lable at ScienceDirect

Social Science & Medicine 111 (2014) 41e49

Contents lists avai

Social Science & Medicine

journal homepage: www.elsevier .com/locate/socscimed

The spillover effect of National Health Insurance on householdconsumption patterns: Evidence from a natural experiment in Taiwan

Ji-Tian Sheu, Jui-fen Rachel Lu*

Department of Health Care Management, College of Management, Chang Gung University, Taiwan

a r t i c l e i n f o

Article history:Received 21 September 2013Received in revised form5 April 2014Accepted 7 April 2014Available online 8 April 2014

Keywords:TaiwanNational Health InsuranceHousehold consumption patternInsurance effectDifferences-in-differences estimation

* Corresponding author. Department of Health CarInstitutie of Business and Management, College ofUniversity, 259, Wen-Hwa 1st Road, Kwei-Shan, Tao-Y

E-mail addresses: jtsheu@mail.cgu.edu.tw (J.-T. Shrachel@mail.cgu.edu.tw (J.-f.R. Lu).

http://dx.doi.org/10.1016/j.socscimed.2014.04.0060277-9536/� 2014 Elsevier Ltd. All rights reserved.

a b s t r a c t

While providing financial risk protection was one of the major aims of introducing a National HealthInsurance (NHI) program in Taiwan in 1995, one may also wonder how the households may exploitresources freed up and available to them as a result of reduced exposure to risk due to health insurance.This paper aims at studying and evaluating the impact of social insurance on these differing householdconsumption patterns. A differences-in-differences estimation model was applied to a sample of 17,899households from the 1993e2000 Taiwan Survey of Family Income and Expenditure to assess the NHI’simpact. This effect was evaluated by the changes in the proportion of the consumption expendituredevoted to medical items and non-medical items in the post-NHI period (1996e2000) compared to thepre-NHI period (1993e1994). Our study found that spending related to the improvement of housingconditions (rental and water bills) had the most significant increase, 1.87% (in the share). Furthermore,examining the NHI impact across socioeconomic status (SES) strata (in terms of income and educationlevels), our study found that households with the lowest SES experienced the largest increase inspending share (2.16%) for rental and water bills, and the least drop (0.64%) on education items.Recognizing how households can exploit the potential benefits associated with NHI provision couldenable the government to devise specific policy tools to facilitate better targeting of investment decisionswith limited resources available for less well-off households.

� 2014 Elsevier Ltd. All rights reserved.

1. Introduction

The value of insurance lies in its inherent purpose to offsetfinancial risk in the face of adverse outcomes. This was one ofTaiwan’s major policy goals in the implementation of the NationalHealth Insurance program. With health insurance reducing theirexposure to risk, households may differ in how they reallocate thenewly available funds, and this paper focuses on the effects of NHIon these household spending patterns.

The potential impact of insurance programs is well evidenced inthe relevant literature. One of the most direct and significant im-pacts of NHI is the enhanced access to health care regardless ofsocioeconomic class and age group (Cheng and Chiang, 1997; Chenet al., 2007). Evidence from around the globe has shown that theavailability of health insurance coverage ameliorates the inequality

e Management and GraduateManagement, Chang Gunguan 333, Taiwan.eu), coolmom01@gmail.com,

of health care services (Broyles et al., 1983; Decker and Remler,2005; Park, 2012). In addition to offering financial risk protection,the extension of insurance coverage also has an impact on the labormarket (Boyle and Lahey, 2010; Chou and Staiger, 2001). Wheninvestigating the impact of a major expansion in health insurancecoverage, in terms of benefit scope as well as population coveragein the U.S. Department of Veterans Affairs health care system in themid-1990s, Boyle and Lahey (2010) found that the availability ofhealth insurance affected employment and retirement decisionssuch as early retirement. Chou and Staiger (2001) found that theavailability of health insurance coverage reduced women’s laborparticipation rate in Taiwan.

Furthermore, when the exposure to risk is constrained by theavailability of health insurance benefits, one can expect to observechanges in how households allocate the resources freed up as aresult. For one, the incentive to save for future uncertainty due tohealth risk is likely to be weakened. Research has found thatcomprehensive health insurance which reduces uncertainty aboutfuture medical expenses will result in a reduction in households’precautious saving (Chou et al., 2003; Chou et al., 2004; Gruber andYelowitz, 1999; Kuan and Chen, 2013). Though most of the

J.-T. Sheu, J.-f.R. Lu / Social Science & Medicine 111 (2014) 41e4942

literature on this topic does find the crowd-out effect on saving,there is no consensus on who experiences more significantlynegative impacts. Chou et al. (2003) found that the crowd-out effectwas greater for households with smallest savings and higher in-come. Furthermore, Chou et al. (2004) found that younger house-holds tended to save less when they were covered by healthinsurance. Although using the same data source as Chou et al.(2004), Kuan and Chen (2013) found that, conversely, higher-saving households tended to have a greater reduction in savingsafter the introduction of NHI, and the crowd-out effect was strongerfor households with higher saving, lower income, and retiringhousehold heads. The inconsistency in empirical results presentedby Chou et al. (2004) and Kuan and Chen (2013) is probably due tosampling issues. Chou et al. (2004) excluded households withnegative saving which made up about one-sixth of the entiresample. Though these two studies disagreed on types of house-holds experiencing a stronger NHI crowd-out effect, both foundNHI had a negative impact on saving. If the households tended tosave less than they would otherwise, what would they spend the“additional”money on? Little research seems to have examined theimpact of insurance on household spending patterns.

Social health insurance (SHI), which is often subsidized byadditional government funds and with some emphasis on incomeredistribution, is likely to produce a further impact on a house-hold’s consumption distribution beyond income protection. Wetermed this effect of social insurance examined here the “spillovereffect” which goes beyond financial protection from unexpectedmedical consumption. This paper hence sets forth to study andevaluate the spillover effect of social insurance on household con-sumption patterns. As economists often regard expenditure oneducation and medical care as important investments in humancapital, we were also particularly interested in examining the effectof social insurance on education expenditure by different socio-economic strata.

In 2005, the World Health Assembly’s policy resolution for theWorld Health Organization (WHO) recommended that low- andmiddle-income countries adopt social health insurance as thehealth care financing strategy. SHI has indeed become a popularmode of financing health care in achieving the goal of universalcoverage in recent years. The major findings of our paper canprovide important policy references for countries considering thesocial insurance approach, to facilitate better targeting of thelimited resources available for less well-off households by under-standing how households may change their consumption behaviordue to the availability of insurance coverage. As 57% of Taiwan’spopulation was covered by some form of social insurance program,with varying scope of coverage, before National Health Insurancewas introduced, it offers a unique laboratory for studying the effectof social health insurance in a natural experiment set-up.

In the following section, we provide a brief overview of Taiwan’sNHI program. In the methodology section, a detailed account ofdata sources and estimation models is presented. Empirical evi-dence is introduced in the results section, followed by results froma robustness check, and we conclude the paper with a discussion ofour findings and research constraints.

2. National Health Insurance program in Taiwan

Aiming to ensure every citizen’s equal access to reasonablehealth care, the Taiwan government implemented its NHI programin 1995 after nearly five years of planning efforts (1988e1993) anda two-year legislative marathon (Lu and Chiang, 2011). Taiwanadopted a single-payer approach in administering the NHI pro-gram, and the National Health Insurance Administration (NHIA,formerly Bureau of National Health Insurance) is the government

agency that oversees the operation of the scheme. The NHI wascreated by merging three then-existing social insurance programs,namely Labor Insurance (LI), Government Employee Insurance(GEI) and Farmer Insurance (FI), which collectively covered 57% ofthe populationdand expanding the insurance coverage to theuninsured, mainly composed of the unemployed, includinghousewives, children and the elderly (Lu and Hsiao, 2003).

Labor Insurance, introduced in 1950, was the first social insur-ance program to provide health and other cash benefits (mainlyretirement benefits) to industrial workers. LI coverage wasextended to almost all labor workers in formal sectors when the1970 Labor Insurance Statute required employers of five or moreemployees to provide LI coverage to their employees between theages of 15 and 60 years. In 1988, LI coverage was extended togovernment employees who were not eligible for GEI, the self-employed, and members of occupational unions who did nothave regular employers. Despite the expansion in coverage over theyears, unlike GEI, LI did not provide family coverage for dependentsof the insured (Chou et al., 2003).

In contrast, GEIdinitially introduced with both health and cashbenefits (mainly pension benefits) in 1958dextended coverage togovernment employees’ spouses (in 1982), parents (in 1989) andchildren (in 1992). Optional coverage was made available to retiredgovernment employees in 1965. In addition, teachers at privateschools were also covered by GEI from 1980 (dependents in 1990).In 1985, the government launched a pilot insurance program tocover agriculture workers, which was fully implemented as FarmerInsurance in 1989. Local council members and town/village heads(0.23% of population) were covered by a government-subsidizedinsurance scheme introduced in 1989. To provide financial pro-tection to vulnerable members of the population, a whollygovernment-funded Welfare Insurance (WI) for low-incomehouseholds was introduced in 1990. By 1994, LI was the largestsocial insurance program in Taiwan, accounting for 69.8% of theinsured population, with GEI, FI, and WI making up 14.73%, 14.28%,and 0.96%, respectively (Lu and Hsieh, 2000).

Despite similarities in the benefits providedby LI, GEI and FI, theirpremium contribution rates and share of premium payments varied.The GEI premiumwas assessed on monthly basic salary with a pre-mium rate of 9% (3.8% for the dependent scheme and 8% for theretiree scheme), of which 65% of the premiumwas paid by the gov-ernment (as an employer) and the rest by the insured. For LI, thecontribution rate was 7% of the monthly insured wage, with em-ployers shouldering 80% of the assessed premium and the insuredpaying 20%. The government subsidized 70% of the FI premium,which was assessed at a premium rate of 6.8% on the weightedaverage of LI-insuredwage in the previous year (Lu andHsieh, 2000).

When the NHI was implemented, the government decided tofinance the scheme with a tax levy on payrolls (premium contri-bution rate set at 4.25% when initially launched, and increased to4.55% in 2005, and 5.17% in 2010, and currently at 4.91% on payrolland 2% on supplementary income), with the employereemployeesplit varying according to occupation category (Lu, 2013). Theschemewas supplemented with government subsidies for the poor(approximately 1% of the insured population were exempted frompremium contributions), veterans, and farmers. The diversity ininsurance-scheme features facilitates effective assessment of theinsurance effects on household consumption patterns.

Compared with pre-NHI insurance schemes, NHI provides aneven more comprehensive benefit package that covers preventiveandmedical services (inpatient and outpatient), prescription drugs,dental services, and home nurse visits as well as hospice care, withmodest co-payment at point of service (Lu and Hsiao, 2003). As of2011, Taiwan devoted 6.62% of its GDP to health, compared to 4.87%in 1995.

J.-T. Sheu, J.-f.R. Lu / Social Science & Medicine 111 (2014) 41e49 43

3. Methods

3.1. Data source

Taiwan Survey of Family Income and Expenditure (SFIE) releasedby the Directorate-General of Budget, Accounting and Statistics(DGBAS) was used to generate empirical results. SFIE is a series ofcross-sectional surveys of national representativehousehold samplesconducted annually in Taiwan from 1972 to the present. In addition,SFIE is a publicly released government household survey datawhichdoes not contain individual identifiers and our study only utilizedhousehold-level data, hence, Institutional Review Board (IRB)approval is not required. Our study exploited data only from 1993 to2000 for two reasons. First, itemized premium paymentsdwithwhich one can identify the type of insurance schemes that the re-spondents were enrolled in, to construct experiment and controlgroups for social insurance schemesdonly became available from1993. Second, we intended only to assess the insurance effect withinfive years of the NHI’s establishment to minimize the potential con-founding effects of external changes in policies introduced or eco-nomic environmental shocks such as the business cycle after 2000.

From 1993 to 2000, the SFIE sampled between 13,701 and16,434 households annually and contains rich, itemized informa-tion on household income, expenditure and assets. In addition,information on the socio-demographic characteristics of the indi-vidual household members (age, gender, educational attainment,marital status, and industrial sector of employment) is also avail-able. Except for consumption expenditure and assets, which werecollected on a household basis, information on itemized non-consumption expenditure (mortgage interest payment, tax pay-ment, premium for social insurance, charity donation, etc) and in-come (full-time and part-time payroll, pension, andentrepreneurial income, etc) was gathered for each member of thesampled households. Our study specifically focuses on the con-sumption expenditure, which was divided into medical and non-medical items. The medical expenditure component is mainlycomposed of household out-of-pocket payments for visits tomedical providers (including traditional healers) in the form of co-payment for insurance-covered services or full payment for non-covered services, and drugs and medical appliances/equipmentpurchased. The non-medical expenditure component covers awhole range of daily spending items, such as food, rental and waterbills (including estimated rental expenditure for home-owners),education, transportation and communication, travel, entertain-ment, and cultural activities, personal wardrobe, utilities, furnitureand household appliances, household services, beverages, tobacco,and miscellaneous at the household level.

To understand the impact of NHI on education spending inhouseholds with different compositions, we further divided thesample into households with school children (more likely to spendon education) and those without. A household with school childrenwas defined as any household with any member still in school(excluding college and graduate school students). The existingresearch regarding Taiwan’s income elasticity of education foundthe income elasticity of education expenditure is less than one anddecreased with income levels (Yeh and Tsai, 2000). This impliesthat education expenditure of households would have been lessimpacted by NHI. In addition, households in higher socioeconomicstrata, such as those with a higher income, will spend money oneducation, with or without the freed income of NHI.

3.2. Sample

As the sampled households were drawn independently eachyear, our study sample was composed of pooled cross-sectional

observations (not panel data). We used data from 1993 to 2000SFIE, except for the 1995 data, as NHI was implemented on March 1of that year. The final analytical sample was composed of 17,899households, with 10,573 in the control group and 7,326 in thetreatment group. For those 10,573 households in the control group,the pre-NHI (1993e1994) sample consisted of 3,766 (35.62%), andof the 7,326 households in the treatment group, 2,457 (33.54%)were surveyed before NHI. Our samplewas restricted to householdswith non-negative disposable income, which is computed by sub-tracting non-consumption expenditure and premiums for life andaccident insurance policies from total income to avoid householdswith obvious reporting errors. As our main purpose was to assessthe effect of NHI on household consumption patterns, we chosehouseholds without any form of insurance in both pre- and post-NHI periods as the treatment group, since they would haveremained uninsured had NHI not been introduced in 1995 andwould be covered by the health insurance after the implementationof NHI. As for the control, households least impacted by NHI wereideal candidates, such as households covered by GEI, LI, WI and FIthat would still have obtained health insurance coverage if therewas no NHI. As enrollment in WI was based on income eligibility,WI recipients are genuinely different from GEI beneficiaries, whoaccount for only 0.96% of the population (Lu and Hsieh, 2000). Wetherefore excluded the welfare recipients from our analyticalsample in order to avoid potential sample selection bias. FI re-cipients are also different from GEI beneficiaries. Most of FIenrollees reside in certain central and southern agriculturalcounties of Taiwan. In addition, farmers’wealth portfolios are quitedifferent in the sense that most of their saving is in the form of non-liquid assets such as land (Chou et al., 2003). We therefore alsoexcluded these agricultural households.

Finally, households with any member insured by GEI in bothpre- and post-NHI periods were used in the control group for tworeasons. First, they could still have obtained health insurancecoverage if there was no NHI. Second, the health insurancecoverage could be extended to their immediate family. In house-holds with any member enrolled in GEI, other members werehighly likely to possess affiliated GEI family coverage, and thus thehouseholds could be regarded as GEI-insured. In other words, theywere the households that were least impacted by the introductionof NHI. As the identification and selection of the experiment andcontrol groups was crucial to the validity of the estimation results,we were careful in our selection of the study groups. One possibleway was to identify the types of social insurance schemes thehouseholds were enrolled in directly from the survey questionsregarding their insurance enrollment status. The 1996e2000 SFIEspecifically asked the respondents whether they had enrolled inany social insurance programs, but there was no correspondingitem in the 1993e1995 surveys. Therefore, we modified our sampleidentification strategy contingent upon whether there was anypremium payment made for the respective social insuranceschemes, which enhanced the consistency in sample selectioncriteria over the years. We therefore identified GEI enrollees byexamining whether they had reported any GEI premium payment.By the same token, the households in the treatment group werethose without any household members reporting any premiumpayment to any of the social insurance programs (including GEI, LI,FI, Fishermen’s Insurance and insurance for military personnel).

To constrain the potential effects of other social insurance pro-grams, we excluded households with any social insurance coverage(LI, FI, Fishermen’s Insurance and insurance for military personnel)from the analytical sample. However, to perform a robustnesscheck on our empirical results, we constructed a quasi-treatmentgroup composed of households with members enrolling in the LIscheme.

J.-T. Sheu, J.-f.R. Lu / Social Science & Medicine 111 (2014) 41e4944

3.3. Estimation model

The differences-in-differences (DID) estimation model wasemployed to assess the impact of health insurance on householdconsumption patterns in our study. The underlying strategy herewas to assess the changes in various household consumptionspending shares between the GEI-insured households (controlgroup) and the uninsured households (treatment group) before andafter the implementation of NHI. The major advantage of the DIDmodel is to exclude the effect of natural growth over the years andit can more appropriately assess the effect of the policy of interestin a pre-post design with a control group.

The differences-in-differences estimation model can beexpressed as the following:

yh;t ¼ aþ b1NHIh;t þ b2Treath;t þ b3NHIh;t � Treath;t þ gXh;t

þ dhh;t þ εh;t

(1)

As we aimed to evaluate the impact on household resource-allocation decisions resulting from NHI implementation, we usedthe household consumption spending share, which reflects the

Table 1Distribution of consumption expenditure by items, 1994 and 1996.a

Economic indicatorsb

GDP per capita in USDNHE as % of GDPNHE per capita in USDFinancing systemb

Financing mixPublicGeneral governmentSocial security

PrivateOut-of-pocketPrivate insuranceOther private funds

Basic household characteristics Amount ($) A

No. of households surveyed 2896Average household size 3.89Average household size (PEAc) 2.27Average annual household income (PEA) 489,618Average annual household disposable incomed (PEA) 410,147Average annual consumption exp (PEA) 266,132Consumption exp./Income 54.36%Consumption patternExpendituree onMedical itemsf 9334 3Nonmedical items 256,798 9Food 58,869 2Rental and water bill 65,262 2Education (including childcare & school books etc.) 15,350 5Transportation and communications 28,596 1Travel, entertainment, and cultural activities 27,003 1Personal wardrobe 14,041 5Fuel and electricity bill 7182 2Furniture and household appliances 8715 3Household maintenance 6800 2Beverage 2644 0Tobacco 1605 0Miscellaneous 20,733 7

Note.a DGBAS Survey of Family Income and Expenditures, 1994 and 1996.b Source: 2012 Taiwan Statistic Data Book, CEPD; 2011 National health account, DOH.c PEA stands for per equivalent adult.d Disposable income is defined as: household income � (non-consumption expenditue All item expenditures are expressed on the basis of per equivalent adult.f Medical consumption includes medical care, medical equipment and appliance.

relative weight of the spending item options, as the dependentvariable in Equation (1). Specifically, yh,t is the consumptionspending share for household h at time t. We first examinedmedical and non-medical consumption spending shares, and thenthe specific spending shares of non-medical consumption, such asfood, rental and water bills, education, transportation andcommunication, travel, entertainment and cultural activities, per-sonal wardrobe, utilities, furniture and household appliances,household services, beverages, tobacco, and miscellaneous. Specificconsumption item spending share is defined as the underlyingconsumption item spending as a proportion of total consumptionexpenditure. NHIh,t is a binary variable indicating the period afterthe implementation of NHI. Treath,t is an indicator variable for thetreatment group. Based on Equation (1), the simple differences-in-differences estimator can be obtained as:

DNHI ¼fEðyjTreat ¼ 1;NHI ¼ 1;XÞ� EðyjTreat ¼ 1;NHI ¼ 0;XÞg� fEðyjTreat ¼ 0;NHI ¼ 1;XÞ� EðyjTreat ¼ 0;NHI ¼ 0;XÞg

¼ ½ðb1 þ b2 þ b3Þ � b2� � ½b1 � 0� ¼ b3

(2)

1994 1996

11,982 13,4284.87% 5.36%589 719

54.29% 64.41%15.07% 10.43%39.23% 53.98%45.71% 35.59%37.87% 24.00%3.75% 3.68%4.09% 7.91%

s a % of consumption exp Amount ($) As a % of consumption exp

22833.732.20507,588417,766276,62954.50%

.51% 8716 3.15%6.49% 267,913 96.85%2.12% 63,233 22.86%4.52% 69,782 25.23%.77% 16,260 5.88%0.74% 30,740 11.11%0.15% 27,679 10.01%.28% 12,899 4.66%.70% 7604 2.75%.27% 8373 3.03%.56% 7518 2.72%.99% 2697 0.97%.60% 1599 0.58%.79% 19,529 7.06%

re þ privately paid insurance premium).

J.-T. Sheu, J.-f.R. Lu / Social Science & Medicine 111 (2014) 41e49 45

If the implementation of NHI causes households to change theirconsumption behavior, the estimated coefficient of b3 should reachstatistical significance.

In addition, we incorporated two sets of control variables intoour models. hh,t represents the characteristics of the household’sresidence, indicated by geographic location (by NHIA regional of-fices, with Taipei regional office as the reference group) and levelsof urbanization (city, town and rural, with city as the referencegroup). Household characteristics are hypothesized to havediffering impacts on consumption patterns so, if one fails to controlfor such characteristics, the estimation results are likely to bebiased. Household characteristics (X) which covered age, genderand educational attainment of the household head; householdswith children under the age of five; households with elders above65; and household income (indicated by income quantiledummies), were also taken into account.

To assess the spillover effect of the implementation of NHI onhousehold consumption spending shares, we estimated the Equa-tion (1) under the regression technique of a generalized linearmodel (GLM), due to the skewed distributions for those con-sumption spending share variables, which violates the basicassumption of Ordinary Least Squares (OLS) regression. One way totackle the distribution problem is to log-transform the dependentvariables, so that the skewed distribution can be normally shaped.However, as mentioned in Manning and Mullahy (2001), OLS es-timators are biased even if the dependent variable is in the log-form. Furthermore, with log-transformed dependent variables,one can no longer assume away the problem of heteroscedasticity,which results in an inefficient and inconsistent estimator

Table 2Household characteristics of the analytical sample.

Unit of observations ¼ household Control group: householdmembers with governmentemployee insurance (GEI) benefits

Trw

Sample size 10,573 73Age (household head)Average of age 43.37 (9.77) 4520e30 (%) 7.07 8.30e40 (%) 30.71 2840e50 (%) 35.02 2750e60 (%) 20.54 1560e65 (%) 6.65 20

Gender (household head)Male (%) 86.24 67Female (%) 13.76 30

Education (household head)Graduate school (%) 9.58 0.College (%) 62.58 13Senior high school (%) 21.73 25Junior high school (%) 2.71 20Elementary school (%) 3.40 40

Household with children under 5Yes (%) 26.54 17

Household with elders above 65Yes (%) 17.53 19

Region (by BNHI branch office)Taipei region (%) 41.35 29Northern region (%) 10.42 9.Central region (%) 15 19Southern region (%) 12.64 15Eastern region (%) 3.99 5.Kao-ping (%) 16.67 20

UrbanizationCity (%) 75.72 67Town (%) 17.54 23Rural area (%) 6.73 9.

Annual household income (mean) 1,211,941.07 (599,669.30) 50Annual household disposable income

(mean)*(per equivalent adult)530,646.21 (278,697.77) 27

(Manning, 1998). In order to solve the above-mentioned potentialestimation problems, we followed the suggestion of Manning andMullahy (2001), and used the GLM with a family of gamma distri-bution and log-link relationship to estimate Equation (1). In addi-tion, the surveyed households were selected through a stratifiedsampling framework based on geographic location (by city/county).We therefore adjusted the estimation by considering the year andcounty clustering effect to tackle the potential problem of under-estimating standard error due to correlation among householdsresiding in the same county in a given year (Bertrand et al., 2004).After performing the GLM estimation, the spillover effect waspresented in the form of the marginal effect due to the nonlinearityof the GLM estimation (Hardin and Hilbe, 2012).

4. Results

As shown in Table 1, on average, non-medical consumption ac-counts for 96.49% (1994) and 96.85% (1996) of total consumptionexpenditure. The bigger-ticket items of the non-medical con-sumption component in 1994 were rental and water bills (24.52%)and food (22.12%), while education accounted for 5.77%. In 1996,both spending shares of household expenditure on rental andwater bills, and food, were slightly increased, by 25.23% and 22.86%respectively, and remained the two biggest spending items in thenon-medical consumption component. The share on education alsoincreased slightly in 1996. As for medical consumption (in the formof a household’s out-of-pocket payment for medical services,medical equipment and appliances), the share fell from 3.51%(1994) to 3.15% (1996).

eatment group: household membersithout any social insurance

Quasi-treatment: household memberswith labor Insurance(LI)

26 33,543

.41 (12.11) 41.26 (10.06)49 11.02.15 36.17.61 33.00.56 13.15.19 6.66

.97 81.38

.03 18.62

60 1.05.19 18.05.83 30.45.09 21.24.29 29.21

.32 23.61

.15 21.37

.55 28.2325 11.37.18 18.44.93 18.4424 3.89.84 19.63

.27 52.94

.61 27.3712 19.691,854.11 (346,178.11) 607,648.15 (353,843.54)2,553.75 (178,811.38) 298,249.9 (178,579.12)

Table 3Comparison of shares of consumption items between treatment group and control group in pre-NHI and post-NHI periods.

Treatment group (households w/o any form of social insurance) Control group (household members with GEIa benefits) DIDc

Pre-NHI(1993e1994)

Post-NHI(1996e2000)

T-diffb Pre-NHI(1993e1994)

Post-NHI(1996e2000)

C-diffb

Medical consumption 4.83% 3.84% �1.00% 3.06% 3.01% �0.05% �0.95%Non-medical consumption 95.17% 96.16% 1.00% 96.94% 96.99% 0.05% 0.95%Food 28.42% 28.15% �0.27% 24.00% 23.22% �0.78% 0.51%Rental and water bill 27.54% 30.31% 2.77% 23.36% 22.96% �0.40% 3.17%Education(including childcare& school books etc.)

5.02% 4.29% �0.73% 6.25% 6.97% 0.72% �1.45%

Transportation and communications 7.19% 8.19% 1.00% 10.52% 11.88% 1.36% �0.35%Travel, entertainment,and cultural activities

6.51% 5.86% �0.65% 9.59% 9.78% 0.19% �0.84%

Personal wardrobe 4.88% 3.85% �1.03% 5.98% 4.93% �1.05% 0.02%Fuel and electricity bill 3.76% 3.82% 0.06% 2.83% 2.68% �0.15% 0.22%Furniture and householdappliances

1.87% 1.67% �0.20% 2.85% 2.67% �0.18% �0.02%

Household maintenance 1.42% 1.66% 0.25% 2.59% 2.83% 0.24% 0.01%Beverage 1.36% 1.21% �0.15% 0.94% 0.94% 0.00% �0.15%Tobacco 1.31% 1.16% �0.14% 0.59% 0.45% �0.14% 0.00%Miscellaneous 5.89% 5.99% 0.10% 7.43% 7.68% 0.26% �0.16%

Note.a GEI, Government Employee Insurance.b T-diff (C-diff) : share in Post-NHI in treatment (control) group � share in Pre-NHI in treatment (control) group.c DID: T-diff � C-diff.

J.-T. Sheu, J.-f.R. Lu / Social Science & Medicine 111 (2014) 41e4946

Table 2 depicts sample characteristics of households in thetreatment, quasi-treatment, and control groups. The householdhead in the control group is 43.37 years old on average, which isyounger than those in the treatment group (45.41), and older thanthose in the quasi-treatment group (41.26). Compared to those inthe treatment and quasi-treatment groups, household heads in thecontrol group appeared to be younger and possessed a higher levelof education and income, whether in terms of annual household

Table 4Marginal effect of NHI on consumption shares.a

Consumption pattern Controlgroup*Pre-NHI

Overall Income level

inc100 inc60 inc20

Medical 3.06% �0.0074*** �0.0068*** �0.0076*** �0.0110Non-medical 96.94% 0.0097*** 0.0097*** 0.0097*** 0.0095Food 24.00% �0.0048 �0.0040 �0.0051 �0.0066Rentaland water bill

23.26% 0.0187*** 0.0176*** 0.0187*** 0.0204

Education (includingchildcare & schoolbooks etc.)

6.25% �0.0101*** �0.0081*** �0.0114*** �0.0084

Education (HHw/school children)c

8.09% �0.0029 �0.0027 �0.0030 �0.0028

Transportationand communications

10.52% 0.0098 0.0114 0.0099 0.0066

Travel, entertainment,and cultural activities

9.59% �0.0073* �0.0102* �0.0067* �0.0039

Personal wardrobe 5.98% �0.0030** �0.0032** �0.0030** �0.0026Fuel and electricity bill 2.83% 0.0005 0.0004 0.0005 0.0007Furniture and householdappliances

2.85% 0.0003 0.0004 0.0003 0.0002

Household maintenance 2.59% 0.0011 0.0014 0.0010 0.0008Beverage 0.94% �0.0014** �0.0011** �0.0014** �0.0018Tobacco 0.59% 0.0004 0.0002 0.0005 0.0009Miscellaneous 7.43% 0.0008 0.0010 0.0007 0.0006

Note.a *p < 0.05,**p < 0.01,***p < 0.001.b High SES ¼ inc100 þ graduate school; middles SES ¼ inc60 & senior high school/voc Households with school children: defined as families with members still in school (e

income or annual household disposable income (per equivalentadult).

Compared to the pre-NHI period, the previously uninsuredhouseholds demonstrated a greater reduction in medical-itemspending share, from 4.83% to 3.84%, while the previously GEI-insured experienced almost no change, from 3.06% to 3.01%(Table 3). The differences-in-differences of the share on medicalconsumption between these two groups is �0.95%. This showed

Education level Income*education level

Graduateschool

Senior highschool/vocationalschool

Elementaryschool

High SESb Middle SESb Low SESb

*** �0.0062*** �0.0077*** �0.0124*** �0.0061*** �0.0074*** �0.0132****** 0.0098*** 0.0097*** 0.0095*** 0.0098*** 0.0097*** 0.0095***

�0.0041 �0.0054 �0.0062 �0.0038 �0.0053 �0.0068*** 0.0175*** 0.0186*** 0.0215*** 0.0171*** 0.0185*** 0.0216***

** �0.0101*** �0.0107*** �0.0064** �0.0086*** �0.0111*** �0.0064**

�0.0025 �0.0028 �0.0030 �0.0024 �0.0029 �0.0029

0.0113 0.0097 0.0063 0.0118 0.0103 0.0057

* �0.0095* �0.0060* �0.0044* �0.0109* �0.0065* �0.0037*

** �0.0031** �0.0030** �0.0025** �0.0031** �0.0031** �0.0025**0.0004 0.0005 0.0007 0.0003 0.0005 0.00080.0004 0.0003 0.0002 0.0005 0.0003 0.0002

0.0018 0.0010 0.0007 0.0019 0.0010 0.0006** �0.0011** �0.0016** �0.0019** �0.0010** �0.0016** �0.0019*

0.0001 0.0007 0.0009 0.0001 0.0006 0.00100.0009 0.0007 0.0006 0.0010 0.0008 0.0006

cational school; low SES ¼ inc20 þ elementary school.xcluding college and graduate school).

J.-T. Sheu, J.-f.R. Lu / Social Science & Medicine 111 (2014) 41e49 47

that without controlling for any other factors, the medical con-sumption decreased after NHI introduction. While examining thechange in non-medical consumption spending shares between thepost-NHI and pre-NHI periods, we found that the previouslyuninsured households showed a declining spending share in edu-cation (�0.73%) and an increasing trend in housing-related items(mainly in the form of rental and water bills, 2.77%) after NHI wasintroduced. However, the change in spending share on non-medical consumption items appeared less pronounced for thepreviously GEI-insured households (Table 3). The descriptive re-sults comparing the control and treatment groups in the pre-NHIand post-NHI periods revealed a spillover effect of a 3.17 percent-age point increase in the spending share for rental and water bills;while the education spending share indicated a drop of 1.45 per-centage points.

Controlling for the exogenous variables, the results from a log-linear GLM regression are shown in Table 4. We transformedparameter estimates based on the method described previously.Overall, the statistically significant negative insurance effect in themedical consumption spending share (a reduction of 0.74 per-centage points in the share) and the positive tendency in non-medical consumption spending share (an increase of 0.97 per-centage points), support our hypothesis. The overall consumptionincreased after the implementation of NHI, providing evidence ofconsumption smoothing.

Our results showed a significant spillover effect of NHI onhousehold consumption patterns. Among the non-medicalspending items, we found that a 1.87 percentage point increase(in the share) was noted for the rental and water bills spendingshare, representing relatively more resources being allocated tohousing condition improvement. However, there was a drop in thespending share for consumption items such as travel, entertain-ment and cultural activities, personal wardrobe, and beverages, butthere was no statistically significant change observed for food.Furthermore, the education spending share experienced the most

Table 5Marginal effect of NHI on consumption shares: quasi-treatment group.a

Consumption pattern Controlgroup*pre-NHI

Overall Income level

inc100 inc60 inc20

Medical 3.06% �0.0062*** �0.0053*** �0.0068*** �0.0080Non-medical 96.94% 0.0064*** 0.0064*** 0.0064*** 0.0063Food 24.00% �0.0095** �0.0075** �0.0107** �0.0130Rental and water bill 23.26% 0.0097** 0.0095** 0.0099** 0.0090Education (includingchildcare & schoolbooks etc.)

6.25% �0.0005 �0.0004 �0.0005 �0.0005

Education (HHw/school children)a

8.09% 0.0031 0.0028 0.0030 0.0031

Transportation andcommunications

10.52% 0.0071 0.0084 0.0068 0.0059

Travel, entertainment,and cultural activities

9.59% �0.0030 �0.0044 �0.0025 �0.0018

Personal wardrobe 5.98% 0.0005 0.0006 0.0005 0.0005Fuel and electricitybill

2.83% �0.0001 �0.0001 �0.0001 �0.0001

Furniture and householdappliances

2.85% �0.0001 �0.0001 �0.0001 0.0000

Household maintenance 2.59% 0.0019* 0.0027** 0.0015* 0.0013Beverage 0.94% �0.0009 �0.0007 �0.0010 �0.0011Tobacco 0.59% 0.0000 0.0000 0.0000 0.0000Miscellaneous 7.43% 0.0029 0.0035 0.0027 0.0023

Note.a See Table 4.

significant drop, �1.01%. The SFIE adopted a broad definition ofeducation spending which covered expenditure on childcare,school books, all items purchased for job training and even anyexpenses related to adult continuous education. In order to un-derstand the impact of NHI on education spending among house-holds with different compositions, we re-estimated this impact byincluding only households with school children (children still inschool, excluding college and graduate school students), making up50.15% of the initial sampled households. As shown in Table 4, therewas no significant change in the spending share on education forhouseholds with school children, which suggested that householdswith school children were less likely to divert resources for edu-cation to other consumption items in the face of an increase in realincome. To assess whether socioeconomic gradient played a role inthe NHI’s impact on consumption patterns, we further divided thehouseholds sample based on socioeconomic status, using thehousehold income and education attainment level of the householdhead. Specifically, we used the 20th percentile and 60th percentileof income distribution as our cut-off thresholds, and divided ouranalytical sample into lower-income households (lowest 20%),middle-income households (between 20% and 60%), and higher-income households (highest 40%). From the estimated marginaleffects (Table 4), it can be seen that, after NHI, the lower-incomehouseholds had a more significant drop in medical expenditureshare, while the overall non-medical consumption showed noobvious difference in spillover effects across the three income-levelgroups. It showed that the availability of NHI freed up a moresubstantial share of resources for the poor than the rich. Uponfurther examination of the distribution of non-medical consump-tion items, the empirical evidence suggested that the householdshad chosen to devote more resources to improving housing andliving conditions, but made less investment proportionally in ed-ucation as a result of NHI, especially for those households in themiddle-income level. In addition, no statistically significantchanges in food and tobacco consumption share were observed.

Education level Income*education level

Graduateschool

Senior highschool/vocationalschool

Elementaryschool

High SESa MiddleSESa

Low SESa

*** �0.0049*** �0.0064*** �0.0086*** �0.0048*** �0.0065*** �0.0090****** 0.0064*** 0.0064*** 0.0063*** 0.0064*** 0.0064*** 0.0063***** �0.0076** �0.0104** �0.0120*** �0.0072** �0.0106** �0.0133**** 0.0093** 0.0097** 0.0096** 0.0089** 0.0099** 0.0093**

�0.0005 �0.0005 �0.0004 �0.0004 �0.0005 �0.0004

0.0026 0.0029 0.0033 0.0026 0.0029 0.0033

0.0080 0.0072 0.0057 0.0086 0.0071 0.0053

�0.0046 �0.0026 �0.0020 �0.0050 �0.0026 �0.0017

0.0006 0.0005 0.0005 0.0006 0.0005 0.0005�0.0001 �0.0001 �0.0001 0.0000 �0.0001 �0.0001

�0.0001 �0.0001 �0.0001 �0.0001 �0.0001 0.0000

* 0.0035* 0.0016* 0.0012* 0.0037* 0.0016* 0.0011*�0.0006 �0.0010 �0.0012 �0.0006 �0.0010 �0.00120.0000 0.0000 0.0000 0.0000 0.0000 0.00000.0034 0.0027 0.0026 0.0036 0.0027 0.0023

J.-T. Sheu, J.-f.R. Lu / Social Science & Medicine 111 (2014) 41e4948

The same estimation model applied to examine the educationgradient showed that household heads with lower education levelshad a more significant drop in medical consumption share afterNHI, and again the results showed no obvious difference in thespillover effect in the non-medical expenditure share acrosshousehold heads with different education levels (Table 4). Uponexamination of the distribution of non-medical consumptioncomponents, household heads with lower education attainmentwere reported to devote a greater share to housing-conditionimprovement but to a lesser extent to reduce their investment ineducation. To further assess the NHI impact along socioeconomicstrata, we divided the sampled households based on the SESinteraction terms constructed (household income*education).Similar patterns persist (shown by the last three columns inTable 4). In particular, no obvious difference in the share changewas noted in non-medical consumption across households withdifferent SES. In contrast, one can see that lower SES householdshad a more significant drop in medial expenditure share. Further-more, after carefully examining non-medical consumption items,households with middle SES tended to experience a more signifi-cant negative spillover effect on education investment.

In order to check whether our results were robust, we re-did allestimations in Table 5, using households with anymember enrolledin labor insurance (and no member with GEI coverage) before NHIas the quasi-treatment group. Labor insurance is a social insuranceprogram that covers only the enrollee him/herself. Thus, comparedto our initial treatment group, the NHI impact on the quasi-treatment group should be less. As shown in Table 5, the magni-tude of the spillover effect for the quasi-treatment group appearedless pronounced than those presented in Table 4, which implied therobustness of our estimation model.

5. Discussion and conclusion

Several potential caveats bear mentioning. First, there are po-tential negative impacts on income resulting from the imple-mentation of a social health insurance program. One possibleexample is income loss due to insurance premium tax as a resultof the insurance program, though this is likely to be modest. Forone, even for the poorest 20% of households, the NHI premium asa percentage of household income is less than 6%, thus the“taxation effect” on income would be modest. On the other hand,the private out-of-pocket health payment as a share of total healthexpenditure declined after the implementation of NHI (Table 1).When one is insured, the medical-cost saving is significant in thecase of health shocks, that is, NHI benefits essentially outweighpremium cost. From a macro-financing perspective, the personaltax burden has not risen since the inception of the NHI program in1995. Despite a rather pronounced and constant structural budgetdeficit since 1990s (Ministry of Finance, 2006), social welfarerelated government expenditure share has significantly increasedfrom 18.75% in 1991 to 30.03% in 2001 at the expense of adecreasing spending share on national defense, dropping from17.8% in 1991 to 10.09% in 2001 (Ministry of Finance, 2003). At thesame time, the tax burden in Taiwan reached its peak of 20% in1990 and then steadily dropped to 13.0% in 2001. Therefore,though there might be a negative impact on income, the impact isconsidered trivial.

Second, potential sample selection issues might be present.There were inherent differences in household characteristicsobserved for those households in the treatment and control groupschosen in this study. The previously GEI-insured householdsappeared to possess higher levels of education and income thanthose previously uninsured households in the treatment group. It ispossible that our model could not totally control for the inherent

differences in these two groups, hence some of the NHI effectsdetected can be attributed to their genuine differences, rather thanto the NHI. However, we have attempted to control for as manyobservable household attributes as possible in the log-linear GLMmodel to reduce the potential bias. In addition, by exploiting theleast well-off households as a control group against the rather well-off GEI-insured group, the NHI effect assessed can be regarded as ahigher bound estimate of the spillover effect. Nevertheless, arobustness check using households with any member enrolled inlabor insurance as the quasi-treatment group did show that ourfindings were consistent and robust.

Third, our study only aims to assess the short-term (1996e2000) effect of NHI, though looking for evidence of policy effects inthe longer termmay be more desirable. However, with the passageof time, one cannot disentangle some exogenous effects (such asmarket economic activities) from the NHI effect in the later years ofNHI. In fact, the potential confounding effect may undermine thepure NHI effect to a greater extent.

Fourth, when evaluating marginal effect by different socialeconomic strata, we implicitly assume that the values of all theother variables remain constant over the years. However, due to thenature of the pooled cross-sectional data, we might have ignoredthe potential impacts of changes in distribution of other variableson our findings, in particular, the distribution of income. None-theless, as the income distribution during the study period did notchange significantly, we believe that the effect should be trivial.

Fifth, estimates on housing expenditures for home owners maybe subject to potential bias. DGBAS (the government agency incharge of SFIE) asked the respondents to estimate the rental price(if their home were put on rental market) and the interviewerwould make adjustment based on current rental market price if thereported estimated rental expenditure seems unreasonable. Thiscommon practice adopted by DGBAS is due to data limitation.

Lastly, our paper mainly aims to assess the impact of NHI onhousehold consumption patterns, and the changes in housing andeducation expenditure shares are viewed as the spillover effectfrom NHI policy. Further discussions on social or education policiescannot be addressed by the data exploited in this study and arebeyond the scope of our paper.

Our study concludes with empirical evidence showing thespillover effect of NHI on consumption patterns. The major findingsof our study are as follows:

First, as expected and evidenced by our results, we found thatafter the implementation of NHI, households spent less on medicalitems and more on non-medical items. Two issues deserve noting:(1) The medical consumption share dropped after NHI was intro-duced but the magnitude was not as great as expected. This couldbe attributed to the medical consumption used in our analysiscovering a broad spectrum of self-pay health services. The imple-mentation of NHI is expected to relieve the insured of somefinancial burden for the NHI-covered services, but the insuredmight divert the freed-up resources to items not covered by theNHI, such as the amenity-enhancing aspect of services (wardupgrading) or cosmetic surgery and so forth. Hence, we observed astatistically significant but modest-in-magnitude drop in medicalconsumption. (2) Overall consumption increased after NHI imple-mentation, which is consistent with the findings in the existingliterature that compulsory health insurance will impact a house-hold’s consumption behavior, and households will re-allocate theirresources after the policy (Chou et al., 2003, 2004; Gruber andYelowitz, 1999; Koch and Alaba, 2010; Kuan and Chen, 2013; Levyand Deleire, 2008).

Second, our results indicate that there was no significantchange in the food spending share in the post-NHI period. Thismight reflect the fact that households in high-income economies,

J.-T. Sheu, J.-f.R. Lu / Social Science & Medicine 111 (2014) 41e49 49

such as Taiwan, in general devote a constant spending share tofood. That is, people who have enough food to eat don’t tend tooverconsume, even if there is an increase in real income (Seale andRegmi, 2006).

Third, our results also showed that disadvantaged households inTaiwan were inclined to divert the resources freed up after theimplementation of NHI to improve their housing conditions. The factthat housing expenses such as rental and water bills account for agreater weight of the consumption expenditure for the poorerhouseholds is consistentwith results by LevyandDeLeire (2008) andKoch and Alaba (2010) which showed that poor households mighthave to forego insurance to pay for basic needs such as housing.

Fourth, we also found that the education spending sharedropped significantly though modestly, after the implementationof NHI. However, the fall in the education spending share was notstatistically significant if we used only the sample of householdswith school children in the estimation model, which suggestedthere was no evidence that households with school childrenforwent investment in education. The drop in the educationspending share might simply be a reflection of the general pref-erence of the adults and could be referred to a decrease in generaleducation expenses, such as continuous education and adult jobtraining. Our findings showed that households with lower socialeconomic status experienced a greater spillover effect of NHI,which presents important policy implications. Authorities plan-ning welfare policies or subsidies should familiarize themselveswith the implications of the greater spillover effect on suchhouseholds, as these could result in necessary adjustments tocurrent subsidy allocation. Recognizing how these householdsexploit the potential benefits associated with NHI provision couldenable the government to devise specific policy tools to facilitatebetter targeting of investment in education for less well-offhouseholds. However, further research is warranted for devisingspecific policies.

Acknowledgments

Grant support from National Science Council (NSC93-2416-H-182-015 and NSC 102-2410-H-182-009) in Taiwan for the workfrom which this paper derived is gratefully acknowledged. Theauthors are also greatly benefited from the valuable commentsreceived from Profs. Willard Manning, Joanna Coast and twoanonymous referees, as well as numerous seminar audiences. Theexcellent research assistance provided by Ching-Hsing Chang, Tzu-Yin Hazel Tseng, and Chao-Chin Sherina Lee, are deeply appreci-ated. Nonetheless, the authors are solely responsible for viewspresented in this paper.

References

Bertrand, M., Duflo, E., Mullainathan, S., 2004. How much should we trustdifferences-in-differences estimates? The Quarterly Journal of Economics 119(1), 249e275.

Boyle, M.A., Lahey, J.N., 2010. Health insurance and the labor supply decisions ofolder workers: evidence from a U.S. Department of Veterans Affairs expansion.Journal of Public Economics 94, 467e478.

Broyles, R.W., Manga, P., Binder, D.A., Angus, D.E., Charette, A., 1983. The use ofphysician services under a National Health Insurance scheme. An examinationof the Canada Health Survey. Medical Care 21 (11), 1037e1054.

Chen, L., Yip, W., Chang, M.C., Lin, H.S., Lee, S.D., Chiu, Y.L., Lin, Y.H., 2007. The effectsof Taiwan’s National Health Insurance on access and health status of the elderly.Health Economics 16 (3), 223e242.

Cheng, S.H., Chiang, T.L., 1997. The effect of universal health insurance on healthcare utilization in Taiwan: results from a natural experiment. Journal of theAmerican Medical Association 278 (2), 89e93.

Chou, S.Y., Liu, J.T., Hammitt, J.K., 2003. National Health Insurance and precau-tionary saving: evidence from Taiwan. Journal of Public Economics 87 (9e10),1873e1894.

Chou, S.Y., Liu, J.T., Huang, C.J., 2004. Health insurance and savings over the life cyclee a semiparametric smooth coefficient estimation. Journal of Applied Econo-metrics 19 (3), 295e322.

Chou, Y.J., Staiger, D., 2001. Health insurance and female labor supply in Taiwan.Journal of Health Economics 20 (2), 187e211.

Decker, S.L., Remler, D.K., 2005. How much might universal health insurance reducesocioeconomic disparities in health? A comparison of the US and Canada.Applied Health Economics and Health Policy 3 (4), 205e216.

Gruber, J., Yelowitz, A., 1999. Public health insurance and private savings. Journal ofPolitical Economy 107 (6), 1249e1274.

Hardin, J.W., Hilbe, J.M., 2012. Generalized Linear Models and Extensions, third ed.Stata Press, College Station, TX.

Koch, S., Alaba, O., 2010. On health insurance and household decisions: a treatmenteffect analysis. Social Science & Medicine 70 (2), 175e2182.

Kuan, C.M., Chen, C.L., 2013. Effects of National Health Insurance on precau-tionary saving: new evidence from Taiwan. Empirical Economics 44 (2),921e943.

Levy, H., DeLeire, T., 2008. What do people buy when they don’t buy health in-surance and what does that say about why they are uninsured? Inquiry 45 (4),365e379.

Lu, J.R., 2013. A single-payer universal social insurance system: the Taiwan expe-rience in financing reform. In: Cheung, F.M., Woo, J.W., Law, C.K. (Eds.), HealthSystems: Challenges, Visions, and Reforms From a Comparative GlobalPerspective. Hong Kong Institute of Asia-Pacific Studies, The Chinese Universityof Hong Kong.

Lu, J.R., Chiang, T.L., 2011. Evolution of Taiwan’s health care system. Health Eco-nomics, Policy and Law 6 (1), 85e107.

Lu, J.R., Hsiao, W.C., 2003. Does universal insurance make health care unaffordable?Lessons from Taiwan. Health Affairs 22 (3), 77e88.

Lu, J.R., Hsieh, C.R., 2000. Health Economics. Pro-Ed Publishing Company, Taipei (inChinese).

Manning, W.G., 1998. The logged dependent variable, heteroscedasticity, and theretransformation problem. Journal of Health Economics 17 (3), 283e295.

Manning, W.G., Mullahy, J., 2001. Estimating log models: to transform or not totransform? Journal of Health Economics 20 (4), 461e494.

Ministry of Finance, 2003, 2006. Yearbook of Tax Statistics. Taiwan’s Ministry ofFinance Press, Republic of China.

Park, J.M., 2012. Equity of access to primary care among older adults in Incheon,South Korea. Asia-Pacific Journal of Public Health 24 (6), 953e960.

Seale, J.L., Regmi, A., 2006. Modeling international consumption patterns. Review ofIncome and Wealth 52 (4), 603e624.

Yeh, K.L., Tsai, C.L., 2000. A study of income elasticity of education expenditures inTaiwan. Educational Research and Information 8 (4), 74e91 (in Chinese).