The effects of mandatory health insurance on equity in access to outpatient care in Indonesia

Transcription

1 doi: /heapol/czh037 Health Policy and Planning 19(5), HEALTH POLICY AND PLANNING; 19(5): Oxford University Press, 2004; all rights reserved. The effects of mandatory health insurance on equity in access to outpatient care in Indonesia BUDI HIDAYAT, 1,2 HASBULLAH THABRANY, 3 HENGJIN DONG 1 AND RAINER SAUERBORN 1 1 Department of Tropical Hygiene and Public Health, University of Heidelberg, Heidelberg, Germany, 2 School of Public Health, University of Indonesia, Indonesia and 3 Center for Health Economic Studies, University of Indonesia, Indonesia This paper examines the effects of mandatory health insurance on access and equity in access to public and private outpatient care in Indonesia. Data from the second round of the 1997 Indonesian Family Life Survey were used. We adopted the concentration index as a measure of equity, and this was calculated from actual data and from predicted probability of outpatient-care use saved from a multinomial logit regression. The study found that a mandatory insurance scheme for civil servants (Askes) had a strongly positive impact on access to public outpatient care, while a mandatory insurance scheme for private employees (Jamsostek) had a positive impact on access to both public and private outpatient care. The greatest effects of Jamsostek were observed amongst poor beneficiaries. A substantial increase in access will be gained by expanding insurance to the whole population. However, neither Askes nor Jamsostek had a positive impact on equity. Policy implications are discussed. Key words: access, equity, concentration index, health insurance, Indonesia, IFLS Introduction Determinants of the effective and equitable use of health care services are multifaceted; among these are financial barriers (Sauerborn et al. 1994; Fabricant et al. 1999) and low quality of care (Lafond 1995; Yip et al. 1998; Akin and Hutchinson 1999). Health insurance has the potential to lower financial barriers, since the financial risk of health care is shared among members and no further costs will arise at the point of health care use (with the exception of cost sharing policies). Therefore, health insurance increases demand for health care through the reduction of effective prices (Feldstein 1993). However, empirical evidence from developing countries is conspicuously scanty. Most studies on demand effects of insurance have been conducted in developed countries (Manning et al. 1987; Kreider and Nicholson 1997; Chiappori et al. 1998; Holly et al. 1998). Some authors studying the demand effect of insurance focused on the probability of use (Vera-Hernandez 1999; Waters 1999; Trujillo 2003) in order to evaluate the effectiveness of such consumer incentives (Schellhorn 2001; van de Voorde et al. 2001). However, few studies deal with the equity effects of insurance on access to health care. Among the few studies on the effect of health insurance on access and equity in access in developing countries is a study in Egypt by Yip and Berman (2001). The authors compared only the predicted probability of health care use and expenditure across income quintiles and insurance status. Furthermore, the authors did not distinguish between public and private health facilities, which may have different implications for access and equity of access. Van Doorslaer et al. (2002) point out that simple quintile distributions as equity measures (e.g. the richest/poorest ratio) produce misleading interpretations of the results, because they are not sensitive to the values of the middle three quintiles. In the present study, we look at income-related inequity in access to public and private outpatient care using the concentration index. During the 1990s, the Indonesian government promoted three health insurance schemes: (1) social health insurance for private employees (Jamsostek), (2) voluntary health insurance (private), and (3) community health maintenance insurance (known as Jaminan Pemeliharaan Kesehatan Masyarakat, JPKM). A social health insurance for civil servants (Askes) has been offered since Notwithstanding these insurance opportunities, about 86% of the Indonesian population is not covered by any form of health insurance scheme (Thabrany 2001). Furthermore, and motivated largely by the expectation that insurance improves access to modern medical care, the government has committed itself to expanding health insurance by proposing a National Social Health Insurance (NSHI) scheme. The proposed NSHI scheme has been incorporated in an academic paper and draft bill for the new national social security by the Task Force for Social Security Reform, established by a Presidential decree (Thabrany 2003). Two questions appear relevant here. First, does mandatory insurance increase access to health care in Indonesia? Second, does it also increase equity in access? The purpose of this paper is two-fold. First, it examines the effects of mandatory health insurance (e.g. Askes and Jamsostek) on access to public and private outpatient care. Second, it investigates the effects of these insurance schemes on equity in access to public and private outpatient care. This study, therefore, will supplement the Task Force simulation on access and equity in access to outpatient care after the introduction of the NSHI in Indonesia.

2 Health insurance and equity 323 Methods Study site With an estimated population of million in 2001, Indonesia is the fourth largest country in the world after China, India and the United States (WHO 2002). The per capita GDP in 1999 was US$723 (UN Statistics Division 2002). The total expenditure on health as a percentage of GDP increased from 1.7% in 1995 to 2.7% in 2000, in which the contribution from private sources of finance accounted for 62.7% in 1995 and increased to 76.3% in Private health care finance came primarily from out-of-pocket payments including user fees (70.1%), with a small proportion from insurance (8.2%) (WHO 2002). The Indonesian health care services represent a mix of public and private providers. The public providers include public hospitals, public health centres and public health subcentres. These facilities provide health services to everyone at heavily subsidized prices. The government has built 873 public hospitals. Regional governments (about 340 districts and municipalities) provide more than 7243 public health centres, which operate as referral point for the district hospitals, and more than public health subcentres (Thabrany 2001). The private providers consist of private hospitals, private clinics and private practitioners. Although these providers are generally available in urban areas (Brotowasisto et al. 1988), doctors in public health centres and public hospitals can offer private services after office hours, and thus private practitioners can be found in most rural areas as well. Private providers are an increasingly important source of health care. Data from the 1998 Social Economics National Survey (SUSENAS) documented that private providers were the most utilized providers of outpatient care (52%) compared with public providers (43%) and traditional practitioners (4%) (Lanjouw et al. 2002). Indonesia has two mandatory health insurance programmes. Their characteristics are summarized in Table 1. Firstly, Askes, mandatory health insurance for civil servants, is the oldest and the largest health insurance scheme in Indonesia, covering all civil-servants, pensioners of civil servants and armed forces, and their families. Eligible dependents include the spouse and the first two children under 21 years of age or 25 years if a child is a full-time student. Mandatory contributions amount to 2% of the monthly salary, and this is deducted automatically from employees by the Ministry of Finance. The benefits are comprehensive health care services provided mainly in public health facilities, in which all members are entitled to the same benefits considered medically necessary. By 1998, this scheme covered approximately 13.8 million people (Thabrany 2001). Secondly, the Social Security Act of 1992 mandates all employers with 10 or more employees, or paying a monthly payroll of more than 1 million rupiah (the Indonesian currency), to register their employees in the social health insurance, Jamsostek. However, the law allows employers who provide better health care than that covered by Jamsostek to opt out. Premiums are 3% and 6% of monthly salary for single and married employees, respectively. This scheme covers employees and dependents up to three children under 21 years of age, who are entitled to comprehensive (but limited) health services. In 1999, coverage of this social security scheme was only 2.3 million workers, less than 5% of those eligible (Thabrany 2001). Data We used data from the second round of the Indonesian Family Life Survey (IFLS2) carried out in 1997/8 by a team from the RAND Corporation in conjunction with Indonesian researchers and various international agencies. The first round of the survey (IFLS1) was concluded in The IFLS1 included interviews with out of individuals of the 7224 households selected from 13 provinces. The IFLS2 was based on the same 1993 sample, and 93.5% of the original households were tracked and interviewed. This survey was described more fully in Frankenberg and Karoly (1995) and Frankenberg and Thomas (2000) for the IFLS1 and IFLS2, respectively. The IFLS questionnaires cover an array of topics that are central to the questions of this study. The survey contains detailed data on socio-economic status, extensive measures Table 1. Characteristics of mandatory health insurance programme in Indonesia Characteristics Mandatory health insurance Askes Jamsostek Organization Para-state agency (PT Askes Indonesia) Para-state agency (PT Jamsostek Indonesia) Beneficiaries Civil servants, pensioners of civil servants and armed Private sector employees forces Eligible dependents Spouse and 2 oldest children <21 years of age (if Spouse and 3 oldest children <21 years of age unemployed, unmarried), or <25 years of age if a full-time student Premium rate 2% payroll deduction (regardless of marital status) 3% payroll deduction for single; 6% payroll deduction for married Premium policy Full contributory (100% paid by employee) Full non-contributory (100% paid by employer) Benefits and provider Outpatient and inpatient care at public providers only Outpatient care at both public and private providers networks Inpatient care at public providers only

3 324 Budi Hidayat et al. of health care utilization, as well as health status measures. Household heads were questioned on health insurance. Statistical methods We used a multinomial logit (MNL) model to identify the effects of insurance on access to public and private outpatient care. In order to measure the equity effects of insurance, we used the concentration index. These techniques are described in detail below. Measuring access to care Access to health care is often measured by the likelihood of using a health care provider (Waters 2000; Yip and Berman 2001). In this study, we measured access by the probability that an individual uses outpatient care (henceforth OPC). We focused on access by individuals who reported being sick during a 4-week recall period of the survey. In Indonesia, when people are ill they have several possibilities to seek care. We considered self-treatment, care from public providers and from private providers as three choices. Since not all mandatory insurance schemes offer both public and private health care providers, we disaggregated OPC into public and private, rather than just focusing on the total OPC use. We therefore model the individual s decision to seek care as a multinomial logit (MNL) problem as follows (Greene 1997): β j e x i Pr( Yi = j) = 2, for j = 0,1 or 2 β ' kxi e k= 0 ' where Y i is a random variable indicating the provider types. This can take a value j = 0, 1 or 2, which represent self-treatment, public and private OPC, respectively. The vector x i represents a set of exogenous variables and β represents regression parameters to be estimated. The estimated equations above provide a set of probabilities for the j + 1 choices for an individual with characteristics x i. Equation (1) was estimated using the maximum likelihood procedure. It is important to note that the MNL model assumes independence of odds ratios of the different alternatives, thus the model requires that the assumption of independence of irrelevant alternatives (IIA) be satisfied (Greene 1997). In order to validate this assumption, we employed the Hausman specification test as well as the Small-Hsiao IIA test. It may be expected that insurance is endogenous. That is, the choice of health insurance could be influenced by many of the same factors that influence health care use (Vera- Hernandez 1999; Waters 1999; Trujillo 2003). 1 Since the insurance status used in this study is mandated related to employment status, we assume that the insurance variable is exogenous. To ensure that health insurance is indeed exogenous, we proceeded as follows. First, we included several (1) covariates in the model and restricted our analyses to the working age group samples (19 60 years of age). This allowed us to minimize any possible endogeneity of insurance. Second, we tested for possible endogeneity of insurance. Following Waters (1999), we first estimated a reduced form of insurance participation using a probit model including all covariates in the health care equation in addition to proposed identifying variables. The predicted and observed values of the insurance variables were then included in the OPC use equation. If the predicted coefficient for insurance participation is not significant, one can assume that insurance is in fact an exogenous variable. 2 Health status is potentially an endogenous variable due to sample selection because we intended to analyze OPC use conditional on an ill sample (Heckman 1979; Dow 1996; Akin et al. 1998). To investigate whether conditional estimates suffer from selection bias, we employed a probit model with sample selection described by van de Ven and van Pragg (1981) and found the correlation between error terms insignificant (Chi 2( 1) = 0.02; p-value = ), ruling out any possibility of sample selection bias. Further, we conducted the analysis based on the ill sample (conditional estimates) only, as well as on the entire sample (unconditional estimates), and also found no difference in the coefficients. The ill sample was derived by grouping together all individuals who reported at least one symptom, one Activity of Daily Living (ADL) impairment, or poor assessment of general health status (GHS). Details about symptom, ADL and GHS measures are given in the variable definition section. To ascertain the pure effects of health insurance on OPC use and to show the magnitude of the effects implied by the coefficients, we predicted the probabilities of using public and private OPC based on the MNL estimations by changing only health insurance status while holding all other variables at their mean. The following scenarios were used to change the value of health insurance status: 3 Scenario 1: assigning all individuals in the sample to uninsured ; Scenario 2: insurance variable (Askes and Jamsostek) at the observed level; Scenario 3: expansion of Askes to all individuals in the sample; Scenario 4: expansion of Jamsostek to all individuals in the sample. In this way, it would be assured that the distribution of the covariates and the sample size for each scenario is the same, and the differences in the probabilities predicted under different scenarios are indeed due to difference in insurance status. Measuring equity in access The conceptual literature on equity in health care distinguishes between horizontal and vertical equity. Horizontal equity assumes that individuals in equal need (in terms of illness) should be treated alike (in terms of utilization) irrespective of income (Wagstaff and van Doorslaer 2000), while

4 Health insurance and equity 325 vertical equity is the unequal, but equitable, treatment of individuals in unequal levels of need (Mooney 1996). However, applied research on equity in health care often concentrates attention on horizontal equity rather than vertical equity (Waters 2000; van Doorslaer et al. 2002). This study used equality of access as an operational definition of equity. The definition used here is in line with the concept of horizontal equity (Waters 2000). To measure equity, we used the concentration index (henceforth C M ). 4 The C M is the extension of the Gini coefficient, the most widely used measure in the economic literature on inequalities. The C M can take values between 1 to +1. If there were no variation in access among income groups, the index would be zero. A positive (negative) score of the C M indicates inequity in access, favouring the richest (poorest) groups. The concentration index was calculated using the formula given by van Doorslaer and Jones (2003) as follows: 2σ 2 y i = α + γri+ ui R µ where R i is the relative fractional rank of the ith individual, N 1 2 σ = R i is the variance of R i and N is the R N i= 1 2 sample size. The R i was derived first by ranking ranking all individuals in the samples according to the consumption per capita and then dividing the rank by total number of observations. The estimate of γ is equal to the concentration index, C M. We calculated the C M for both public and private OPC use based on the actual data and the prediction results of the MNL estimation under different policy scenarios, as described above. Equation (2) needs to be estimated in order to get γ. This could be done with ordinary least squares (OLS) regression. However, as we expect heteroskedasticity due to inclusion of only one regressor (R i ) in the model, as well as possible intrahousehold correlation among individual observations, a correction in simple OLS regression will be required. This could be done by using Newey-West regression that produces Newey-West standard errors for coefficients estimated by OLS regression (Newey and West 1994). Note that the coefficients estimated by Newey-West would be the same as OLS, but since Newey-West assumes that the error structure is heteroskedastic and possibly autocorrelated up to some lag, standard errors are corrected for these problems. All estimation was carried out using STATA version 7 (StataCorp 2001). To illustrate inequity empirically, we also drew the concentration curve which shows the cumulative proportion of OPC use against the cumulative proportion of the sample population ranked by income (Wagstaff and van Doorslaer 2000). If everyone, irrespective of the income level, has exactly the same value of OPC use, the concentration curve will coincide with the diagonal (equality line). If the concentration curve lies above (below) the equality line, it indicates inequity in (2) access favouring individuals in the poorest (richest) groups. The further the concentration curve lies from the equality line, the greater the degree of inequity in OPC use across income groups. To see empirically the effects of insurance on equity, we drew the concentration curve for the insured and uninsured on the same graph. Independent variables Health insurance This study includes two types of mandatory health insurance schemes: Askes and Jamsostek. The survey asked the head of the household if s/he was enrolled in one of these insurance schemes. Even if the answer was yes, this does not imply that all individuals in the household were insured. This is because the eligible dependents covered by Askes, for example, are spouse and two children under 21 years of age or under 25 years if child is a full-time student. In order to obtain a valid coverage, we therefore adjusted the eligible dependents based on the policy for the respective scheme (see Table 1), using other information such as employment status, individual relationship with the household head, individual s age and education. The insurance types were included in the model as dummy variables. Further, in order to capture the effects of insurance status on equity, we included the interaction of insurance with per capita household consumption in the model. Health The use of health services, conditional upon need, has for some time been accepted as a practical indicator of equity of access in a health system, measurable in the context of a household survey (Makinen et al. 2000; Waters 2000). In this paper, the need for health care was derived based on individuals responses to the health-related questions. We used three measures of self-reported illness with a recall period of 4 weeks: morbidity (symptom), activity of daily living (ADL) and self-rated general health status (GHS). The questions on self-reported morbidity data were based on illness symptoms experienced and their duration. Adult respondents were asked whether they had experienced any of 17 common illness symptoms, while children s respondents (carer or mother) were asked to report whether the children had experienced any of 19 common symptoms. This study measured the presence of an illness symptom by a dummy for whether the individual reported any of the asked symptoms. The questions on GHS were based on self-reported ordinal scales: very good, good, bad and very bad. Due to the low number of individuals reporting very bad GHS, observations from this category were aggregated with those reporting bad into one category labelled poor. Thus, the GHS consists of three groups (e.g. very good, good and poor), included in the model as dummy variables with very good GHS treated as the reference. The ADL measures were based on individuals self-ratings of

5 326 Budi Hidayat et al. ability to engage in specific activities, collected from each individual above 5 years of age. Interviewees were asked, whether they had difficulty in: (1) carrying a heavy object 20 metres, (2) climbing stars, (3) walking, (4) bending, kneeling or stooping, (5) drawing water from a well, (6) dressing without assistance, (7) rising from a sitting position in a chair, (8) toileting, and (9) rising from a sitting position on the floor. These ADL questions have been appropriately modified to fit into the local Indonesian cultural context. These measures have also been tested extensively for reliability (consistency between tests and interviewers) and validity (consistency between individual assessments of different skills). All the three health measures were included in the model as covariates. In addition, a dummy indicating whether individuals had a serious illness in the last 4 years was also included. The severity of the disease was self-reported. Income Income, which reflects people s ability to pay, is considered an important determinant of health care use. Since information related to income is subject to bias and is difficult to assess, particularly in households of subsistence farmers, we used household consumption as a proxy to measure income; the validity of this approximation is not in question (Gertler et al. 1987; Hjortsberg 2003; Trujillo 2003). To correct for price differences in various locations, we adjusted the consumption with the consumer price index data in 1997, with Jakarta serving as reference. To control for the effect of household size, we used per capita household consumption. However, as household size can have effects on the demand for health care and on the equity index, we also included household size as a predictor variable. Other covariates Additional covariates included socio-demographic variables such as Female (1/0), Married (1/0), Education (dummies indicating No school (reference), Elementary, Junior, Senior and High), Electricity (1/0), age (years), travel cost (Rp), and travel time (minutes). We also included residence variables such as urban (1/0) and the eight regions 5 of the survey site (dummies indicating: Jakarta (reference), West Java, Central Java, Bali & West Nusa Tenggara, Kalimantan and Sulawesi). Results Descriptive sample characteristics Table 2 gives the distribution of mandatory health insurance coverage (left column) and OPC use (right column). The proportion of the individuals covered by Askes was 9%, compared with only 4.1% covered by Jamsostek. The distribution of the insured by per capita consumption quintile was clearly concentrated among the rich (p < 1%). The proportion of the insured ranged from 1.5% and 0.9% of those in the lowest quintile to 21.2% and 8.9% of those in the highest quintile for Askes and Jamsostek, respectively. The overall utilization rate for public OPC was 15.42%, compared with only 7.11% for private. Significant differences abound between the quintile groups for private OPC use (p < 1%). The utilization of private OPC had a clear trend in favour of the wealthier. Only 2.6% of individuals in the poorest quintile who reported being sick used private OPC, whilst for the richest quintile the figure reached 14.6%. That is, the rate of private OPC use for the richest quintile is 5.7 times larger than for the poorest one. For public OPC use, this ratio was only To examine whether health insurance increases OPC utilization due to its ability to reduce the financial barriers to access, Table 3 reports average total health expenditures (aggregating cost of provider treatment and medicines) on public and private OPC by insurance types. Individuals covered by Askes and Jamsostek spent on average 55% and 25%, respectively, less than uninsured people for public OPC. For private OPC, Jamsostek beneficiaries spent 27% less than uninsured people, while Askes members spent on average 90% more than uninsured. The table also indicates Table 2. Distribution of mandatory health insurance coverage and outpatient care use (overall and by per capita consumption quintiles) Mandatory health insurance Outpatient care use Askes (se) Jamsostek (se) Public (se) Private (se) Overall 9.04% (0.67) 4.09% (0.53) 15.42% (0.73) 7.11% (0.54) By per capita consumption quintiles 1 st (n = 3297) 1.50 (0.35) 0.85 (0.49) (1.36) 2.57 (0.59) 2 nd (n = 3298) 3.54 (0.63) 1.99 (0.48) (1.60) 3.83 (0.72) 3 rd (n = 3298) 9.62 (1.38) 4.60 (0.92) (1.17) 9.49 (1.46) 4 th (n = 3296) (1.48) 5.64 (1.00) (1.44) 7.19 (0.97) 5 th (n = 3296) (1.65) 8.88 (1.74) (1.40) (1.64) Ratio (Q5/Q1) F (4370) a 44.23*** 9.81*** *** Notes: Total sample (n) = ; a F(4370) adjusted Wald test for testing the null hypothesis of equality between per capita consumption quintiles; (se) = standard errors; *** significance at the 1% level.

6 Health insurance and equity 327 that overall health expenditure on public OPC was substantially lower than that on private OPC. Model estimates Table 4 presents the estimates of the multinomial logit (MNL) models based on the ill sample (conditional estimates). 6 The last row of the table shows that the model passes the IIA assumption based on both Hausman and Small- Hsiao specification tests. The tests for potential endogeneity of the health insurance variable confirm that the insurance variable is in fact exogenous, and hence no corrections for endogeneity of insurance are needed. The predicted values of the insurance variable 7 when inserted in the OPC use equation were insignificantly different from zero. That is, the coefficient estimates of the predicted value for Askes were (p = 0.666) in public and (p = 0.523) in private use. For Jamsostek, the coefficient estimates of the predicted value were (p = 0.179) in public and (p = 0.461) in private use. Using an instrumental variable approach, the exogeneity test also could not reject the null hypothesis in all cases (see Appendix A). To investigate whether the instrumental variables selected are appropriate, we employed the following tests. First, we tested whether the instruments have no impact on the suspected endogenous variable using the F-test. The joint test that the coefficients of the instrumental variables in the insurance participation equation are equal to zero was rejected in all cases at the p < The F-test statistics for Askes and Jamsostek were and 79.48, respectively. 8 Second, we tested the assumption that the instruments are uncorrelated with the error term of the OPC model, using a Sargan s statistic test of the overidentifying restrictions (Baum et al. 2003). We could not reject this hypothesis in both public OPC (p = 0.36) and private OPC (p = 0.11) models. As the conditional estimates are in line with the concept of measuring horizontal equity (Waters 2000), and there was no evidence of sample selection bias, we opted to use the results of conditional estimation. The most important result of the MNL estimates shows that the estimated coefficient for Askes insurance was positive in both public and private OPC use, but only significant in public OPC. For Jamsostek insurance, the estimated coefficient was significantly positive in both public and private OPC use. The nature and magnitude of the interaction coefficients are the key parameters in evaluating the effects of insurance on equity. The negative estimate for the coefficient of the interaction term indicates that the effect of insurance on the probability of using OPC is smaller for wealthier people. Results of most covariates were consistent with expectations and as such will not be discussed here. However, in both submodels of public to self-care and private to self-care, one way travel time and cost had positive influences on public and private OPC use. One possible reason for this counterintuitive result is that in the data on the self-care group (the base category), the values of travel time and cost were zero (no variation), whilst the values of one way travel time and cost of public and private OPC were always positive. Simulating the effect of insurance on access The results indicate that having health insurance increased the use of OPC, but since we used a MNL model, the coefficients in Table 4 do not readily tell us much about the magnitude of the effect of each variable. This section evaluates the effects of simulating insurance participation on OPC use based on the model estimations. We estimated the magnitude of the effect by performing a within-sample forecast of OPC use for alternative restrictions on the values of insurance variables (Askes and Jamsostek). Table 5 indicates that moving all individuals into the uninsured group (Scenario 1) reduced the probability of using public OPC from the predicted mean of (observed level) to This simulation also slightly reduced private OPC use from to Taking the enrolment of Askes from the observed level to a hypothetical 100% (Scenario 3) resulted in a large increase in public OPC use from to (62.6% increase), but reduced private OPC use by about 2%. Scenario (4) would lead to a modest increase (24.7%) in public OPC use and substantially increased (160.9%) private OPC use. The predicted probabilities of using OPC by consumption quintiles under different policy scenarios are presented in Table 6. As compared with the uninsured, the first panel of the table shows that expanding Askes (Scenario 3) led to a higher probability of using public OPC across all the consumption quintile groups. This pattern was also observed for expanding Jamsostek (Scenario 4). Table 3. Average health expenditure on public and private outpatient care in the last month by insurance types Health expenditure (Rp) Public OP (se) Private OP (se) Overall (812) (2 314) By insurance type Uninsured (968) (1 848) Askes (650) (13 511) Jamsostek (1 877) (4 287) Note: US$1 = Rp.8500; (se) = standard errors.

7 328 Budi Hidayat et al. Table 4. Multinomial logit (MNL) a estimations (conditional on ill sample, n = ) Public OPC use Private OPC use b (se) c b (se) c Askes 0.661*** (0.102) (0.143) Askes per capita consumption 0.071* (0.041) (0.050) Jamsostek 0.502* (0.276) 1.322*** (0.190) Jamsostek per capita consumption 0.758*** (0.244) 0.375*** (0.113) Whether has any symptoms in last 4wks 1.577*** (0.207) 3.410*** (0.716) Whether has any ADLs limitations 0.250*** (0.060) 0.402*** (0.080) General health status (GHS): Very good d GHS is good 0.355*** (0.117) 0.446*** (0.150) GHS is poor 1.389*** (0.128) 1.682*** (0.165) Whether has serious illness in last 4yrs 0.479*** (0.073) 0.836*** (0.085) Whether female 0.582*** (0.057) 0.257*** (0.075) Number household members (0.011) 0.048*** (0.013) Whether married 0.644*** (0.101) 0.215** (0.103) Schooling: No school d Elementary school (0.081) 0.391*** (0.142) Junior school (0.109) 0.480*** (0.169) Senior school (0.109) 0.549*** (0.165) High school 0.317** (0.152) 0.745*** (0.187) Age (years) (0.003) (0.004) Log per capita consumption (0.040) 0.412*** (0.051) Whether household uses electricity 0.500*** (0.084) 1.147*** (0.198) One way travel cost to health facilities (Rp) 0.005* (0.003) (0.004) One way travel time to health facilities (minutes) 0.026*** (0.009) 0.026** (0.012) Whether urban resident 0.383*** (0.061) 0.172** (0.085) Region : Jakarta d Sumatra 0.317*** (0.120) 0.255** (0.128) West Java 0.308*** (0.119) (0.118) Central Java 0.233* (0.122) (0.124) East Java 0.527*** (0.131) 0.612*** (0.138) Bali & WNT 0.836*** (0.128) 0.300** (0.146) Kalimantan 0.666*** (0.150) 0.916*** (0.256) Sulawesi 0.553*** (0.153) 0.588** (0.241) Constant 6.608*** (0.541) *** (1.015) Observations Pseudo R 2 = ; Wald x 2 (58) = , sign level Hausman tests of IIA assumption: x 2 (30) = (omitted public), sign level 1.000; x 2 (30) = (omitted private), sign level Small-Hsiao tests of IIA assumption: x 2 (30) = (omitted public), sign level 0.310; x 2 (30) = (omitted private), sign level a The comparison group are those who choose self-treatment; b The estimated parameters (βs) and asterisks indicate significance at the 1% (***), 5% (**) and 10% (*) level; c Robust standard errors in parentheses; d Omitted groups. OPC = outpatient care. Table 5. Predicted probability of outpatient care use under different policy scenarios Scenario Public outpatient care use Private outpatient care use (1) No insurance (0.0008) (0.0006) (2) Insurance at observed level (0.0008) (0.0007) (3) Expanding Askes (0.0011) (0.0006) (4) Expanding Jamsostek (0.0008) (0.0012) Note: Sample size (n) = ; standard errors in parenthesis. It is clear that the distribution of private OPC use for all scenarios has a clear trend in favour of the wealthier (Table 6, second panel). In all consumption quintile groups, the insured, except for Askes, had a higher probability of using private OPC than the uninsured. For individuals in the lowest quintile, expanding Jamsostek increased the probability of using private OPC from to (190% increase). A sharp increase was also observed in other quintile groups, but

8 Health insurance and equity 329 Table 6. Predicted probability of outpatient care use ( 100) under different policy scenarios by per capita consumption quintiles Scenario Per capita consumption quintiles Ratio (Q5/Q1) 1st 2nd 3rd 4th 5th Public outpatient care use (1) No insurance (2) Insurance at observed level (3) Expanding Askes (4) Expanding Jamsostek Private outpatient care use (1) No insurance (2) Insurance at observed level (3) Expanding Askes (4) Expanding Jamsostek to a lesser extent. The differential effects of insurance across quintile groups reduced the accessibility gap to private OPC between the highest and lowest quintile individuals who were covered by insurance. For the uninsured (Scenario 1), the ratio was 3.21, but decreased to a value as low as 2.68 for Scenario 4. However, this ratio does not imply that insurance schemes reduced inequity. We defer a discussion of whether expanding insurance improves equity until the next section where we present the concentration indices. Note that the estimated coefficient of the interaction term between Jamsostek and per capita consumption was significantly negative. Empirically this corresponds with Figure 1 showing that the greatest effects of Jamsostek on both public and private OPC use were found in the poor, and the effects decline as the quintile level increases. Simulating the effect of insurance on equity The estimated concentration index (C M ) and the t-statistics are presented in Table 7 for both public (left column) and private (right column) OPC use. Based on the actual data, the significantly negative C M in public OPC use indicates that poor people used public OPC more often (pro-poor inequity) than the rich. The picture is quite different from that of private OPC use. The C M in private OPC use was significantly positive, indicating that the utilization of private OPC is concentrated in the rich (pro-rich inequity). Figure 2 depicts the concentration curves for both public and private OPC use derived from the actual data. It shows that the concentration curve for private OPC use lies far below the equality line, while for public OPC use it lies very close to and slightly above the equality line. Public OP use Private OP use 210 OPC use (% increase compared to the uninsured) st (lowest) 2nd 3rd 4th 5th (highest) Per capita consumption quintiles Figure 1. The effects of Jamsostek insurance on public and private outpatient care use

9 330 Budi Hidayat et al. 100 Line of equality Public OP use Private OP use Cumulative % outpatient care use Cumulative % samples, ranked by per capita consumption Figure 2. Concentration curves of public and private outpatient care use Based on our simulations, health insurance programmes have a negative impact on equity. Table 7 shows that the magnitude of a pro-poor inequity (in absolute value) in public OPC use for the insured was higher than for uninsured people (Scenario 1). In this case, the members that are using public OPC are disproportionately from the poorest groups, hence insurance increases inequity that favours the poorest. Empirically this finding corresponds with the concentration curve illustrated in Figure 3, showing that the curve for the uninsured is closer to the equality line than that of the insured. For private OPC use, the magnitude of pro-rich inequity for the insured was also higher than for the uninsured. For further clarity, this is also presented using the concentration curves in Figure 4. Discussion The ability of this study to disaggregate private and public utilization is a strong point in comparison with previous related studies (Waters 2000; Yip and Berman 2001). As the provider networks are different for different insurance schemes in Indonesia (e.g. Jamsostek offers both public and private provider networks while Askes offers only public), Table 7. Concentration indices of public and private outpatient care use Public outpatient care use Private outpatient care use C M (se) t-stat C M (se) t-stat Actual mean of dependent variable (0.0102) 2.01** (0.0082) 12.53*** Predicted mean of (scenario): (1) No insurance (0.0028) 13.83*** (0.0023) 40.09*** (2) Insurance at observed level (0.0030) 6.73*** (0.0025) 42.56*** (3) Expanding Askes (0.0039) 15.42*** (0.0022) 42.24*** (4) Expanding Jamsostek (0.0030) 24.98*** (0.0041) 50.03*** Note: Standard error in parenthesis; *** significant at the 1% level, ** significant at the 5% level.

10 Health insurance and equity 331 Line of equality Uninsured (Scenario 1) Insured (Scenario 4) 100 Cumulative % public OPC use Cumulative % samples, ranked by per capita consumption Figure 3. Concentration curves of public outpatient care use for the insured and the uninsured Line of equality Uninsured (Scenario 1) Insured (Scenario 4) 100 Cumulative % private OPC use Cumulative % samples, ranked by per capita consumption Figure 4. Concentration curves of private outpatient care use for the insured and the uninsured

11 332 Budi Hidayat et al. not being able to disaggregate the choice of provider would give misleading results. Further, this study not only uses the richest-poorest ratio but also uses the concentration index in examining equity in health care. Focusing only on the richest/poorest ratio as a measure of equity may lead to overestimation, and could result in wrong conclusions. For instance, if we look at the distribution of private OPC use across per capita consumption quintiles, expanding Jamsostek insurance reduced differentials in visits rates between the highest and lowest quintiles, suggesting that insurance helps to reduce the inequity in access (Yip and Berman 2001), e.g. reduced the richest/poorest ratio. However, using the C M measure, which takes into account the visit rates of the entire population, the equity effect was not detected. The results of this study confirm that health insurance programmes had a strongly positive impact on access to OPC, a finding consistent with previously published works (e.g. Kreider and Nicholson 1997; Vera-Hernandez 1999; Waters 1999; Trujillo 2003). 9 Given that Askes insurance offers access to only public providers, it is not surprising that this scheme had no significant impact on the use of public OPC. On the other hand, the Jamsostek insurance programme had a positive and significant impact on the use of both public and private OPC. It is also important to note that the effects of Jamsostek on OPC use in both public and private sectors were significantly higher among the poor. The simulations indicate that expanding Jamsostek increased the utilization of both public and private OPC use, with the increase more pronounced among individuals in the lowest quintile. One possible reason for this finding is that the poor have a higher price elasticity of demand, hence the reduction in the effective price of health services due to insurance coverage increases utilization to a greater extent among poorer than among richer individuals. This finding is consistent with empirical evidence from Burkina Faso and Egypt. In Burkina Faso, Sauerborn et al. (1994) found that although demand for health care appeared inelastic ( 0.79) overall, disaggregated analyses by income quartiles revealed that elasticities were substantially higher in the first quartile ( 1.44) than in the fourth ( 0.12). In Egypt, Yip and Berman (2001) reported that children in poor families who were covered by the School Health Insurance Program (SHIP) had the greatest advantage in terms of access to health care. This study also reveals that income-related inequity in access to OPC exists in Indonesia. The distribution of OPC use showed inequity favouring the rich for private providers and inequity favouring the poor for public ones. Higher public OPC use among the poor may be particularly the case with regard to the extensive subsidization of medical care by the government, and hence prices in public health facilities are generally low. Gertler and Gruber (2002) reported that household spending on medical care in Indonesia (not overall spending) was quite low, averaging less than 1% of non-medical consumption. As the estimate was derived from the product of reported health care use and prices, the authors concluded that low medical care spending reflected both low levels of utilization and the extensive subsidization of medical care costs. With regard to general knowledge on the ability of health insurance schemes to reduce inequity, this study suggests they have a negative impact on equity. Our simulation indicates that expanding Askes to the whole population increased inequitable (favouring the poor) use of public OPC, whilst expanding Jamsostek increased inequitable (favouring the rich) use of private OPC. This is in line with the study in Ecuador by Waters (2000), who found that expanding insurance to all eligible members under the General Health Insurance (GHI) programme had a negative impact on equity. However, unlike our study, Waters was not able to distinguish between the use of public and private health care facilities among the insured. Policy implications Expanding health insurance programmes can increase the utilization of medical care. Although increasing health care use among insured people is typically viewed as a moral hazard effect, 10 this effect may be desirable from a public health point of view, given that insurance improves utilization among poorer individuals effectively. At the same time, poorer people in low-income countries typically have a larger unmet need for health care services. Hence, it seems plausible that providing the population with health insurance would increase access to modern health care effectively, and it might have a positive impact on health outcomes. It has often been argued that provision of insurance could lead to better health outcomes in developing countries (Sachs 2001; Pokhrel and Sauerborn 2003). How health insurance can really improve health status is an important topic for future research. However, a great challenge lies in how insurance could be expanded so that it reaches the poor. In Indonesia, mandatory health insurance schemes cover only civil servants and a limited group of private sector employees. We have found that the distribution of insurance was concentrated among the rich even in these schemes. Although the Jamsostek scheme is a mandatory insurance for private sector employees, the coverage of this scheme was only 4%. One possible option to enlarge membership among private sector employees is to remove the conditional mandatory provision of Jamsostek, which contributes a very low enrolment (Thabrany 2001). For the poor, the introduction of welltargeted subsidies to join an insurance scheme is worth considering. There are two reasons for this. First, supply-side subsidies in Indonesia have not been shown to be an effective way to reach the poor. The majority of health subsidies are received by the non-poor (World Bank 1993). 11 Secondly, given that Indonesia might look for external support in expanding insurance for the poor, external resources could be better directed to distributional actions, such as encouraging people to buy insurance at subsidized rates, rather than backing undertakings from the supply side (Pokhrel and Sauerborn 2003).

12 Health insurance and equity 333 Conclusion This study has analyzed the effects of mandatory health insurance for government employees (Askes) and private employees (Jamsostek) on access and equity in access to outpatient care. It has shown that health insurance had a positive impact on access to outpatient care. While the distribution of both insurance schemes was concentrated among the rich, the largest effects of insurance (particularly Jamsostek) were observed among individuals in the lowest income groups. There was significant inequity in access to both public (pro-poor) and private (pro-rich) outpatient care, and the magnitude of this was smaller for public than for private care. However, neither Askes nor Jamsostek had a positive effect on equity in access. An important policy implication of this study is that expanding health insurance would increase access to health care. To enlarge membership, repealing the conditional mandatory provision of Jamsostek and introducing a demand-side subsidy for the poor to take up insurance should be considered. Endnotes 1 Previous studies on the effects of health insurance on health care demand have addressed the endogeneity problem of insurance choice, which depends largely on the availability of the identifying variables as well as the nature of dependent variable used to measure the demand. Waters (1999), for example, has used and developed the bivariate probit estimator to correct endogeneity of regressors resulting from selection bias in the context of discrete binary choice. For a count data model, the Generalized Method of Moment (GMM) estimator has been used by Vera-Hernandez (1999) to address endogeneity of insurance choice. 2 Additionally, we employed the augmented Hausman specification test using Instrumental Variable (IV) estimation. Grogger (1990) discussed testing for exogeneity of the regressors and proposed the use of a Hausman specification test after estimation of the model by IV. Baum et al. (2003) discussed how to implement an exogeneity test using a Hausman specification test in the IV estimation. The consistency of this estimation, however, depends largely on the availability of instruments. If the instruments are only weakly related to the endogenous variable replaced by the instruments, the resulting parameter estimates will be biased, even if the instruments are not correlated with the error term of the health care equation (Bound et al. 1995; Staiger and Stock 1997). Therefore, we tested the relevancy of the instruments using an F-test of the joint significance of the instruments in the reduced form of insurance participation estimates (Bound et al. 1995). The validity of the instruments was tested using a test for overidentification restrictions (Windmeijer and Santos-Silva 1997). 3 We are grateful to an anonymous referee for this point. 4 We also present the richest/poorest quintile ratio, a simple and useful equity measure. However, this measure could hide important features outside the extreme, since it is not sensitive to the values of the middle three quintiles (van Doorslaer et al. 2002). 5 In this study, we grouped the 13 IFLS sampling provinces into eight regions: (1) Jakarta; (2) Sumatra (West Sumatra, North Sumatra and South Sumatra); (3) West Java; (4) Central Java (Jogjakarta and Central Java); (5) East Java; (6) Bali and West Nusa Tenggara; (7) Kalimantan; and (8) Sulawesi. Jakarta is the capital and the most developed city in Indonesia, and characteristics very specific to Jakarta preclude its being classified into other regions. Thus, although located in the island of West Java, Jakarta was not grouped into West Java. 6 For comparison purposes, we estimated identical models based on the entire sample (unconditional estimates). The findings confirm that both unconditional and conditional estimates yield similar results with respect to the direction and significant level of the insurance variables as well as some other covariates. The only difference was that the covariates in the unconditional estimates explained 14% variation in the models, compared with only 11% in the conditional ones. The results of the unconditional estimates are available on request from the authors. 7 These were generated from a probit model of insurance participation, estimated separately for Askes and Jamsostek, using the following identifying variables: (1) employment status of the household head (whether government employee, whether private employee), (2) relationship to household head (whether spouse), (3) whether individual involved in community meeting activity, (4) whether individual involved in water organization activity. These variables have been selected as the appropriate instrumental variable since they turned out to be insignificant in the OPC use but were highly correlated to the insurance participation. The probit R 2 values for the insurance equation were for Askes and for Jamsostek. 8 As a conservative rule of thumb, an F-statistic below 10 is an indicator of a weak instrument (Bound et al. 1995). 9 In the United States, Kreider and Nicholson (1997) concluded that providing homeless people with health insurance would significantly increase their demand for health care. After correcting for selection bias, the General Health Insurance (GHI) programme in Ecuador was found to have a strongly positive association with the use of curative health care services more than twice (Waters 1999). In Catalonia, duplicate insurance coverage status significantly increased the number of visits to specialists by about 27% (Vera-Hernandez 1999). Likewise, Trujillo (2003) concluded that participation in the Colombian social health insurance scheme increased health care use. 10 Moral hazard refers to the tendency for insured individuals to increase their consumption of health services, and has been subject to considerable theoretical and empirical work (Zweifel and Manning 2000). However, Vera-Hernandez (1999) point out that increasing health care use among the insured is not always correctly labelled as a moral hazard effect. 11 The World Bank (1993) reported that in Indonesia in 1990, the richest utilized hospitals, which receive the largest per capita subsidy, and the poorest used health centres and subcentres, which receive much lower per capita subsidies. In addition, only 12% of the government health care subsidy was consumed by the poorest households, while the wealthiest consumed 29% of the government subsidy in the health sector. References Akin JS, Guilkey DK, Hutchinson PL, McIntosh MT Prices elasticities of demand for curative health care with control for sample selectivity on endogenous illness: an analysis for Sri Lanka. Health Economics 7: Akin JS, Hutchinson P Health-care facility choice and the phenomenon of bypassing. Health Policy and Planning 14: Baum CF, Schaffer ME, Stillman S Instrumental Variables and GMM: Estimation and testing. Working Paper No Boston, MA: Boston College. Bound J, Jaeger DA, Baker R Problems with instrumental variables estimation when the correlation between the instruments and the endogeneous explanatory variable is weak. Journal of the American Statistical Association 90: Brotowasisto, Gish O, Malik R, Sudharto P Health care financing in Indonesia. Health Policy and Planning 3: Chiappori PA, Durand F, Geoffard PY Moral hazard and the demand for physician services: first lessons from a French natural experiment. European Economics Review 42: Dow WH Unconditional demand for health care in Cote d Ivoire: does selection on health status matter? LSMS Working Paper No Washington DC: The World Bank.