/j W“ "

I \}

 

LIBRARY
Michigan State
University

u.- «fllp-v— "~—

 

 

This is to certify that the
dissertation entitled

Two Essays on Taiwan’s Health Care Reform in the 19903

presented by

Kang-Hung Chang

has been accepted towards fulfillment
of the requirements for the

Ph.D. degree in Economics

 

 

7 W 91 14W; ,
U

Major Professor‘s Signature

 

Q/tX’IM
T .I
Date

MSU is an Affirmative Action/Equal Opportunity Employer

 

-.-.—.-t-«-.-.---v-.—.—._._

 

PLACE IN RETURN Box to remove this checkout from your record.
TO AVOID FINES return on or before date due.
MAY BE RECALLED with earlier due date if requested.

 

DATE DUE DATE DUE DATE DUE

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

5/08 KzlProlecc8-PreleIRCIDateDue.indd

TWO ESSAYS ON TAIWAN’S HEALTH CARE REFORM IN THE 1990s
By

Kang-Hung Chang

A DISSERTATION

Submitted to
Michigan State University
in partial fulfillment of the requirements
for the degree of

DOCTOR OF PHILOSOPHY

ECONOMICS

2009

ABSTRACT
TWO ESSAYS ON TAIWAN’S HEALTH CARE REFORM IN THE 19905
By
Kang-Hung Chang

This dissertation consists of two essays examining major health care reforms in
Taiwan in the 19905 and their effects on the elderly. Both reforms are analyzed as natural
experiments, comparing the experience of an affected group with a control group that was
not affected.

The first essay investigates the National Health Insurance (N HI) program that
provided universal coverage to Taiwan’s 21 million citizens, including 8 million
previously uninsured. The essay finds that NHI largely increased utilization for the
previously uninsured elderly relative to their continuously insured counterparts.
Corresponding to the increase in utilization, the mortality hazard ratio of the previously
uninsured to the continuously insured elderly significantly dropped from 1.3 in the pre-
NHI period to close to 1 in the post-NH] period, suggesting that NHI improved the health
of the previously uninsured elderly through medical care.

The second essay inspects a special medicine culture in Taiwan—both physicians
and pharmacists prescribed and dispensed drugs—which was profoundly changed in the
19903 by NHI and a separation policy (SP) that forbade physicians from dispensing and
pharmacists from prescribing. The essay finds that NHI made the previously uninsured
elderly more likely to visit physicians, including some who previously only visited
pharmacists. Following this, the SP made all elderly patients more likely to only visit

physicians and buy drugs from on-site pharmacists who were hired by physicians.

ACKNOWLEDGEMENTS

This dissertation cannot be completed without the guidance of Dr. John Goddeeris,
a dedicated Health Economist whom I dub as “Captain” to acknowledge his essential role
in directing me through my exploration in the field of health economics. His rigorous
training, insights and inspirations have certainly nourished my capability of conducting
independent research.

I am also very grateful to have another four brilliant scholars on my committee.
Dr. Steven Haider and Dr. Todd Elder have supported me with their constructive critics
and suggestions. Dr. Scott Swinton (Agricultural, Food and Resource Economics) and Dr.
Joseph Gardiner (Epidemiology) also offered me many helpful comments. Without their
inputs, this dissertation would have been lackluster.

During my career at Michigan State, I have received the F ulbright-Hayes Study
Grant sponsored by the US. Department of State and the Doctoral Dissertation
Fellowship from the Chiang Ching-Kuo Foundation in Taiwan. Along the way, I was also
funded by the assistantship and travel grants from the Department of Economics, College
of Social Science and Office for International Students and Scholars at Michigan State.
Without their generous financial supports, I could never accomplish my Ph.D. in
Economics. Moreover, I am indebted to the Department of Health Promotion in Taiwan
for providing me good-quality data.

Writing a dissertation is like cooking a dish, it takes a recipe and many
ingredients. Without a doubt, the people and institutions mentioned above have provided

me all I need. I hope I have cooked a good dissertation in return.

iii

PROLOGUE

Financial incentives and institutional constraints are both important factors in
determining people’s health care utilization behavior and thus can further affect their
consequent health outcomes. This dissertation consists of two chapters studying effects of
two major health care reforms in Taiwan in the 19903 that profoundly changed financial
incentives faced by and institutional constraints imposed upon the elderly.

In 1994, about 60% of Taiwan’s 21 million citizens were insured by three major
social insurance programs—Government Employees Insurance, Labor Insurance and
Farmer’s Health Insurance—which all provided comprehensive medical benefits. As
suggested by their names, these programs were employment-based and had covered most
of the healthy working population at the time. Besides, comprehensive private health
insurance did not exist then. The rest of the population—mostly the elderly, women and
children—were thus left as uninsured and had to pay their medical care expenditure
completely out-of-pocket.

On March 1, 1995, the government integrated the three major social insurance
programs into a single-payer third-party universal program entitled as National Health
Insurance (N HI). This new program was not only universal but also compulsory—it
required the previously uninsured to enroll regardless of their health, while others who
were previously insured by the three major social insurance programs were automatically
transferred to NH]. The implementation of NH] was a rare large-scale natural experiment
on health insurance in recent decades. It greatly reduced the price of medical care faced

by the previously uninsured elderly when ill. More importantly, its universal and

iv

compulsory feature generated arguably exogenous variation in insurance status among
the previously uninsured.

Exploiting the arguably exogenous variation in insurance status, the first chapter
examines the NHI effects on health care utilization and mortality of the previously
uninsured elderly. It adopts a difference-in-differences (DD) strategy and uses a
longitudinal survey of the elderly in Taiwan. In particular, the previously uninsured
elderly are treated as the treatment group and the continuously insured elderly as the
control group. The first chapter finds that NHI largely increased utilization of health care
of the previously uninsured elderly relative to their continuously insured counterparts.
Corresponding to the increase in utilization, the mortality hazard ratio of the previously
uninsured to the continuously insured elderly dropped from 1.3 in the pre-NHI period to
close to 1 in the post-NHI period, suggesting that NHI improved the health of the
previously uninsured elderly through health care.

The second chapter is about a special medicine culture in Taiwan—physicians
and pharmacists both prescribed and dispensed drugs. This culture had a long history
dating back to its Japanese colonial period (1895-1945) and remained unchanged until the
mid-19903. In the eyes of many elderly who were accustomed to this culture, physicians
and pharmacists were to some extent synonymous, i.e. close substitutes, at least in the
context of minor illness, because they both prescribed and dispensed drugs.

However, physician and pharmacist services differed in two major aspects. In
general, physician services were of higher quality because physicians were well trained to
diagnose and treat medical conditions, while pharmacists were not. On the other hand,

physician services were also more expensive because physicians typically charged not

only for dispensing drugs but also for diagnosing, prescribing and other related fees,
while pharmacists charged only for dispensing drugs. The price difference could be
substantial depending on medical conditions. For example, to treat a common cold, an
uninsured patient would have to pay NT$ 300 (US$ 10) if visiting a physician but only
NT$ 100 (US$ 3.3) if visiting a pharmacist. It is thus not surprising that the uninsured
elderly were more likely to visit pharmacists at that time.

The Special medicine culture started to change because of two health care reforms
in the mid-199OS--NHI in 1995 and a separation policy (SP) in 1997 called yi yiaofeng
ye (in Chinese), meaning separation of drug prescribing and dispensing. In particular,
NHI provided an economic incentive for the previously uninsured elderly to visit
physicians by paying for their visits and drugs. The SP basically forbad physicians from
dispensing and pharmacists from prescribing. Patients were forced to visit a doctor in
order to get a prescription, which was required when they purchased drugs from a
pharmacist.

The second chapter investigates the effects of NHI and SP on elderly patients’
choice between physician and pharmacist services. It also adopts the DD method to study
both policies. For NHI, again, the previously uninsured elderly were treated as the
treatment group and the continuously insured elderly as the control group. For SP,
because the policy was gradually phased in throughout the country, the elderly who lived
in counties that had adopted SP were treated as the treatment group, and others who lived
in counties that had not adopted as the control group.

This chapter finds that by providing an economic incentive to the previously

uninsured elderly, NHI raised the probability that they would visit physicians, relative to

vi

their continuously insured counterparts. In particular, some previously uninsured elderly
who once only visited pharmacists were more likely to also visit physicians after NHI
was implemented. Following this, the SP made it more likely that all elderly patients
would only visit physicians and buy drugs from on-site pharmacists hired by physicians.
The SP effect was influenced by two critical concessions that the government made to
physicians—physicians were allowed to hire their own on-site pharmacists in their clinics
and to dispense drugs to elderly aged 65 and older if they could not go to a community
pharmacy to purchase drugs by themselves. These two concessions saved the elderly
patients a separate trip to a community pharmacy by allowing them to buy drugs from the
physician or the on-site pharmacist in the clinic.

The lessons learned from Taiwan’s experience in the 19905 provide an insight for
other countries—financial incentives and institutional constraints matter when it comes to

designing health care policies.

vii

TABLE OF CONTENTS

LIST OF TABLES ................................................................................. ix
LIST OF FIGURES ................................................................................ xi
CHAPTER 1

COVERAGE MATTERS: IMPACTS OF NATIONAL HEALTH INSURANCE ON
HEALTH CARE UTILIZATION AND MORTALITY OF THE PREVIOUSLY

UNINSURED ELDERLY IN TAIWAN ......................................................... 1

1.1 . Introduction ....................................................................................... 1

1.2. Institutional Background ....................................................................... 5
1.3. Empirical Framework ........................................................................... 9
1.4. Data .............................................................................................. 17
1.5. Estimation Results ............................................................................. 24
1.6. Conclusion ...................................................................................... 32
CHAPTER 2

THE HEALER OR THE DRUGGIST: EFFECTS OF TWO HEALTH CARE
POLICIES IN TAIWAN ON ELDERLY PATIENTS’ CHOICE BETWEEN

PHYSICIAN AND PHARMACIST SERVICES ............................................ 34
2.1 . Introduction ...................................................................................... 34
2.2. Two Health Care Reforms in the 19905 ..................................................... 37
2.3. Empirical Framework ......................................................................... 40
2.4. Data and Descript Statistics ................................................................... 47
2.5. Estimation Results ............................................................................. 52
2.6. Conclusion ...................................................................................... 57
APPENDICES ....................................................................................... 59
Appendix 1. Insurance Status in 1993 and Survey Questions ................................. 59
Appendix 2. Differences Between Chapter 1 and Chen et al. (2007) ........................ 62
Appendix 3. Availability of Data and Full Estimation Results ............................... 67
BIBLIOGRAPHY .................................................................................. 68

viii

Table 1.1.

Table 1.2.

Table 1.3.

Table 1.4.

Table 1.5.

Table 1.6.

Table 1.7.

Table 1.8.

Table 1.9.

Table 2.1.

Table 2.2.

Table 2.3.

Table 2.4.

Table 2.5.

Table 2.6.

Table 2.7.

LIST OF TABLES

Elderly Insurance Status in 1993 ..................................................... 18
Characteristics of the Elderly in 1993 ................................................ 20
Summary of Utilization Variables ................................................... 21

NHI Effect on Probability of Utilization ............................................ 25

NHI Effect on Probability of Utilization by Subgroups ........................... 27
NHI Effect on Mortality Hazard ...................................................... 28
Alternative Mortality Hazard Models ................................................ 30
NHI Effect on Mortality Hazard by Subgroups .................................... 31
Validity Test of The First DD Assumption ......................................... 32
Timetable of the SP implementation ................................................ 41

Baseline Characteristics of the Elderly .............................................. 48
Summary of Outcome Variables ...................................................... 49
Comparison of Time Trends in Pharmacy Visits ................................... 52
NHI and SP Effects on Utilization ................................................... 53
NHI and SP Effects on Self-reported Health ....................................... 55
SP Effect on Utilization by Insurance Group ....................................... 56

ix

Table 2.8. SP Effect on Self-reported Health by Insurance Group ........................... 57

Table A2.1. Comparison of One Year Mortality Rates (%) ................................... 63

Table A2.2. NHI Effect on Hazard Ratio (Analysis Time: 1993-1999) ..................... 63

LIST OF FIGURES

Figure 1.1. Uninsurance Rate: 1985-2005 ......................................................... 9
Figure 1.2. Yearly Mortality Rate by Insurance Group ........................................ 23
Figure 2.1. Map of Taiwan ........................................................................ 45

xi

Chapter 1
Coverage Matters: Impacts of National Health Insurance on Health Care Utilization
and Mortality of the Previously Uninsured Elderly in Taiwan
1.1. Introduction

The effects of health insurance have long been an interest among researchers and
policy makers. In particular, how health insurance affects people’s utilization behavior
and health outcomes is an important question when a government considers providing
insurance coverage to the public. Since the famous RAND Health Insurance Experiment
in the 19705 (Manning et al. 1987), a large body of studies investigating insurance effects
has accumulated in the literature. Researchers, however, have not yet reached a
unanimous conclusion of the insurance effects on health outcomes despite their
agreement that insurance affects utilization.

There are numerous reasons why there is a lack of consensus among researchers
on the insurance effects on health outcomes (Levy and Meltzer 2004). One of the major
reasons is that causation between insurance status and health may run in both directions.
For example, while insurance provides people access to medical care that might improve
health, it is also true that—at least in the United States—insurance companies ofien
screen the health of their applicants and relatively healthy applicants are more likely to be
approved. This could create a positive association between insurance and health that does
not reflect a causal effect of insurance.

To tackle this problem, more and more studies have exploited exogenous
variation in insurance status generated by so-called “natural experiments”, which usually

refer to a policy or an institutional change that is essentially uncorrelated with the health

of its targeted subjects (Card et a1. 2004; Chen et al. 2007; Cheng and Chiang 1997;
Currie and Gruber 1996a, 1996b, 1997; Haas et al. 1993a, 1993b; Hanratty 1996;
Lichtenberg 2002; Polsky et a1. 2006; Wen et al. 2008). Typically, a natural experiment
involves two groups: one group is subject to the policy and called the treatment group
and another group is not subject to the policy and called the control group. If the policy-
generated variation in insurance status among the treatment group is uncorrelated with
the subjects’ health, it is feasible to identify the insurance effect on health by comparing
the difference in differences in outcomes between the two groups before and after the
experiment under some regular assumptions. Following this line of research, this paper
examines the effects of a rare large-scale “natural experiment” on health insurance in
Taiwan in the mid-19905.

On March 1, 1995, Taiwan initiated its National Health Insurance (N HI) program
that mandated insurance coverage to its 21 million citizens, including 8 million who had
been previously uninsured. The previously uninsured were required to enroll in NHI
regardless of their health.1 In addition, the strict mandate eliminated the possibility for the
previously uninsured to opt out. These two facts together generated arguably exogenous
variation in insurance status. On the other hand, others who had been previously insured
by various social insurance programs were automatically transferred to NHI and enjoyed
a similar range of medical benefits as before. Exploiting these features, this paper adOpts
a difference-in—differences (DD) method to examine the NHI effects on utilization

behavior and mortality. Due to data availability, this paper focuses on the elderly.

 

I In fact, no matter when the previously uninsured actually signed up for NHI, their coverage automatically
took effect on March I, 1995.

Several studies such as Cheng and Chiang (1997), Chen et a1. (2007), Wen et a1.
(2008), Zimmer et a1. (2002) and Zimmer et a1. (2005) have also investigated the effects
of Taiwan’s NHI on utilization and health. The majority consensus among these studies is
that NHI increased utilization for the previously uninsured. Nevertheless, while some of
them claim that NHI also improved the health of the previously uninsured, others
disagree.

In particular, Chen et a1. (2007) used the same survey of the elderly that is also
used in this chapter.2 They find that NHI largely increased utilization for the previously
uninsured elderly but did not improve their one-year mortality relative to their
continuously insured counterparts.3 They attribute the lack of evidence on mortality to
several possible reasons. For example, they argue that one-year mortality may not be a
sensitive measure of health and that their sample period—from 1989 to 1999, which
includes only covers 4 years after 1995—may not be long enough to detect any NHI
effect. Other explanations they provide are related with effectiveness and quality of
health care in producing health.

In our view, however, there are two other possibilities. First, they did not fully
utilize the mortality information. They only used two pre-NHI (1989 and 1993) and two
post-NHI yearly mortality rates (1996 and 1999), while mortality information from other
years was simply ignored. In this study, we are able to use more recent and detailed
monthly mortality data up to the end of 2003. Our study period thus covers 8 years after

NHI was implemented. Moreover, instead of looking at limited discrete points, we utilize

 

2 Chen et al. (2007) was published after this paper had begun in 2006.
3 They also examined the NH 1 effect on self—reported health but did not find evidence of an effect on it,
either.

every piece of monthly mortality information throughout our entire analysis period from
1993 to 2003.

Second, the demarcation between the previously uninsured and continuously
insured elderly in Chen et al. (2007) was problematic. They imputed insurance status for
the elderly in the pre-NHI period and assumed their status remained unchanged until
1995 when NHI was initiated.4 However, we examine the pre-NHI insurance history of
the elderly using both administrative data and a telephone follow-up in 1991 of the same
elderly survey, which they did not use; we find that many of the elderly changed their
insurance status in the pre-NHI period. In particular, of the elderly in the survey whose
insurance status was available in both 1989 and 1993, about 26% of them changed their
insurance status, either from uninsured to insured or vice versa.5 This implies that their
imputed insurance status very likely has led to wrong demarcation between the treatment
and control group, which is essential in their difference-in-differences analysis. (See
Section 2 and Appendix 2 for more details.)

Therefore, to get a more accurate demarcation, we use insurance status in 1993,
which is the available insurance information closest to 1995, to reduce the possibility of
changing insurance status before the implementation of NHI.

The major findings of this chapter show that NHI largely increased the utilization
of the previously uninsured elderly relative to their continuously insured counterparts.
Correspondingly, the mortality hazard ratio of the previously uninsured elderly to their

continuously insured counterparts significantly dropped from 1.3 in the pre-NHI period to

 

4 They imputed insurance status because no insurance information was available in the initial survey in
1989. Another wave of the survey in 1993 does have insurance information. However, only a subset of the
initial sample appeared in 1993.

5 There were 2,887 elderly whose insurance status in 1989 and 1993 were both available. Among them, 672
changed from uninsured to insured, 75 from insured to uninsured and the rest did not change.

4

close to one in the post-NHI period, suggesting that NHI did improve the health of the
previously uninsured elderly through health care.

The rest of this chapter is organized as follows. The second section introduces the
institutional background. The third section illustrates the empirical method. The fourth
section describes the data. The fifth section reports the estimation results. The major
findings are then concluded in the last section. Appendix 1 demonstrates how the insured
and uninsured elderly in the pre-NHI period are defined. A detailed comparison between
this paper and Chen et a1. (2007) is provided in Appendix 2.

1.2. Institutional Background
Health Insurance Before 1995

At the end of 1994, about 2.3 million people in Taiwan (11% of the entire
population) were aged 60 and older. Among them, about 75% were insured by four social
insurance programs: Farmers Health Insurance (FHI), Government Employees Insurance
(GEI), Labor Insurance (LI) and Veterans Insurance (VI).6 As suggested by their names,
FHI covered farmers; GEI covered govemment employees and school teachers; L1
covered private sector workers; VI covered veterans. In 1994, FHI, GEI, LI and VI
respectively accounted for 45%, 20%, 18% and 17% of the insured elderly.7 Besides
social insurance, there was virtually no comprehensive private insurance at the time.8

These programs share six common features. First, they were employment-based

and generally covered only active workers except for GEI and VI, which also provided

 

6 Several other social insurance programs also existed at the time but only accounted for a very small
portion of the elderly population.

All these numbers are my calculations using administrative statistics from Bureau of Central Trust,
Bureau of Labor Insurance, Bureau of National Health Insurance, Veteran Affairs Commission and
Ministry of the Interior.

8 More precisely, private insurance provided only very limited coverage and generally did not cover any
outpatient care (Liu and Chen 2002; Cheng 2003).

coverage to retirees and dependents.9 Second, they were mandatory and all eligible
individuals were required to enroll in one of these programs. Third, they provided
comprehensive medical benefits including outpatient, inpatient, emergency care and
prescription drugs. 10 Fourth, they all adopted a single-payer fee-for-service (F FS)
system.ll Fifth, there was virtually no copayment for any service.'2 Sixth, the insurance
premium was generally shared among government, employers (if applicable) and
insurees.

Given the nature of the pre-NHI insurance market, the remaining 25% of the
elderly was left uninsured. Many of them were women and not working. In general, the
uninsured elderly had to pay out-of-pocket for their health care expenditures and utilized
less health care than their insured counterparts (Cheng and Chiang 1997).

National Health Insurance

After a decade of planning, the government decided to consolidate all social
insurance programs into a universal and compulsory National Health Insurance (N H1).
The National Health Insurance Act (N HIA) was passed in the legislature on July 19, 1994
and took effect on March I, 1995. Meanwhile, a new government agency called the

Bureau of National Health Insurance was established to administer NHI.

 

9 Since F HI and LI were supposed to cover only active farmers and private sector workers, it seems
puzzling at first glance that many elderly were insured by the two programs in 1994. This was actually a
result of loose law enforcement. According to the law, farmers and workers could sign up PH! and LI by
becoming members of some farmers’ association or association of some occupation or union. By law, these
associations were supposed to screen and track their members’ work status. In practice, however, screening
and tracking work status is difficult. Therefore, many elderly retained their insurance through this loophole
even afier they had retired.

'0 In addition to medical care, they also provide maternity, injury, disability, death, and retirement benefits.
H Each program had its own contracted provider network. According to Cheng and Chiang (1997), roughly
85% of hospitals and 70% of clinics in Taiwan were contracted with these social insurance programs in
1994. Care providers within each network were reimbursed solely by a government agency based on a
fixed fee schedule. Moreover, insurees had the right to choose any provider within the network without any
referral.

'2 Patients typically only paid a fixed “registration fee” NT$ 50 (USS 1.7) for each outpatient visit.

6

NHIA required all of the previously uninsured to enroll in NHI and insurees of the
various social insurance programs to be automatically transferred to NHI. '3 Due to its
mandatory and universal nature, NHI extended its coverage quickly. By the end of 1996,
about 93% of the population had enrolled in NHI. Among the elderly aged 60 and older,
the coverage rate was near 100%.

In general, NHI provides a similar wide range of medical benefits as previous
social insurance programs did. M The insurance premium is also shared among
government, employers (if applicable) and insurees. For low-income people and veterans,
their premium is entirely paid by the government.

One feature that distinguishes NHI from previous programs is its copayment
scheme. Typically, one pays an amount of NT$ 100 to NT$ 200 (roughly USS 3.3 to US$
6.7) copayment for each outpatient visit. For a hospital stay, the coinsurance rate varies
from 5% to 30% depending on the length of stay and type of room, and the total
copayment is subject to an upper limit. Copayments are waived for low-income people,
veterans, and people with some major illnesses such as cancer, end-stage renal disease,
etc.

NHI also has its own contracted provider network. According to the Department
of Health, about 98% of hospitals and 91% of clinics contracted with NHI in 1996. In its
early stage, NHI adopted a FF S system to reimburse its contracted care providers.
Unfortunately, this system was unable to contain medical care expenditures, and NHI

faced its first budget deficit two years after its inception.

 

’3 Active military personnel were not covered by NHI until 2001. These social insurance programs still
exist today and continue to provide non-medical benefits.
'4 Except that NHI has more preventive care and home care benefits.

7

To tackle its growing budget woe, NHI adopted a series of measures. For example,
it imposed a new copayment on prescription drugs in 1999 and raised the insurance
premium by 7 percent in 2002. It also raised copayments for patients who have
“excessive” doctor visits.15 In addition, a Global Budget System (GBS) that uses a
predetermined budget to control expenditures has been applied to all medical services
since 1998 (Hsueh, Lee and Huang 2002; Lee and Jones 2004). MoSt recently, a
Diagnosis-Related Group (DRG) system that reimburses hospital service based on
classifications of cases that use similar hospital resources was launched in 2008.
Uninsurance Trend: 1985-2005

While it is well—known that uninsurance in Taiwan was mostly eliminated by NHI,
it is relatively less noticed that the uninsurance rate actually had been declining fast as
early as a decade before. Figure 1.1 shows the uninsurance rate in Taiwan from 1985 to
2005.16 The solid line illustrates the uninsurance rate of the whole population and the
dashed line is that of the elderly 60 and older. Both lines show a downward trend in the
pre-NHI era (1985-1994). The downtrend in the pre-NHI era Was mainly resulted from a
series of expansions in the major social insurance programs. For example, GEI extended
its coverage to private school staff in 1980, to dependents of government employees in
1982, to retired private school staff and their spouses in 1985 and to dependents of

private school staff in 1990; FHI was initiated in 1985 and became a mandate for farmers

 

'5 For example, the copayment is raised by NT$ 50 starting on one’s 49th visit within a year and by NTSIOO
starting on one’s 157‘h visit (Liu and Chen 2002).

16 Numbers are my own calculations using statistics from the following sources. Population statistics are
from Statistical Yearbook of Interior published by the Ministry of Interior, which is available at
http://www.moi.gov.tw/stat/english/year.asp. Numbers of GEI insurees are from GE] Statistics published
by Bank of Taiwan (formerly by Bureau of Central Trust), which is available at
http://www.bot.com.tw/GESSl/Statics/default.htm. Numbers of FHI and LI insurees are from Annual
Report published by Bureau of Labor Insurance, which is available at http://www.bli.gov.tw/en/. Numbers
of VI insurees are kindly provided by Veteran Affairs Commission (VAC). They are not available on-line
but can be requested from the VAC.

 

 

o4 ~e——o——-*~-o-—-o—

 

1985 1987 1989 1991 1993 1995 1997 1999 2001 2003 2005

 

-—0— Whole population --O-- 60 & older

 

 

 

Notes:the uninsured before 1995 refers to people who were not covered by GEI, Ll FHI and VI; the
uninsured starting in 1995 refers to people who were not enrolled in NHI.

Sources: Bureau of Central Trust; Bureau of Labor Insurance; Bureau of National Health Insurance;
Veteran Affairs Commission; Ministry of Interior. See note 16 for more details.

Figure 1.1. Uninsurance Rate: 1985-2005
in 1989 (Peabody et al. 1997; Chiang 1997). The decline was particularly sharp for the

elderly between 1987 and 1989 when FHI was quickly expanded. F HI had a particularly
large impact on the elderly because farmers constituted a major component of the elderly
population at the time. The downtrend of uninsurance in the pre-NHI period has an
important implication—many elderly people changed their insurance status in the pre-
NHI period. This suggests that we should use the insurance status right before the
initiation of NHI in 1995 to demarcate between the previously uninsured and the
continuously insured elderly.

1.3. Empirical Framework

Health Care Utilization

Our main identification strategy relies on the fact that all previously uninsured
elderly were required to enroll in NHI regardless of their health. In practice, no matter
when they actually enrolled, their coverage automatically took effect on March 1, 1995.
This mandatory feature of NHI eliminated the possibility for them to opt out and
generated exogenous variation in their insurance status before and after 1995.

Suppose we observe the average utilization outcomes of the previously uninsured
elderly before and after NHI was implemented, we could have used the observed
difference in outcomes to measure the NHI effect on utilization. However, this difference
is likely to be confounded, especially when the two outcomes are far apart in time, by
other factors such as aging, environmental change, medical technology improvement
etc. '7 The challenge is thus how to filter out the confoundedness.

One natural choice is to use the difference in outcomes experienced by their
continuously insured counterparts and adopt a difference-in-differences (DD) strategy
under two standard assumptions. The first assumption is that the continuously insured
elderly were not affected in terms of utilization outcomes by transferring from previous
social insurance programs to NHI. '8 The second assumption is that on average both the

previously uninsured and the continuously insured elderly would have experienced the

 

'7 Another concern is that the previously uninsured elderly may have anticipated the coming of NHI and
delayed their pre-NHI utilization until 1995. This anticipatory effect could possibly lead to an overestimate
of the NHI effect on utilization. However, the earliest time that the public in Taiwan learned that NHI
would be implemented in March 1995 was in July 1994 after the National Health Insurance Act was passed
in the legislature. In our later analysis, we compare utilization of the elderly in 1993, 1996, 1999 and 2003.
Back in 1993, nobody knew if and when NHI will come. In addition, if the previously uninsured elderly
really repressed their utilization in 1994, this pent-up demand was more likely to be satisfied in 1995 right
after NHI was implemented than 1996 or later. Therefore, the anticipatory effect should be of less concern
in our analysis.

'8 Or, loosely speaking, the NHI effect on the continuously insured elderly was so minimal that changes in
their outcomes mainly reflect non-NHI factors.

10

same confoundedness in the absence of NHI.l9 By these DD assumptions, we aim to
estimate the NHI effect on the previously uninsured elderly, or the treatment effect on the
treated.20 We discuss the validity of the first assumption in more detail below.

The continuously insured elderly could possibly be affected by NHI in two
aspects. First, while under NHI they still enjoyed a similar wide range of medical benefits
as before, they now had to pay copayments, which were virtually absent under the
previous programs. Cheng and Chiang (1997) compare the utilizations of the
continuously insured adults before and after 1995 and find that the new copayment
scheme seemed to have only insignificant effect on curbing utilization. They suggest that
the effect of copayment on utilization was likely offset by more contracted medical care
providers under NHI and relatively inexpensive cost sharing.

Second, the inclusion of the previously uninsured into NHI may have caused an
adverse effect on the continuously insured, if there was not sufficient medical care supply
to serve the sudden rise in demand. Nevertheless, medical care supply actually outgrew
the population in Taiwan from 1986 to 1995. Specifically, while the population increased
by 10% during this period, the numbers of medical personnel, hospitals and clinics grew
by 70%, 34% and 37% respectively?"

Under these assumptions, we use the following two DD models to estimate the

NHI effect on utilization.

 

'9 This assumption is generally not testable unless there was a group not subject to NHI. However, if such a
group really existed, we would have used it as a control group to measure the common confoundedness.

0 If we want to estimate the general treatment effect (the NHI effect) on the entire population using this
DD approach, we would have to additionally assume that the treatment (N HI) would have the same effect
on the control group (the continuously insured elderly), should they become treated.

2' Numbers are calculated based on official statistics released by the Department of Health and available on
its website http://www.doh.gov.tw/EN2006/index_EN.aspx. Part of this growth in supply was resulted from
the govemment’s planning in anticipation of NHI, and another part was from the response of the private
medical care providers who foresaw the potential profits brought by NHI.

11

(1) Y}, = .30 + ,BIUIi + ,BZYR, + ,B3U1ix YR, + XiB + git,
where i indexes individuals; I indexes time; Y is a dummy variable indicating one’s
utilization of certain type of care in a given time period; UI is a dummy variable
indicating the previously uninsured elderly; YR is a year dummy indicating the year in the
post-NHI period; X is a row vector of pre-NHI controls including age, sex, education,
ethnicity, marriage, residence region, living arrangement, employment history, income,
self-reported health and health behavior; 22 B is a column vector of coefficients
corresponding to the variables in X; 8 is an unobserved error term. Equation (1) is a linear
probability model and estimated by pooled OLS. 23

Conditional on pre-NHI characteristics, ,6] measures the outcome difference
between the previously uninsured and the continuously insured elderly in the pre-NHI

period; ,62 picks up the common confoundedness; ,63 assesses the difference in outcome

differences between the elderly groups before and after 1995, that is, the NHI effect.

To reduce the possible bias caused by non-random attrition, an issue will be
discussed in more detail in the data section, we estimate equation (1) using three balanced
two-year pooled panels with one year in the pre-NHI period and another in the post-NHI
period—1993 versus 1996, 1993 versus 1999 and 1993 versus 2003. Every elderly

subject is observed in both years of each pair. We pool in pairs and run separate

 

22 Living arrangement is to capture the possible non-monetary resources provided by their adult children
with whom they live; employment history is highly correlated with their pre-NHI insurance status as well
as their lifestyle which is further correlated with their health and utilization. For example, both farmers and
government employees tend to be insured in the pre-NHI period but they also tend to have different
lifestyles. In particular, farmers are more likely to involve in strenuous work, while government employees
tend to live a sedentary lifestyle.

23 We also estimated a probit model. However, the estimated NHI effects are similar to the linear
probability model.

12

regressions in order to retain as many observations as possible.24 In addition, we use
cluster-robust standard errors for statistical inferences because of the pooling.

In equation (1), the sample compositions of the three separate regressions are
partially different. In other words, the NHI effects across years are estimated based on
partially different samples. Alternatively, we estimate the following pooling model with
the sample confined to the same group of elderly.

(2) Y,-, = ,60 + fllUIi + flZYR96, + ,63YR99, + ,B4YRO3, + ,65Uli x YR96, + ,66U1ix YR99,
+ ,67UI,- x YRO3, + XiB + a“,

where YR96, YR99 and YR03 are year dummies indicating 1996, 1999 and 2003
respectively; other notations are the same as before.

Equation (2) is estimated using a balanced panel pooled across all 4 years: 1993,
1996, 1999 and 2003. The three interaction terms measure the NHI effect in 1996, 1999
and 2003 relative to 1993 using the identical sample. Again, cluster-robust standard
errors are used for inferences. Because we restrict the pooled sample to be balanced
across all four years, the sample is much smaller, compared with the three pooled pairs
used in equation (1), and constitutes probably only the healthiest elderly who were able to
survive to 2003. However, this pooling allows us to compare the NHI effect in the short-
term (say, in 1996) to that in the long-term (say, in 1999 and 2003) at least for the
subgroup who survived to the end of 2003.

Moreover, we devise the following more general DD model to provide suggestive
evidence of the validity of the first assumption of our DD strategy. We divide the

continuously insured elderly into five previous social insurance programs.

 

24 . . . . .
The more years we Include In one regressron, the more observations we lose m order to keep a balanced
pooled panel.

13

(3) Yr: = .50 + fllUIi + .3205]: + .53L1i + fl4V1i+ .5501: + A6”: + fl7U1rx YR:
+flsGEIi XYRI +fl9Lli XYR, +fl10VII-XYRI +flHOII-XYR1 +XiB+gi19

where GEI, L], V1, and 01 are dummy variables indicating the elderly who are previously
covered by GEI, LI, VI and 01 (meaning other social insurance programs) respectively;
all other notations are the same as in equation (1); note that the reference group here is
FHI, the largest insurance group in the pre-NHI period. Again, equation (3) is estimated
by OLS using three separate two-year pooled panels.

If we maintain the common time trend assumption but relax the first assumption,

,68 in equation (3) actually measures the “NHI effect” for those who transferred from

GEI to NHI relative to the “NHI effect” for those who transferred from FHI to NHI.

Likewise, ,69 , .510 and ,6“ measure the “NHI effect” for those who shifted from L1, VI

and 01 to NHI respectively. If the continuously insured elderly groups were really
affected by NHI, it is likely that the “NHI effects” for these five different groups were
also different because these programs were different from each other.25 This suggests a
test of the following null hypothesis.

HypotheSiSI .38 = 159 = filo = .1311 = 0-

This null hypothesis states that either the continuously insured elderly were not
affected by transferring from former insurance programs to NHI or they were affected to
the same extent. However, the latter possibility seems less probable.

Mortality Hazard Ratio
In general, we follow the same identification strategy and assumptions as in the

utilization analysis. However, in contrast to the linear DD model of utilization, we

 

25 Cheng and Chiang (1998) point out the three major programs, GEI, LI and F HI, differ in their medical
availability.

14

incorporate the DD design into an exponential hazard function in order to fully exploit
the mortality data. Let us define the hazard function as a conditional instantaneous rate of
dying at any given time point as follows (throughout this chapter, “hazard” always refers
to mortality hazard).

(4) ’70): lim Pr(tST<t+At|T2t)’

At —-)0 At

 

where h is the hazard function; T is survival time that takes on only non-negative values, I;
Pr (-) is the probability that an individual dies at some point between t and t+At
conditional on having survived for t units of time.

Further, let us model the hazard function as an exponential function with a DD
design as follows.
(5) h(t;UI, X) = exp(fl0 + AU] + ,BZPOST(t)+ ,63UI x POST(t)+ XB),
where t is the number of months of survival since the beginning of our analysis period
from 1993 to end of 2003; POST is an indicator for the post-NHI period which is
therefore a function of t; other notations are the same as in equation (1). In the set of
baseline controls, X, we additionally add age-squared and age-cubic. Based on this
hazard function, an individual’s survival time in months, t, is assumed to be exponentially
distributed. We estimate the parameters of the model by the maximum likelihood method.

Based on the hazard function, the log-likelihood function can be constructed
accordingly as In L = 227:] {d1 In f,- (ti)+ (1 — di)ln 51(0)} , where d,- is an indicator that

equals on if the individual has died by the end of our analysis period; f,-(t,-) is the

density function of t,- and equals hi(t,-)exp{— I: hi(u)du} ; S,-(t,-) is the survivor function

15

and equals exp{— I: h,-(u)du} . Note that if an individual was still alive at the end of the

analysis period, her I,- equals the length of the analysis period and her contribution to the
log likelihood is In Si(t,-); otherwise, her contribution is In f,(t,-).

In equation (5), exp(fl0 + [32 POST (t)) represents the baseline hazard common to
both the previously uninsured and continuously insured elderly. Note that the baseline
hazard can change across periods because of the time-varying variable, POST. Equation
(5) can be viewed as a piecewise constant hazard model because it allows the baseline
hazard to vary across the pre- and post-NHI periods but remain constant within each
period. Assuming a constant hazard in each period may not be sensible in the context of
elderly. However, for our DD strategy, we only need to estimate the change in baseline
hazard as a result of NHI. The exact shape of the baseline hazard function is not our main
concern. To allow additional flexibility in the way the hazard changes with time, we also
estimate a version of the model with individual year effects that are common to all
individuals.

In equation (5), conditional on the pre-NHI controls, exp(fl]) measures the ratio
of the hazard of the previously uninsured elderly to the hazard of the continuously

insured elderly in the pre-NHI period, while exp(fll + ,63) assesses that hazard ratio in

the post-NHI period, and exp(fl2) is the pre- versus post-NHI within-group

confoundedness which is assumed to be common for both groups.

Alternatively, we can interpret the ,Bs in terms of log hazards. By taking natural
logarithm of equation (5), we turn the hazard function into a log-hazard function and ,6]

is the difference in log-hazards in the pre-NHI period, ,62 is the common

16

confoundedness, and ,63 is the difference in differences in log hazards. However, hazard

ratios are more interpretable than log hazards.
1.4. Data
A Longitudinal Elderly Survey

We use a longitudinal survey of the elderly called the Survey of Health and Living
Status of the Middle-aged and the Elderly in Taiwan (referred to as the survey hereafter)
administered by the Bureau of Health Promotion (BOH) in Taiwan (formerly known as
Institute of Family Planning).26 The survey contains rich information regarding
demographics, health care utilization, health outcomes, health insurance, income,
employment history etc. The survey drew a nationally representative elderly sample of
4,049 who were aged 60 and older in 1989. The elderly were then followed in 1991, 1993,
1996, 1999 and 2003.27

In our analysis of utilization, we only use the 1993, 1996, 1999 and 2003 surveys
for the following two reasons. First, the 1989 survey contains no health insurance
information and its utilization questions are different from those in latter surveys (see
Appendix 1 for more details). Second, the 1991 survey lacks health utilization
information. In addition, we drop the privately insured elderly and those with unknown
insurance status in 1993.28 In the end, our final sample consists of 785 uninsured and

2,351 insured elderly in 1993.

 

26 Although the middle-aged are mentioned in the title of the survey, we do not use the middle-aged
because they were added into the survey in 1996 and we do not know their pre-NHI insurance status.

27 All waves were conducted in person except for the 1991 survey, which was done by telephone. Every
wave was generally administered from March to June; the completion rate for each wave is over 90%. In
1996, the survey added in a middle-aged sample of 55 to 66 years old. The middle-aged sample is not used
in this study because it has no pre-NHI insurance information.

28 There are 16 privately insured and one with unknown insurance status. See appendix 1 for more details.

17

 

Table 1.1. Elderly Insurance Status in 1993

 

n %
Government Employee Insurance Programs (GEI) 483 15.40
Labor Insurance (LI) 164 5.23
Farmer’s Health Insurance (F HI) 1,227 39.13
Veteran Insurance Programs (V I) 391 12.47
Other Insurance Programs (01) 86 2.74
Uninsured 785 25.03
Total 3,136 100.00

 

Notes: GEI includes Government Employee Insurance, Government Retiree Insurance, Government
Employee’s Dependents Insurance and Government Retiree’s Dependents Insurance. VI includes
Veteran Insurance and Veteran’s Dependents Insurance. 01 includes Councilor Insurance, Welfare
Insurance, Fishermen Insurance, Military Insurance, Military Personnel’s Dependents Insurance.

As shown in Table 1.1, FHI, GEI and VI amount to 39%, 15% and 12% of the
sample respectively, while LI accounts for only 5% of the sample.29 The insurance
structure reflects the agriculture-oriented economy for this generation as well as a
massive influx of troops from mainland China to Taiwan around 1949 due to a civil war.

The survey is supplemented with three sources of mortality data. The survey itself
collects information about death month and year of each deceased elderly up to the end of
2003. To confirm with administrative data, the survey uses national ID numbers of the
elderly to link them to mortality records at the Department of Health (DOH) and Ministry
of the Interior (MOI). However, for various reasons, about 6% of the DOH death records
are inconsistent, either in death month or year, with the survey’s own collection, and only
a rather small number of deaths are successfully linked to MOI data.’0 Instead of

dropping the mismatched cases, we choose to use both the survey’s collection and the

 

29 In other words, Ll only accounts for 7% of the insured elderly in our sample in 1993. Earlier on in the
institutional background section, we mentioned that LI, based on administrative statistics, accounted for
about 18% of the insured elderly aged 60 and older in 1994. However, in our sample, the elderly were aged
at least 64 and older in 1993. Elderly in our sample were older and very few of them were still working in
the private sector in 1993. And, recall that LI covered the private sector workers.

30 According to the Bureau of Health Promotion, some national ID numbers provided by the elderly are
wrong and some others’ ID numbers are not even available. However, we are not told about how many
cases have this problem.

18

DOH data to estimate the hazard model and present both sets of results. From the
beginning of 1993 to the end of 2003, 1,367 and 1,389 out of the total 3,136 elderly had
died according to the survey’s collection and the DOH data, respectively.
Characteristics of the Elderly in the Pre-NHI Period

Table 1.2 summarizes the characteristics of the uninsured and the insured elderly
in 1993. Compared with their insured counterparts, the uninsured elderly are more likely
to be women, unmarried or widows, Minnan people, and live with their children in an
urban area in the north region of Taiwan. They also tend to be less educated, less likely to
be working and receive less monthly income. Meanwhile, they are more likely to live
with their adult children.

As expected, their insurance status is correlated with their current or past
occupation as the insured elderly are much more likely to have ever worked as farmers,
government employees and military personnel than the uninsured elderly. In terms of
health outcomes, the uninsured elderly are less healthy and more likely to die than the
insured elderly. Surprisingly, the uninsured elderly also live slightly healthier in terms of
smoking and betel nut chewing.3 I
Utilization Variables

To measure their utilization behavior, we use information about whether they
have ever foregone doctor visits due to lack of money in a 3-month period, their
outpatient visits in a month, hospitalizations and emergency room (ER) visits in a year.
See Appendix 1 for original survey questions in Chinese and their English translations.

Table 1.3 summarizes these variables across years. Column (1) and (2) in the first

panel show that the probability of forgoing doctor visits due to lack of money drops by 4

 

3' Betel-nut chewing is found to be associated with mouth cancer by the medical profession.

19

 

Table 1.2. Characteristics of the Elderly in 1993

 

Insured Uninsured

Age (mean) 71 72
Female (%) 39 57
Married with spouse present (%) 67 47
Ethnicity (%)

Minnan 56 74

Hakka 17 10

Mainlander 24 14

Aboriginal 2 1
Region (%)

East 9 6

North 23 38

Central 36 29

South 32 26
Urban area (%) 59 77
Live with adult children (%) 64 70
Years of education (%)

0 Year 45 60

1-6 Years 33 31

7 Years and more 22 9
Currently employed/ self-employed (%) 23 11
Monthly income > NT$10,000 (%) 45 31
Ever worked as (%)

Farmer 51 26

Government employee 15 5

Military personnel 21 6
Self-reported health (%)

Very good/ good 41 36

Fair 33 31

Bad/ very bad 21 24

No response 5 9
Functional limitation (ADL) (mean) 0.38 0.54
Have one or more chronic condition(s) (%) 75 74
One Year Mortality (%) 3.3 4.8
Health behavior (%)

Smoking 29 27

Drinking 13 13

Betel nut chewing 6 4
Total 2,351 785

 

Notes: ADL is an average score of 12 functional limitation items of daily activities. Each item ranges
from 0 (no difficulty) to 3 (cannot do it at all). Chronic conditions include arthritis, asthma, cancer,
cataract, diabetes, hypertension, heart problem, kidney problem, liver problem, stroke, and ulcer.
Mortality is based on survey’s collection.

20

 

Table 1.3. Summary of Utilization Variables

 

Probability Average Positive Visits
( 1) (2) (3) (4)
Insured Uninsured Insured Uninsured

F orego Visits

1993 0.01 (2,351) 0.06 (785)

1996 0.01 (1,928) 0.02 (584)

1999 0.01 (1,649) 0.03 (503)

2003 0.01 (1,249) 0.02 (3 62)
Outpatient Visits

1993 0.50 (2,347) 0.34 (784) 2.9 (1,153) 2.8 (261)

1996 0.63 (1,926) 0.60 (583) 2.6 (1,052) 2.6 (286)

1999 0.71 (1,648) 0.63 (503) 3.1 (1,154) 3.0 (312)

2003 0.74 (1,248) 0.70 (361) 2.4 (923) 2.3 (252)
Hospitalizations

1993 0.20 (2,351) 0.10 (785) 1.4 (463) 1.2 (79)

1996 0.22 (1 ,928) 0.20 (584) 1.6 (419) 1.3 (114)

1999 0.24 (1,649) 0.23 (502) 1.6 (391) 1.6 (1 16)

2003 0.20 (1,249) 0.24 (362) 1.8 (254) 1.4 (86)
ER Visits

1993 0.1 1 (2,351) 0.07 (783) 1.4 (240) 1.2 (52)

1996 0.13 (1,928) 0.1 1 (584) 1.9 (254) 1.9 (64)

1999 0.15 (1,63 7) 0.15 (502) 1.8 (245) 1.7 (76)

2003 0.18 (1,249) 0.17 (362) 1.7 (225) 1.7 (62)

 

Notes: observations in parentheses; majority of non-responses (over 70%) is due to death; percentages of

non-responses due to death are roughly equal between insured and uninsured; about 9% of the insured and

1 1% of the uninsured respondents in 1996 fail to recall their exact number of outpatient visits; for other
Jean and visits, percentages of failed recalls are less than 2%.

percentage points for the previously uninsured elderly from 1993 to 1996 and remains at
roughly 2% afterwards, whereas it remains stable at 1% for the continuously insured
elderly throughout this period. This suggests that NHI seems to have removed the
financial barrier to visiting doctors for some of the previously uninsured elderly.

In the following three panels, we observe a similar pattern across outpatient visits,
hospitalizations and ER visits. In 1993, the then-uninsured elderly have a much lower
probability of utilization than the insured elderly. However, in 1996, the probability of
utilization for the previously uninsured elderly rises dramatically. For example, the

probability of outpatient visits increases by 26 percentage points, the probability of

21

hospitalizations by 10 percentage points and ER visits by 4 percentage points for the
previously uninsured elderly.

On the other hand, the changes for the continuously insured elderly are not as
dramatic as for the previously uninsured elderly. Furthermore, the increase in 1996 brings
the previously uninsured elderly to par with their continuously insured counterparts
afterwards.

In contrast to probability of a visit, the two elderly groups have very similar
numbers of positive visits conditional on having at least one visit across years (column (3)
and (4)), suggesting that NHI affects the extensive margin but not the intensive margin of
utilization.

Sample Attrition Problem

Table 1.3 obviously shows that our sample depletes across years. The sample
attrition is mostly a result of death and could possibly bias our DD estimation of the NHI
effect on utilization if the attrition is not random. For example, if the uninsured elderly
tend to die faster than the insured and the deceased elderly are more likely to be high
users than their living counterparts, then the DD method with an unbalanced panel tends
to underestimate the NHI effect on utilization. We verify this by comparing utilization in
1993 of those who only appeared in the 1993 survey with that of those who appeared in
1993 and later surveys. We find that the former actually had higher utilization in 1993
than the latter. To reduce this possible attrition bias, we use balanced panels in all our
regressions.

Yearly Mortality Rate

22

To get a general picture of the crude mortality, we use the yearly mortality rate
defined as deaths occurred during a 12-month period divided by those alive in the
beginning of that period. To correspond to the implementation date of NHI, each period
begins in March and ends in February. Figure 1.2 plots the yearly mortality rates for the
two elderly groups from March 1993 to February 2003 using the survey’s collection of

death records and the DOH mortality data.

I

I

l

 

l

.02 .03 .04 .05 .06 .07 .08
J

 

 

 

5 _
o _.
l I F l I I I I I I
1993 1994 1995 1996 1997 1998 1999 2000 2001 2002
Year
—O—- Uninsured-Survey --O-- Insured-Survey
---O--'- Uninsured-DOH ~~~~0-~-- Insured-DOH

 

 

 

Notes: every year period starts in March and ends in February.
Figure 1.2. Yearly Mortality Rate by Insurance Group
As shown, whereas the yearly mortality rate of the continuously insured elderly
rises steadily from 3% in 1993 to 4.5% in 1996, that of the previously uninsured elderly
first rises from 5% in 1993 to 7% in 1994 and then quickly declines to 6% in 1995 and to
4.7% in 1996. After 1996, the two groups basically merge and move together until 1999

when they start to oscillate up and down again. However, the mortality gaps between the

23

two groups after 1999 are never as big as in the pre-NHI period. At first glance, the
pattern in yearly mortality rate seems to be associated with the pattern in utilization.
1.5. Estimation Results

NHI Effect on Utilization

Estimation results of NHI effects on probabilities are reported in Table 1.4.32 Note
that model 1 is the model stated in equation (1) using three separate panels pooled across
only a pair of years and model 2 is the pooling model stated in equation (2) using a panel
pooled across all four years. In general, the results show that NHI removes the financial
barrier and largely increases utilization for the previously uninsured elderly relative to
their continuously insured counterparts. More precisely, in panel A (1993 versus 1996),
model 1 results show that NHI decreases the probability of foregoing doctor visits by 4
percentage points, increases the probability of outpatient visits by 13 percentage points
and the probability of hospitalizations by 9 percentage points for the previously uninsured
elderly relative to their continuously insured counterparts; and these results are
statistically significant. In contrast, we only find an insignificant 1.7-percentage-point
change in their probability of ER visits.

Model 1 results in panel B (1993 versus 1999) and C (1993 versus 2003) are
basically similar to panel A except that there is a significant 5 percentage-point increase
and an insignificant 3.4 percentage-point increase in the probability of ER visits in the
panel B and C respectively. This similarity across Waves suggests that the NHI effect on

utilization was realized soon after the NHI was implemented.

 

32 We also estimate NHI effect on numbers of visits conditional on having at least one visit. However, we
do not find any significant evidence and the results are thus not reported.

24

 

Table 1.4. NHI Effect on Probability of Utilization
F orego Visits Outpatient Visits Hospitalizations ER Visits

(1) (2) (3) (4) (5) (6) (7) (8)
Model] Model 2 Model] Model 2 Model] Model 2 Model] Model 2

A: 1993 versus 1996
-0.041 -0.043 0.131 0.143 0.093 0.042 0.017 0.000
(—3 .66) (-2.79) (4.51) (3 .68) (4.18) (1.62) (0.9) (0.02)
[2,512] [1,469] [2,504] [1,463] [2,512] [1,469] [2,510] [1,458]

B: 1993 versus 1999
-0.032 -0.038 0.09 0.09 0.095 0.062 0.049 0.036
(-2.74) (-2.58) (2.88) (2.37) (3.75) (2.15) (2.27) (1.48)
[2,152] [1,469] [2,148] [1,463] [2,151] [1,469] [2,137] [1,458]

C: 1993 versus 2003

-0.046 -0.043 0.1 17 0.11 0.126 0.139 0.034 0.045

(-3.32) (-2.95) (3.21) (2.88) (4.20) (1.62) (1.33) (1.64)

[1,611] [1,469] [1,607] [1,463] [1,611] [1,469] [1,609] [1,458]
Notes: dependent variables are dummy variables indicating each type of utilization; only the
interaction terms are reported; all pooled panels are restricted to balanced panels; model I only pools a
pair of years; model 2 pools all 4 years; cluster-robust t in parentheses; subjects in brackets; all
regressions include pre-NHI demographic, economic and health controls.

Model 2 results are similar to model 1 results with the largest differences in the
cases of hospitalizations and ER visits. In particular, NHI effects on hospitalizations in
1996 and 1999 estimated by model 2 are smaller than those estimated by model 1. Recall
that model 2 uses only the elderly who appeared in all four years, while model 1 only
requires the elderly to appear in two years. More specifically, some of those who
appeared, for example, in the 1993 versus 1996 sample died soon after 1996, and these
deceased elderly were supposedly the most fragile ones who most likely would have been
hospitalized in 1999 were they still alive. However, these deceased elderly are excluded
from the model samples. Therefore, this finding suggests that differences in NH] effects

are likely resulted by attrition due to death.

Subgroup Analysis of Utilization

25

Since the previously uninsured elderly and their continuously insured counterparts
were different in many aspects in their baseline characteristics, e.g. age, gender, ethnicity,
health, etc., it is interesting to see if the NHI effects also differ among subgroups.
Especially, one might expect that the sick elderly, e.g. elderly with chronic conditions,
would have benefited more from NHI than healthy elderly.

Panel A in Table 1.5 shows that the relatively younger uninsured elderly (aged
64-73 in 1993) benefit from NHI more than the very old uninsured elderly (aged 74 and
older in 1993) in terms of foregoing doctor visits and outpatient visits, but the very old
uninsured benefit more in hospitalizations and ER visits. In panel B, uninsured women
tend to benefit more than uninsured men except for outpatients visits. In panel C,
uninsured non-Minnan people seem to benefit more than uninsured Minnan people.

As for chronic conditions, in our sample, diabetes (10%), hypertension (30%) and
heart diseases (21%) are the three most prevalent chronic conditions among the elderly.
These three conditions typically require constant medical attention and medications. We
thus compare people with these conditions to those with no chronic conditions. The
results suggest that (in panel D) uninsured elderly with these chronic conditions
experienced larger increases in utilization than those without chronic conditions.

NHI Effect on Mortality Hazard

In Table 1.6, the top panel reports the original estimates of coefficients in
equation (5), while the bottom panel reports pre- and post-NHI hazard ratios based on the
estimated coefficients. Columns (1) to (3) use the survey’s collection and column (4) to
(6) use the DOH mortality data. Column (1) and (4) do not control for pre-NHI

characteristics; column (2) and (5) control for pre-NHI demographic and economic

26

 

.comEmEoQ mzocmnsm some 5 3885 Co 032:; 2: .5: Epoxo 295:8 E Z

-8: 25.2: 2232mm: :5 :28: :28: go. 323 m9..— wEms _ .25.: Eat 08 3.3.2 ”88—85 5 £8.33 608553: E mutate: ~ 538-5620 ”882

 

as: was: was: :s: E: :m: as: as:
3.3 38s as. c and s. 3 2a.: 6.... 3 A33
N85- m as- 85 _ _ 3 Need 83 ~85- Sod-
02 86:35:»: 02 communion»: oZ :oficotoma 02 86:38:»:
8m: 5: :8: Gm: .8: 7.8: :8: GM:
and 833 fine so: 2 _. c and 81$ £23
_ 3 $3. No 83 see 23 weed- Rod-
ommvmmn— two: £509me QmMOEQ “Maui wagon—ED 093me “.803 mononmma @9339 two: mouobmma
252280 3530 tam ”Q
:3: 8%.: as: 8%.: as: #8.: as: 8%.:
God: Gas Gas 5.9 Gad seed 93.3 Game
mood- a2; 83 £3 93 a; god. God-
SMEEIGOZ Sufism—Z ENE—21:02 5:52 Cmcmzrcoz 3:52 «352-952 52532
5335 Am Ho
7.3.: :2: $3.: 82.: $3.: 32.: not: 82.:
A83 as. : :sd can 5.0 3.9 Ea: 5.3
coed- e85 ass Sec 9.3 2 3 Rod. s3?
252 26th 2m: 0380.» 2&2 Bantam 2m: 0380....—
xom hm ”m
as: 6%.: :se: 33.: as: fine: as: 33.:
5e. : $23 84.9 and am. : ass 30.3 as?
weed 23.. mm; 83. sod 9.3 23. N86-
is ms 2 S is ms s S is s e S is ms 9 S
82 a a? E a.
a: 5 6V 5 as 5 O 3
3.4.3 mm ESBNQSESI EMS “5.333% 3.5; owesok

 

eaaeaweam 3 5.35:5 a 5:332.— ae .85 52 .2 as;

27

 

Table 1.6. NHI Effect on Mortality Hazard

Data Source Survey DOH
(1) (2) (3) (4) (5) (6)
UI 0.449 0.311 0.246 0.447 0.321 0.273
(3.48) (2.34) (1.84) (3.37) (2.36) (1.99)
POST 0.307 0.438 0.55 0.381 0.492 0.576
(3.71) (5.26) (6.56) (4.49) (5.79) (6.75)
UIXPOST -0.334 -0.318 -0.268 -0.293 -0.282 -0.248
(-2.29) (-2.17) (-1.83) (-1.97) (-l .89) (-1.66)
Pre-NHI controls:
Demographic & No Yes Yes No Yes Yes
Economic
Health No No Yes No No Yes
Subjects 3,136 3,136 3,136 3,136 3,136 3,136
Deaths 1,367 1,367 1,367 1 ,3 89 l ,3 89 1 ,3 89
Hazard Ratio:
Pre-NHI 1.57 1.36 1.28 1.56 1.38 1.31
Post-NHI 1.12 0.99 r 0.98 1.17 1.04 1.03

 

Notes: I statistics are in parentheses; POST is an indicator for period after March 1995 when NHI was
implemented; hazard ratio is calculated as exp(Ul) in the pre-NHI period and exp(Ul+Ul><POST) in the
post-NHI period.

variables; column (3) and (6) further control for pre-NHI health variables. The more
controls we add in, the more conservative the estimates become.

Our main interest lies in the interaction term U1 ><POST, which represents the
difference in differences in log hazards between the two groups across the pre and post-
NHI periods.

The results show that the DD in log hazards range from -0.334 to -0.248, which
implies the log hazard of the previously uninsured elderly decreased in the post-NHI
period relative to their continuously insured counterparts. The more controls we add, the
smaller the absolute values of the DD estimates. However, they are all significant at least
at the 10% level.

To get more comprehensible interpretations, it is helpful to look at the hazard

ratios at the bottom panel in Table 1.6. In general, pre-NHI hazard ratios are all well

28

above one and post-NHI hazard ratios are close to one. More precisely, the hazard ratio is
about 1.3 in the pre-NHI period, indicating the risk of dying of the then-uninsured elderly
is about 30% higher than that of the insured elderly. It then drops to one in the post-NHI
period, implying the risk of dying for the two elderly groups merge together after NHI
has been implemented.

Instead of dichotomously cutting time into the pre and post period, we modify
equation (5) by estimating two alternative models. Model 1 includes a full set of year
dummies but remains one interaction term of U1 XPOST . Model 2 include a full set of
year dummies and interaction terms of year dummies and U1.33 In Table 1.7, the
interaction terms of UIXPOST in model 1 are similar to the results in Table 1.6. On the
other hand, the interaction terms in model 2 suggest that the NHI effect on mortality
gradually increased from 1996 and reached its peak in 1999 and then decreased towards
the end.

Subgroup Analysis of Mortality Hazard

Our subgroup analysis results suggest that NHI effect on mortality hazard differs across
subgroups. As shown in the top panel in Table 1.8, NHI seemed to have larger effect on
the relatively younger uninsured elderly, women and non-Minnan people, compared to

the very old uninsured elderly, men and Minnan people, respectively. In the bottom panel,
NHI seemed to have a much larger effect on the uninsured elderly with diabetes or
hypertension than those without chronic conditions. However, the NHI effect on elderly
with heart diseases seemed minimal, although we previously find that NHI increased

their utilization.

 

33 All regressions still include a full set of pre-NHI controls.

29

 

Table 1.7. Alternative Mortality Hazard Models

 

Survey
(1) (2) <3) (4)
Model 1 Model 2 Model 1 Model 2
UI 0.24 0.24 0.27 0.27
(1.8) (1.8) (1.96) (1.96)
U1 ><POST -0.26 -0.24
(-1.77) (-1.61)
UIX1995 0.06 0.08
(0.26) (0.33)
UI><1996 -0.29 -0.25
(-1.17) (-l.00)
UI><1997 -0.33 -0.39
(-1.46) (-1.67)
UI><1998 -0.31 -0.27
(-1 .24) (-1 .06)
UI><1999 -0.58 -0.60
(-2.33) (-2.42)
UI><2000 -0.13 -0.12
(-0.55) (-0.50)
UI><2001 -0.42 -0.39
(-l .64) (-1.64)
UI><2002 -0.31 -0.21
(-1.21) (-0.83)
Ul><2003 0.00 0.04
(0.00) (016)
Subjects 3,136 3,136 3,136 3,136
Deaths 1,367 1,367 1 ,3 89 1 ,3 89

 

Notes: all regressions include a full set of year dummies; model 1 includes one interaction term of U1
and POST, which indicates the entire post-NHI period; model 2 includes a full set of interaction terms
of U1 and year dummies; a full set of pre-NHI controls are included; estimates of year dummies are
not reported.

30

 

Table 1.8. NHI Effect on Mortality Hazard by Subgroups

 

A: By Age in 1993 B: By Sex C: By Ethnicity
64 to 73 74+ Female Male Minnan Non-
Minnan
UI 0.334 0.215 0.422 0.027 0.104 0.591
(1.62) (1.23) (2.16) (0.14) (0.63) (2.55)
POST 0.621 0.494 0.597 0.537 0.571 0.57
(5.02) (4.31) (4.14) (5.2) (5.22) (4.35)
UIXPOST -0.422 -0.169 -0.5 -0.008 -0.191 -0.587
(-1 .88) (-0.87) (-2.32) (-0.04) (-1.07) (-2.21)
Hazard Ratio: ,
Pre-NHI 1.4 1.24 1.52 1.03 1.1 1 1.81
Post-NHI 0.92 1.05 0.92 1.02 0.92 1
Subjects 2,183 953 1,365 1,771 1,901 1,235
Deaths 710 657 542 825 874 493
D: By Chronic conditions
Diabetes Hypertension Heart Diseases No chronic
condition
UI 0.438 0.602 -0.18 0.418
(1.43) (2.51) (-0.67) (1.26)
POST 0.603 0.904 0.226 0.757
(2.86) (5.6) (1.48) (3.46)
UIXPOST -0.599 -0.734 0.165 -0.334
(-1.7) (-2.8) (0.56) (-0.95)
Hazard Ratio:
Pre-NHI 1.55 1.83 0.836 1.52
Post-NHI 0.85 0.87 0.986 1.09
Subjects 328 946 669 790
Deaths 197 43 5 329 276

 

Notes: 1 statistics in parentheses; all regressions include pre-NHI controls except for the variable of
interest in each subgroup comparison. Use survey data. Results using DOH data are similar but not

reported.

Suggestive Evidence of the Validity of the F irst DD Assumption

Panel A in Table 1.9 reports F statistics and p-values to test the null hypothesis:

[1’3 = ,69 = ,610 = ,6] 1 = 0 in equation (2). In general, the results do not reject the null

hypothesis except for two cases.34 This suggests that in general NHI had minimal impact

on utilization for the continuously insured elderly.

 

34 Outpatient visits in the case of 1993 versus 1999 and 1993 versus 2003.

31

 

Table 1.9. Validity Test of The First DD Assumption
A: Utilization

F orego Outpatient Hospitalization ER
(1) (2) (3) (4)
1993 versus 1996
1.16 1.2 0.96 1.7
(0.33) (0.31) (0.43) (0.15)
1993 versus 1999
1.4 2.51 1.01 0.52
(0.23) (0.05) (0.4) (0.72)
1993 versus 2003
1.77 3.13 1.06 0.63
(0.13) (0.01) (0.37) (0.64)
B: Mortality Hazard
Survey DOH
1.16 0.88
(0.89) (0.93)

 

Notes: F statistics and p-values (in parentheses) are reported; FHI is the reference group; all regressions
include pre-NHI demoggphic, economic and health controls.

We also modify equation (4) by including 4 insurance dummy variables
indicating GEI, LI, VI and OL and interaction terms of these dummies with the POST
dummy. Then we test the same hypothesis. As show in panel B, we do not reject the null
hypothesis, which implies that our first DD assumption is also valid for the mortality
hazard model.

1.6. Conclusion

One common argument of providing universal insurance coverage to people is
that insurance will improve people’s health by giving them access to medical care and
increasing their utilization. This paper finds that Taiwan’s NHI removed the financial
barrier for the previously uninsured elderly and largely increased their utilization relative
to the continuously insured elderly. Corresponding to the increase in utilization, the
mortality hazard ratio of the previously uninsured elderly to the continuously insured

elderly dropped from 1.3 in the pre-NHI period to close to one in the post-NHI period.

32

Our findings suggest that NHI improved the health of the previously uninsured elderly
through health care.

Our subgroup analyses provide a closer look at the NHI effects. We find that the
NHI effects seem to differ across age, sex, ethnicity and health groups. While we find
that the elderly with a larger increase in medical care utilization—due to NHI—generally
also seemed to experience a larger decline in their mortality hazard, this is not always the
case. For example, we find that NHI seemed to increase the utilization for the elderly
with heart diseases but this increase in utilization did not seem to translate into a lower
mortality hazard. This suggests that whether and how much insurance can improve
people’s health may still depend on the need and effectiveness of medical care.

While several previous studies had investigated Taiwan’s NHI effects, the major
contribution of this chapter is the finding of the NHI effect on mortality, which is
different from the conclusion of Chen et al. (2007). As we mentioned in various places in
the main text and Appendix 2, the difference between our findings can be attributed to
several reasons. One of the main possible reasons is that Chen et al. did not fully utilize
the mortality data. In addition, as detailed in Appendix 2, we find different mortality rates

in 1999 than theirs.

33

Chapter 2
The Healer or the Druggist: Effects of Two Health Care Policies in Taiwan on
Elderly Patients’ Choice between Physician and Pharmacist Services

2.1. Introduction

Until the mid-19903, both physicians and pharmacists prescribed and dispensed
drugs in Taiwan, a nation with a special medicine culture that dates back to the Japanese
colonial period of 1895—1945 (Unschuld 1976; Peabody et al. 1995; Hsieh 1999, 2003;
Chou et al. 2003; Huang, Lee, and Huang 2004).35 In the eyes of many elderly people
accustomed to this culture, physicians and pharmacists were to some extent synonymous
(i.e., close substitutes), generally in the context of minor medical conditions, in that they
both prescribed and dispensed drugs.36 This chapter studies two Taiwanese health care
policies mandated in the 19905 that affected elderly people’s choice between physicians
and pharmacists: National Health Insurance (N HI), initiated in 1995, which provided
insurance coverage for all citizens, and a separation policy (SP), gradually phased in
since 1997, which forbad physicians from dispensing and pharmacists from prescribing
drugs.

Prior to the mid-1990s, many elderly people in Taiwan, especially the uninsured,
directly sought care at pharmacies because it was usually less expensive than seeing
physicians. Pharmacists generally charged only for dispensing drugs, while physicians

charged not only for dispensing drugs but also for diagnosing, prescribing, and other

 

35 Although not legally allowed to prescribe in the colonial period, pharmacists oflen provided free medical
consulting and “recommended” drugs to patients (Unschuld 1976). The line between “prescribing” and
“recommending” was usually vague. Afler the Chinese Nationalist Party took over Taiwan in 1945,
pharmacists were legally allowed to prescribe drugs for numerous minor medical conditions based on two
official prescription books.

36 It is doubtful that pharmacists treated major illnesses such as cancer or broken limbs.

34

related services. The price difference could be substantial, depending on medical
conditions. For example, to treat a common cold, an uninsured patient would have to pay
NT$300 (US$10) out of pocket if visiting a physician but only NT$100 (US$33) out of
pocket if visiting a pharmacist.37

Not only was their price different, but the quality of physician and pharmacist
services was different. While physicians were well trained to diagnose and treat medical
conditions, pharmacists were not. In the postwar period (after 1945), due to scarce
medical resources and loose law enforcement, many pharmacists were neither well
trained nor licensed (Hsieh 2003). Therefore, elderly patients who visited pharmacists
could have had received better diagnoses and care from physicians.

In 1995 Taiwan initiated its NHI program, which provided compulsory coverage
to the nation’s 21 million citizens, including 8 million who were previously uninsured.
The implementation of NHI lowered an economic barrier for the previously uninsured
elderly to visit physicians because NHI paid for their physician visits and prescription
drugs.

In 1997 a policy called yi yiao fen ye, meaning the separation of drug prescribing
and dispensing, was launched in the two biggest cities and gradually phased into all other
cities and counties over six years. The combination of drug prescribing and dispensing is
believed to provide an economic incentive for care providers to prescribe drugs more
than medically necessary and is associated with high drug expenditure. Meant to abate
drug expenditures, the SP prohibits physicians from dispensing and pharmacists from

prescribing drugs (Chou et al. 2003). In practice, pharmacists are not allowed to sell

 

37 Estimates based on personal communications with several clinic physicians and pharmacists who have
been practicing for at least 15 years.

35

drugs to patients without a physician’s prescription. Under this legal constraint, patients
are forced to first visit physicians in order to get a prescription.

However, the government made an exception in the case of physicians dispensing
drugs to patients aged 65 and older. Also, physicians at clinics are allowed to hire their
own on-site pharmacists.38 These exceptions have had critical effects on patients’ choice
on where to purchase drugs after receiving a prescription from a doctor.

To avoid confusion, throughout this article the term pharmacists refers to those
who practice at community pharmacies, while those who are hired by physicians and
practice at clinics are specifically referred to as on-site pharmacists.

NHI and the SP are two major health care reforms that affected elderly people’s
choice between physicians and pharmacists in different ways. NHI eliminated the
difference in out-of-pocket expenses between visiting physicians and pharmacists for the
previously uninsured elderly. The SP forced all elderly patients to first visit physicians in
order to get prescriptions. This paper empirically studies the effects of these two policies
on elderly people’s choice between physicians and pharmacists.

These two polices can be seen as natural experiments because their
implementation was not contingent on the health or health care utilization of the elderly.
Exploiting this feature, we adopt a difference-in-differences (DD) method to identify
their effects on elderly outpatient and pharmacy visits. We also examine the policies’
effects on self-reported health.

The first finding of this paper is that NHI induced more previously uninsured

elderly patients to visit physicians, including those who were accustomed to only visiting

 

38 According to the data presented by Chou et al. (2003), about 64 percent of clinics have hired on-site
pharmacists in the post-SP period.

36

pharmacists. Moreover, NHI raised overall health care utilization for the previously
uninsured elderly, which in turn is associated with an improvement in their self-reported
health.

On the other hand, the SP did not increase overall utilization but only redirected
elderly patients from pharmacists to physicians. More precisely, the probability of
pharmacy visits was reduced by almost the same amount as the increase in the probability
of only outpatient visits, which means that once elderly patients visited physicians, they
did not come back to buy drugs at pharmacies. A probable reason for this result is that the
government allowed clinic physicians to hire on-site pharmacists who dispensed drugs to
the elderly.

Moreover, the SP’s effect on self-reported health is trivial and not statistically
significant, which suggests that before the SP was initiated, patients used to substitute
pharmacist services for physician services only for minor medical conditions for which
the difference in care quality did not matter much.

The rest of this chapter is organized as follows. The second section gives a
detailed introduction to NHI and the SP and discusses their implications for the behavior
of physicians, pharmacists, and elderly patients. The third section lays out our empirical
framework. The fourth section summarizes the data. The fifth section reports estimation
results. We discuss our findings and conclude in the last section.

2.2. Two Health Care Reforms in the 1990s

On March 1, 1995, Taiwan initiated its National Health Insurance (N HI) program,

which provides compulsory insurance coverage to roughly 21 million people including 8

million previously uninsured. At the end of 1994, about 2.3 million people in Taiwan

37

(11% of the entire population) were aged 60 and older. Among them, about 75% was
insured by four social insurance programs: Farmers Health Insurance (FHI), Government
Employees Insurance (GEI), Labor Insurance (LI) and Veterans Insurance (VI).39 As
suggested by their names, FHI covered farmers; GEI covered government employees and
school teachers; LI covered private sector workers; VI covered veterans. These programs
all provided comprehensive medical benefits including inpatient, outpatient, dental, and
emergency service as well as prescription drugs. In 1994, F HI, GEI, LI and VI
respectively accounted for 45%, 20%, 18% and 17% of the insured elderly.40 Besides
social insurance, there was virtually no comprehensive private insurance at the time.“

In the pre-NHI period, if the uninsured elderly chose to visit physicians, they had
to pay entirely out of pocket for all physician services and prescription drugs. In contrast,
if they chose to visit pharmacists, they only needed to pay for drugs. The out-of-pocket
expenditure difference between visiting physicians and pharmacists made the uninsured
elderly more likely to choose the latter.

The implementation of NHI largely eliminated the out-of-pocket expenditure
difference between visiting physicians and pharmacists for the previously uninsured
elderly because NHI now paid for almost all physician services, including diagnosing,
prescribing, and dispensing drugs. In fact, NHI provides a range of benefits similar to that

of the former social insurance programs. See the previous chapter for more details about

NHI.

 

39 Several other social insurance programs also existed at the time but only accounted for a very small
portion of the elderly population.

40 All these numbers are my own calculations using administrative statistics from Bureau of Central Trust,
Bureau of Labor Insurance, Bureau of National Health Insurance, Veteran Affairs Commission and
Ministry of the Interior.

4' More precisely speaking, private insurance only provided very limited coverage and generally did not
cover any outpatient care (Liu and Chen 2002; Cheng 2003).

38

Unfortunately, NHI has been besieged by fast-growing budgetary woes since its
inception.42 Cheng (2003) points out that one of the biggest contributions to the budget
crisis are growing drug expenditures, and many believed that physicians’ ability to
continue to both prescribe and dispense drugs was one of the major causes of these
soaring costs. The government thus initiated the SP to split the two practices.43 The SP
was first initiated in 1997 in the two biggest cities, Taipei and Kaohsiung, and then
gradually expanded to include the remaining 21 counties and cities over 6 years.

The law that requires drug prescribing and dispensing to be separated is called
yao shi fa in Chinese, or the Pharmaceutical Affairs Act (PAA) in English, which was
amended in 1993. Article 50 of the PAA forbids pharmaceutical sellers to sell
prescription drugs to consumers without a physician’s prescription, with some
exceptions.44 The exceptions include the sale of prescription drugs among pharmaceutical
sellers and medical care and research institutions. In addition, drugs listed in two official
collections of prescriptions—zhong hua yiao dian (Chinese medicine book) and guo ming
chu fang xuanji (National collection of prescriptions)——can be sold to patients without a
physician’s prescription.

In general, pharmacists will be fined between NT$3,000 and NT$15,000
(US$1,000—US$5,000) if they are caught selling drugs to patients without a physician’s
prescription. On the other hand, to encourage community pharmacists to fill prescriptions,

NHI pays them NT$20 (US$07) for each prescription they fill (Chou et al. 2003).45

 

42 According to Cheng (2003), over the period 1995—2001, NHI revenues increased at an average rate of
4.26 percent, while expenditures increased at 6.26 percent.

43 South Korea launched a similar policy in 2000 (Kwon 2003; Park et al. 2005).

44 The law does not apply to over-the-counter (OTC) drugs.

45 This prescription filling fee was changed several times afierwards.

39

For physicians, Article 102 of the PAA clearly states that physicians are not
allowed to dispense drugs except in some special conditions.46 For example, physicians
can still dispense drugs to patients over age 65 and under age 3 who are unable to go to
pharmacies by themselves. Meanwhile, to encourage physicians to release prescriptions
to patients, NHI pays a NT$25 (US$08) “prescription release fee” for each prescription
that they release to be filled by community pharmacies (Lee, Huang, and Huang 2007).47

The implementation date of the SP in each county was determined based on some
criteria set up by the Department of Health (DOH). Two of the most important criteria are:
( 1) the ratio of community pharmacies to clinics contracted with NHI must be at least one
to three within any given county, and (2) the outpatient department of a DOH-run county
hospital has to release at least 3 percent of its prescriptions to patients. Table 2.1 lists the
exact implementation dates in all 23 counties or cities in Taiwan.

2.3. Empirical Framework

To evaluate the impacts of NHI and the SP, it is helpful to first envision the
triangular relationship of the physician, pharmacist, and patient. The analysis time
involves three periods: (1) before 1995, (2) 1995—1997, and (3) after 1997. These periods

are separated by the implementation of NHI and the SP, respectively.

 

46 Conditions include: (1) the patient is too young (under 3) or too old (over 65) to go to pharmacies by
himself; (2) the patient has some emergency condition and need drugs immediately; or (3) the patient is in
some officially announced remote area that has no licensed pharmacists. The latest definition of the third
type of exception is that if there is no licensed pharmacist within a radius of 1.8 kilometers (1.1 miles). If so,
the physician in the clinic is allowed to dispense drugs.

47 This “prescription release fee” was canceled in 2006. The NHI reimbursement for medications was
adjusted in 2002 so that the payment for a prescription filled by an on-site pharmacist is lower than by a
community pharmacist. In response to this adjustment, some clinic physicians set up their own pharmacies,
which may be registered under someone else’s name, just in front of or next door to their clinics and refer
patients there. This type of pharmacy is called a “gateway pharmacy” (Lee et al. 2007). To tackle this
problem, NHI simply canceled the “prescription release fee” in 2006 and now strictly requires clinic
physicians to release prescriptions.

40

 

Table 2.1. Timetable of the SP implementation

 

Date City/County
03.01.1997 Taipei City; Kaohsiung City
03.10.1998 Taichung City; Chiayi County; Chiayi City; Keelung City
04.20.1998 Changhua County; Miaoli County
06.06.1998 Kaohsiung County; Hsinchu County; Yunlin County
07.06.1998 Taichung County; Pingtung County
11.05.1998 Nantou County; Tainan County; Tainan City
12.09.1998 Taipei County
01.28.1999 Taoyuan County; Hsinchu City
06.21.1999 Yilan County
04.01.2001 Hualien County
05.01.2001 Taitung County
12.10.2002 Penghu County

 

Source: Department of Health,
http://drug.doh.gov.tw/admin/new_file_download.php?Pact=FileDownLoad&Pval=297.

In the first period, both the physician and the pharmacist prescribed and dispensed
drugs. In principle, the patient would have received better diagnosis and treatment from
the physician than the pharmacist. But the cost of visiting the physician was also higher
than visiting the pharmacist if one were uninsured and had to pay completely out of
pocket. Therefore, the uninsured patient faced a trade-off between price and care
quality.48

In the second period, everyone was insured by NHI. The physician and the
pharmacist still both prescribed and dispensed drugs. However, the out-of-pocket
expenditure difference between visiting the physician and the pharmacist was largely
eliminated for the previously uninsured, because NHI paid for most physician services
and prescription drugs. This provided an economic incentive for the previously uninsured
patient to visit the physician, who would provide better care than the pharmacist.

Moreover, the patient could still purchase drugs at the pharmacy without a prescription.

 

48 It is also possible that the uninsured patient could choose to see the physician first and then buy drugs at
the pharmacy. But, in practice, this was less likely to happen because he or she would have to pay the
additional cost of traveling between the clinic and the pharmacy.

41

But he or she would have to pay out of pocket for drugs without a prescription, because
NHI only paid for drugs with a prescription.

In the third period, the pharmacist was not allowed to dispense drugs to the
patient without a physician’s prescription. But the physician could still dispense drugs to
the elderly patient. Besides, many physicians hired their own on-site pharmacists. Keep
in mind that every patient was now covered by NHI. Because of the SP, the patient was
now forced to see the physician in order to get a prescription. Then, the key question is
where the patient chose to buy drugs. Most likely, the patient—whether elderly or not—
would want to buy drugs from an on-site pharmacist so that he or she would not have to
travel to a pharmacy outside the clinic. In other words, extraneous costs discouraged the
patient, elderly or not, from visiting an outside pharmacy once the patient had visited the
physician in the clinic.

Empirical Strategy

To assess the NHI effect, we adopt the same DD strategy used in the previous
chapter. We consider the previously uninsured elderly as the treatment group and the
elderly who were covered by the major social insurance programs as the control group.
The validity of the DD method relies on the following two assumptions. First, shifting
from social insurance to NHI had little effect on the control group. Second, the two
groups would have experienced a similar time trend in the absence of NH1. We estimated

the following DD model.
(1) yi, = ,60 + AU]; + ,BZYR, + ,63UI,- x YR, +controlsi +£,-,,
where i indexes observations; t indexes time; y is an outcome variable; UI is a dummy

variable indicating the previously uninsured elderly based on their insurance status in

42

1993; YR is a dummy variable indicating the latter year of a pooled two-year panel;
controls is a set of baseline control variables including sex, age, age-squared, education,
ethnicity, marriage, employment, and income evaluated at the baseline year of 1993. To
account for county level differences, we also include a full set of county dummies; s is
the error term. In equation (1), the two interaction terms measure the difference in
differences in outcomes.

We also use the same longitudinal survey of the elderly in Taiwan used in the
previous chapter. To estimate equation (1), we run three separate regressions with three
separate pooled two-year panels—1993 versus 1996, 1996 versus 1999 and 1999 versus
2003. In each pooled panel, we retain only those who appear in both years. We do so
because of a similar sample attrition issue as discussed in the previous chapter. The
purpose is to reduce a possible non-random attrition bias as well as to retain as many
observations in each pooled panel as possible.

To estimate the SP effect, we mainly exploit the gradual phasing-in of the SP. As
mentioned above, the SP was first initiated in 1997 in the two biggest cities and then
gradually applied to other counties and cities over six years. Because every wave of the
survey was conducted between March and June in each year, we thus use June 1999 as a
cutoff point.

By June 1999, 18 counties and cities on the west coast of Taiwan had
implemented the SP, while the remaining three counties on the east coast—Yilan County,
Hualien County, and Taitung County—had not (Table 2.1).49 Therefore, we use the

elderly people who lived on the west coast as the treatment group and those who lived on

 

49 None of the elderly in our sample lived in Hsinchu County and Penghu County.

43

the east coast as the control group. Then we adopt the DD method to estimate the SP
effect by assuming that the two groups would have experienced a similar time trend in
the absence of the SP.

Moreover, the three counties on the east coast also adopted the SP during 1999
and 2001 (Table 2.1). Therefore, afier 1999, we can further use the three counties as the
new treatment group and the 18 counties on the west coast that had previously adopted
the SP as the control group.

It is legitimate to question whether some of the elderly people from the west coast
traveled the east coast, or vice versa, to purchase drugs without a prescription. If many
elderly people did so, our identification strategy is problematic. But this chance is
minimal because, as shown in Figure 2.1, the east coast counties are geographically
separated from the west by the Central Mountain Range, which is 1,000—3,000 meters
above sea level on average. This geographic segregation imposes a very high transaction
cost for the elderly travel between the coasts.50

Specifically, to estimate the SP effect, we estimate the following DD model by
OLS.

(2) yi, = :30 + ,BISPI- + flzYR, + ,6'3SPl- x YR, + controlsi + a”,

where SP is a dummy variable indicating the elderly who lived on the west coast if it is
before June 1999 and the elderly who lived on the east coast if it is after June 1999; all
other notations are the same as in equation (1). In equation (2), again, the interaction term
measures the difference in differences in outcomes. To estimate equation (2), we also run

three separate regressions with the same three pooled panels used for equation (1).

 

50 For example, it would take 1.5 hours to travel from Taipei City to Yilan County by train, one of the most
common ways to commute between the two places.

44

 

 

 

 

 

Figure 2.1. Map of Taiwan
It is worth noting that in the case of 1993 versus 1996 and 1996 versus 1999, the

treatment group refers to those who lived on the west coast and the control group on the
east coast. On the contrary, in the case of 1999 versus 2003, the treatment group refers to

those who lived on the east coast and the control group on the west coast.

45

Under the DD assumptions mentioned above, we only aim to estimate the
treatment effects, i.e. the NHI effects and the SP effects, on the treated, i.e. the previously
uninsured elderly, the elderly who lived on the west coast before 1999 and the elderly
who lived on the east coast afier 1999. As mentioned in the previous chapter, estimating
the treatment effects on the entire population require a stronger assumption—the control
group would have experienced the same treatment effect if they were treated. However,
we do not have a fully randomized treatment and control group in both the NHI and the
SP case. Plus, we have observed that the two groups are different in many aspects in the
baseline. Therefore, it is hard to believe that such a strong assumption would hold.

Furthermore, the previously uninsured elderly were more likely to seek care at
pharmacies in the pre-NHI period because they paid less out-of-pocket than seeking care
at clinics. It thus would be interesting to further examine if the SP had different effects on
the previously uninsured elderly than their continuously insured counterparts. Therefore,
after estimating the SP effects, we re-estimate equation (2) by insurance groups to see if
the SP had different effects on the insurance groups.

Key outcome variables

In each wave of the elderly survey, respondents are asked to recall if they made

any pharmacy and outpatient visits in the month prior to the interview.5 ' Two dummy

variables are constructed accordingly to the answers, pharmacy and outpatient.

 

5 ' There are actually two questions asking about outpatient visits to practitioners of (1) Western medicine
and (2) Chinese medicine. I define outpatient visits as visits to either a Western or Chinese outpatient
service. In addition, outpatient visits can be visits to the outpatient department in a hospital or a clinic
because the survey does not make a distinction between the two. The question regarding pharmacy visits is
stated as follows: “In the past month, did you or your family member ever purchase drugs at yao fang for
your use?” Here, yao fang is “pharmacy” in Chinese. In particular, it generally refers to pharmacies outside
hospitals and clinics. One may worry about whether respondents may have interpreted yao fang to include
so-called gateway pharmacies which are outside clinics but actually owned by physicians. But gateway
pharmacies only started to emerge after 2000 (Lee et al. 2007), which is outside our analysis period.

46

To get a clearer picture of the substitution of physicians for pharmacists and vice
versa, we construct four mutually exclusive variables: outpatient only, pharmacy only,
both, and neither. Outpatient only is a dummy variable indicating that the interviewee
made only outpatient visits but no pharmacy visits in the past month; pharmacy only is a
dummy variable indicating only pharmacy visits but no outpatient visits in the past month;
both is a dummy variable indicating both outpatient and pharmacy visits in the past
month; neither is a dummy variable indicating neither outpatient nor pharmacy visits in

the past month. These four variables together exhaust all possible combinations.

Um

To measure health, we use a 5-point scale of self-reported health, with very good
at the highest level and very bad at the lowest. Because of relatively small proportions at
the two tails of the distribution, we combine very good and good into one group and bad
and very bad into another.

2.4. Data and Descriptive Statistics

This chapter also uses the Survey of Health and Living Status of the Middle-aged
and the Elderly in Taiwan. We mainly use the elderly sample in four waves of the survey:
1993, 1996, 1999 and 2003.52 We do not use the 1989 survey because its utilization
questions are different from other waves.53 Note that, in the analysis of NHI, the
treatment and control group are defined by the insurance status self-reported by the
elderly in the 1993 survey. In the analysis of the SP, the treatment and control groups are

defined by residence region.

 

52 We are not able to use the middle-aged sample for two reasons. First, it was drawn in 1996 and the pre-
NHI insurance status of the middle-aged was thus unknown. Second, the number of the middle-aged who
lived in the area where SP was implemented afier 2000 and appeared in both the 1996 and 1999 survey,
was too small (only 28) to constitute a reliable control group in analyzing the SP effect.

53 For example, the 1989 survey asks the elderly to recall their pharmacy visits in a year, while all other
waves ask the elderly to recall their visits in one month.

47

 

Table 2.2. Baseline Characteristics of the Elderly

 

By Insurance

Uninsured
(1)

Age in 1993 (mean) 72
Female (%) 57
Married with spouse present (%) 47
Ethnicity (%)

Minnan 74

Hakka 10

Mainlander 14

Aboriginal 1
Education (%)

No education 60

1-6 years 31

7 years and more 9
Region (%)

North 40

East 6

Middle 28

South 27
Employed/ self-employed (%) 1 1
Have enough money for living 75
expenses (%)
Total 784

Insured

(2)
71
39
67

56
I7

24
2

45
33
22

24
8
36
32
23
85

2,369

By Region
West East
(3) (4)
74 74
44 40
62 60
62 52
16 l l
22 21

0 16
49 44
32 39
1 9 l 7
30 0

0 100
36 0
33 0
l 3 9
83 79

2,909 242

 

Notes: categorizations based on insurance status and residence region in 1993.

Table 2.2 summarizes the baseline characteristics of the elderly by insurance and

residence region in 1993. In 1993, there were 784 uninsured and 2,369 insured elderly as

well as 2,090 living on the west coast and 242 on the east coast. Compared to their

insured counterparts, the previously uninsured were more likely to be women, unmarried

or widows, Minnan, less educated, living in the North, not working and poorer.54 On the

other hand, compared to their east coast counterparts, the elderly lived on the west coast

are more likely to be women, Mnnan, working and richer. It is worth noting that there

 

54 There are four major ethnic groups in Taiwan: the Minnan, Hakka, Mainlanders, and aboriginals. The

first three groups originated from China but migrated to Taiwan at different times. In particular,
Mainlanders generally refer to those who migrated from mainland China in 1949.

48

were a relatively high proportion of the aboriginal people living on the east coast. This is
because that the east coast is their traditional inhabitation area.

Table 2.3 summarizes their utilization and self-reported health. For probability of
having outpatient visits, we find that, from 1993 to 1996, the probability for the
uninsured elderly increased (about 26 percentage points) twice as large as the insured
elderly (about 13 percentage points). However, from 1996 t 2003, the two groups
experienced a similar increase of about 14 percentage points in probability. These
findings suggest that NHI, initiated in 1995, increased the probability of outpatient visits

for the previously uninsured elderly relative to their continuously insured counterparts.

 

Table 2.3. Summary of Outcome Variables

 

By Insurance By Region
(1) (2) (3) (4)
Uninsured Insured West East
Outpatient % n % n % n % n
1993 33.59 783 50.02 2,365 46.04 2,904 45.04 242
1996 59.79 582 62.66 1,939 61.68 2,333 66.13 186
1999 63.15 502 70.64 1,659 69.02 2,001 67.09 158
2003 73.15 324 76.88 1,146 76.49 1,374 69.15 94
Pharmacy
1993 28.83 784 21.41 2,368 23.18 2,908 23.97 242
1996 32.25 583 23.70 1,941 25.68 2,336 25.81 186
1999 16.17 501 14.79 1,656 14.47 1,997 23.42 158
2003 16.98 324 15.49 1,143 16.24 1,373 9.78 92
Very Good/
Good
1993 35.59 784 41.07 2,369 40.53 2,909 29.75 242
1996 26.76 583 29.11 1,941 28.72 2,336 26.34 186
1999 24.10 502 26.69 1,660 26.37 2,002 22.78 158
2003 26.59 361 26.80 1,261 27.25 1,512 20.37 108
Bad / Very Bad
1993 24.36 784 20.81 2,369 20.94 2,909 30.99 242
1996 29.50 583 31.12 1,941 30.31 2,336 36.02 186
1999 44.42 502 39.04 1,660 39.86 2,002 45.57 158
2003 45.98 361 42.51 1,261 42.92 1,512 47.22 108

 

Notes: categorizations based on insurance status and residence region in 1993.

 

49

In the long run, the two groups simply experienced a similar time trend, which
strengthens the validity of our DD strategy.

By region, we find that from 1993 to 1999, the elderly living on both west and
east coast experienced a similar increase in the probability of outpatient visits by about 23
percentage points. From 1999 to 2003, the west seemed to have a larger increase than the
east, reason of which is not clear.

As for pharmacy visits, we find that the previously uninsured elderly were always
more likely to visit pharmacies than their continuously insured counterparts. Both
experienced a dramatic drop in probability of pharmacy visits from 1996 to 1999, which
suggests the influence of the SP initiated in 1997. However, the drop for the previously
uninsured elderly was much larger than that for the continuously insured elderly. Besides,
their probabilities in 2003 remained similar to the 1999 level.

Comparing the west with the east, we find that both groups were very similar in
the pre-SP period (before 1997), which also strengthens the validity of using DD strategy.
From 1996 to 1999, we find that the west—already subject to SP—had a significant 11-
point drop in probability of pharmacy visits, while the east—not subject—only had a
small 2-point drop. From 1999 to 2003, on the contrary, the east—newly subject to SP—
had a large 13.6-point drop in probability, while the west—still subject SP—had a small
1.8-point increase. These preliminary findings all suggest the SP affected the elderly in
terms of their pharmacy visits.

Table 2.3 also summarizes the self-reported health of the elderly. Overall, the
probability of reporting very good or good health decreased over time; the probability of

reporting bad or very bad health increased over time. One interesting finding is that, from

50

1993 to 1996, while both previously uninsured and continuously insured elderly were
having higher probabilities of reporting bad or very bad health, the increase for the
previously uninsured elderly (5 percentage points) was only half of that for the
continuously insured elderly (10 percentage points.)

For our DD strategy, one important assumption is that the treatment group and the
control group share a common time trend. Comparing the time trends of both groups
before the treatment, i.e. NHI or SP, could provide us some information. Such
comparisons would take at least two points before the treatment. Unfortunately, for NHI
which was initiated in 1995, we have only one point in 1993. However, for SP which was
initiated in 1997, we have 1993 and 1996 and thus can look at the trend before 1997.

We calculate the trend in the probability of pharmacy visits from 1993 to 1996 for
the west and the east by taking the difference between the two points. Further, since the
uninsured elderly were affected by the 1995 NHI while the insured elderly were not, we
separate the comparison by insurance groups. Then, we use t test to test if the difference
across years is the same for the west and the east. Beside, in order to take difference, we
only use those who appear in both 1993 and 1996. In Table 2.4, we find that among the
uninsured elderly, the probability of pharmacy visits increased 3.4 percentage points for
the west, while there was no increase for the east. However, the t statistic shows that the
two differences for both groups are not statistically different. On the other hand, among
the insured elderly, the west increased by 2.7 percentage points and the east by 1.3 points.
Again, the test result shows that these two increases over time are not statistically
different from each other. These findings suggest the common time trend is reasonable at

least for the SP analysis.

51

 

Table 2.4. Comparison of Time Trends in Pharmacy Visits

 

Uninsured Insured
(1) (2) (3) (4)
West East West East
Pharmacy (%)

1993 28.9 30 20.8 23.7
1996 32.3 30 23.5 25
Difference 3 .4 0 2.7 1.3

n 551 30 1,777 156

t (p-value) -0.31 (0.75) -0.32 (0.75)

 

Notes: the null hypothesis is that the difference in the west equals to the difference in the east.

2.5. Estimation Results
NHI Effects on Outpatient and Pharmacy Visits

Estimated NHI effects are reported in Panel A in Table 2.5. Note that all results
are from separate regressions and only the interaction terms are reported. In the case of
1993 versus 1999, we find a significant 13.2-p0int increase in probability of outpatient
visits (column 1), which is resulted from a 6.5-point increase in probability of making
only outpatient visits (column 3) and a 6.6-point increase in probability of making both
outpatient and pharmacy visits (column 5). Meanwhile, there was almost no change in
probability of pharmacy visits (column 2). However, we do find a significant 5.9-point
drop in probability of making only pharmacy visits. These findings suggest that NHI
made the previously uninsured elderly more likely to have outpatient visits, including
those who previously only visited pharmacies. In addition, there was a 7.3-point decrease
in probability of making neither outpatient nor pharmacy visits, suggesting NHI raised

the overall utilization for the previously uninsured elderly.

52

 

63:38 8: 95% .828 :5 80:58: 2: .m 35m E moon 5%? 32 mo 88 2: .5“—
dn8w _8::8 2: 3:58 :58 3.8 2: :_ :3: oz? 3.2: :5 :38w 20:58: 2: 8: 3:58 :58 :33 2: :_ :o>__ 8:5 >520
2: .m 35: :_ «coo— mseo> :02 :5 0:2 .m> moi no 050 2.: .5“— .mo_:=::: .058 no :8 :3..— m :5 50» 0:205: 2: 5 95.53
068:: :5 £55838: 6835.: 565:: .5526: €853-05 .ow: £8 5 :85. £828 3:05: 0:205 82383..

=< .5888 8: £83 5:535 3:0 52389: 025:8 58.: 8: 3:5: 38:55: E 8: 8558 8 638-3820 ”882

 

as. 8 5.3 5.3 5. 8 84.3 :2: 53.8
83 2 3- 5.8 388 8:8- 5.8 88 a, 88
5.3 83-8 5.3 5.8 5. E 5. : $5.88
:88. $5- ”8.8 88 N88- 85 88 as 88
5.8 5.8 5.8 so. 3 8.8 5.3 83.3:
58 85 85 03.8 85 38.8 88 .s 88
8:... mm H: as:
8 5.8 8.2.8 88: c _.8 5.8 53".:
o 53 48.8 85- 3o: 48.8 88 .s 88
5.: 8.3 :s8 5. 8 85.8 5.3 55”.:
488 38.8 ms: 888 88. 8.5- 88 .s 88
83-8 5.8 5.3 5.8 5.8 5.8 63.85
«88. 85 $5- $5 85 ~88 88 a. 38
8:0 52 H< .25

.8532 89% A3: A3: 8:35:83 3:28:30

. 55:33 383839 .
8 8 A8 8 5 :8

 

8:58: .5 :35 :m 83 82 .2 28:.

53

For the case of 1996 versus 1999 and 1999 versus 2003, we find no statistically
results except for a significant 7-point drop in probability of pharmacy visits from 1996
to 1999. This finding may be related to the implementation of the SP.
SP Eflects 0n Outpatient and Pharmacy Visits

Estimated SP effects are reported in Panel B in Table 2.5. There was no
statistically significant findings in the case of 1993 versus 1996, which is not surprising
since the SP was absent in this period. However, we find a 92-point drop (when the SP
was applied to the west) and 18.9-point drop (when the SP was applied to the east) in
probability of pharmacy visits in the case of 1996 versus 1999 and 1999 versus 2003
respectively (column 2). Both drops result from decreases in probability of making only
pharmacy visits and probability of making both outpatient and pharmacy visits. These
findings suggest that the implementation of the SP made the elderly less likely to visit
pharmacies and more likely to make only outpatient visits. In addition, it is interesting to
note that the SP effect applied to the east was twice as large as that to the west. This
suggests that the cast was much more responsive to the SP than the west. However, it is
not clear why this occurred, given our limited information about the two sides.
Effects on Self-reported Health

The NHI effect on self-reported health is reported in Panel A of Table 2.6. In the
case of 1993 versus 1996, we find a significant 5-point drop in the probability of
reporting bad or very bad health (column 2) for the previously uninsured elderly relative
to their continuously insured counterparts. This suggests that although the health of both
groups was deteriorating, the NHI made the previously uninsured elderly deteriorate at a

slower rate than the continuously insured elderly, which was presumably resulted from

54

the increase in utilization.55 However, other than this finding, we do not find any other
significant result.

Estimated SP effects are reported in Panel B. We do not find any statistically
significant results, which is somewhat surprising because one would expect the elderly
should have received better care once they stopped visiting pharmacists and started to

visit physicians.

 

Table 2.6. NHI and SP Effects on Self-reported Health

(1) (2)
Very Good/ Good Bad/ Very Bad

Panel A: NHI effect

1993 vs. 1996 0.024 -0.051
(n=2,516) (0.92) (-1.99)
1996 vs. 1999 -0.007 0.069
(n=2,044) (-0.24) (2.23)
1999vs. 2003 -0.022 0.003
(n=1,466) (-0.61) (0.07)
Panel B: SP effect

1993 vs. 1996 -0.061 0.039
(n=2,514) (-1.45) (0.92)
1996 vs. 1999 —0.003 -0.028
(n=2,042) (-0.06) (-O.6)
1999 vs. 2003 -0.007 0.007
(n=1,464) (-0.14) (0.12)

 

Notes: See Table 2.5.

 

Differential SP Eflects between the Insurance Groups

Table 2.7 reports the SP effects on utilization by insurance group. Panel A is the
case of 1996 versus 1999 and Panel B is the case of 1999 versus 2003. Most of the results
are not statistically significant. However, there are some interesting findings that suggest

the SP seem to have differential effect on the previously uninsured and the continuously

 

55 Of course, NHI also affected other types of care. For example, the previous chapter finds that NHI
largely increased the probability of hospitalization for the previously uninsured elderly. Therefore, the
improvement in self-reported health could have also come from increases in other types of care.

55

insured elderly. In the case of 1996 versus 1999, we find that the SP effect on pharmacy
visits was dominated by the continuously insured (column 2), which is not surprising
because the continuously insured elderly account for 77% of the sample. However, the
SP did not seem to make the previously uninsured less likely to visit pharmacies. Instead,
the SP seemed to make them much more likely to only make outpatient visits (20-point in
column 3, although not statistically significant. Moreover, the SP also seemed to make
the previously uninsured elderly more likely to make both outpatient and pharmacy visits
in the case of 1996 versus 1999 (21-point in column 5). Consequently, the SP seemed to
largely reduce their probability of making neither outpatient nor pharmacy visits.

In the case of 1999 versus 2003, we find that the SP effect on pharmacy visits
seemed to be similar for the previously uninsured and continuously insured elderly

(column 2).

 

Table 2.7. SP Effect on Utilization by Insurance Group

 

(1 ) (2) (3) . (4) (5) (6)
Outpatient Pharmacy Outpatient Pharmacy Both Neither
only only
Panel A: SP effect—1996 vs. 1999
Uninsured 0.221 0.022 0.2 0.001 0.021 ~0.222
(n=471) (1.6) (0.21) (1.33) (0.01) (0.22) (-1.99)
Insured 0.03 -0.107 0.102 -0.035 -0.073 0.005
(n=l ,571) (0.5) (-1.95) (1.67) (-0.92) (-1.52) (0.09)
Panel B: SP effect—1999 vs. 2003
Uninsured 0.017 -0.193 0.1 1 1 -0.098 -0.095 0.081
(n=324) (0.1) (-1.94) (0.62) (-l .2) (-1.47) (0.54)
Insured -0.041 -0.188 0.077 -0.07 -0.1 18 0.11 1
in=1,140) (-0.64) (-2.99) (1.1) (-1.75) (-2.06) (1.64)

 

Notes: results are from separate regressions among the previously uninsured elderly and continuously
insured elderly. See Table 2.5 for other notes.

56

 

Table 2.8. SP Effect on Self-reported Health by Insurance Group

 

(1) (2)
Very Good/ Good Bad/ Very Bad

Panel A: SP effect—1996 vs. 1999

Uninsured 0.177 -0.013
(n=471) (1.39) (-0.09)
Insured -0.036 -0.038
(n=1,571) (-0.77) (-0.78)
Panel B: SP effect—1999 vs. 2003

Uninsured 0.046 0.098
(n=324) (0.36) (0.74)
Insured -0.02 -0.014
(n=l,l40) (-0.35) (-0.23)

 

Notes: See Table 2.7.

SP effects on self-reported health by insurance group are reported in Table 2.8.
Again, we do not find any statistically SP effects. However, the SP seemed to have a
positive effect on the previously uninsured elderly in the probability of reporting very
good or good health (column 1), especially for the case of 1996 versus 1999 (17.7
percentage points, although not statistically significant.)
2.6. Conclusion

This paper studies the effects of NHI and the SP on elderly people’s choice
between physicians and pharmacists and reaches the following conclusions. First, NHI
induced more previously uninsured elderly patients—including those who previously
only visited pharmacists—to visit physicians by providing them an economic incentive.
Meanwhile, NHI also raised their overall utilization and there was some evidence
suggesting the increase in utilization is associated with an improvement in self-reported
health.

Second, the SP shifted the elderly patients from pharmacists to physicians. In

particular, the SP made them more likely to make only outpatient visits, which suggests

57

that once the elderly patients had visited physicians, they did not come back to
pharmacies to buy drugs. Instead, they bought drugs from on-site pharmacists who were
hired by the clinic physicians.

Third, there is little evidence on the SP effect on self-reported health. This is
somewhat surprising because one would expect that elderly patients would have received
better care by visiting physicians instead of visiting pharmacists. A possible explanation
for this finding is that in the past elderly patients only substituted pharmacist services for
physician services for minor medical conditions for which quality did not matter much.

Fourth, there was some evidence suggesting that the SP had differential effects
between the previously uninsured and continuously insured elderly. Specifically, while
the SP made both insurance groups more likely to only make outpatient visits, the effect
was much larger for the previously uninsured elderly than their continuously insured
counterparts. In addition, in the case of the 1996 versus 1999, the SP seemed to largely
raise the probability of reporting very good or good health for the previously uninsured
elderly relative who lived on the west coast, although the effect was not statistically
significant.

Last, it is worth noting that the two policies are actually intertwined. In particular,
the SP would have had a much different impact were it not preceded by NHI. For
example, the out-of-pocket expenditure difference would still have been substantial for
the uninsured in the absence of NHI. It is thus possible that implementing the SP without
NHI could have led some poor uninsured patients to simply forego the care they needed
because they could not afford to visit a physician to get a prescription, which is needed in

order to buy drugs from a pharmacist.

58

Appendix 1
INSURANCE STATUS IN 1993 AND SURVEY QUESTIONS
Insurance status in I 993

In the 1993 survey, the elderly respondents are first asked if they are insured. If
yes, they are further asked what kind of insurance they have. In the second question, there
are 15 insurance programs of which 14 are social insurance programs and 1 is private
insurance. In the raw data, 787 report uninsured, 2,367 insured and 1 unknown. However,
among the 787 uninsured, 4 report being insured by some insurance program in the
following question. To check these conflicts, I use another question asking them who
paid for most of their medical expenses. If they answer health insurance paid for most of
their medical expenses, they are recoded as insured. In the end, 2 are recoded and thus
785 are uninsured in 1993.

For the insured, I drop the 16 privately insured because private insurance is not
comprehensive and they are likely to be affected by NHI. On the other hand, they are not
completely uninsured and should not be categorized as uninsured.

In addition, I drop one case that actually died before 1993 but surprisingly
appeared in the 1993 survey. I also drop another 2 insured cases without knowing their
pre-NHI insurance program. Finally, there are 2,351 insured elderly who are covered by
various social insurance programs.

Survey Questions about Utilization
In each wave of survey, the elderly respondents were asked a series of utilization

questions in Chinese. Below are the original questions in Chinese and their English

59

translations. We do not use the 1989 survey because its questions are generally different
from other waves of the survey.
Hospitalization
In the 1993, 1996, 1999 and 2003 survey, the elderly respondents were asked the
following question regarding hospitalization.
fi-ZEHEE ’ {mfiififififiifi? (In the last year, have you ever been
hospitalized?)
In the 1989 survey, however, the question was stated as:
fi-ifififlibfik ’ {711;sz Efi’t’éf’. If LEE? (@iéfirfi t. a/ lama... ifimig
If???) (In the past half year, have you ever been hospitalized? (Including visits to
an emergency room, stays in a hospital for a day or two for observation. ))
ER
In 1993, 1996, 1999 and 2003 survey, the elderly respondents were asked the
following question regarding emergency room visits.
EEK—‘5; ’ @1355??? JEEE Ufifi All}??? (In the last year, have you ever visited
the emergency room in a hospital?)
There was no separate question about emergency room visits in the 1989 survey.
Western Outpatient Visits
In 1993, 1996, 1999 and 2003 survey, the elderly respondents were asked the

following question regarding Western outpatient visits.
TEE Zr‘flfl H E Ilfli’sififiifi Eggfif’éiéfflf‘t) (In the last month. have

you ever had western medicine visits?(Excluding hospitalizations. ))

This question refers to western medicine outpatient visits.

60

In the 1989 survey, there was no western outpatient visits question. Instead, there
were questions stated as:

EE'Z“$$ ’ WEE ‘Fiéfi’i‘ifi/Aflgélfifififi? (In the last year, have you

ever visited a public hospital or clinic?)

Efif—“EE ’ I315; [33 éfifififbflfilfifififi? (In the last year, have you

ever visited a private hospital or clinic?)

From these two questions, we do not know if the patients were hospitalized or just
visited the outpatient department in a hospital or clinic.
F oregoing Doctor Visits

In 1993, 1996, 1999 and 2003 survey, the elderly respondents were asked the
following questions.

E‘ZEEEIIH E ' Iffififfifiélfi‘fflffifili , t§£fi%$ . IEIIUiQ-E? (In the Past

three months, have you ever felt ill and would like to see a doctor but eventually

did not do it?)

If one answered yes, he or she was further asked to choose from a list of reasons.
One of the listed reasons was because ifififi’é (lack of money). Respondents were
allowed to have multiple choices.

We thus use these two questions to construct a dummy variable indicating one did
not visit a doctor because lack of money. There were no such questions in the 1989

survey.

61

Appendix 2
Differences between Chapter 1 and Chen et al. (2007)
Approach on Estimating NHI Eflect on Mortality

As we briefly discussed before, Chen et al. use a linear DD model to estimate the
NHI effect on one-year mortality. More precisely, they compare the one-year mortality
rates in 1996 and 1999 to the pre-NHI rate which is an average of the 1989 and 1993
rates. In this sense, they only look at 4 annual data points and ignore information outside
those years. In their Table III, their DD estimates show 1.16 and 0.42 percentage points
of NHI effect on one-year mortality in 1996 and 1999 respectively, although the effects
are not statistically significant. We replicate their one-year mortality rates in 1996 and
1999 and find similar rates in 1996 but not in 1999: their 1999 rate is 1.87% for the
continuously insured elderly and 3.08% for the previously uninsured elderly, which are
much lower than what we find. We find them to be at least 5.95% and 4.69%.56 In
addition, their rates decrease from 1996 to 1999, while we find them to increase over
time. See Table A2.] for comparison of one year mortality rates.

In their paper, they argue that their analysis time (1989-1999) may not be long
enough to detect any NHI effect on mortality. In contrast to their approach, we
incorporate the DD design into a hazard model that allows us to utilize all mortality data
without having to give up any piece of information we have. This approach makes a
major difference even without extending the analysis beyond 1999. We can verify this by
limiting the analysis time to 1993-1999 and re-estimating equation (4) in chapter 1. The

results are reported in Table A2.2. From 1993 to 1999, there are 901 and 875 deaths

 

56 It is not clear which data they use. Besides, we are not able to replicate their pre-NHI rates which should
be averages of 1989 and 1993 rates, unless we know their imputation of insurance for those who appeared
only in the 1989 survey.

62

 

Table A2.]. Comparison of One Year Mortality Rates (%)
Previously Uninsured Elderly Continuously Insured Elderly

1996 1999 1996 1999

Chen et al. (2007) 4.85 3.08 4.38 1.87
Survey 4.75 4.69 4.63 5.95
DOH 4.71 4.98 4.36 6.47

 

Notes: yearly mortality rates in the first row are adopted from Table 111 in Chen et al. (2007); in their
table, the previously uninsured were called “No Pre-NHI Insurance Group” and the continuously
insured “Pre-NHI insurance group.”

 

Table A2.2. NHI Efl'ect on Hazard Ratio
Analysis Time: 1993-1999

Data Survey DOH
(1) (2) (3) (4) (5) (6)
UI 0.449 0.304 0.241 0.447 0.315 0.257
(3.48) (2.25) (1.77) (3.37) (2.27) (1.84)
POST 0.26 0.346 0.451 0.288 0.363 0.451
(2.94) (3.91) (5.07) (3.19) (4.01) (4.96)
UIXPOST -0.33 -0.326 -0.272 -0.315 -0.313 -0.269
(-2.09) (-2.07) (-1.73) (-1.96) (-1.95) (-1.67)
Pre-NHI controls:
Demographic & No Yes Yes No Yes Yes
Economic
Health No No Yes No No Yes
Subjects 3,136 3,136 3,136 3,136 3,136 3,136
Deaths 901 901 901 875 875 875
Hazard Ratio:
Pre-NHI 1.57 1.35 1.27 1.56 1.37 1.29
Post-NHI 1.13 0.98 0.97 1.14 1 0.99

 

Notes: I statistics are in parentheses; POST is an indicator for period afier March 1995 to end of 1999;
hazard ratio is calculated as exp(U1) in the pre-NHI period and exp(U1+U1><POST) in the post-NHI period.
according to the survey’s collection and the DOH data respectively. Even with this much
shorter time period, we still find that the pre-NHI hazard ratio is about 1.3 and the post-
NHI hazard ratio is about 1. And the pre-NHI ratio is statistically different than I, while
the post-NHI ratio is not.

Defining the Insured and Uninsured in the Pre-NHI Period

63

Chen et al. try to use insurance status of the elderly in 1989 and assume they
remained unchanged throughout the pre-NHI period.57 However, since there is no
insurance information in the 1989 survey, they have to impute insurance status of the
elderly using the following two rules: 1) for those who appeared in the 1993 survey, they
assume their insurance status in 1989 is identical to 1993 status; 2) for those who
appeared in the 1989 survey but did not appear in the 1993 survey, they assume those
who were government employees, farmers and soldiers in 1989 were insured in 1989. In
their final sample, they have 2,990 insured (called pre-NHI insurance group) and 909
uninsured (called no pre-NHI insurance group) in 1989.

However, their imputed insurance status is problematic for two main reasons.
First, they assume the 1989 and 1993 insurance statuses are identical which is unlikely to
be true. As we discussed in section 2, there was a series of expansions of GEI and FHI in
the late 19803 and early 19903. This implies that some elderly who were insured in 1993
could have been uninsured in 1989. We verify this by comparing the 1989 insurance
information contained in the 1991 survey with the 1993 insurance status.58 Of the 2,887
elderly whose insurance statuses in both 1989 and 1993 were available, 672 changed
from uninsured to insured and 75 changed from insured to uninsured. In other words,
about 26% of them changed their insurance status. This implies that the 1989 and 1993
insurance statuses of the elderly are not identical.

Second, they assume that government employees, farmers and soldiers were

insured in 1989. This assumption obviously ignores the private sector workers who might

 

57 In fact, they only use 3,899 of the initial 4,049 elderly. They are not able to assign pre-NHI insurance
status to 150 elderly due to incomplete information about employment status.

58 The 1991 survey is a telephone follow-up focusing on only health and life insurance. It has insurance
history of the elderly including when they gained and stopped insurance.

64

be covered by LI. In addition, only 20% of those who appeared in the 1989 survey but
did not appear in the 1993 survey reported that they were working in 1989. It is thus not
clear how they imputed insurance status for the other 80%.

Our sample consists of only those who appeared in the 1993 survey and is thus
smaller. However, we have a more accurate demarcation between the insured and the
uninsured in the pre-NHI period by using the 1993 insurance status.59 In the context of
estimating NHI effects on the previously uninsured elderly, an accurate demarcation
between the treatment and control group is essential and probably more important than
having a larger sample of the elderly population.

Why Our Estimates of NHI Eflects on Utilization Are Similar?

Although they use the initial 1989 sample and we use the 1993 sample, we
happen to have very similar estimates of NHI effects on utilization. This happens simply
because comparable information about utilization is not available in the 1989 survey and
both of us use utilization information starting in 1993. As a result, those who appeared in
the 1989 survey but did not appear in the 1993 survey do not enter their utilization
analysis. In addition, since we both adopt DD method, our results on utilization are thus
similar.

However, we still find one thing puzzling in their Table 111. As indicated in note 3
in their Table 111, they claim that “only the 1993, 1996 and 1999 surveys offered
information on outpatient and inpatient care utilization.” Therefore, the columns under
“before NHI” in their Table 111 should contain only 1993 utilization information and the

sample in that year. However, in the first two rows, probability of using outpatient and

 

59 It is possible that the elderly may still change their insurance status between 1993 and 1995. However,
the chance is much lower.

65

inpatient care, the sample size (indicated in parenthesis) of the ‘No pre-NHI insurance
group’ (correspond to our previously uninsured) is 598 and the sample size of the ‘Pre-
NHI insurance group’ (correspond to our continuously insured) is 2,557. These two
numbers are not consistent with the data. As mentioned in Appendix 1, the original data

show that there are 787 uninsured, 2,367 insured and 1 with unknown status in 1993.

66

Appendix 3
Availability of Data and Full Estimation Results

For brevity reasons, only key estimation results are reported in the previous
chapters. For readers who are interested in learning more, the full estimation results are
available upon request. Please contact me at changkanghung@gmail.com.

The availability of the elderly data requires permission from the Department of
Health Promotion in Taiwan. Therefore, I do not have the right to distribute the data to
readers who are interested in replicating the results or in doing further analysis. To

acquire the data, please visit their website at http://www.bhp.doh.gov.tw/NHPnet/Portal/.

67

BIBLIOGRAPHY

Card, D., C. Dobkin and N. Maestas (2004). “The Impact of Nearly Universal Insurance
Coverage on Health Care Utilization and Health: Evidence from Medicare.” Working
paper #10365, National Bureau of Economic Research.

Chen, L., W. Yip, M. C. Chang, H. S. Lin, S. D. Lee, Y. L. Chiu and Y. H. Lin (2007).
“The Effects of Taiwan’s National Health Insurance on Access and Health Status of
the Elderly.” Health Economics, 16: 223-242.

Cheng, T. M. (2003). “Taiwan’s New National Health Insurance Program: Genesis and
Experience So Far.” Health Aflairs, 22: 61-76.

Cheng, S. H., and T. L. Chiang (1997). “The Effect of Universal Health Insurance on
Health Care Utilization in Taiwan: Results from a Natural Experiment.” Journal of the
American Medicine Association, 278: 89-93.

--------- (1998). “Disparity of Medical Care utilization among Different health Insurance
Schemes in Taiwan.” Social Science and Medicine, 47: 613-620.

Chiang, T. L. (1997). “Taiwan’s 1995 Health Care Reform.” Health Policy, 39: 225-239.

Chou, Y. J ., Yip, W. C., Lee, C. H., Huang, N., Sun, Y. P., Chang, H. J. (2003). “Impact
of separating drug prescribing and dispensing on provider behavior: Taiwan’s
experience.” Health Policy and Planning, 18(3): 316—329.

Currie, J ., and J. Gruber (1996a). “Health Insurance Eligibility, Utilization of Medical
Care and Child Health.” Quarterly Journal of Economics CXI: 432-466.

--------- (1996b). “Saving Babies: The Efficacy and Cost of Recent Changes in the
Medicaid Eligibility of Pregnant Women.” Journal of Political Economy, 104: 1263-
1296.

--------- (1997). “The Technology of Birth: Health Insurance, Medical Interventions and
infant Health.” Working paper #5985, National Bureau of Economic Research.

Haas, J. S., S. Udvarhelyi, and A. M. Epstein (1993a). “ The Effect of Providing Health
Coverage to Poor Uninsured Pregnant Women in Massachusetts.” Journal of the
American Medical Association, 269: 87-91.

--------- (1993b). “The Effect of Health Coverage for Uninsured Pregnant Women on

Maternal Health and the Use of Cesarean Section.” Journal of the American Medical
Association, 270: 61-64.

68

Hanratty, M. (1996). “Canadian National Health Insurance and Infant Health.” The
American Economic Review, 86: 276-284.

Hsieh, H. Y. (1999). “Institutionalization of medicine production: Interaction of
institution, organizational ecology and consumption patterns.” Doctoral dissertation,
Department of Sociology, National Taiwan University (in Chinese).

Hsieh, H. Y. (2003). “The impact of chain pharmacy on pharmacists’ professional status.”
The National Chengchi University Journal of Sociology, 35: 109—132 (in Chinese).

Hsueh, Y. S., S. Y. Lee and Y. T. Huang (2004). “Effects of Global Budgeting on the
Distribution of Dentists and use of Dental Care in Taiwan.” Health Services Research,
39: 2135-2153.

Huang, K. H., Lee, Y. C., Huang, Y. T. (2004). “Impact of the separation of drug
prescribing and dispensing policy on the release rate of prescriptions from clinics.”
Taiwan Journal of Public Health, 23(6): 469-478 (in Chinese).

Kwon, S. (2003). “Pharmaceutical reform and physician strikes in Korea: Separation of
drug prescribing and dispensing.” Social Science & Medicine, 57: 529—53 8.

Lee, M. C., and A. M. Jones (2004). “How Did Dentists Respond to the Introduction of
Global Budgets in Taiwan? An Evaluation Using Individual Panel Data.”
International Journal of Health Care Finance and Economics, 4: 307-326.

Lee, Y. C., Huang, K. H., Huang, Y. T. (2007). “Adverse pharmaceutical payment
incentives and providers’ behavior: The emergence of GP-owned gateway pharmacies
in Taiwan.” Health Policy and Planning, 22: 427—435.

Levy, H., and D. Meltzer (2004). “What Do We Really Know about Whether Health
Insurances Affects Health?” in Health Policy and the Uninsured: Setting the Agenda,
Catherine McLaughlin (ed.), Urban Institute Press: Washington DC.

Lichtenberg, F. R. (2002). “The Effects of Medicare on Health Care Utilization and
Outcomes.” Frontiers in Health Policy Research, 5: Article 3.

Liu, T. C., and C. S. Chen (2002). “An Analysis of Private Health Insurance Purchasing
Decisions with National Health Insurance in Taiwan.” Social Science and Medicine,
55: 755-74.

Manning, W. G., J. P. Newhouse, N. Duan, E. B. Keeler, A. Leibowitz and M. S.

Marquis (1987). “Health Insurance and Demand for Health Care: Evidence from a
Randomized Experiment.” The American Economic Review, 77(3): 251-277.

69

Park, S., Soumerai, S. B., Adams, A. S., Finkelstein, J. A., Jang, S., Ross-Degnan, D.
(2005). “Antibiotic use following a Korean national policy to prohibit medication
dispensing by physicians.” Health Policy and Planning, 20(5): 302—309.

Peabody, J. W., J. C. Yu, Y. R. Wang, and S. R. Bickel (1995). “Health System Reform
in the Republic of China: Formulating Policy in a Market-Based Health System.”
Journal of the American Medical Association, 273: 777-781.

Polsky, D., J. A. Doshi, J. Escarce, W. Manning, S. M. Paddock, L. Cen and J. Rogowski
(2006). “The Health Effects of Medicare for the Near-Elderly Uninsured.” Working
paper #12511, National Bureau of Economic Research.

Unschuld, P. U. (1976). “The social organization and ecology of medical practice in
Taiwan.” In C. Leslie (Ed.), Asian medical systems: A comparative study. Berkeley
and Los Angeles: University of California Press: 300—3 16.

Wen, C. P., S. P. Tsai and W. S. Chung (2008). “A 10-Year Experience with Universal
Health Insurance in Taiwan: Measuring Changes in Health and Health Disparity.”
Annals of Internal Medicine, 148: 258-267.

Zimmer, Z., L. G. Martin and M. C. Chang (2002). “Changes in Functional Limitation
and Survival among Older Taiwanese, 1993, 1996, and 1999.” Population Studies, 56:
265-276.

Zimmer, Z., L. G. Martin and H. S. Lin (2005). “Determinants of Old-age Mortality in
Taiwan.” Social Science and Medicine, 60: 457-70.

70

II.“
"
.lllll
"II
H
“I
”ll
“
l‘ll
“
Ill
A“
Full]
HI

3 1293 03062 9376