogeneity by regressing cost sharing on all

PAGE 69

69 exogenous variables, then adding the residual as a new variable into the initial structural equati on. The significant coefficient for the residuals indicate s the existence o f endogeneity. The Hausman test compare s the OLS and two stage least square s (2SLS) estimates and determine whether these differences are statistically significant. Both OLS and 2SL S are consistent if the cost sharing variable is exogenous. If OLS and 2SLS differ significantly, then the suspectible regressor must be endogenous ( Wooldridge, 2005 ) since now the 2SLS estimate is consistent while the OLS estimate is inconsistent. In thi s case, as a solution, instrumental variables (IV s ) should be introduced that only affect cost sharing level but not utilization and expenditure. An of pocket premium within each plan type can be an eligible IV. As an of pocket premium increases, some enrollees tend to drop o ut o f plan s In order to maintain enrollees, insurers will reduce cost sharing level to make their plan competitive in the market. Therefore out of pocket premiums influence cost sharing level s but on ce their levels are deter mined there is no way to influence actual utilization and expenditure s The IV approach wa s used to account for this endogenous problem because it is unlikely to identify and measure all the unobserved h eterogeneity. With the aid of IV s the predicted cost sharing level for each person was obtained, replacing the actual ones for the main effect and interaction term. Specifically, two stage least square s techniques we re em ployed. T he analysis use d the first stage by regress ing the e ndogenous independent variable on IV s and other exogenous variable s, and get ting the predicted cost sharing value to replace the actual value. This study also needs to test the goodness and effectiveness of IV s A good instrumental variable must have two properties: first, it must be correlated with the endogenous regressor; second, it should not be correlated with the error term of structural equation (Wooldridge, 2005).

PAGE 70

70 In the first stage, an IV is usually valid so long as the F statistic as a rule of t humb is greater than 10 Meanwhile, this study directly regress ed the error term of utilization or expenditure on the IV, with an IV being deemed valid if its coefficient wa s not s ignificant, indicating that it wa s not related to dependent variables. In the final analysis, a nalytic model selection should account for characteristics of count data and expenditure data M easured by number of visits u tilization is count data. Count data are non negative values, with the possibility of a substantial number of zeros. As a random variable, small value counts may have higher probability than large value count s. In addition the variance may not be equal to its mean. A negative binomial model may best reflect these characteristics for the count data than any other model. Expenditure is a continuous variable; accordingly, a generalized linear model could be employed. Specifically, the two part model is suitable for analysis, using logit in the first part to estimate the probability of any expenditure. For the secon d part, conditional on any expenditure, non linear regression could be employed to account for the skewed distribution. This usually means using log transformation with residual diagnosis to check the model fit. Specifically the statistical analysis or mod el set up for each hypothesis is as follows: 1. Cost sharing level directly affects physician care utilization change differently by health status; specifically, high cost sharing reduces physician care utilization less for the severely ill group than for the healthier group. A n egative binomial regression was used. ln E( primary care visits ) = 0 + 1 cost sharing + 2 *severity + 3 cost sharing *severity + 4

PAGE 71

71 ln E( MD visits ) = 0 + 1 cost sharing + 2 *severity + 3 cost sharing *severity + 4 E is the expected value. Based on theory prediction and variables coding, it is expected that 1 will be negative, indicating that the high cost sharing group has lower mean utilization value; 2 will be positive, indicating that the s everely ill group has higher mean utilization value; and 3 will be positive, indicating that the severely ill group reduces their utilization amount less than the reference group. Note the interaction term is a reduced form that only represents the simple st case where both cost sharing and severity are dichotomous variables. In the final analysis where cost sharing and severity, alone or in combination, can be multiple groups, this reduced form of interaction may represent a series of dummy interactions. 2. Cost sharing level directly affects physician care expenditure change differently by health status; specifically, high cost sharing reduces physician care expenditure less for the severely ill group than for the healthier group. A two part model was us ed. In the first part logit wa s used to estimate the probability of any expenditure ln odds ( MD expenditure ) = 0 + 1 cost sharing + 2 *severity + 3 cost sharing *severity + 4 Odds are the ratio of probability of having any physician c are expenditures relative to that of not having any expenditure Odds = P / (1 P) For the second part:

PAGE 72

72 ln MD expenditure = 0 + 1 cost sharing + 2 *severity + 3 cost sharing *severity + 4 The coefficients should follow the same pattern as that in hypothesis 1. It is expected that 1 will be negative, indicating that the high cost sharing group has l ower mean expenditure value; 2 will be positive, indicating that the severely ill group has higher mean expenditure value; and 3 will be po sitive, indicating that the severely ill group reduces their expenditure amount less than the reference group. 3. There is an indirect differential impact of cost sharing in physician care on ER and inpatient care utilization change by health status; speci fically, high cost sharing in physician care leads to increased ER and inpatient care utilization for severely ill people but little change for the reference group. A n egative binomial regression w as used. ln E( ER / inpatient care visits ) = 0 + 1 cost sharing + 2 *severity + 3 cost sharing *severity + 4 It is expected that 1 will be positive, indicating that the high cost sharing group has overall higher mean increased offset ER or inpatient care utilization; 2 will be po sitive, indicating that the severely ill group has higher mean increased offset ER or inpatient care utilization; and 3 will be positive, indicating that the severely ill group increased ER or inpatient care utiliz ation more than the reference group; the latter can be around zero. 4. There is an indirect differential impact of cost sharing in physician care on ER and inpatient care expenditure change by health status; specifically, high cost sharing in physician care leads to increased ER and inpatient ca re expenditure for severely ill people but little change for the reference group.

PAGE 73

73 A two part model w as used. In the first part logit wa s used to estimate the probability of any expenditure. ln odds ( ER and inpatient care expenditure ) = 0 + 1 cost sharing + 2 *severity + 3 cost sharing *severity + 4 F or the second part: ln ER / inpatient care expenditure = 0 + 1 cost sharing + 2 *severity + 3 cost sharing *severity + 4 It is expected that 1 will be po sitive, indicating that the high cost sharing group has overall higher mean increased offset ER or inpatient care expenditure; 2 will be positive, indicating that the severely ill group has higher mean increased offset ER or inpatient care expenditure; an d 3 will be positive, indicating that the severely ill group increased ER or inpatient care expenditure more than the reference group ; the latter can be around zero. 5. Cost sharing level in physician care affects expenditure change in overall service ty pe differently by health status; specifically, high cost sharing reduces physician care expenditure less for the severely ill group than for the healthier group. A t wo part model was used. In the first part logit wa s used to estimate the probability of an y expenditure ln odds (total care expenditure ) = 0 + 1 cost sharing + 2 *severity + 3 cost sharing *severity + 4 F or the second part: ln total expenditure = 0 + 1 cost sharing + 2 *severity + 3 cost sharing *severity + 4

PAGE 74

74 It is expected that 1 will be negative, indic ating that the high cost sharing group has lower mean expenditure value; 2 will be positive, indicating that the severely ill group has higher mean expenditure value; and 3 will be positive, indicating that the severely ill group reduces their expenditur e amount less than the reference group. For any main effect and interaction, the predicted expenditure amount was re transformed to obtain estimates on the original scale, since it has been transformed to satisfy OLS assumptions. As heteroskedasticity wi ll be a potential issue, subgroup specific smearing factors w ere used to obtain unbiased estimates to quantify the magnitude of the interaction effect if the differential impact wa s significant (Manning, 1998). Meanwhile, t he bootstrap technique was employ ed to get standard e rror and significance for the differential impacts. This study also calculate d price elasticity and cross price elasticity in physician care for the severely ill and the reference group. Sensitivity analysis w as also conducted to test finding robustness for the following model s pecifications: First, if there wa s a sizable proportion of individuals whose cost sharing wa s exactly 20% for Medicare and private plans or the mean value for Medicaid, the results may be influenced when they are categorized into high or low cost sharing groups. This study examine d both situations. Second, this study may also consider different cut off points for cost sharing since there could be some con cer n about whether it is really a high or low level if cost sharing is a little higher or lower than 20% for Medicare and private plans or the mean value for Medicaid. This study can also test 25% and 15% as cut off points for Medicare and private plans, and 5% away from either side of the mean value for Medicaid. This way can produce three categories in cost sharing : low, middle and high cost sharing groups. Third, this study also include d the middle tertile subjects of severity level and create d another severity dummy variable to examine

PAGE 75

75 the differential impact for this moderately ill group within the whole picture. Fourth, alternative ways of categorizing PCS were sides of two standard deviations from the PCS mean as cut off points.

PAGE 76

76 Table 4 1. S ummary of outcome measures Variable Type Utilization Number of primary care visits Count Number of MD visits Count Number of ER visits Count Number of hospitalizations Count Expenditure Physician care expenditures Continuous Emergency care expenditures Continuous Inpatient care expenditures Continuous Total expenditures Continuous

PAGE 77

77 Table 4 2. Explanatory variables Variable Type Categor y Cost sharing level Categorical 0 = L ow 1 = H i gh Disease severity level Categorical 0 = T op tertile PCS 1 = Middle tertile PCS 2 = B ottom tertile PCS Interaction Categorical Cost sharing *D isease severity Age (yrs) Categorical 0 = 18 34 1 = 35 64 2 = 65 and older Gender Categorical 0 = F emale 1 = M ale Race Categorical 0 = Other 1 = Non Hispanic white 2 = Non Hispanic African American 3 = Hispanic Geographic area Categorical 0 = Midwest 1 = Northeast 2 = South 3 = West Rural/urban location Categorical 0 = R ural 1 = U rban Marital status Categorical 0 = Not married 1 = M arried Education level (yrs) Categorical 0 = 0 8 1 = 9 12 2 = 13 17 Income Categorical 0 = P oor 1 = Near poor 2 = Low income 3 = Middle income 4 = High income Co morbidities Categorical 0 = N o 1 = Y es Gatekeeper status Categorical 0 = N o gatekeeper 1 = G atekeeper

PAGE 78

78 CHAPTER 5 RESULTS Overview The results of the study are presented in the following section s. First, the demographic and socioeconomic characteristics of selected sample are described. Second, before any analysis variable operationalization is reported. Specifically, the missing cost sharing values were imputed by insurance type, and the variab le selection issue was addressed due to the concern about multicollinearity. Third, the potential endogenous problem between cost sharing and health care utilization or expenditure was tested and confirmed, then an instrumental variable was introduced to a ddress the endogeneity. Fourth, the results from the multivariate analysis of the differential impact are presented. This section is organized by outcome measures in the sequence of health care utilization and expenditures by service types. Specifically, p hysician visits, primary care visits, ER visits, hospitalization admissions, physician visit expenditures, ER visit expenditures, hospitalization expenditures, and total expenditures. The of Stata 10.0 ( StataCorp 2007 ) were employed for all the statistical inference tests. Description of the Sample Table 5 1 presents the number and composition of individuals eligible for the study. The final study sample consisted of 13,020 individuals who m et the inclusion criteria (Table 5 1) Altogether there were 60,595 observations, since each individual may have had several instances of health care utilization with different service types or doctor specialties that could not be combined into a single value when the service type variable was kept in the data For example, if and cardiology service, these service types could not be combined into a composite value, thus total observations may be more than the number of individu als.

PAGE 79

79 T he age of the study sampl e ranged between 18 and 85 years, with a mean age of 50.19 years and a standard deviation of 17.75 years. T he study sampl e consisted of 3,482 (26.7%) Medicare beneficiaries, 1,630 (12.5%) Medicaid beneficiaries, and 9,747 (7 4. 9 %) private plan enrollees. There were some overlaps, since some individuals may have a full year of insurance coverage with Medicare and Medicaid dual eligibility, plus Medicare supplemental plans. The study sampl e had a mean PCS score of 48.19 ranging from 5.68 to 70.98 with a standard deviation of 11.35. In Table 5 2, the characteristics of the study sampl e are described. The majority of this study sample are middle age d individuals ( 54.7% ) female ( 55. 4% ) non Hispanic White ( 61.7 % ) living the S outh ( 3 5. 4 % ) and are urban ( 83.0 % ) married individuals ( 60.1% ) with college education ( 50. 1 % ) and high income ( 41. 1 % ) Variable Operationalization Missing Cost sharing Values I mputation In this study cost sharing values in physician care were calculated by dividing out of pocket expenditures by total expenditures for each event by each person Because of how cost sharing was calculated, some missing cost sharing values would be created in STATA for individuals because of their zero medical expenditure in th e denominator. Eliminating these missing values may cause potential sample selection bias if the remaining observations are unrepresentative of the population T o avoid this problem these missing values were replaced with mean values within corresponding insurance types. Insurance types can be separated into seven subgroups in this study: Medicare, Medicaid private plans Medicare Medicaid, Medicar e private plans, Medicaid private plans, and Medicare Medicaid private plans Since 9,906 individuals the maj ority of this study sample ha ve cost sharing values, this imputation approach should not cause substantial bias. In this way all the missing values were imputed.

PAGE 80

80 Test for the M ulticollinearity between D isease Severity and Priority C onditions The disease s everity dummy variables and priority condition variables may be correlated, because the former are self perceived summary health status health conditions. The covariance matrix of disease severity and priority conditions indicated all of them are consistently highly correlated ( p <0.01 Table 5 3). Andersen Andersen applies to statistical models t hat contain either perceived or evaluated needs but not both. Thes e two types of needs can be substitutes when available data only contains measures of one or the other type s since disease severity is the main variable of intere st and represents were dropped from the model. Keeping both group variables w ould hurt the model by inflating the variance and underestimat ing the significance of coefficients. Test for P oten tial E ndogeneity and IV V alidity The Hausman T est The Hausman test can be employed to test for endogeneity by regressing cost sharing on all exogenous variables, then adding the residual as a new variable into the initial structur al equation. The significa nt coefficient for the residuals indicate s the existence of endogeneity. In this case, as a solution, an instrumental or identifying variable would be introduced that affect s cost sharing level but not utilization and expenditure s In the first step of th e Hausman test the physician care cost sharing variable was used as a dependent variable and regressed on all exogenous variables, then its error term was obtained. In the second step when the error term of the cost sharing in physician care was entered into the initial structural equation this error term was highly significantly associated with physician care utilization and expenditure variables ( p < 0.0 1 for both tests).

PAGE 81

81 Test for IV R elevance and E xogeneity As planned, an instrumental variable w as the n introduced to address endogeneity. T he F statistic was examined i n the first stage of 2SLS. Meanwhile, this study examined the association between the error term of utilization or expenditure and IV The results a re reported in Table s 5 4 and 5 5 In Tab le 5 4 the F s tatistic i n the first stage of 2SLS was much higher than 10 ( p <0.01) indicating the i dentifying variable pre mium was relevant to the endogenous regressor of the cost sharing variable. Next, the association between the identifying variable and the error terms in the structural equation of health care utilization and expenditure was tested, with p = 0.23 and p = 0.27 respectively, indicating the identifying variable was not significantly associated with these two dependent variables. Therefo re, the premium variable met the requirement of a good IV. Use IV to G et P redicted Cost sharing V alues Using the first stage regression of 2SLS, predicted cost sharing values can be obtained, and the initial cost sharing values can be replaced by these p redicted ones. Based on this continuous cost sharing variable, a dichotomous cost sharing variable was then created with 20% as a cutoff point in Medicare and private plan s and the mean cost sharing value of 2.9% as a cutoff point for Medicaid. This dummy cost sharing variable examine s whether high cost sharing relative to low cost sharing in physician care reduce s health care utilization and expenditure s differently by health status Next, two interaction terms were further generated as products of the du mmy cost sharing variable and two disease severity variables severe and moderate conditions. Captured by the interaction term of severe condition and cost sharing dummy variables, the differential impact can be examined. In the final analysis, the interac tion term was dropped in both physician care expenditur e and total expenditure models the findings for tw o important outcome dimensions due to

PAGE 82

82 vary by the pred ictor interaction term. In order to get complete findings, this study used 19% as a proxy for the 20% cutoff point to divide individuals into high and low cost sharing groups, which reclassified 2,416 observations into high cost sharing group, 3.98% of the total observations. The results remain ed similar, but with a gain of complete results. The d escription of cost sharing variable i s then reported in Table 5 6. Note that in this table Medicaid insurance refers to both Medicaid coverage alone and Medicaid i n combination with Medicare or private plans, as do Medicare and private plans. When a 19% cutoff point in private and Medicare plan s is determined first, this may include some Medicaid enrollees due to their dual eligibility. Thus, the above mean cost sha ring value of 2.9% for Medicaid refers to the remaining beneficiaries with Medicaid coverage alone, so this mean value may differ slightly from the 2.7% mean value reported in Table 5 6. The M ultivariate A nalysis R esults The H ealth C are U tilization R esults Physician c are u tilization The hypothesis is that high cost sharing reduces physician care utilization less for the severely ill group than for the healthier group. The regression is expressed as ) = 0 + 1 cost sharing + 2 *severity + 3 cost sharing *severity + 4 C oefficient could be interpreted as the difference between the log of expected counts x+ 1 ) x x+ 1 x ), meani ng change in a dependent variable given each unit change or group difference in the independent variable x from x to x+1 Since the counts of an event occur in a given period by default, the count can then be considered an inciden ce rate, thus the can be in terpreted as the log inciden ce rate ratio. Intuitively,

PAGE 83

83 the incidence rate ratio is the exponentia t ed coefficient In this study the coefficient could be interpreted as the log incidence rate ratio of seeing a physician in 2007 between groups The diff erential impact in physician visits was analyzed in a negative binomial regression. The results a re reported in Table 5 7. Holding covariates constant, the high cost sharing group had an insignificant 15% less incidence rate of seeing a physician for indiv iduals relative to the low cost sharing group ( p = 0.09) T he severely ill had significant ly higher incidence rate (156%) of physician care visits relative to healthier individuals ( p <0.01 ). The incidence rate difference in physician care visits between h igh and low cost sharing groups was 8% different between the severely ill and the healthier group which was not significant ( p = 0.67). In this model significant results were also found for other covariates. Compared with the reference group, the signif icant higher physician care visits were found for the White individuals ( p = 0. 0 2) living in urban areas ( p <0.01 ) with low income ( p = 0.03) with moderate conditions ( p <0.01 ) and primary care physicians ( p = 0.04). In summary, the high cost sharing gro up was not different in physician care visits from the low cost sharing group, the severely ill had more physician care visits than the healthier group and the physician care visit difference between high and low cost sharing groups was not different betw een the severely ill and the healthier group Primary c are p hysician u tilization The hypothesis is that high cost sharing reduces p rimary care utilization less for the severely ill group than for the healthier group. The differential impact in primary ca re physician visits was analyzed in a negative binomial regression. The results a re reported in Table 5 8. Holding covariates constant, the high cost sharing group had an insignificant 6% higher incidence rate of seeing a primary care physician relative to the low cost sharing group ( p = 0.16). The severely ill had a significant 9%

PAGE 84

84 l ower incidence rate of primary care physician visits relative to healthier individuals ( p = 0.05). The incidence rate difference in primary care physician visits between high an d low cost sharing groups was 16% different between the severely ill and the healthier which was not significant ( p = 0.09). In this model significant results were also found for other covariates. Compared with the reference group, the significant high er physician care visits were found for the middle aged ( p < 0.0 1 ) and older ( p < 0.01) African American individuals ( p = 0.05) living in the West ( p = 0.02) and rural areas ( p < 0.01) unmarried ( p < 0.01) and with primary care physicians ( p < 0.01) Meanwhil e, the significant lower physician care visits were found for the high income individuals ( p = 0.0 2 ) and those with college education ( p < 0.01) In summary, the high cost sharing group was not differ ent in primary care physician visits from the low cost sh aring group. T he severely ill had fewer primary care physician visits than the healthier group since most of the time they may go to see specialty physicians based on their diagnosis and condition The primary care physician visit difference between high a nd low cost sharing groups was not different between the severely ill and the healthier Although insignificant, the positive sign of cost sharing and interaction term indicated t hese anomal ous which will be further explored in the later part of this study. E R c are u tilization As stated in the hypotheses the direct differential impact in MD visit s was not significant, which may suggest that the severely ill perhaps use more downstream services to mee t their ne eds but suppressed health care need due to high cost sharing pressure for MD visit. Thus, this study also examined and reported the results in the downstream services. The hypothesis wa s

PAGE 85

85 that high cost sharing in physician care leads to more incr eased ER care utilization for severely ill people than for the healthier group. The differential impact in ER visits was analyzed in a negative binomial regression. The results a re reported in Table 5 9. Holding covariates constant, the h igh cost sharing g roup had an insignificant 4% l ower incidence rate of ER visits relative to the low cost sharing group ( p = 0.80). The severely ill had an insignificant 19% higher incidence rate of ER visits relative to healthier individuals ( p = 0.08). The incidence rate difference in ER visits between high and low cost sharing groups was 122% significantly different between the severely ill and the healthier ( p = 0.02) T he positive coefficient and over 100% incidence rate ratio means the severely ill reduced ER visits le ss than the healthier In this model significant results were also found for other covariates. Compared with the reference group, significant ly higher ER visits were found for unmarried individuals ( p = 0.01), while cost sharing in ER was significantly associated with fe wer ER visits ( p < 0.01 ). In summary, the two main effects of cost sharing and sever it y variables did not affect ER care utilization. Individuals with high physician care cost sharing had similar ER visits to those with low physician care cost sharing the severely ill had similar ER visits to the healthier The ER visit difference between high and low cost sharing groups was less for the severely ill than the healthier These results were consistent with those for physician care visits bu t did not demonstrate an expected offset effect. At this point it is hard to judge whether these results were reasonable or not I t is better to examine the expenditure results in ER visit as well and see if there is a systematic pattern. If those results are inconsistent with this one, further analysis would be needed.

PAGE 86

86 Inpatient c are u tilization The hypothesis wa s that high cost sharing in physician care leads to more increased inpatient care utilization for severely ill people than for the healthier grou p. The differential impact in hospitalizations was analyzed in a negative binomial regression. The results a re reported in Table 5 10. Holding covariates constant, the h igh cost sharing group had an insignificant 12 % higher incidence rate of hospital admis sions relative to the low cost sharing group ( p = 0. 31 ). The severely ill had a significant 40 % higher incidence rate of hospital admissions relative to healthier individuals ( p < 0.0 1 ). The incidence rate difference in hospital admissions between high and low cost sharing groups was 9 % different between the severely ill and the healthier which was not significant ( p = 0. 68 ). In this model significant results were also found for other covariates. Compared with the reference group, the significant higher inpatient care visits were found for the middle aged ( p = 0.0 1 ) male ( p = 0.03) African American individuals ( p = 0.03) who were in the near poor group ( p = 0.0 1 ) w hile cost sharing in hospitalizations was significantly associated with lower inpatient car e ( p = 0.03). In summary, the high physician care cost sharing group was not different in hospital admissions from the low physician care cost sharing group, the severely ill had more hospital admissions than the healthier the hospital admission differenc e between high and low physician care cost sharing groups was not different between the severely ill and the healthier The H ealth C are Expenditure R esults Physician c are e xpenditure The hypothesis wa s that high cost sharing reduces physician care expen ditures less for the severely ill group than for the healthier group. The differential impact in physician visit expenditures was analyzed in a two part model. In part one a logit regression was used to predict

PAGE 87

87 the probability of any expenditure reduction difference between the severely ill and the reference group I n part two a log transformed regression was employed to indicate the difference in magnitude in expenditure reduction given those who had any medical expenditures. The results of the regressio ns are presented in Table s 5 11 and 5 12. Part I. H igh cost sharing policy holders insignificant ly had 0.91 times the odds of having any physician care expenditures relative to the low cost sharing group ( p = 0.46), the severely ill had 13.80 times the odd s of having any physician care expenditures relative to healthier individuals ( p < 0.01 ), and their probability of reducing any physician care expenditures was also significantly less than the reference group, with 10.81 times the odds ( p = 0.02). In this model significant results were also found for other covariates. Compared with the reference group, the significant higher odds of having any physician care expenditures were found for the middle aged ( p < 0.01 ) and older ( p < 0.01 ) White ( p < 0.01 ) female individuals ( p < 0.01 ), who had high school ( p < 0.01 ) and college education ( p < 0.01 ), moderate conditions ( p < 0.01 ) and primary care physicians ( p < 0.01 ). However, the near poor and low income individuals both had significant ly lower odds of having any ph ysician care expenditures than the poor ( p = 0.02 and 0.00 respectively). Since commands for goodness of fit assessments, such as a likelihood ratio test, Hosmer Lemeshow (H L) test, R2, or this study directly evaluated the logit model fit with the area under a ROC curve as a proxy approach. An area under a ROC curve greater than 0.7 is a predict or of good model fit according to common practice. The area under a ROC cur ve for the logit regression wa s 0. 8450, indicating the logit model fit was fairly good

PAGE 88

88 In summary, i ndividuals with high cost sharing had similar odds of having any physician care expenditures than those with low cost sharing the severely ill had a highe r probability for having any physician visit expenditures than the healthier and their odds of having lower physician care expenditures w ere significantly less than for the healthier Part II. H igh cost sharing policy holders had significant ly lower phy sician care expenditures c ompared to the low cost sharing group ( p = 0.01) T he severely ill had significant ly higher physician care expenditures than the healthier ( p < 0.01 ), but expenditure difference between high and low cost sharing groups was not diff erent between the severely ill and the healthier ( p = 0.70). In this model significant results were also found for other covariates. Compared with the reference group, the significant higher physician care expenditures were found for the middle aged ( p = 0.01) and older ( p = 0.01) White ( p = 0.04) individuals, who were married ( p < 0.01 ) living in urban area s ( p < 0.01 ) with moderate health conditions ( p < 0.01 ) and near poor income ( p = 0.02), mid dle income ( p < 0.01 ), and high income ( p < 0.01 ). However, those living in the South had significant lower physician care expenditures than those in the Midwest ( p = 0.01). The log transformation model of physician care expenditures for those who had some expenditures yielded an error term with a skewness of 0.05 and kurtosis of 3.16. Figure 5 1 is the standardized normal probability (P from normality in the middle range of data. Figure 5 2 is the quantiles of residuals against quantiles of normal distributio normality near the tails. These f igure s show that the residuals were close to a normal distribution. The p = 0 .95). Model goodness of fit statistics, such as the Hosmer Lemeshow test and AIC, cannot be

PAGE 89

89 prediction plot (Figure 5 3) show s that the model fit was sticity for the overall model ( p < 0.01 ) but no heteroskedasticity for the two main effects of cost sharing and severity variables and their interaction variable ( p = 0.76, 0.17 and 0.46 respectively), which are the The marginal m agnitude would be calculated if the differential impact was significant in the log transformed regression. Smear ing estimators for each severity group would then be calculated after OLS regression retransformation to correct for heteroskedasticity, since t a In summary, in dividuals with high cost sharing had less physician care expenditures than those with low cost sharing the severely ill had higher physician care expenditures than the healthie r, but expenditure difference s between high and low cost sharing groups w ere not different between the severely ill and the healthier The two parts were in consistent in the signs of the interaction variables; specifically the differential impact was sig nificant in part I but not in part II. The difference in physician care expenditures between high and low physician care cost sharing groups was less for the severely ill than the healthier group for the probability of having any expenditure but not in t heir actual expenditures. ER v isit e xpenditure The hypothesis wa s that high cost sharing in physician care leads to more increased ER care expenditures for severely ill people than for healthier people The differential impact in ER visit expenditures was analyzed us in g a two part model. In part one a logit regression was used to predict the probability of any expenditure reduction difference between the severely ill and the reference group I n part two a log transformed regression was employed to indica te the

PAGE 90

90 difference magnitude in expenditure reduction given those who had any medical expenditures. The results of the regressions are presented in Table s 5 13 and 5 14. Part I. H igh cost sharing policy holders had insignificant 3.41 times the odds of havin g any ER care expenditures relative to the low cost sharing group ( p = 0.16), the severely ill had insignificant 0.47 times the odds of having any ER care expenditures relative to healthier individuals ( p = 0.23), and their probability of increasing any ER expenditures was also insignificantly higher than the reference group, with 1.99 times more odds ( p = 0.60). In this model significant results were also found for other covariates. Compared with the reference group, significant higher odds of having any ER care expenditures were found for middle income individuals ( p = 0.04). However, individuals with moderate conditions had significant 0.08 times lower odds for having any ER care expenditures ( p = 0.04). The area under the ROC curve for the logit regre ssion wa s 0. 7726, indicating logit model fit was fairly good In summary, in dividuals with high physician care cost sharing did not have a different probability of having any ER expenditures than those with low physician care cost sharing the severely ill did not have a different probability of having any ER expenditures than the healthier and the probability of having any ER expenditure difference between high and low physician care cost sharing groups was not different between the severely ill and the h ealthier Part II. H igh cost sharing policy holders had insignificant ly higher ER care expenditures relative to the low cost sharing group ( p = 0.48), the severely ill had insignificant ly higher ER care expenditures relative to healthier individuals ( p = 0. 41 ), and the ER expenditure difference between high and low physician care cost sharing groups was not different between the severely ill and the healthier ( p = 0.64).

PAGE 91

91 In this model significant results were also found for other covariates. Compared wi th the reference group, the significant ly higher ER care expenditures were found for middle aged individuals ( p = 0.04), while cost sharing in ER was significantly associated with lower ER care expenditures ( p < 0.01 ). The log transformation model of physic ian care expenditures for those who had actual expenditures yielded an error term with a skewness of 0.21 and kurtosis of 3.35. Figure 5 4 is the standardized normal probability (P from normality in the middle range of data. Figure 5 5 shows the quantiles of residuals against the normality near the tails. These f igure s show that the residuals were close to a normal distrib ution. The residual prediction plot (Figure 5 6) show s that the model fit was well. Model goodness of fit statistics, such as the Hosmer good linearity with an insignificant quadratic term of model fit ( p = for the overall model ( p = 0.54). In summary, in dividuals with high physician care cost sharing did not have different ER expenditu res than those with low physician care cost sharing the severely ill did not have different ER expenditures than the healthier and ER expenditure difference s between high and low physician care cost sharing groups was not different between the severely i ll and the healthier N either of the two part s demonstrated a significant differential impact. Inpatient c are e xpenditure The hypothesis wa s that high cost sharing in physician care leads to more increased inpatient care expenditures for severely ill peo ple than for the healthier group. The differential

PAGE 92

92 impact in inpatient care expenditures was analyzed in a two part model. In part one a logit regression was used to predict the probability of any expenditure reduction difference between the severely ill and the reference group I n part two a log transformed regression was employed to indicate the difference magnitude in expenditure reduction given those who had any medical expenditure The results of the regressions are presented in Table s 5 15 and 5 16. Part I. H igh cost sharing policy holders had 0.64 times the odds of having any inpatient care expenditures relative to the low cost sharing group ( p = 0.05), the severely ill had 3.37 times the odds of having any inpatient care expenditures relative to h ealthier individuals ( p < 0.0 1 ), and the probability of having any inpatient care expenditure difference between high and low physician care cost sharing groups was not different between the severely ill and the healthier with 1.00 times the odds ( p = 1.00 ). In this model significant results were also found for other covariates. Compared with the reference group, significant ly higher odds in inpatient care expenditures were found for individuals with moderate conditions ( p = 0.01). However, the middle ag ed ( p < 0.0 1 ) Hispanic ( p = 0.04) individuals with middle income s ( p < 0.01 ) and high income ( p < 0.0 1 ) had significant ly lower odds of having any inpatient care expenditures. The area under the ROC curve for the logit regression wa s 0. 6816, near 0.7, indica ting logit model fit was basically satisfactory In summary i ndividuals with high physician care cost sharing had lower odds of having any inpatient care expenditures, the severely ill had higher odds of having any inpatient care expenditures, and the pro bability of having any inpatient care expenditure difference between high and low physician care cost sharing groups was not different between the severely ill and the healthier

PAGE 93

93 Part II. H igh cost sharing policy holders had insignificant ly higher inpati ent care expenditures relative to the low cost sharing group ( p = 0.43), the severely ill had significant ly higher inpatient care expenditures relative to healthier individuals ( p < 0.01 ), and their expenditure increase was not significantly different from the reference group ( p = 0.45). In this model significant results were also found for other covariates. Compared with the reference group, the significant ly higher inpatient care expenditures were found for the middle aged ( p = 0.01) and older ( p = 0.05) married ( p < 0.01 ) individuals, who were White ( p = 0.0 1 ) or African American ( p < 0.01 ). However, those living in the South had significant ly lower inpatient care expenditures than in the Midwest ( p = 0.01) Cost sharing in inpatient care was significantl y associated with lower inpatient care expenditures ( p = 0.01). The log transformation model of physician care expenditures for those who had actual expenditures yielded an error term with a skewness of 0.86 and kurtosis of 6.90. Figure 5 7 is the standar dized normal probability (P from normality in the middle range of data. Figure 5 8 is the quantiles of residuals against normality near the tails. Alth ou gh a kurtosis of 6.9 is not near normal t hese f igure s show that the residuals were basically close to a normal distribution. The residual prediction plot (Figure 5 9) showed that the model fit was test indicated good linearity with an insignificant quadratic term of model fit ( p = overall model ( p = 0.07), and no heteroskedasticity for the two main effects of cost sharing and severity va riables and their interaction variable ( p = 0.85, 0.69 and 0.41 respectively). In summary, i ndividuals with high cost sharing in physician care did not have different inpatient care expenditures from those with low physician care cost sharing the sever ely ill had

PAGE 94

94 higher inpatient care expenditures than the healthier and inpatient care expenditure difference between high and low physician care cost sharing groups was not different between the severely ill and the healthier Similar to the consistent re sults of insignificant differential impact o n ER care expenditures, neither of t he two parts in inpatient care expenditures demonstrated a significant differential impact. However, in a utilization model the differential impact was significant for ER visit s, but insignificant for hospital admissions. This may suggest that ER and inpatient care are complements that together serve as the downstream services for physician care. If the ER service can meet the unmet medical care needs for physician care for the severely ill, then hospitalizations would be redundant It may also be possible that the offset effect was evenly distributed between ER and inpatient care, so these two options attenuate the significant differential impact of either of them. Total c are e xpenditure The hypothesis wa s that high cost sharing in physician care reduces total expenditures for the severely ill group less than for the healthier group. The differential impact in total care expenditures was analyzed in a two part model. T otal care refers to the sum of physician care, ER visits and inpatient care. In part one a logit regression was used to predict the probability of any expenditure reduction difference between the severely ill and the reference group, and in part two a log transfor med regression was employed to indicate the difference magnitude in expenditure reduction given those who had any medical expenditures. The results of the regressions are presented in Table s 5 17 and 5 18. Part I. H igh cost sharing policy holders had simi lar odds of having any medical care expenditures relative to the low cost sharing group ( odds ratio = 0.95, p = 0.65), the severely ill had 15.05 times the odds of having any medical care expenditures relative to healthier

PAGE 95

95 individuals ( p < 0.01 ), and their odds of having lower total care expenditures was significantly less than the healthier with 8.85 times the odds ( p = 0.04). In this model significant results were also found for other covariates. Compared with the reference group, the significant ly hig her odds of having any medical care expenditures were found for the middle aged ( p < 0.01 ) older ( p < 0.01 ) White ( p < 0.01 ) female individuals ( p < 0.01 ), who had high school ( p < 0.01 ) or college education ( p < 0.01 ), moderate conditions ( p < 0.01 ) and prim ary care physicians ( p < 0.01 ) living in an urban area ( p = 0.04). However, the near poor ( p = 0.03) and low income ( p = 0.03) individuals and individuals who live in the West ( p = 0.02) had significant ly lower odds of having any medical care expenditures t han the reference groups. The area under the ROC curve for the logit regression wa s 0. 8466, indicating logit model fit was fairly good In summary, i ndividuals with high physician care cost sharing did not have a different probability of having any total medical care expenditures relative to low cost sharing group the severely ill had higher odds of having any total medical care expenditures relative to healthier individuals and their odds of having lower total medical expenditures associated with high p hysician care cost sharing was less than the healthier individuals Th ese results w ere essential ly consistent with the hypotheses. Part II. H igh cost sharing policy holders had significant ly lower total medical care expenditures relative to the low cost s haring group ( p = 0.03), the severely ill had significant ly higher total medical care expenditures relative to healthier individuals ( p < 0.01 ), but total medical care expenditure difference between high and low physician care cost sharing groups was not di fferent between the severely ill and the healthier ( p = 0.55).

PAGE 96

96 In this model significant results were also found for other covariates. Compared with the reference group, significant ly higher total medical care expenditures were found for married individu als ( p < 0.01 ) those living in an urban area ( p < 0.01 ) individuals with moderate conditions ( p < 0.01 ) and high income ( p < 0.01 ). However, those living in the South and West had significant ly lower total medical care expenditures than those in the Midwest ( p = 0.01), while cost sharing in ER visits was significantly associated with lower total medical care expenditures ( p < 0.01 ). The log transformation model of physician care expenditures for those who had actual expenditures yielded an error term with a s kewness of 0.11 and kurtosis of 2.75. Figure 5 10 is the standardized normal probability (P from normality in the middle range of data. Figure 5 11 is the quantiles of residuals against quantiles of normality near the tails. These figur es showed that the residuals were close to a normal distribution. The residual prediction plot (Figure 5 12) showed that the model fit was well. The Pre test indicated good linearity with an insignificant quadratic term of model fit ( p = 0.96). ( p = 0.07), and no heteroskedasticity for the two main effects of cost sharing and severity variables and their interaction variable ( p = 1.00, 0.68 and 0.10 respectively). In summary in dividuals with high cost sharing in physician care had less total medical care expenditures relative to the low cost sharing group the severely ill had higher total medical care expenditures relative to healthier individuals but the expenditure difference between high and low physician care cost sharing groups was not different between the severely ill and the healthier

PAGE 97

97 The differential impact was significant in part I but not in part II. The probability of having any total medical expenditures, rather than act ual total medical expenditures in the difference between high and low physician care cost sharing groups was less for the severely ill indiv iduals than the healthier individuals. Since the differential impact in expenditure models in physician care and total medical care was split, this study further examined an overall effect that incorporated these two parts using the bootstrap technique. S amples were replicated 1,000 times. The results indicated that for physician care the differential impact was $3 052.36, with a 95% confidence interval ranging from $5 180.53 to $924.20. Similarly, for total medical care the differential impact was $12 853.23, with a 95% confidence interval ranging from $17,582.86 to $8,123.60. Both results were significant, meaning severely ill individuals actually reduced expenditure in physician care and total medical care more than the general health population ; the sick population was even more sensitive to high cost sharing pressure than the general health population. The results further indicated a significant physician care reduction within the severely ill group ( p < 0.01 for both utilization and expenditure s).

PAGE 98

98 Table 5 1. Description of the study sample (N = 13,020) Variable Mean SD Minimum Maximum Age 50.19478 17.75128 18 85 PCS 48.19457 11.35009 5.68 70.98

PAGE 99

99 Tabl e 5 2. Study sample characteristics (N = 13,020) Variable Number % Age (yrs) 18 34 2,832 21.75 35 64 7,122 54.70 3,066 23.55 Gender Male 5,809 44.62 Female 7,211 55.38 Race Non Hispanic white 8,035 61.71 Non Hispanic African American 2,021 15.52 Hispanic 1,987 15.26 Other 977 7.51 Geographic area Northeast 2,203 16.92 Midwest 3,006 23.09 South 4,606 35.38 West 3,205 24.62 Rural/urban location Urban 10,804 82.98 Rural 2,216 17.02 Marital status Married 7,827 60.12 No t married 5,193 39.88 Education level (yrs) 0 8 1,047 8.04 9 12 5,453 41.88 13 17 6,520 50.08 Income Poor 1,503 11.54 Near poor 551 4.23 Low income 1,648 12.66 Middle income 3,973 30.51 High income 5,345 41.05 Co morbidities Hypertension 4,661 35.89 No hypertension 8, 326 64.11 Coronary heart disease 744 5.73 No coronary heart disease 12,238 97.47 Angina 407 3.13 No angina 12,578 96.87 Heart attack 536 4.13 No he art attack 12,457 95.87

PAGE 100

100 Table 5 2. Continued Variable Number % Other heart disease 1,188 9.15 No other heart disease 11,799 90.85 Stroke 552 4.25 No stroke 12,445 95.75 Diabetes 1,543 11.87 No diabetes 11,452 88.13 Arthritis 3,565 27.48 No arthritis 9,406 75.52 Asthma 1,284 9.89 No asthma 11,705 90.11 Gatekeeper status Gatekeeper 4,854 37.28 No gatekeeper 8,166 62.72 Cost sharing level Low 6,939 53.29 High 6,081 4 6 71 Disease severity level Top tertile PCS 4,369 33.56 Middle tertile PCS 4,316 33.15 Bottom tertile PCS 4,335 33.29

PAGE 101

101 Table 5 3 Association between disease severity and priority conditions ( p < 0.01 ) Pearson Chi square df = 1 Severe Moderate Hypertensio n 3.8e+03 189.64 Coronary heart disease 2.4e+03 26.81 Angina 2.3e+03 88.05 Heart attack 1.8e+03 28.02 Other heart disease 2.3e+03 10.43 Stroke 2.2e+03 70.55 Diabetes 3.5e+03 16.03 Arthritis 7.5e+03 50.55 Asthma 1.2e+03 55.66

PAGE 102

102 Table 5 4 T he first stage of IV 2SLS results with c ost sharing as dependent variable Variable Coef. S.E. p value Age (yrs) 18 34 Reference 35 64 0.01 0.00 0.00 0.09 0.00 0.00 Gender Female Reference Male 0.00 0.00 0.00 Race Other Reference Non His panic white 0.02 0.00 0.00 Non Hispanic African American 0.01 0.00 0.00 H ispanic 0.01 0.00 0.00 Geographic area Midwest Reference Northeast 0.01 0.00 0.00 South 0.03 0.00 0.00 West 0.03 0.00 0.00 Rural/urban location Rural Reference Urban 0.01 0.00 0.00 Marital status Not married Reference Married 0.03 0.00 0.00 Education level (yrs) 0 8 Reference 9 12 0.01 0.00 0.00 13 17 0.01 0.00 0.00 Income Poor Reference Near poor 0.00 0.00 0.01 Low income 0.00 0.00 0.57 Middle income 0.01 0.00 0.00 High income 0.02 0.00 0.00 Gatekeeper status No gatekeeper Reference Gatekeeper 0.00 0.00 0.65 Disease severity level Top tertile PCS Reference Middle tertile PCS 0.03 0.00 0.00 Bottom tertile PCS 0.07 0.00 0.00 Premium 2.45e 06 0.00 0.00

PAGE 103

103 Table 5 5 The association between IV and the error terms in the structural equation of health care utilization and expenditure Variable Coef. S.E. p value Physician care utilization e rror 0.00 0.00 0.23 Physician care expenditure error 0.11 0.10 0.27

PAGE 104

104 Table 5 6. Cost s haring value description by insurance types Insurance type Low level ( % ) High level (%) Mean SD Medic are 91.67 8.33 0. 086 0.111 Medicaid 90.44 9.56 0.027 0.088 Private plans 56.59 43.41 0.174 0.064 Total sample 67.00 33.00 0.137 0.107

PAGE 105

105 Table 5 7. Negative binomial regression estimates for physician visits Variable Coef. S.E. IRR S.E. p value Cost sharing Low Reference High 0.16 0.09 0.85 0.08 0.09 Disease severity level Top tertile PCS Reference Middle tertile PCS 0.51 0.09 1.66 0.14 0.00 Bottom tertile PCS 0.94 0.08 2.56 0.21 0.00 Cost sharing Bottom tertile PCS 0.08 0.19 1.08 0.20 0.67 Middle tertile PCS 0.03 0.12 0.97 0.12 0.82 Age (yrs) 18 34 Reference 35 64 0.06 0.08 1.06 0.08 0.48 0.15 0.09 1.16 0.1 1 0.10 Gen der Female Reference Male 0.07 0.05 0.93 0.05 0.16 Race Other Reference Non Hispanic white 0.22 0.10 1.25 0.12 0.02 Non Hispanic African American 0.17 0.13 1.19 0.16 0.20 Hispanic 0.24 0.13 1.28 0.17 0.07 Geographic area Midwest Reference Northeast 0.09 0.08 1.09 0.08 0.25 South 0.05 0.07 0.95 0.06 0.45 West 0.09 0.07 0.91 0.06 0.18 Rural/urban location Rural Reference Urban 0.30 0.06 1.36 0.08 0.00 Marital status No t married Reference Married 0.01 0.05 0.99 0.05 0.87 Education level (yrs) 0 8 Reference 9 12 0.23 0.13 0.80 0.10 0.08 13 17 0.02 0.12 0.98 0.12 0.88 In come Poor Reference Near poor 0.21 0.20 1.23 0.24 0.30

PAGE 106

106 Table 5 7 Continued Variable Coef. S.E IRR S.E. p value Low income 0.20 0.09 0.82 0.08 0.03 Middle income 0.03 0.10 0.97 0.09 0.72 High income 0.09 0.09 1.10 0.10 0.30 Gatekeeper status No gatekeeper Reference Gatekeeper 0.11 0.05 1.12 0.06 0. 04

PAGE 107

107 Table 5 8 N egative binomial regression estimates for primary care physician visits Variable Coef. S.E. IRR S.E. p value Cost sharing Low Reference High 0.07 0.05 1.0 7 0.05 0.16 Disease severity level Top tertile PCS Reference Middle tertile PCS 0.03 0.05 0.97 0.04 0.50 Bottom tertile PCS 0.10 0.05 0.91 0.04 0.05 Cost sharing Bottom tertile PCS 0.15 0.09 1.16 0.10 0.09 Middle tertile PCS 0.00 0.06 1.00 0.06 0.96 Age (yrs) 18 34 Reference 35 64 0.33 0.04 1. 39 0.06 0.00 0.33 0.04 1. 39 0.06 0.00 Gender Female Reference Male 0.04 0.02 1.04 0.0 3 0.10 Race Other Reference Non Hispanic white 0.01 0.05 0.99 0.0 5 0.79 Non Hispanic African American 0.1 2 0.06 1.12 0.07 0.05 Hispanic 0.06 0.06 1.06 0.06 0.33 Geographic area Midwest Reference Northeast 0.08 0.04 0.92 0.04 0.05 South 0.04 0.04 0.97 0.04 0.36 West 0.09 0.04 1.09 0.04 0.02 Rural/urban location Rural Reference Urban 0.11 0.03 0.89 0.03 0.00 Marital status No t married Reference Married 0.08 0.03 0.92 0.02 0.00 Education level (yrs) 0 8 Reference 9 12 0.05 0.05 0.95 0.05 0.33 13 17 0.20 0.05 0.82 0.04 0.00 Income Poor Reference Near poor 0.07 0.09 0.93 0.08 0.44

PAGE 108

108 Table 5 8 Continued Variable Coef. S.E. IRR S.E. p value Low income 0.00 0.05 1.00 0.05 0.96 Middle income 0.02 0.04 0.98 0.04 0.72 High income 0.10 0.04 0.90 0.04 0.02 Gatekeeper status No gatekeeper Reference Gatekeeper 0.10 0.03 1.11 0.03 0.00

PAGE 109

109 Table 5 9 N egative binomial regression estimates for ER visits Variable Coef. S.E. IRR S.E. p value Cost sharing Low Reference High 0.04 0.17 0.96 0.17 0.80 Disease severity level Top tertile PCS Reference Middle tertile PCS 0.12 0.11 1.13 0.12 0.28 Bottom tertile PCS 0 .17 0.10 1.1 9 0.12 0.08 Cost sharing Bottom tertile PCS 0.80 0.32 2.22 0.72 0.01 Middle tertile PCS 0.13 0.18 0.87 0.16 0.45 ER cost sharing 0.66 0.14 0.51 0.07 0.00 Age (yrs) 18 34 Reference 35 64 0.16 0.13 1.17 0.15 0.23 0. 10 0.13 0.91 0.1 1 0.44 Gender Female Reference Male 0.01 0.07 1.01 0.07 0.87 Race Other Reference Non Hispanic white 0.08 0.10 1.09 0. 11 0.40 Non Hispanic African American 0.13 0.16 1.14 0.19 0.43 Hispanic 0.03 0.11 0.97 0.11 0.78 Geographic area Midwest Reference Northeast 0.06 0.14 1.06 0.15 0.66 South 0.06 0.12 0.94 0.11 0.60 West 0.14 0.11 0.87 0.10 0.22 Rural/urban location Rural Reference Urban 0.02 0.07 0.98 0.07 0.82 Marital status No t married Reference Married 0.25 0.09 0.78 0.07 0.01 Education level (yrs) 0 8 Reference 9 12 0.02 0.12 1.02 0.12 0.87 13 17 0.02 0.12 1.02 0.12 0.88 Inco me Poor Reference Near poor 0.18 0.11 0.84 0.09 0.11

PAGE 110

110 Table 5 9 Continued Variable Coef. S.E. IRR S.E. p value Low income 0.06 0.11 0.94 0.10 0.57 Middle income 0.04 0.14 1.04 0.15 0.78 High income 0.09 0.11 1.09 0.13 0.45 Gatekeeper status No gatekeeper Reference Gatekeeper 0.03 0.09 0.97 0.09 0.76

PAGE 111

111 Table 5 10 N egative binom ial regression estimates for hospital admissions Variable Coef. S.E. IRR S.E. p value Cost sharing Low Reference High 0.11 0.11 1.12 0.12 0.31 Disease severity level Top tertile PCS Reference Middle tertile PCS 0.19 0.10 1.21 0.12 0.06 Bottom tertile PCS 0.34 0.07 1.40 0.10 0.00 Cost sharing Botto m tertile PCS 0. 09 0. 21 1. 09 0. 23 0. 68 Middle tertile PCS 0.29 0.13 0.75 0. 10 0.03 Inpatient care cost sharing 0.56 0.26 0.5 7 0.15 0.03 Age (yrs) 18 34 Reference 35 64 0.18 0.07 1 .20 0.09 0.01 0.11 0.09 1.12 0.10 0.22 Gender Female Reference Male 0.16 0.07 1.17 0.09 0.03 Race Other Reference Non Hispanic white 0.16 0.09 1.17 0.10 0.07 Non Hispanic African American 0.30 0. 14 1.35 0.19 0.03 Hispanic 0.17 0.10 1.19 0.12 0.09 Geographic area Midwest Reference Northeast 0.01 0.10 1.01 0.11 0.91 South 0.01 0.08 0.99 0.08 0.90 West 0.07 0.08 0.93 0.08 0.43 Rural/urban location Rural Reference Urban 0.00 0.09 1.00 0.09 0.99 Marital status No t married Reference Married 0.01 0.06 0.99 0.06 0.86 Education level (yrs) 0 8 Reference 9 12 0.14 0.09 1.16 0.10 0.11 13 17 0.16 0.09 1.17 0.10 0.08

PAGE 112

112 Table 5 10 Continued Variable Coef. S.E. IRR S.E. p value Income Poor Reference Near poor 0.42 0.15 1.52 0.23 0.01 Low income 0.06 0.09 1.06 0.10 0.53 Middle income 0.02 0.10 0.98 0.10 0.86 High income 0.02 0.08 1.02 0.08 0.81 Gatekeeper status No gatekeeper Reference Gatekeeper 0.03 0.07 0.97 0.06 0.67

PAGE 113

113 Table 5 11 Logit regression predicting probability of having any physician care expendi ture s Variable Coef. S.E. O R S.E. p value Cost sharing Low Reference High 0.09 0.12 0.91 0.11 0.46 Disease severity level Top tertile PCS Reference Middle tertile PCS 1.02 0.13 2.76 0.3 5 0.00 Bottom tertile PCS 2.62 0.18 13.80 2.54 0.00 Cost sharing Bottom tertile PCS 2.38 1.04 10.81 11.27 0.02 Middle tertile PCS 0.12 0.17 1.13 0.19 0.49 Age (yrs) 18 34 Reference 35 64 0.55 0.08 1.73 0.14 0.00 2.02 0.14 7.51 1.09 0.00 Gender Female Reference Male 1.18 0.07 0.31 0.02 0.00 Race Other Reference Non Hispanic white 0.57 0.15 1.77 0.27 0.00 Non Hispanic African American 0.08 0.17 1.08 0.19 0.66 Hispanic 0.23 0.17 1.26 0.21 0.17 Geographic area Midwest Reference Northeast 0.09 0.13 1.10 0.14 0.46 South 0.07 0.10 0.93 0.10 0.51 West 0.21 0.11 0.81 0.09 0.05 Rural/urban location Rural Reference Urban 0.17 0.10 1.1 9 0.12 0.09 Marital status No t married Reference Married 0.03 0.07 1.03 0.07 0.65 Education level (yrs) 0 8 Reference 9 12 0.59 0.12 1.81 0.22 0.00 13 17 1.06 0.13 2.88 0.38 0.00

PAGE 114

114 Table 5 1 1 Co ntinued Variable Coef. S.E. O R S.E. p value Income Poor Reference Near poor 0.50 0.22 0.60 0.13 0.03 Low income 0.48 0.15 0.62 0.09 0.00 Middle income 0.22 0.15 0.81 0.12 0.16 High income 0.07 0.16 0.94 0.15 0.68 Gatekeeper status No gatekeeper Reference Gatekeeper 0.61 0.08 1.85 0.15 0.00

PAGE 115

115 Table 5 12 Log transformed OLS regression estimates for physician care expenditures Variable Coef. S.E. p value Cost sharing Low Reference High 0.28 0.10 0.01 Disease severity level Top tertile PCS Reference Middle tertile PCS 0.46 0.09 0.00 Bottom tertile PCS 1.08 0.09 0.00 Cost sharing Bottom tertile PCS 0.08 0.19 0.70 Middle tertile PCS 0.07 0.11 0.49 Age (yrs) 18 34 Reference 35 64 0.20 0.07 0.01 0.21 0.08 0.01 Gender Female Reference Male 0.08 0.05 0.11 Race Other Reference Non His panic white 0.18 0.09 0.04 Non Hispanic African American 0.10 0.12 0.41 Hispanic 0.03 0.12 0.81 Geographic area Midwest Reference Northeast 0.01 0.07 0.89 South 0.17 0.07 0.01 West 0.13 0.08 0.12 Rural/urban location Rural Reference Urban 0.18 0.06 0.00 Marital status No t married Reference Married 0.15 0.05 0.00 Education level (yrs) 0 8 Reference 9 12 0.01 0.12 0.96 13 17 0.23 0.12 0.05

PAGE 116

116 Table 5 1 2 Con tinued Variable Coef. S.E. p value Income Poor Reference Near poor 0.42 0.17 0.02 Low income 0.09 0.09 0.31 Middle income 0.32 0.09 0.00 High income 0.46 0.09 0.00 Gatekeeper status No gatekeeper Reference Gatekeeper 0.01 0.04 0.87

PAGE 117

117 Table 5 13 Logit regression predicting probability of having any ER visits expenditure Variable Coef. S.E. OR S.E. p value Cost sharing Low Reference High 1.23 0.88 3.41 3.00 0.16 Disease severity level Top tertile PCS Reference Middle tertile PCS 0.45 0.72 1.56 1.13 0.54 Bottom tertile PCS 0.75 0.63 0.47 0.30 0.23 Cost sharing Bottom tertile PCS 0.69 1.31 1.99 2.59 0.60 Middle tertile PCS 2.57 1.26 0.08 0.10 0.04 ER cost sharing 3.27 2.45 26.28 64.49 0.18 Age (yrs) 18 34 Reference 35 64 0.87 0.60 2.38 1.43 0.15 0.27 0.63 1.30 0.82 0.67 Gender Female Reference Male 0.16 0.51 1.17 0.60 0.75 Race Other Reference Non Hispanic white 1.19 0.66 3.28 2.15 0.07 Non Hispanic African American 0.60 0.73 1.82 1.33 0.41 Hispanic 1.14 0.72 3.13 2.25 0.11 Geographic area Midwest Reference Northeast 0.87 0.64 0.42 0.27 0.17 South 0.38 0.49 0.68 0.33 0.43 West 0.54 0.75 0.58 0.44 0.47 Rural/urban location Rural Reference Urban 0.54 0.43 1.71 0.74 0.21 Marital status No t married Reference Married 0.10 0.59 0.90 0.53 0.86 Education level (yrs) 0 8 Reference 9 12 0.25 0.60 1.28 0.77 0.68 13 17 0.30 0.61 0.74 0.46 0.63

PAGE 118

118 Table 5 1 3 Continued Variable Coef. S.E. OR S.E. p value Income Poor Reference Near poor 0.87 0.74 0.42 0.31 0.24 Low income 0.60 0.63 1.83 1.15 0.34 Middle income 1.44 0.68 4.23 2.89 0.04 High income 0.19 0.66 1.20 0.79 0.78 Gatekeeper status No gatekeeper Reference Gatekeeper 0.46 0.40 1.58 0.62 0.25

PAGE 119

119 Table 5 14 Log transformed OLS regression estimates for ER care expenditures Variable Co ef. S.E. p value Cost sharing Low Reference High 0.16 0.22 0.48 Disease severity level Top tertile PCS Reference Middle tertile PCS 0.15 0.18 0.42 Bottom tertile PCS 0.16 0.19 0.41 Cost sharing Bottom tertile PCS 0.26 0.54 0.64 Middle tertile PCS 0.06 0.25 0.81 ER cost sharing 1.38 0.31 0.00 Age (yrs) 18 34 Reference 35 64 0.33 0.16 0.04 0.16 0.18 0.37 Gender Female Reference Male 0.11 0.10 0.28 Race Other Reference Non Hispanic white 0.12 0.22 0.59 Non Hispanic African American 0.05 0.24 0.85 Hispanic 0.26 0.28 0.36 Geographic area Midwest Reference Northeast 0.09 0.13 0.50 South 0.24 0.13 0.06 West 0.28 0.16 0.08 Rural/urban location Rural Reference Ur ban 0.03 0.12 0.79 Marital status No t married Reference Married 0.20 0.11 0.07 Education level (yrs) 0 8 Reference 9 12 0.15 0.20 0.45 13 17 0.21 0.21 0.32

PAGE 120

120 Table 5 14. Continued Variable Coef. S.E. p value Income Poor Reference Near poor 0.04 0.22 0.84 Low income 0.24 0.17 0.16 Middle income 0.06 0.18 0.75 High income 0.35 0.18 0.06 Gatekeeper status No gatekeeper Reference Gatekeeper 0.04 0.12 0.77

PAGE 121

121 Table 5 15 Logit regression predicting probability of having any inpatient care expenditure Variable Coef. S.E. OR S.E. p value Cost sharing Low Reference High 0.44 0.22 0.64 0.14 0.05 Disease severity level Top tertile PCS Reference Middle tertile PCS 0.58 0.22 1.78 0.39 0.01 Bottom tertile PCS 1.21 0.19 3.37 0.65 0.00 Cost sharing Bottom tertile PCS 0.00 0.31 1.00 0.31 1.00 Middle tertile PCS 0.31 0.29 1.37 0.40 0.28 Inpatient care cost sharing 0.07 1.85 1.07 1.99 0.97 Age (yrs) 18 34 Reference 35 64 0.51 0.15 0.60 0.09 0.00 0.30 0.16 0.74 0. 12 0.07 Gender Female Reference Male 0.10 0.10 0.91 0.10 0.36 Race Other Reference Non Hispanic white 0.28 0.19 0.76 0.14 0.14 Non Hispanic African American 0.40 0.23 0.67 0.16 0.09 Hispanic 0.50 0.24 0.60 0.14 0.04 Geographic area Midwest Reference Northeast 0.14 0.15 0.87 0.13 0.35 South 0.14 0.13 1.15 0.15 0.30 West 0.14 0.14 0.87 0.13 0.34 Rural/urban location Rural Reference Urban 0.22 0.13 1.25 0.17 0.10 Marital status No t married Reference Married 0.16 0.10 1.17 0.12 0.14 Education level (yrs) 0 8 Reference 9 12 0.07 0.18 0.93 0.17 0.70 13 17 0.06 0.19 1.06 0.20 0.76

PAGE 122

122 Table 5 1 5 Continued Variable Coef. S.E. OR S.E. p value Income Poor Reference Near poor 0.01 0.25 1.01 0.25 0.97 Low income 0.14 0.17 0.87 0.15 0.39 Middle income 0.48 0.16 0.62 0.10 0.00 High income 0.48 0.17 0.62 0.10 0.00 Gatekeeper status No gatekeeper Reference Gatekeeper 0.09 0.10 1.09 0.11 0.38

PAGE 123

123 Table 5 16 Log transformed OLS regression estima tes for inpatient care expenditures Variable Coef. S.E. p value Cost sharing Low Reference High 0.15 0.19 0.43 Disease severity level Top tertile PCS Reference Middle tertile PCS 0.18 0.1 9 0.35 Bottom tertile PCS 0.52 0.16 0.00 Cost sharing Bottom tertile PCS 0.23 0.30 0.45 Middle tertile PCS 0.15 0.23 0.52 Inpatient care cost sharing 1.46 0.58 0.01 Age (yrs) 18 34 Reference 35 64 0.28 0.10 0.01 0.24 0.12 0.05 Gender Female Reference Male 0.18 0.11 0.11 Race Other Refe rence Non Hispanic white 0.35 0.14 0.01 Non Hispanic African American 0.65 0.18 0.00 Hispanic 0.24 0.21 0.26 Geographic area Midwest Reference Northeast 0.16 0.12 0.18 South 0.24 0.11 0.04 West 0.08 0.13 0.57 Rural/urban location Rural Reference Urban 0.06 0.10 0.56 Marital status No t married Reference Married 0.34 0.09 0.00 Education level (yrs) 0 8 Reference 9 12 0.01 0.16 0.94 13 17 0.05 0.17 0.76

PAGE 124

124 Table 5 16. Continued Variable Coef. S.E. p value Income Poor Reference Near poor 0.35 0.30 0.25 Low income 0.03 0.14 0.83 Middle income 0.06 0.14 0.65 High income 0.06 0.14 0.6 6 Gatekeeper status No gatekeeper Reference Gatekeeper 0.01 0.10 0.95

PAGE 125

125 Table 5 17 Logit regression predicting probability of having any medical care expenditure Variable Coef. S.E. OR S.E. p value Cost sharing Low Reference High 0.06 0.13 0.95 0.12 0.65 Dis ease severity level Top tertile PCS Reference Middle tertile PCS 1.07 0.12 2.90 0.36 0.00 Bottom tertile PCS 2.71 0.20 15.05 3.01 0.00 Cost sharing Bottom tertile PCS 2.18 1.05 8.85 9.27 0.04 Middle tertile PCS 0.10 0.17 1.10 0.18 0. 57 ER cost sharing 0.18 0.33 0.83 0.27 0.58 Inpatient care cost sharing 0.75 0.53 0.47 0.25 0.16 Age (yrs) 18 34 Reference 35 64 0.50 0.08 1.65 0.14 0.00 1.97 0.15 7.17 1.09 0.00 Gender Female Reference Male 1.19 0.07 0.30 0.02 0.00 Race Other Reference Non Hispanic white 0.54 0.15 1.72 0.26 0.00 Non Hispanic African American 0.09 0.17 1.09 0.19 0.61 Hispanic 0.21 0.17 1.24 0.21 0.21 Geographic area Midwest Reference Northeast 0.09 0.12 1.10 0.13 0.46 South 0.08 0.10 0.93 0.09 0.44 West 0.25 0.10 0.7 8 0.08 0.02 Rural/urba n location Rural Reference Urban 0.20 0.10 1.23 0.12 0.04 Marital status No t married Reference Married 0.02 0.07 1.02 0.07 0.74 Education level (yrs) 0 8 Reference 9 12 0.56 0.13 1.75 0.22 0.00 13 17 1.03 0.14 2.81 0.39 0.00

PAGE 126

126 Table 5 17. Continued Va riable Coef. S.E. OR S.E. p value Income Poor Reference Near poor 0.52 0.23 0.60 0.14 0.03 Low income 0.58 0.16 0.56 0.09 0.00 Middle income 0.30 0.16 0.74 0.12 0.07 High income 0.17 0.17 0.84 0.14 0.31 Gatekeeper status No gatekeeper Reference Gatekeeper 0.60 0.08 1.82 0.15 0.00

PAGE 127

127 Table 5 18 Log transformed OLS regression estimates for total medical care expenditures Variable Coef. S.E. p value Cost sharing Low Reference High 0.25 0.11 0.03 Disease severity level Top te rtile PCS Reference Middle tertile PCS 0.58 0.10 0.00 Bottom tertile PCS 1.35 0.10 0.00 Cost sharing Bottom tertile PCS 0.12 0.21 0.55 Middle tertile PCS 0.08 0.13 0.52 ER cost sharing 1.13 0.39 0.00 Inpatient care cost sharing 0.88 1.01 0.39 Age (yrs) 18 34 Reference 35 64 0.04 0.07 0.52 0.03 0.09 0.76 Gender Female Reference Male 0.09 0.05 0.10 Race Other Reference Non Hispanic white 0.17 0.10 0.09 Non Hispanic African American 0.14 0.13 0.30 Hispanic 0.00 0.13 0 .99 Geographic area Midwest Reference Northeast 0.05 0.08 0.52 South 0.14 0.07 0.04 West 0.16 0.08 0.04 Rural/urban location Rural Reference Urban 0.19 0.07 0.00 Marital status No t married Reference Married 0.19 0.05 0.00 Education level (yrs) 0 8 Reference 9 12 0.02 0.12 0.83 13 17 0.21 0.11 0.06

PAGE 128

128 Table 5 18. Continued Variab le Coef. S.E. p value Income Poor Reference Near poor 0.33 0.18 0.08 Low income 0.04 0.09 0.63 Middle income 0.12 0.10 0.22 High income 0.24 0.09 0.01 Gatekeeper status No gatekeeper Reference Gatekeeper 0.01 0.05 0.88

PAGE 129

129 Figure 5 1. P P plot for log transformed OLS regression on physician care expenditures Figure 5 2 Q Q plot for log transformed OLS regression on physician care expenditur es

PAGE 130

130 Figure 5 3. Residual fitted plot for log transformed OLS regression on physician care expenditures Figure 5 4 P P plot for log transformed OLS regression on ER care expenditures

PAGE 131

131 Figure 5 5 Q Q plot for log transformed OLS regression on ER c are expenditures Figure 5 6 Residual fitted plot for log transformed OLS regression on ER care expenditures

PAGE 132

132 Figure 5 7 P P plot for log transformed OLS regression on inpatient care expenditures Figure 5 8 Q Q plot for log transformed OLS r egression on inpatient care expenditures

PAGE 133

133 Figure 5 9 Residual fitted plot for log transformed OLS regression on inpatient care expenditures Figure 5 10 P P plot for log transformed OLS regression on total medical care expenditures

PAGE 134

134 Figure 5 11 Q Q plot for log transformed OLS regression on total medical care expenditures Figure 5 12 Residual fitted plot for log transformed OLS regression on total medical care expenditures

PAGE 135

135 CHAPTER 6 DISCUSSION Summary of Basic Findings The overall r esults have l aid a framework to address the research question. This study summarize s systematic patterns within each service type across health care utilization and expenditure models. I n physician care t h e differential impact was not significant in the utilization model and the actual expenditure model, but significant in the probability expenditure model indicating less expenditure reduction (Figure 6 1) In ER care this differential impact was significant in the utilization model (Figure 6 2) but was not significant in either part of the expenditure model. T he inpatient care results demonstrated consistent patterns ; no differential impact was found in either the utilization or the expenditure model As a component of physician care, the primary care p hysician visit model also produced a n insignificant differential impact Last, in total medical care t his differential impact was not significant in the actual expenditure model, but significant in the probability expenditure model (Figure 6 3) In summar y, the differential impact was less pronounced since it was only demonstrated in the ER utilization model and the expenditure probability model in physici an care and total medical care (Table 6 1) The integrated medical care expenditure models revealed op posite differential impacts indicating severely ill individuals actually reduced expenditure s in physician care (Figure 6 4) and total medical care (Figure 6 5) more, thus were even more sensitive to high cost sharing pressure than the general health popu lation. A ll the significant differential impacts can be summarized by s ervice t ypes as follows (Table 6 2): ER visits the probability of having any physician care and total care expenditures, integrated physician care and total care

PAGE 136

136 expenditures. As ER ut ilization result is an intermediate outcome for total care results, so this study will focus on discussing physician care and total care results. Th e negative significant differential impact s in integrated physician care and total care expenditure models we re opposite to expectations. As discussed in this study, severe diseases and a high cost sharing financial burden compete against each other and p resent the severely ill with a dilemma. Under high cost sharing pressure, either choice retaining needed phy sician care or sav ing money will come by sacrificing the other. Thus, it is possible that th is opposite differential impact may be due to a poverty constraint or fear of excessive debt burden in the severely ill group, and this reasoning was confirmed by t he study results. Poverty was measured by family income categories and Medicaid enrollment status. Chi square test results consistently indicated a significantly higher proportion of poor individuals within the severely ill group than the healthier group ( p < 0.01 and 0.01 respectively). The results further indicated a significant physician care reduction within the severely ill. S ince the severely ill group contained a substantial proportion of poor individuals, their significant care reduction in both ut ilization and expenditures in response to high cost sharing clearly indicated that high cost sharing policies were harmful ; it push ed them to forgo substantial physician care. A natural concern would arise as to whether this substantial physician care redu ction was needed or essential to maintain their health, or whether this loss would have detrimental consequence s The answer would be demonstrated in their downstream ER service utilization or direct health outcome change Th e result s indicate d within the severe ly ill group, relative to the low MD cost sharing group the high MD cost sharing group had significant ly higher ER utilization and expenditures and significant ly lower PCS The ir sizable offset effects

PAGE 137

137 in downstream ER service and worse health stat us, in conjunction with their substantial care reduction suggest ed that the severely ill may have forgone essential physician care. Thus, h igh physician care cost sharing policies coul d deprive the severely ill of essential physician care thus worsening their health conditions and push ing them in to the undesired downstream ER service, a more expensive service for worse clinical conditions, and thus may hurt both their health and finances Recall that in the physician care probability expenditure model, the differential impact was positively significant, and in the physician care integrated expenditure model, the differential impact was negatively significant. Considered together, this means that severely ill individuals had a strong desire to reduce phys ician care expenditures less than the general health population H owever, they actually reduced more because high cost sharing policies, reinforced by their financial difficult ies resulted in more elastic demand Thus, this finding in combination with of fset ER utilization and worse health status demonstrated that high cost sharing policies could greatly hurt and penalize severely ill individuals. Physician Care Price Elasticity Calculat ion of price elasticity of physician care also contributes to the un derstanding of the research questions. Price elasticity is calculated with the midpoint approach. E = { (Q 1 Q 0 ) / [ (Q 1 +Q 0 ) / 2 ] } / { (P 1 P 0 ) / [ (P 1 +P 0 ) / 2] } = [ (Q 1 Q 0 )*(P 1 +P 0 ) ] / [ ( Q 1 + Q 0 )* (P 1 P 0 ) ] The price el asticity for the severely ill wa s 0.139 in physician care utilization and 0. 399 in physician care expenditures. The cross price elasticity of physician care for the severely ill was 0.381 in ER care utilization, and 0.183 in ER care expenditures. These results are comparable to those found in the literature, with a range of price elasticity from 0.14 to 1.9 for physician care. The severely ill had the upper limit price

PAGE 138

138 elasticity of 0.14, which show s that their physician care ha d little room to be cut back and wa s uniquely valuable and essential to maintain their health. Meanwhile, these positive cross price elasticity estimates further indicate that physician care and ER are substitutes rather than complements. ER care can act as a backup service whe n physician care is unavailable to the severely ill due to high cost sharing but ER care a more expensive service for worse clinical conditions, is by no means a better choice for the severely ill. Other Sensitivity Analyses This study also examined ot her s ensitivity analys es to test finding robustness. First, if there are a number of individuals whose cost sharing is exactly 20% or 19% for Medicare and private plans the results may be influenced when they are categorized into high or low cost sharing groups. Th e results indicated that there were no observations whose cost sharing wa s exactly 20% or 19% for Medicare or private plans. Second, this study plan also consider ed different cut off points for cost sharing since there could be some doubt abou t whether it is really a high or low level if cost sharing is a little higher or lower than 20% for Medicare and private plans or the mean value for Medicaid. This study considered 25% and 15% as cut off points for Medicare and private plans, and 5% away f rom either side of the mean value for Medicaid. However, the proxy treatment of 19% for 20% due to the model specification of cut off point extension. Nevertheless this study did adopt the same cut off point of 19% across insurance plan types including Medicaid to examine the result robustness. B asically the u niversal cut off point did not change the results, since only 597 observations out of 60 595 transferred from high to low cost sharing group s. As a result, all the results ke pt the same significance and signs except p rimary c are p hysician visit and total care expenditure part I result s, where significance changed slightly from marginally significant to significant, or vice versa. Thus, the res ults are robust.

PAGE 139

139 Third, this study also planned and had include d the middle tertile subjects in the severity level as a dummy variable in the model, and its interaction with the cost sharing variable produced a less pronounced differential impact than the interaction by the severely ill and cost sharing robustness, such as using both sides of two standard deviations from the PCS mean as cut off points. This approach howev er, did not apply to the skewed distributi on of PCS, which yielded only a few observations in the severely ill group.

PAGE 140

140 Table 6 1 Differential i mpact s ummary b y s ervice t ypes Service t ype Coefficient S.E. p value Physician care visits 0.08 0.19 0.67 Physician care expenditure part I 2.38 1.04 0.02 Physician care expenditure part II 0.08 0.19 0.70 Primary care visits 0.15 0.09 0.09 ER visits 0.08 0.32 0.01 ER expenditure part I 0.69 1.31 0.60 ER expenditure part II 0.26 0.54 0.64 Hospital admis sions 0.09 0.21 0.68 Inpatient care expenditure part I 0.00 0.31 1.00 Inpatient care part II 0.23 0.30 0.45 Total care expenditure part I 2.18 1.05 0.04 Total care expenditure part II 0.12 0.21 0.55 Table 6 2 Significant d ifferential i mpact s ummary b y s ervice t ypes Service t ype Differential i mpact sign Probability of physician care expenditure s Integrated physician care expenditures + Emergency room care visits + Probability of total care expenditure s + Integrated total care expenditu res

PAGE 141

141 Figure 6 1 The differential impact by health status in probability of having any physician care expenditures Figure 6 2 The differential impact by health status in ER care visits Price of medical care Low cost sharing High cost sharing Utilization of ER care Severely ill group Reference group Price of medical care Low cost sharing High cost sharing Severely ill group Reference group Probability of having any physician care expenditures

PAGE 142

142 Figure 6 3 The differential impact by health status in probability of having any total medical care expenditures Figure 6 4. The differential impact by health status in integrated physician care expend itures Price of medical care Low cost sharing High cost sharing Probability of having any total medical care expenditures Severely i ll group Reference group Price of medical care Low cost sharing High cost sharing Severely ill group Reference group Integrated physician care expenditures

PAGE 143

143 Figure 6 5. The differential impact by health status in integrated total medical care expenditures Price of medical care Low cost shar ing High cost sharing Severely ill group Reference group Integrated total medical care expenditures

PAGE 144

144 CHAPTER 7 CONCLU SION S Summary and Interpretation In reality, severely ill individuals most usually have low incomes A s was identified in this study th e severely ill group had a significantly higher proportion of poor patients than the healthier group Thus, the sick and poor group finan cial difficult ies made them sensitive and vulnerable to high cost sharing policies t hat constrained them from maintaining adequate necessary health care, and their utilization reduction magnitude was likely an involuntary behavior rather than an active choice. Thus in response to high cost sharing pressure, severely ill individuals could experience substantial physician care reduction ER care increase and worse clinical conditions. Their price elasticity of demand for physician care was small indicating that physician care was essential to main tain their health. However, high cost sharing policies actually thwar ted their need for more frequent physician care Negative C onsequence of H igh Cost sharing P olicies S evere diseases and a high cost sharing financial burden compete against each other fo r severely ill and poor individuals and present the sick and poor group with a dilemma. Either choice maintaining health or lowering cost will come at the price of the other. Under this condition, t he existence of positive differential impact in the expend iture probability models for severely ill individuals indicated that they had a strong desire and tendency to reduce less needed medical care. However, the negative differential impact in the integrated expenditure models indicated that severely ill indivi duals actually reduce d more needed medical care. As a whole, severely ill individuals demonstrated that although they wished to reduce less needed medical care because i t was esse ntial to maintain their health, they actually reduced it more than the

PAGE 145

145 gener al population because high cost sharing policies, in conjunction with their financial difficult ies distorted their desire s and hindered their voluntary behaviors; in short, they were more sensitive and vulnerable to high cost sharing pressure. The positi vely significant differential impact in the probability expenditure model and the negatively significant differential impact in the integrated expenditure model indicated an even more harmful effect of high cost sharing policies for this especially vulnera ble subpopulation. High cost sharing policies acted as a counter force to undo efforts by imposing h igh financial pressure eroding their bottom line of affordability, and then forc ing them to delay or forgo needed phy sician care. In fact, h igh cost sharing policies for physician care had already pushed the severely ill away from needed physician care and into the less desir able downstream ER service, a more expensive service for worse clinical conditions, and thus may hurt them both clinically and financially. Moreover, high cost sharing policies also may have In summary, opposite differential impacts in expenditure models in combination with offset ER uti lization and worse health status results consistently demonstrated that high cost sharing policies could greatly distort to pay for needed care. Policy I mplication s A high cost sharing policy will naturally directly r educe health care utilization, regardless of subpopulations. In t heory, severely ill individuals should be price inelastic because care is necessary to maintain their health However high cost sharing policies acted as a counter force to undo the severely ills health maint enance efforts by imposing a financial barrier, and these policies were especially harmful to sick and poor individuals In reality, severely ill individuals are usually associated with low incomes Because of this th is sick and poor gr oup was even more vulnerable to high cost sharing policies than the general health population T he opposite

PAGE 146

146 differential impact for the sick and poor group highligh ted their financial difficulties or even risk, which distorted their willingness to maintain their health by reducing needed physician care to a large r extent than health ier individuals. Th us current high cost sharing policies should be replaced with low cost ones to reflect and match situation. Recall that the existen ce of a positive differential impact in the expenditure probability models for severely ill individuals indicated that they had a strong tendency and desire to reduce needed medical care less than the general health population in response to high cost shar ing pressure no t being contaminated by their practical financial difficult ies This phenomenon may reveal the potential underlying truth that the efficient and inefficient moral hazard fraction s differ by disease severity Sever ely ill people usually have less inefficient moral hazard and more efficient moral hazard than healthy people. They have a smaller fraction of discretionary care and a larger fraction of needed care than the general health population, so they waste less medical care. Cost sharing wa s originally introduced in insurance policies to f cost. Based on the moral hazard level difference between a c ost sharing levels should be designed differently and uniquely for subpopulations. Specifically, since severely ill individuals have a smaller fraction of discretionary care they should be treated differently with a low cost sharing policy that should c ut down smaller inefficient health care utilization As the results indicated, high cost sharing policies for physician care cut back needed care for the severely ill to a greater extent, which highlight s ies and vulnerab ility to cost pressure. Consequently, high cost sharing policies for physician care could hurt and penalize the severely ill both clinically and financially. A low cost sharing policy would

PAGE 147

147 remove the financial threat from the severely ill population ision making so as to provide them with access to needed care. Considering the se facts and results together, low cost sharing is not only necessary in theory, but also feasible and beneficial in practice for severely ill individuals It protect s and preve nt s them from exacerbating their poor health status and financial burden, and save s cost s for the society at the aggregate level. In addition to copayment, coinsurance and deductible, cost sharing can also take the form of an out of pocket maximum. An ap propriate out of pocket maximum level can effectively cost sharing categories. RAND HIE found no differential impact in health care expenditures O ne possible explanation discussed in a paper by Manning et al. ( 1987) may be due to the low out of pocket maximum level, so that not only the sick, but also healthy people could easily exceed this boundary and enjoy free inpatient care. That level may be appropriate for sick individuals, but questionable for healthy people. In the 1970s when the RAND HIE was conducted, the out of pocket maximum level wa s only $1,050. The findings of this present study call for a low out of pocket maximum level for the severely ill, and today this level should be different from several decades ago and adjusted based on the current situation. More importantly, this level should not be high so that it can provide the severely ill with adequate access to medical care to maintain their health and minimize their fin ancial burden. This study highlight ed the necessity and importance of value based insurance design in terms of differentiation and specification of its target population, so as to best protect and prevent them from exacerbating their poor health status an d financial burden, and to save cost for the society at the aggregate level. Furthermore, this study will contribute to the current debate on

PAGE 148

148 health care reform, especially specialized plans for subpopulations instead of a universal plan, may want to be co nsidered to complement the existing public and private designs Limitation s There are some limitations in the data, measurement s, and operationalization. MEPS data only represents non external va lidity. In add ition, self report ing system in MEPS data may cause measurement error concern. Moreo ver, the data do not contain information from providers if induced demand exists. There may be concerns that health care utilization is also influenced by sup ply side dynamics, cost sharing from the demand side. In order to compensate and maintain their revenue level, providers will increase utilization by inducing demand so that th e healthier group may not reduce medical use substantially B y doing so they may demonstrate a similar ly less er amount of care reduction as the severely ill group. In this situation, no differential impact will be observed. Fortunately, the results indica ted a significant physician care reduction within each group the severely ill ( p < 0.01) and the reference group ( p < 0.01 ) for both utilization and expenditures thus eliminat ing the supply side information concern. In add ition, the PCS measure focuses mor e on health status and functional status that indirect ly reflect a so it is an imperfect measure of severity of illness This study examined service types including physician care, ER visits and inpatient care. However, me dication is another service type that might influence the relationships among these service types and the results Inclusion of medication s will be one of the future research directions. Despite these limitations, this study significantly adds to the lite rature by focusing on cost sharing only in physician care, using a nationally representative sample and more precise

PAGE 149

149 measures, and adopting more rigorous methods to improve the understanding of the research question. havior, specifically their health care utilization change in response to high cost sharing pressure, this study allows researchers to evaluate the effect of high cost sharing policies, and whether they are necessary to cut down health care utilization and expenditures for the severely ill. Severely ill individuals demonstrated a strong desire to reduce necessary medical care to a less er extent than the general health population, however, high cost sharing policies went against their wishes and pushed the s everely ill away from needed physician care and into the less desired downstream ER service, and may hurt them both clinically and financially. Therefore, th ese vulnerable people should be treated differently. Specifically, a low cost sharing policy should be designed to reflect and match their situation. This study contributes to inform the necessity and importance of insurance policy design in terms of differentiation and specification for its target population, so as to best protect and prevent them from exacerbating their poor health status and financial burden, and to save cost s for society at the aggregate level. Beside adding service types, future research plan s to explore the differential impact of high cost sharing in physician care on health care u tilization and expenditures by age (elderly and non elderly populations) and income (poor and non poor populations), which correspon d s to Medicare (relative to non Medicare) and Medicaid (relative to non Medicaid) plans. The purpose and significance of thi s series of inquiries is to determine whether insurance policy design, specifically the cost sharing level should be unique to Medicare and Medicaid or subpopulations, which will contribute to the insurance policy debate on universal and specialty health plans.

PAGE 150

150 LIST OF REFERENCES Agency for Healthcare Research and Quality. Medical Expenditure Panel Survey: Household Component Weighting and Estimation Issues Agency for Healthcare Research and Quality. 2007. Journal of Health and Social Behavior 1995. 36: 1 10 Renal function as a risk indicator for cardiovascular events in 3216 patients with manifest arterial disease Atherosclerosis 2008. 200 (1): 184. Cartwright, W. S., T. W. Hu, and L. on the use Applied Economics 1992. 24: 529 39. Chandra, Amitabh, Jonathan Gruber, and Robin McKnight. Patient Cost Sharing, Hospitalization Offsets, and the Design of Optimal Health Insurance for the Elderly Working Paper 12 972. 2007. http://www.nber.org/papers/w12972 Accessed 09/04/2009. Me dical Care 1989. 27(7): 669 78. Health Affairs 2008. 27: 103 12. Cohen, S.B., R. D Methodology Report #5: Estimation Procedures in the 1996 Medical Expenditure Panel Survey Household Component Agency for Healthcare Research and Quality. 1999. Characteristics of Persons w ith High Medical 2002. Statistical Brief 73. 2005. expenditures, 2002. Health Care Financing Review 2004. 25: 4. Cutler, David M., and Richard J. Zeckhauser. Working Paper 7176. 1999. http://www.nber.org/papers/w7176.pdf Accessed 09/0 6 /2009. Daniels Justice, health, and healthcare. American Journal of Bioethics 2001. 1(2): 2 16 Dormuth Colin R., Peter Neumann, Malcolm Maclure, Robert J. Glynn, and Sebastian Effects of prescription coinsurance and income based deductibles on net health Medical Care. 2009. 47: 508 516.

PAGE 151

151 Doty, Michelle M., Sara R. Collins, Sheila D. Rustgi, and Jennifer L. Kriss. Seeing Red: The Growing Burden of Medical Bills and Debt Faced by U.S. Families Commonwealth Fund pub. 1164, vol. 42. Eppig, Franklin Medicare current beneficiary survey: Cen ters for Medicare & Medicaid Services. http://aging.senate.gov/award/cms3.pdf Accessed 09/04/2009. term follo w up study of the effect of a three tier prescription drug copayment system on Clinical Therapeutics 2003. 25(12): 3147 61. directed hea lth plans: New Inquiry 2007. 44: 26 40. Biometrics 2009. 65: 885 893. Freeman, J.D ., S. Kadiyala, J.F. Bell, and D.P. Martin Medical Care 2008. 46(10): 1023 32. American Journal of Managed Care. 2005. 11:730 40. Goldman, Dana P., Geoffrey F. Joyce, and Pinar Karaca Mandic. with clinical status: The case of cholesterol American Journal of Managed Care 2006. 12: 21 28. Goldman, Dana P., Geoffrey F. Joyce, Grant Lawless, William H.Crown, and Vincent Willey. Health Affairs 2006. 25(5): 1319 1331. Goldman, Dana P., Geoffrey F. Joyce, Jose J. Escarce, Jennifer E. Pace, Matthew D. Solomon, Marianne La Pharmacy benefits and the use Journal of the American Medical Association 2004. 291: 2344 2350. Goldman, Dana P., Geoffrey F. Jo Prescription drug cost sharing: Associations with medication and medical utilization and spending and health T he Journal of the American Medical Association 2007. 298(1): 61 69. Grossman The Journal of Political Economy 1972. 80(2): 223 55.

PAGE 152

152 Gruber, Jonathan. The Role of Consumer Copayments for Health Care: Lessons from the RAND Health Insurance Experiment and Beyond The Kaiser Family Foundation. 2006 ncreased patient cost sharing, weak American Health & Drug Benefits 2009. 2(3): 1 34 41. Harman, J.S., S.H. Scholle, R.E. Mardon, A.S. Bierman, S.C. Haffer, J.H. Ng, S. Shih, and L.G. Medical Care 2010. 48(3): 217 223. Hausman, J.A. Econometrica 1978.46 (6):1251 71. Helms, Jay, Joseph New Bell Journal of Economics 1978. 9: 192 208. effective Medical Care Research and Review 2008. 65(4): 437 49. Hildebrand, D.K., J.D. Laing, and H. Rosenthal. 1977. Prediction Analysis of Cross Classifications New York: John Wiley and Sons, Inc. Hsu, J., M. Price, R. B sharing for emergency care and unfavorable clinical events: Findings from the safety and financial ramifications of ED Health Services Research. 2006. 41: 1801 20. ey. Health Insurance Premiums and Cost Sharing: Findings from the Research on Low Income Populations Kaiser Family Foundation 2003. impact of increasing patient p rescription drug cost sharing on therapeutic classes of drugs Health Services Research 1997. 32(1): 103 122. Kaiser Family Foundation. H ealth Coverage for Low Income Americans: An Evidence Based Approach to Public Policy 1. Kaiser Family Foundation. 2007. Kaiser Family Foundation. Medicare at a Glance Kaiser Family Foundation 2008, 11. Kaiser Family Foundation. The Uninsured: A Primer. Key Facts about Americans without Health Insurance 10. K aiser Family Foundation. 2008. of pocket costs and diabetes preventive services: The Translating Research into Action for Diabetes (TRIAD) Diabetes Care. 2003. 26: 2294 99.

PAGE 153

153 Landsman, P.B ., W. Yu X. Liu, S.M. Teutsch, and M.L. Berger Impact of 3 tier pharmacy penefit design and increased consumer cost American Journal of Managed Care 2005. 11(10): 621 28. Health Care Financing Revi ew 1980. 2(2): 25 31. Lwe, Bernd, Levke Willand, Wolfgang Eich, Stephan Zipfel, Anthony D. Ho, Wolfgang Herzog, and Christoph Fiehn. Psychiatric Comorbidity and Work Disability in Patients with Psychosomatic Medicine 2 004. 66: 395 402. Journal of Health Economics 1998. 17(3): 283 295. Manning, Willard G., Joseph P. Newhouse; Naihua Duan; Emmett B. Keeler; and Arleen L Health insurance and the demand for medical care: Evidence from a randomized The American Economic Review 1987. 77(3): 251 277. Evide Inquiry 1991. 28(3): 276 87. McCauley, J., M.W. Jenckes, C. Saeger Procter, M.J. McGuire, D.E. Ford, and A.M. Warwick. Frequent fliers: Managed care patients with multiple emergency room visits Association for Health Services Research Annual Meeting. 1998. 39 40. The effect of a copay increase on American Journal of Managed Care 1999. 5: 138 3 1394. Health Affairs 2004. 23(5): 194 199. driven health plan on medica Health Services Research. 2004. 39: 1189 210. The American Economic Review 1968. 58(3): 531 537. Pearte, C.A., C.D. Furberg, E.S. O'Meara, et al. Characteristics and baseline clinical predictors of future fatal versus nonfatal coronary heart disease events in older adults: The Cardiovascular Health Study Circulation 2006. 113 (18): 2177 85. Annual Review of Public Health 2009. 30: 293 311.

PAGE 154

154 y of cost sharing responsiveness: How do sick people respond to cost Health Economics 2003. 12: 269 280. the literature and implications for health Medical Care Research and Review 1994. 51: 235. Medical Care. 1986. 24:873 76. Rukavina, Mark. The Financial Burden of Health Care Boston: The Access Project. 2009. Medical Care 1984. 22(3): 247 54. Scitovsky, A.A., and N.M. S Effect of coinsurance on use of physician services Security Bulletin 35. 1972. (6): 3 19. New Eng land Journal of Medicine 1996. 334: 635 41. Cost sharing in Medicaid and CHIP: How does it affect out of Health Affairs 2009.:28(4): w607 w619. Shi, Leiyu, and Douglas A. Singh. 2005. Essentials of the US Health Care System Sudbury, MA: Jones and Barlett. copayments on outpatient mental health utilization by members of a health maintenance American Journal of Psychiatry 1996. 153: 331 38. Models of Man, Social and Rational: Mathematical Essays on Rational Human Behavior in a Social Setting New York: John Wiley. 1957. The effects of copayments and generic substitution on the use and costs of Inquiry 1993. 30: 189 198. Smith, D.G., and D.M. Kirking Pharmaco Econ omics 1992. 2(4): 335 42.

PAGE 155

155 sharing and the utilization of clinical preventive American Journal of Prevention Medicine 1999. 17: 127 33. high concentration of U.S health care expenditures Rockville, MD: Agency for Healthcare Research and Quality. 2005. Research in Action Issue 19. AHRQ Pub. No. 06 0060. StataCorp. Stata Statistical Software: Release 10 College Station, TX: StataCorp LP. 2007. Stuart, Bruce, an d Christopher Zacker Who bears the burden of Medicaid drug copayment Health Affairs 1999. 18(2): 201 212. Tamblyn, Robyn, Rejean Laprise, James A. Hanley, Michael Abrahamowicz, Susan Scott, Nancy Mayo, Jerry Hurley, Roland Grad, Eric Latimer, Robert Perreault, Peter McLeod, Allen Adverse events associated with prescription drug cost Journal of the American Medical As sociation 2001. 285: 421 29. Effect of cost sharing Archives of Pediatrics & Adolescent Medicine 2008. 162(2): 104 10. item short form health survey (SF 36). I. Medical Care 1992. 30(6): 473 83. Wharam, J.F., B.E. Landon, A.A. Galbraith, K.P. Kleinman, S.B. Soume rai, and D. Ross Journal of the American Medical Association 2007. 297: 1093 102. Wooldridge, J.M. Introductory Econometrics: A Modern Appr oach Third Edition. Mason, OH: Thomson South Western. 2005. American Journal of Managed Care 2007. 13: 335 4 6. Inquiry 2002. 39(1): 76 86.

PAGE 156

156 BIOGRAPHICAL SKETCH Haichang Xin was born in Beijing, China. His parents always encourage d him to pursue as much education as possible, especially in medicine and the health profession s to treat diseases and save lives. His undergraduate study in the field of medicine and public health at Shanghai Medical University laid a foundation for his future career. Durin g this period, he realized that each physician can treat only a limited number of patients in a certain area, while an effective health policy would benefit thousands of people in a community and even a society. B earing this philosophy in mind, he later ac h ealth m anagement and s ocial m edicine from Capital University of Medical Sciences, Beijing In 2006, he came to the United States to pursue a PhD degree in health services research at the University of Florida. At UF, he receive d systematic training in the area of health policy and system, health care utilization and quality, health outcome, health economics, and health insurance He focused on health services research methods and applied them in this research area. He graduated in summer 2010.