The Impact of Price Subsidies on Child Health Care Use: Evaluation of the Indonesian Healthcard

Financial barriers to seeking care are frequently cited as one of the main causes of underutilization of child health care services. This paper estimates the impact of Indonesia's healthcard on health care use by children. Evaluation of the healthcard effect is complicated by the fact that card allocation was non-random. The analysis uses propensity score matching to control for systematic differences between treatment and control groups. A second potential source of bias is related to contemporaneous, exogenous influences onhealth care use unrelated to the healthcard itself. Using panel data collected prior to and after the introduction of the healthcard, a difference-in-differences estimator is constructed to eliminate the effects of exogenous changes over time. The author finds that although health care use declined for all children during the crisis years of 1997-2000, use of public sector outpatient services declined much less for children with healthcards. The protective effect of the healthcard on public sector use was concentrated among children aged 0-5 years. The healthcard had no significant impact on use of private sector services. The results highlight the need to provide adequate protection against the financial burden of health care costs, particularly during economic crises.


Introduction
Demand-side subsidies are considered an effective means of tackling financial barriers to access and expanding the demand for health services, especially among the poorest and most vulnerable groups in the population [1,2]. Yet, little empirical evidence exists on how utilization of child health care services responds to price subsidies, particularly in the demand constrained settings that exist in developing countries. An important underlying question is whether reducing financial barriers to care is indeed the magic pill needed to expand the demand for child health services, or whether physical barriers to access and a lack of information and knowledge about services would continue to limit the expansion of demand, despite the availability of price subsidies. This paper examines the impact of the Indonesian healthcard on health care use by children. The healthcard was the key component of a targeted price subsidy program in the public health sector in Indonesia. It was introduced in a context where, not only was utilization of child health services already low by regional standards, but utilization had been depressed as a consequence of the economic crisis that affected South-East Asia in the 1990s. This study goes beyond a previous evaluation of the Indonesian healthcard by Pradhan et al [3] in several ways. It is concerned with the impact of the healthcard on children's use of services, whereas the earlier study examined outpatient care use for the population as a whole, making no distinction between adults and children. In addition to evaluating outpatient care utilization, this study looks at the effects of the healthcard on self-medication among children which has not been considered previously. Finally, 2 Pradhan et al's analysis reflects the experiences of the first five months of the healthcard program while data used for this paper reflect a longer time interval of 2-3 years.
Healthcard ownership nearly doubled after the first year of program implementation, making it quite critical to evaluate its effects beyond the initial time period. I find that, although health care use declined for all children during the crisis years of 1997-2000, the use of public sector outpatient services declined less for children with healthcards. The protective effect of the healthcard on public sector use was concentrated among children aged 0-5. These findings are consistent with those of Pradhan et al's results for the general population. Contrary to Pradhan et al, I find that the healthcard had no significant impact on use of private sector services. This suggests that with respect to children at least, the healthcard program was not associated with any substitution from private to public health care. My findings also suggest that there was less recourse to selfmedication among healthcard owning children, although these results are not statistically significant.
Evaluation of the healthcard effect is complicated by the fact that the allocation of the healthcards to households was non-random. I use propensity score matching to control for systematic differences between treatment and control groups that could potentially bias the estimation of the program effect. A key advantage of propensity score matching over regression methods is that it does not require a parametric model of health care use involving assumptions about functional forms and error distributions. A second potential 3 source of bias is related to contemporaneous, exogenous influences on health care use unrelated to the healthcard. The availability of panel data collected just prior to the introduction of the healthcard in 1997 and again in 2000 makes it possible to use a difference-in-differences estimator. The combination of propensity score matching with difference-in-differences eliminates selection bias as well as bias caused by timeinvariant unobservables from the estimated healthcard effect. This represents another source of deviation from Pradhan et al's study, which used a post-intervention treatmentcontrol group design. They evaluated the healthcard impact as the difference between utilization outcomes of healthcard owners and observationally similar non-healthcard owners in 1999. I find that failure to control for these sources of bias significantly underestimates the program effect.
I begin by describing health care utilization trends in the aftermath of the economic crisis and provide an overview of the Indonesian healthcard program. This is followed by a review of the literature. I then lay out the hypotheses and analytical framework, explain the empirical strategy and describe the data. Thereafter, I present and discuss the results.

Background
Low levels of use of modern health services in Indonesia are largely attributed to financial barriers to access. The large household out-of-pocket share in overall health financing provides justification for such claims. Households accounted for 60-70% of total health expenditures during 1995-2000 [4]. Out-of-pocket payments are incurred 4 when seeking care from private providers as well as public providers, where official and unofficial user fees are widespread. User fee exemptions exist at public facilities but are rarely effective. Two insurance schemes exist for civil servants and private sector employees but account for less than 5% of total health financing. Recent work on the distribution of household health expenditures has shown that out-of-pocket payments for health care impose a considerable burden on household budgets, especially for lowincome groups [5,6].
The Indonesian economic crisis began in the second half of 1997 when the rupiah depreciated rapidly and plunged the country into economic recession through 1998. Food prices increased by an estimated 80% during the same year. The devaluation of the rupiah increased levels of debt faced by private companies resulting in bankruptcies. This led to a reduction in labor demand, rising unemployment and a corresponding loss of social security coverage [7]. Much of the impact was felt in the manufacturing, construction and finance sectors, which were concentrated in the provinces of Java and Bali. Rural provinces experienced declines in real agricultural wages. Real GDP decreased by roughly 15% in 1998 [8]. Poverty rates, estimated at 11.3% prior to the crisis rose to around 18 to 20%.
The severity of the crisis inevitably affected households' health care utilization and expenditures. Frankenberg et al. [9] found that household consumption fell by 20% in 1998, with investments in human capital (health and education) decreasing by 37%. 5 Utilization of modern health care services fell sharply during 1997-98 and remained constant during 1998-99. Much of this decline was due to a significant decrease in the use of outpatient services, particularly public sector services as shown in Figure 1 [7,9,10].
Trends in child health care use mirrored overall trends, with public sector outpatient care experiencing the greatest decline [11]. The price of treatment at Indonesian government health centers, where the bulk of public sector outpatient care is provided, increased by an estimated 67% during the crisis due to higher input prices of drugs [7]. To households, already faced with diminishing purchasing power, this was a significant increase in health care costs. The price increase, combined with supply outages experienced by cash-strapped government facilities, may explain why public sector utilization rates fell in Indonesia.
Yet in neighboring Thailand, public sector use increased during the same period despite virtually identical macroeconomic conditions and comparable increases in costs of inputs and services.
Different levels of social protection in the two countries underlie this different performance. In Thailand, insurance coverage was already around 77% when the crisis began and public assistance programs were expanded during the crisis. By contrast, a large majority of Indonesians remained uninsured during the first year of the crisis.
The Indonesian health subsidy program -JPS-BK -examined in this paper, was part of a series of measures introduced by the Indonesian government in 1998 to improve social 6 protection for the population. The Social Safety Net (SSN) programs included workfare, subsidized rice sales, targeted scholarships and subsidized health services. Starting in 1998, the JPS-BK initiative financed a range of reproductive and child health services, supplementary feeding for young children and pregnant mothers, other basic health services, and outpatient and inpatient care for poor patients who were referred to hospitals [12]. Utilization of public sector outpatient care services, which saw the greatest decline immediately after the crisis, rebounded faster after 1998 than other types of services. This improvement in the use of public sector services has been attributed in part to JPS-BK [10,11].

The Indonesian healthcard
Under JPS-BK, the most vulnerable households in each community were allocated healthcards, which entitled them to the price subsidy. Government health facilities that provided subsidized care received extra budgetary support to compensate for the increased volume of services they provided. The scheme was financed through the central government using both GOI and donor sources of funding. Village and municipality level committees were responsible for allocating the healthcards to the eligible population in each community. Eligibility was determined on the basis of the prosperity measure described above, although local leaders used their discretion in identifying who was poor and vulnerable. Healthcards were usually distributed through local health centers or village midwives. The healthcard entitled the owner and family members to free services at public health care providers for outpatient and inpatient care, contraceptives for women in child bearing age, pre-natal care and assistance at birth [3].
Ownership of the healthcard among Indonesian households grew to 10.6% in 1999 and to 20% in 2000 [14,15]. The healthcard was not very well targeted, especially initially when less than 5% of the poorest quintile was covered. Pritchett et al. [14] note that as the program went from being a "regular" program to a crisis program, it moved from the typical pattern of middle class capture to being quite pro-poor in the early phase of 8 expansion. However, as the program expanded further it became less targeted once again (Table 1). On average, however, healthcard owning households were more likely to be poor and employed in agriculture, and have a head of household who was less educated and was female [3].  [20][21][22][23][24][25][26]. Another strand demonstrates that prices have important negative effects on utilization [27][28][29][30][31]. Demand studies have also noted that the price elasticity of demand is greater for poorer households than richer households [32,33]; Sahn et al. 2003), and that price effects are positive when they are accompanied by quality improvements [34]. Of these studies, Ex post studies, which examine the impact of price-related policy interventions avoid having to rely on out-of-sample predictions to measure the impact on health care use and outcomes. However, a key challenge for these studies is the need to control for selection bias inherent to many public interventions such as health insurance and fee waivers. It is well-established in the literature that individuals do not always take up the public assistance or benefits that they are eligible [36]. Individuals who are poorer and less well informed may be less likely to enroll in public assistance programs, while sicker individuals may be more likely to enroll. To the extent that these variables also influence utilization rates and outcomes, failure to control for selection effects would produce a biased estimate of the program effect. Currie and Gruber exploit variations in the extent 10 to which Medicaid eligibility was increased in different states to identify the impact of Medicaid expansions. They find a significant positive impact on child health utilization and mortality.
In the developing country literature, only a limited amount of ex-post evaluations of health care utilization has controlled for selection bias. Yip and Berman [37] analyzed the impact of a school-based insurance scheme in Egypt on child health care use, treating participation as selection on observables. They found that school health insurance significantly increased visit rates among children and reduced differentials in utilization between low and high income socioeconomic groups. Wagstaff and Pradhan examined the impact of the introduction of Vietnam's health insurance (VHI) program using a combination of propensity scores and panel data to control for selection effects [38]. They found that among young children, VHI increased use of primary care facilities. It led to substitution away from the use of pharmacists as a source of non-prescribed medicine towards the use of pharmacists as a supplier of medicines prescribed by a health professional. Pradhan et al examined the impact of the Indonesian healthcard on outpatient care utilization, also using propensity score matching to control for nonrandom healthcard allocation. A key finding of their work is that the price subsidy resulted in a higher level of utilization of outpatient services among poor beneficiaries, and a substitution from private to public providers for non-poor beneficiaries.

Hypotheses
The following hypotheses are tested with regard to changes in the use of health services between 1997 and 2000, the period before and after the introduction of the healthcard.
(1) Following the economic crisis, the probability of contacting a public provider for inpatient or outpatient care declined less for children in healthcard owning households than for others, so that the net impact of the healthcard on use of government health services was positive.
(2) The contact rate and number of visits for public sector outpatient care declined less for children in healthcard owning households than for others, so that the net impact of the healthcard on use of public sector outpatient care services was positive.
(3) The impact of the healthcard on private sector outpatient care contacts was negative, as healthcard households increased their use of government services.
(4) Children in healthcard owning households were less likely than others to self-treat in 2000 relative to 1997, so that the net impact of the healthcard on self-treatment was negative.
(5) The poorest recipients of the healthcard and healthcard owners in rural areas were less likely to experience large gains in use of government health services relative to more affluent recipients and those in urban areas.

Analytical framework
This section sets out the estimation problem, discusses potential sources of bias and proposes ways of controlling for them.

The estimation problem
The impact of the healthcard for a given individual, i on utilization, Y is the difference between the value of Y after the individual was exposed to the healthcard and the value of 13 Y that would have been observed, had the individual not been exposed to the healthcard.
More formally, the mean impact (β) on healthcard owners is: where Y 0 it = the level of utilization when not exposed to the healthcard

Non-random healthcard allocation
The Indonesian healthcard was not distributed randomly within communities. Allocation was based on a range of socioeconomic and demographic attributes of the household.
Many of these attributes also influence health care use. A direct comparison of health 14 care use by children with healthcards and those without would produce a biased estimate of the program effect. It cannot be assumed that, absent exposure to the healthcard the outcome would have been the same for both healthcard and non-healthcard owners. In short, the following cannot be assumed: It is reasonable to expect that had there been no healthcard program, healthcard owners who are poorer and less educated would have used less health care, relative to nonhealthcard owners. To obtain unbiased estimates of the healthcard effect, it is, therefore, important to adequately control for systematic differences between healthcard and nonhealthcard owners, so that the two groups are comparable in terms of their preintervention characteristics.
This paper uses propensity score matching (PSM) methods to address the problem of non-random healthcard allocation. PSM relies on matching each healthcard owner with a non-healthcard owner with the same observed characteristics in order to reduce the differences between the treatment and control groups. Critical to this method is the assumption of conditional independence, which implies that, conditional on a set of observed characteristics, allocation of the card can be treated as random [39]. The more variables used to match households, the more likely it is that systematic differences between treatment and control groups will be reduced. Rosenbaum and Rubin [40] have 15 shown that if it is valid to match on all of the selected variables separately, it is equally valid to match on a score estimated from those variables called the propensity score. The probability that an individual is allocated a healthcard is estimated as a function of all relevant pre-intervention socioeconomic and demographic characteristics as follows: The predicted value H * i = Φ(Xiγ*) from equation (3), referred to as the propensity score can be interpreted as the probability an individual i is allocated a healthcard.
An appropriate comparison group is created by matching households that own a healthcard to households without a healthcard, based on the propensity score. The treatment and control groups will yield unbiased estimates of the healthcard impact if the following assumption holds: That is, conditional on the propensity score, treatment and control groups would have had the same outcomes in the absence of the healthcard. The extent to which propensity score matching reduces the bias in estimation depends on the specification of the propensity score model and the quality of the control variables [41]. It is thus critical to include 16 sufficient information about the allocation procedure in the estimation of the propensity score.
An alternative method is to run a regression of the health care use variable on dummy variables, indicating ownership of the healthcard. Observable covariates enter in the regression as linear controls. In the health economics literature, program participation of this nature is generally regarded as endogenous, leading to the use of instrumental variable (IV) estimators within a regression framework. Typically, a subset of variables is found that is highly correlated with program allocation, but does not have a direct impact on the likelihood of using health services. The set of variables forms the IV and is used to predict program participation. An alternative IV technique is the Heckman "two-stage" estimator. In the first stage, the probability of participation is estimated. In the second stage, the first stage results are used to statistically adjust the disturbance term in the outcome regression so that the impact estimate will be unbiased. Typically, only one variable that both predicts participation and is uncorrelated with the outcome is needed to construct a good instrument.
Regression methods require the same, un-testable assumptions about conditional independence, which underpin PSM methods. In IV regressions, this assumption is implicit in the exclusion restriction that the IV is independent of the outcomes, given participation in the program [42]. The IV estimator is consistent, provided the independence assumption holds and the IV is highly correlated with program 17 participation. While there are many good examples of IV based identification strategies from the developed world [43], evaluation work in the developing world has generally failed to produce convincing IV's that satisfy these assumptions. It is particularly difficult to find a suitable IV in the context of the evaluation of the healthcard because almost all of the criteria used to assess healthcard eligibility (low income, inadequate access to health care) also affect health care utilization.
An advantage of PSM over regression methods is that the former does not require a parametric model linking program allocation to outcomes. It thus allows estimation of mean impacts (including impacts conditional on income or area of residence) without arbitrary assumptions about functional forms and error distributions [42]. Both OLS and IV regression methods impose functional form assumptions about the treatment effects and control variables.
PSM allows for the inclusion of a wider range of control variables than regression methods. Jalan and Ravallion [42] note that, in the regression context, there is a bias towards control variables that are exogenous predictors of the outcome variable. Rubin and Thomas' [44] analysis and simulations showed that variables with weak predictive ability for outcomes can still help reduce bias in estimating causal effects using PSM. In the evaluation of the healthcard for instance, variables such as the number of rooms in the house which is a poor predictor of health care use, or health care use by other family members which may be endogenous, would be excluded from the regression analysis.
They are, however, instrumental in reducing systematic differences between treatment and control groups and, when included in the PSM model, contribute to reducing bias in the estimated impact of the healthcard.
A further advantage of PSM over regression methods is to do with the sample used for the analysis. Regression methods use the full sample of households or individuals for the analysis. Estimation based on PSM is restricted to the matched group of treatment and controls only; unmatched comparison households are dropped. Rubin and Thomas [44] have shown that impact estimates based on full (unmatched) samples are generally more biased, and less robust to mis-specification of the regression function, than those based on matched samples.

Exogenous influences on utilization
Estimation of the healthcard effect may be biased, even after propensity score matching, ( ΔYi M ). Using the notation developed above, changes in outcomes for matched intervention and control groups may be defined as: Following Wagstaff and Pradhan [38], the change in outcomes pre and post intervention can be decomposed as follows: where Δ denotes the change between 1997 and 2000, Y * it is the counterfactual outcome for healthcard owners, HC it is the impact of the healthcard program, θ t is the unobserved effect specific to time t, and ε it is a white noise measurement error term associated with the outcome variable. The data are observed for N individuals in households owning healthcards. For N matched control individuals, change in outcomes may be written: where μ it is the white noise error term. Difference-in-differences in outcomes are obtained by subtracting (8) from (7): As the program effect is only relevant for the post-intervention period, ΔHC it = HC i,2000 and Equation (9) may be re-written as: Taking expectations of Equation (10) results in: assuming ε and μ distributed with zero mean. Examining pre and post observations for intervention and control groups has one further advantage. The assumption in (4) is most likely to hold when selection bias is largely a result of observable differences, which can be controlled for using propensity scores.
Systematic unobservable differences between intervention and control groups, if present, could still bias the impact estimate. To the extent that unobservable differences are time invariant, using multiple observations for each participant will difference out the unobservable factors, thus controlling at least partially for this source of bias.
The average treatment effect of the healthcard on healthcard owners is estimated by comparing the change in utilization for the intervention (ΔY it H ) and control groups ). The advantage of this estimation approach is that it is not necessary to specify a model for the outcome variable. It is necessary, however, to specify a model for the allocation of the healthcard itself, which is described in the next section. to ensuring high re-contact rates. Households that moved between one wave and the next are followed up even when they move to a different province. Of the original IFLS-1 households, around 95% were re-interviewed in IFLS-2 and IFLS-3. also collected on the type of provider sought, frequency of visits and in most instances, the cost of each visit. The instrument used in all three waves followed a similar structure and repeated the same questions, in order to allow comparison across different waves.
In addition, I obtained data on BKKBN sub-district poverty rates, which are used by the government to distribute healthcards across districts and sub-districts.

Specification of the propensity score function
The first step is to estimate the propensity scores which will be used as a basis for matching each child in a healthcard household with a child in a household without the healthcard. The propensity score function in Equation (3) above was specified with a binary outcome variable indicating healthcard ownership. Its purpose is to model the likelihood of an individual belonging to a household that was allocated a healthcard on the basis of observable characteristics of the individual, household and community. 24 Understanding the design of the healthcard program and ensuring that the propensity score specification reflects all aspects of the card allocation procedure are vital for reducing the bias in estimation. The following aspects of program design are pertinent in this respect: Criteria for allocating subsidies: the amount of health subsidy and the number of healthcards allocated by the central government to districts and then to sub-districts were based on the pre-intervention (1999) poverty rate in each district or sub-district.
Criteria for allocating healthcards to households: at the sub-district level, village committees distributed healthcards to households on the basis of the prosperity measure described above, additional criteria they received from the government and their own village-specific information. Special attention was paid to households that were severely affected by the crisis. Although village committees were expected to identify eligible households based on information available to them, in practice they faced significant informational barriers. The information and criteria used to identify eligible households were not well defined. This meant that households that were better able to signal their socio-economic situation and needs to the village committees faced a higher probability of receiving a card.
Healthcard distribution mechanism: as healthcards were distributed through village midwives and health centers, households' familiarity with and proximity to health providers would have influenced card ownership. 25 The propensity score function, therefore, includes a range of household and community level variables that would have influenced the allocation criteria above.

Estimation of propensity scores
The propensity score was estimated using a probit model. When the explanatory variables are not balanced between treatment and control groups, the balancing property fails and the model is mis-specified. I tested the necessary conditions for the balancing property using an algorithm developed by Becker and Ichino [45] and re-calibrated the model to ensure the balancing property was satisfied. With the exception of some of the province dummies, the explanatory variables were balanced between treatment and control groups in the final specification for each of the two regions.

Creation of matching control groups
The propensity score is a continuous variable. No two observations in the sample are likely to have the same propensity score. Several matching methods exist for creating a control group with similar propensity scores to the treatment group [46,47]. Rubin's [48] nearest neighbor matching involves choosing the most similar household based on the propensity score for each household in the treatment group. The potential control group is re-weighted so that households that are not matched receive a weight of zero, those that are matched receive a weight of one, and those matched more than once receive a weight greater than one. Inevitably, some of the matches may be quite poor because for some 30 treated units, the nearest neighbor may have a very different propensity score; s/he would still contribute to the estimation of the treatment effect independently of the difference.
Using the alternative radius matching method, each treated unit is matched only with control units whose propensity score falls within a predefined neighborhood of the treated unit [45]. The smaller the size of the radius, the higher the quality of the match but the higher the probability of some treated units not being matched because the specified neighborhood does not contain any control units. I have chosen to use nearest neighbor matching with replacement for this analysis after having experimented with both the nearest neighbor and radius matching methods.
In addition, the common support restriction was imposed in order to ensure high quality matches. The estimator was calculated only where the propensity score overlapped for treatment and control groups. The drawback of imposing this restriction was that high quality matches were lost at the boundaries of the common support and the sample was reduced as a consequence.

Estimation of average treatment effect
Average treatment effect on the treated (ATT) was estimated for changes in health care utilization, ΔY it between 1997 and 2000. The following utilization outcomes were examined.
Child received any public sector care (admissions or outpatient visits) during past 12 months. 31 Child received any outpatient care during the past 4 weeks.
Child received public sector outpatient care from a hospital, health centre or subhealth centre during the past 4 weeks.
Child received private sector outpatient care from a hospital or doctor during past 4 weeks.
Number of public sector outpatient visits for child during past 4 weeks.
Number of private sector outpatient visits for child during past 4 weeks.
Any modern medicines purchased from pharmacy or shop for child during past 4 weeks.
Estimation of ATT was carried out for the entire sample and separately for the following sub-groups: age-groups: all children were grouped into three groups based on age in 1997 -0-5 years,

Healthcard allocation
A propensity score function was estimated first. The full propensity score models are provided in Annex Table A1.
Household socio-economic status is an important predictor of healthcard ownership.
Physical characteristics of the house in which the child lives prove to be as important as more conventional measures of socio-economic status. Having good quality floors, water supply inside the house and own toilets was associated with a lower probability of healthcard ownership, while dirt floors and poor quality walls had a positive effect.
Higher rent or house values and the number of rooms also had a negative impact. It is 33 quite likely that local authorities considered the physical features of each family's living environment when evaluating social and economic deprivation in villages. The low levels of statistical significance may be due to multicollinearity between the variables.
Ownership of assets such as fridges and gas or electric stoves had a negative effect and was statistically significant. Household consumption also had a negative impact. It is worth noting that interpretation of the consumption and asset coefficients is complicated by the fact that the former may reflect expenditures on the latter. Finally, the coefficient on the dummy for purchasing goods at subsidized prices is strongly significant in both models, confirming that households which participated in existed poverty alleviation programs stood a greater chance of being selected for a healthcard.
Other important household level predictors include education of the head of household and the proportion of children in the household. A head of household with primary schooling was more likely than one without any education to secure a healthcard for his or her family. Households with a large number of children stood a greater chance of getting a healthcard, which is consistent with targeting rules used to allocate the cards.
Head of household's employment is not statistically significant, but it is notable that the coefficients for manufacturing, finance and retail are positive, while the coefficient for agriculture and mining are negative. Not an unexpected finding given that the former were the worst affected sectors of employment during the crisis. As it was pointed out earlier, village committees were instructed to target households that were severely 34 affected by the crisis, which would, no doubt, have included households involved in these three sectors.
Prior access to and familiarity with the The urban dummy is positive and statistically significant in the propensity score model.
Urban areas were worse affected by the economic crisis than rural areas because the sectors of employment that were directly affected were more likely to be in urban areas.
To the extent that the healthcard program was introduced as a social safety net to help households cope with the crisis, a higher probability of healthcard ownership among worse affected urban households is not surprising. On the other hand, it may be a consequence of the increase in the number of health centers and midwives -the main vehicle for distributing healthcards -being more accessible in urban or semi-urban areas. 35 The sub-district poverty rate coefficient is positive in both models, indicating that households living in sub-districts with higher poverty rates were more likely to receive healthcards.
Province dummies were jointly significant in both models. The pseudo R-squared was 0.0889 for the model as a whole. Table 4 provides a description of the estimated propensity scores for the two regions. The mean propensity score is 0.30, with little variability between treatment and control groups. The propensity score model was calibrated so that the balancing property was satisfied at significance level of p < 0.005 but without loss of the key individual, household and community level explanatory factors described above. The region of common support between treated and control groups is relatively high, covering 90.7% of all children. Children with propensity scores outside the common support were not considered for the analysis. Table 4 about here The quality of the match is best illustrated by the distributions of the propensity scores for treatment and control groups, before and after matching (Figure 2-3). As Figure 2-3 shows, the distributions are almost identical for treatment and control groups after matching on propensity scores. There are few matches at high values of the propensity score, or high probabilities of being allocated a healthcard. Examining average values of 36 observed variables for matched treatment and control groups provides further evidence of the degree of comparability between the two groups ( Table 5). The first two columns show descriptive statistics for healthcard owners and all others, while the third and fourth columns present the same information for the matched pairs. Differences in covariate means are considerably smaller for the matched pairs. As the last two columns show, the differences in means are not significantly different from zero in most cases.  Table 5 about here To summarize, the matched groups are generally well balanced across observed characteristics. The results of the propensity score estimations suggest a high degree of overlap between treatment and control groups, which should eliminate most of the bias due to the non-random allocation of the healthcards and enable reliable estimation of average treatment effects. Table 6 reports both the average treatment effect on the treated (ATT) estimated by comparing healthcard owners and their matched controls, and the unmatched differencein-difference effect obtained by comparing healthcard owners and all potential controls.

Use of public sector services
Household survey data indicate that public sector utilization rates declined substantially for children during 1997-2000 [11,49]. The healthcard, which was introduced during this period, entitled household members to free outpatient and inpatient care services at public facilities. I tested the hypothesis that the net impact of the healthcard on the change in public sector use rates was positive from 1997 to 2000. I find that the healthcard resulted in a net increase of 3.6% in public sector utilization for healthcard children relative to others. The difference-in-difference estimates for the unmatched groups are not statistically significant and smaller in magnitude than estimates of ATT.
Results presented earlier in Table 3 point to a significant decline in public sector use rates for children between 1997 and 2000, a finding consistent with earlier work. The unmatched difference-in-difference results presented in Table 6 suggest that the decline in public sector use rates was less dramatic for children in healthcard owning households compared to others, although the estimates are not significant. ATT estimates obtained by comparing healthcard children with similar children without healthcards confirm that use of government health facilities fell significantly less for the former. In the aftermath of the economic crisis, the healthcard evidently helped maintain public sector utilization rates among those who received it. Estimates of the healthcard's impact on the number of public sector outpatient contacts shown in Table 6 largely mirror those for outpatient contact rates above. The total number of visits declined for healthcard children and others between 1997 and 2000, but less so for the former as indicated by the positive ATT estimates. 39 The results confirm earlier findings that public sector outpatient use generally fell between 1997 and 2000. They also show that the healthcard "protected" public sector outpatient care use in the post-crisis period. These findings are in line with Pradhan et al's [3] results that the healthcard had a positive and statistically significant effect on public sector outpatient contact rates in Indonesia. It is not possible to directly compare the magnitude of the impacts as the previous study is a one period analysis and ATT results reported for adults and children are combined.

Use of private sector outpatient care services
It was shown earlier that the probability of receiving private outpatient care declined in both regions between 1997 and 2000 (Table 3). In contrast to public sector outpatient care, the magnitude of the decline in contact rates was not statistically different for children with healthcards and a group of observationally identical children without healthcards. As Table 6 shows, estimates of the treatment effect are not significantly different from zero. While the lack of statistical significance means it is not possible to draw any firm conclusions, these results do point to substitution away from private to public outpatient care use among healthcard owners.
The healthcard did not have any significant impact on the number of private outpatient care contacts, which declined for both healthcard children and the control group during the crisis. The hypothesis that the healthcard had a dampening effect on private outpatient 40 care is not confirmed by these results, contrary to Pradhan et al who found a substitution effect between public and private outpatient care.

Purchase of over-the-counter medicines
Households increasingly resorted to self-treatment as prices at health facilities rose sharply following the economic crisis. This trend was reported for children as well as for adults [11,49]. If the healthcard made public sector care more affordable, it would have reduced households' reliance on medicines purchased at pharmacies and shops. I tested the hypothesis that the healthcard had a negative impact on the use of over-the-counter (OTC) medicines between 1997 and 2000.
The last row in Table 6 shows the change in the probability of consuming OTC medicines during 1997-2000. It was shown in Table 3 that use of OTC medicines increased for children without healthcards but decreased for healthcard owners. After matching based on propensity scores, ATT estimates are consistently negative but not statistically significant for any of the groups examined. The ATT estimates are also smaller in magnitude than the simple difference-in-difference estimates. The results point to some differentials in OTC use between healthcard owning children and observationally similar children without healthcards. Children in households without healthcards were more likely to have experienced a net increase in OTC use between the two years. It is not possible, however, to reject the null hypothesis of no-effect based on these results.
The healthcard did not have a significant negative effect on OTC use during the crisis. 41

Distribution of the healthcard impact by age-group
As discussed above, use of public sector services declined less for children in households with healthcards than for observationally similar children without healthcards. This effect was driven largely by a positive and statistically significant ATT effect of 7.0% on children aged 0-5 years (Table 7). Treatment effects on older children are not significantly different from zero. The finding that the protective effect of the healthcard was limited to very young children is not surprising, given that the program targeted mothers and young children. The absence of any statistically significant effect, particularly for the 10-15 age-group may be due in part to small sample sizes. With less than 10% of all children reporting any contact with public sector providers on average, there was insufficient variation within small sub-groups, each consisting of 100-400 treated and control individuals to reliably estimate treatment effects.  Table 8 shows  No such trends are evident for outpatient care in the private sector ( Table 9) or use of OTC medicines (Table 10). It is worth noting that the net impact on private care use by children aged 0-5 years was negative, implying that the fall in utilization was much larger for this age group. This result is not statistically significant. However, when examined in conjunction with the results for public sector outpatient care, it does suggest substitution from private to public sectors for the 0-5 age-group.
Tables 9 and 10 about here

Distribution of the healthcard impact by socio-economic status and geography
The healthcard was a targeted price subsidy program, designed to benefit poor and vulnerable households during the economic crisis. If the poor are more sensitive to changes in prices as the health care demand literature suggests, the healthcard should have had a larger impact on poor children relative to others. This section examines whether the price response varied by household consumption level and urban or rural setting.  ATT estimates for children living in rural and urban areas are presented in Table 12.
Results for urban areas in Eastern-Sumatera region are not reliable, as they are based on a small sample of 185 treated children and 144 controls. With average outpatient care utilization rates of less than 10%, the number of children who experienced any type of care were too few to produce meaningful estimates. I discuss results for the other three groups for which there were 250-500 observations in each treated and control group.
The impact of the healthcard on use of public sector care generally and public outpatient care was positive, but not significant in rural areas in both regions and urban Java-Bali.
As described above, this reflects the protective effect the healthcard had on public sector care use during the crisis. Within Java-Bali, the magnitude of the health care effect was greater in urban areas compared to rural areas. ATT estimates for private outpatient care are negative in urban Java-Bali and rural areas of the Eastern-Sumatera region, implying that private use fell more for children with healthcards than for the control group. The healthcard also had a negative impact on use of OTC medicines for children in the same two regions, although it was not statistically significant.
The findings for urban and rural areas are consistent with Pradhan et al for public sector outpatient care use. They found that the healthcard had a positive impact in both rural and urban areas, with the latter impact relatively larger. The impact on private outpatient care 45 use, however, was negative and significant for rural and urban areas. In my analysis, small sample sizes and a lack of statistical significance mean that it is not possible to draw any firm conclusions about the regional and urban/rural distribution of the healthcard effect.

Discussion
This paper set out to examine the impact of a targeted price subsidy, the Indonesian healthcard program on child health care use. The availability of panel data collected before and after the introduction of the healthcard made it possible to eliminate any confounding effects associated with improvements to the macroeconomic environment that took place alongside the introduction of the healthcard. Combining propensity score matching with a double-difference estimator meant that the estimation was purged of any bias due to selection on observables as well as on time-invariant unobservables.
However, other potential sources of bias remain that could not be controlled for. Timevarying unobservable differences between healthcard owners and their matched controls could bias the estimation of the treatment effect. This would be the case, for instance, if changes in living standards made healthcard owners' expectations with respect to health care utilization to increase at a faster rate, relative to the control group. I acknowledge the existence of this source of bias but argue that, in the absence of experimental data, the study design used here represents the best approach to the problem. The estimation method is also attractive because it avoided having to specify a regression model for the 46 utilization outcomes, which could potentially be subject to incorrect specification of the functional form.
The first objective of this analysis was to determine the impact of the healthcard on My analysis has also shown that failure to control for the non-random allocation of the healthcard could lead to substantial underestimation of the program impact. Estimates obtained by comparing healthcard children with all other children in the sample were found to be consistently lower than the ATT estimates for matched treatment and control pairs. The change in utilization for the matched control group is essentially the counterfactual change in utilization -that is, the change in utilization that would have been experienced by healthcard owners had they not received the healthcard. The fact 48 that the actual program impact (ATT) was much larger than the simple difference in means implies the following: while healthcard owners experienced a smaller decline in utilization relative to the rest of the sample, their reduction in use is considerably smaller than what it would have been had they not received a healthcard.
The second objective of this analysis was to examine the distribution of the impact of the healthcard. The healthcard was a targeted subsidy program and there is evidence that the allocation of the healthcard was somewhat pro-poor. Conditional on being allocated a healthcard, however, richer children were more likely to benefit from the healthcard. In effect, the distribution of the impact of the healthcard favours the economically privileged. While ATT results by consumption groups are mostly statistically insignificant, the magnitude of the impact is found to be larger for the richest group. ATT results by urban and rural groups are not statistically significant and the sample sizes within each regional group were too small to draw any meaningful conclusions about the geographic impact of the healthcard.
In conclusion, the healthcard played an important role in protecting health care utilization in the aftermath of the economic crisis, particularly among very young children, but its The evaluation of the healthcard carried out in this paper suggests that the program could potentially be used to significantly expand children's use of modern health services.
However, to improve the distribution of the healthcard impact, the government would have to address significant non-price barriers to care that exist in Indonesia.
50       Notes: Standard errors in parentheses; ATT standard errors bootstrapped with 100 replications + significant at 10%; * significant at 5%; ** significant at 1% (a) Simple difference-in-difference effect estimated by comparing treatment group with all potential controls, prior to matching based on propensity scores Notes: Standard errors in parentheses; ATT standard errors bootstrapped with 100 replications + significant at 10%; * significant at 5%; ** significant at 1% (a) Simple difference-in-difference effect estimated by comparing treatment group with all potential controls, prior to matching based on propensity scores Standard errors in parentheses; ATT standard errors bootstrapped with 100 replications + significant at 10%; * significant at 5%; ** significant at 1% (a) Simple difference-in-difference effect estimated by comparing treatment group with all potential controls, prior to matching based on propensity scores Notes: Standard errors in parentheses; ATT standard errors bootstrapped with 100 replications + significant at 10%; * significant at 5%; ** significant at 1% (a) Simple difference-in-difference effect estimated by comparing treatment group with all potential controls, prior to matching based on propensity scores