WV5 ;1 7 POLICY RESEARCH WORKING PAPER 2775 Patterns of Health Care Utilization in Vietnam Analysis of 1997-98 Vietnam Living Standards Survey Data Pravin K. Trivedi The World Bank Development Research Group Macroeconomics and Growth February 2002 l POLICY RESEARCH WORKING PAPER 2775 Abstract Trivedi provides an econometric analysis of health care emphasized. Econometric models are estimated for use utilization in Vietnam based on individual and household probability and frequency of contact for all major level data from the 1997-98 Vietnam Living Standards categories of care, and for individual and household Survey. The author focuses on the major features of medical expenditure. Econometric results reveal health care utilization patterns, including the differential responses to income changes at different determinants of largely self-prescribed use of levels of income. Commune health centers and self- pharmaceutical drugs, and the use of government medication are normal goods at lower income levels but hospitals, commune health centers, and private health inferior goods at higher income levels. The author facilities. The role of income and health insurance is discusses the policy implications of these results. This paper-a product of Macroeconomics and Growth, Development Research Group-is part of a larger effort in the group to study household welfare and poverty reduction in Vietnam. Copies of the paper are available free from the World Bank, 1818 H Street NW, Washington, DC 20433. Please contact Rina Bonfield, room MC3-354, telephone 202-473- 1248, fax 202-522-3518, email address abonfieldCa worldbank.org. Policy Research Working Papers are also posted on the Web at http://econ.worldbank.org. The author may be contacted at trivedi(EJ indiana.edu. February 2002. (55 pages) The Policy Research Working Paper Series disseminates the . of work in progress to encourage the exchbage of ideas abouit development issues. An objectiie of the series is to get the findings out quickly, even if the presentations are less than fully polished. The papers carry the names of the authors and should be cited accordingly. The ,, '. . interpretations, anzd conclusions expressed in this paper are enitirely those of the auithors. They do not necessarily represent the view of the World Bank, its Executive Directors, or the countries they represent. Produced by the Research Advisory Staff Patterns of Health Care Utilization in Vietnam: Analysis of 1997-98 VLSS Data Pravin K. Trivedi Department of Economics Wylie Hall Indiana University Bloomington, [N 47405, U.S.A Email: trivedi@indiana.edu * Acknowledgments: I am very grateful to Paul Glewwe for guidance throughout the preparation of this paper. His extensive comments on specific issues helped to improve the paper. I also thank Fwu-Ranq Chang for useful discussions especially about the issues concerning self- medication. Murat Munkin and Jie Qun Guo provided excellent research assistance at various stages. 1 Introduction The changes to the Vietnam health sector initiated by doi moi have been described and analyzed in previous World Bank empirical studies based on the 1992-93 Living Standards Survey. These studies noted, in particular, the decline of the traditional public sector provider of the health care to the poor, the commune health center, and the incipient rise in the early 1 990s of private sector health providers. The deregulation of the pharmaceutical industry was followed by a dramatic growth of private pharmacies as the single most important source of drugs for self-medication. Contacts with pharmacies, both public and private, became the most important type of contact between the provider and patient, while at the same time the role of the commune health center declined. Deregulation permitted the emergence of private care facilities provided by doctors and nurses, some of whom are simultaneously employed in the government hospitals and other public facilities. Another major new feature of the health care sector, largely absent in 1992, is the Vietnam Health Insurance (VHI) program, with mandatory coverage for some sections of the community and voluntary coverage for others. As discussed in a later section, this is an important new development that has significant implications about the rate at which the relative importance of different providers in the health sector has changed and is likely to change in future. These emergent trends, also highlighted and explored in the 1999 Vietnam Health Sector Review (VHSR, 1999), have raised important analytical and policy questions about the access to and utilization of health services by different socio-economic groups. This article provides a further analysis of the direction and scope of changes in the health sector. Here the emphasis is on econometric modeling of demand for different types of health care in Vietnam, and we have relatively little to say about supply aspects. In contrast, the VHSR (1999) surveys all aspects of the health sector. In modeling the demand for health care we use 2 both individual and household level data from the 1997-98 Vietnam Living Standards Survey (VLSS). The focus is on four major features of health care utilization patterns, including: * the determinants of the largely self-prescribed use of pharmaceutical drugs; * the use of government hospitals; * the (declining) use of commune health centers, and * the increasing use of private health care facilities. In each case, our major interest concerns the role of price and income variables. Given the high rate of economic growth, accompanied by major structural changes in health care delivery, it is useful to consider how these changes have impacted on the pattem of health care utilization, and what further changes we might expect to see in future if these trends persist. Our major tool of analysis is multivariate regression analysis using models that respect the discrete nature of many of the outcome variables that are reported in the VLSS. The analysis of this paper complements that provided in the comprehensive 1999 Vietnam Health Sector Review (VHSR, 1999). The latter provides a great deal of descriptive statistical information on the structure and organization of health care delivery and on the broad pattern of health care use in Vietnam. The VHSR (1999) also provides many points of comparisons between the 1992-93 VLSS and its 1997-98 counterpart. Although the tabular summary information provided there is both rich and informative, and also highly suggestive of factors that influence health care utilization, no formal modeling of the data is undertaken. Relationships between variables are studied or interpreted on a bivariate basis, e.g., the relation between health care usage and household income. Because this method cannot control for the presence of other relevant factors, and can at most only establish informal associations, there is a danger of misleading interpretations due to the neglected confounding variables. This article 3 addresses the task of measuring and testing various hypotheses about health care utilization within a multivariate econometric framework. The remainder of this paper is organized as follows. In Section 2 we take a preliminary look at data and outline the substantive issues considered later in the paper. We also summarize the changes and trends affecting the health sector since the 1991-92 VLSS. Section 3 surveys the main empirical issues considered in the paper. Section 4 considers the modeling framework and related statistical issues. Section 5 analyzes the determinants of enrollment into the VHI program. We then tackle the task of modeling the utilization of the major components of health care, ignoring the remaining providers who collectively account for only a small share of the health care budget. Econometric methods used in this section are those appropriate for binary- valued and count variable. These are used to model the probability and/or number of visits to commune center, pharmacies, government hospitals, private providers, and public hospitals. Section 6 discusses results for individual and household health care expenditures. Section 7 discusses the policy implications of the results. Section 8 summarizes and concludes. 2 1997/98 VLSS Survey 2.1 Coverage of health care questionnaire The health component of the 1997-98 VLSS is the main source of the data used here. Utilization data are collected for seven types of providers (government hospital (GOVHOSP); commune health center (CHC); regional polyclinic; other government health facility; private health facility (PHF); traditional eastern medical practitioner; pharmacy visits or self-medication (PHARVIS). For each type of provider, the questionnaire seeks information on number of contacts, total expenditure, the amount spent on medicines, and transportation and other costs associated with 4 the visits. For the self-medication part the questionnaire also sought information on whether the visit to the drug vendor was self-initiated or by another provider. The survey also provides information on government hospital admissions (HOSPADM) in the twelve month period preceding the survey, and the number of nights spent in hospitals (HOSPNITE). The responses to questions about expenditure on health care refer to a period of 12 months. The data make it possible to calculate this figure for the household both inclusive and exclusive of health insurance expenditure. Finally, the questionnaire collected information on whether the respondent had health insurance (HLTHINS) and the amount spent on the same in the previous 12 months. The survey includes information on the current health status such as occurrence of illness (ILL) or injury (INJ) in the preceding 4 weeks, the number of days of illness (ILLDAYS), and number of days of restricted activity (ACTDAYS). 1 However, there are some gaps in the information available that affect econometric modeling of health care utilization. For example, there is still no information on long term health status such as the presence of limiting and nonlimiting chronic conditions. The general health status of an individual is a very important conditioning variable in most analyses of demand for health care. Further, the direct information on HLTHINS lacks some necessary detail as to which l There is a suspicion, based on information collected in the field, that the responses to the question about recent illness or injury may be biased. The bias is thought to arise from the greater propensity to report illness or injury of those in higher socio-economic categories. 5 of the several levels of coverage of insurance applied to a survey respondent; this issue is elaborated further in the section on health insurance.2 The survey also provides information on various socio-demographic variables such as gender (SEX), years of schooling (EDUC), age (AGE), marital status (MARRIED), as well as total household expenditure (INC). Table 1 provides the data definitions and descriptive summary statistics. These data can support empirical investigations both at the level of an individual and the household. The available frequency-of-use data can be used in regression modeling of the probability of contact between provider and patient, and also in modeling the frequency of such contact. Models of probability of contact attempt to explain the factors that distinguish those who received some care from those that did not receive any care. Such models only distinguish between zero and positive levels of utilization. Count data models distinguish between different levels of usage, but do not distinguish between high quality contact that may have cost more and those of lower quality. Models of expenditure, on the other hand, allow us to take into account expenditure variations that may be due to variation in the quality of care, but they do so indirectly. Count models are analogous to models that explain quantities, but expenditure models attempt to explain the product of quantity and price of service. Aggregated health care expenditure data can be used to develop total household expenditure models. The two modeling approaches are largely complementary and mutually reinforcing. 2By combining data sets it may be feasible in future to impute the type of insurance of each individual, and thereby make it possible to study the differential impact of each on utilization measures. 6 3 Survey of Main Issues The VHSR (1999: 49-50) considers five factors in health utilization: income, price, quality of care, access (especially by income levels), and the role of education. In this section we outline qualitatively the main issues concerning the role of these factors. Regression models are considered in Section 5. We begin by summarizing the main features of health care utilization data provided in the 1997-98 VLSS. Table 2 below summarizes the contact rates for the main types of providers in 1993 and 1998. The data show that other govermment health facilities and traditional providers are a small part of the total number of visits. Pharmacies and drug vendors, government hospitals, private health facilities and commune health centers account for bulk of the total usage. All types of utilization have grown since 1993, but the use of drug vendors has shown the fastest rate of growth - - it has more than tripled. This attests to the overwhelming importance of self- medication. Pharnacies are both private and public. Private pharmacies are more prevalent in the urban areas and public ones more so in the rural areas. The VHSR (1999, p. 56) points out that the increase in average number of pharmacy visits is accompanied by a decline in out of pocket expenditure on drugs. Also notable is the growth of utilization of private health facilities, which has grown rnore than 2.5 times. Previous analyses of the 1992-93 data have attributed the impressive rise in pharmacy visits to a combination of factors. The first is the improvement in the supply and availability of pharmaceutical drugs between 1993 and 1998, following the deregulation of the retail markets and liberalization of the pharmaceutical industry in 1989. Evidence suggests that the real price of drugs may have declined over the 1993-98 period by as much as 30 per cent or more. The second factor is the persistence of self-medication induced in part by the ease of access of medicines 7 relative to the alternatives. In rural areas, especially, distance from government health facilities as well as poor quality of health services at commune health centers have been cited as possible reasons for the continued growth of self-medication. The growth of private health providers is another important facet of health care utilization. There are two main types of private health providers: (i) full time providers who own private facilities, and (ii) part-time providers who are employed by the public health facilities but engage in private practice during off-hours, (VHSR, 1999, p. 101). Both licensed and unlicensed practices are included in this category, and hence the quality of care in this sector may be variable. Nearly 70% of PHFs are estimated to be in urban areas. 3.1 Access and costs A rough measure of access is the annualized health service contact rate by provider type. These are shown in Table 3. This measure can be misleading if not supplemented with other information. In any sample survey, one is likely to observe zero contact frequency for some respondents, in part because the respondent was healthy in the survey period and did not need health care. Table 4 compares the overall contact rates with the subset consisting of those classified as sick or injured, and extends the same comparison across different income levels. The figures show that commune health care is sought at a 3-4 times higher rate by the sick in the lowest income quartile, compared with the sick in the highest income quartile. The situation is reversed in the case of government hospitals. The contact rate in that case is 3-4 time higher for the sick rich than for the sick poor, indicating that these hospitals are more important providers of health care for the relatively better off. This differential usage is further in evidence for private health providers, but the difference multiple is closer to 2 than 3 or 4. The differential 8 usage is the smallest for pharmaceutical providers. This appears to indicate that access to private drug vendors is roughly equal for the sick, whether low- or high-income. However, this needs a caveat because government hospitals that are accessible to the VHI enrollees, and are favored by the high-income groups, also dispense pharmaceutical drugs, and act as a substitute for pharmacies. Table 4 shows the average contact rates with providers by health status. This shows that at all income levels pharmacy is the most frequently contacted care provider for the sick. Private health providers and government hospitals, respectively, are the next most frequently contacted by those in the top income quartile. For the sick in the lowest income quartile, private providers are relatively more frequently contacted than are commune health centers. The contact rates for the latter are very low for high income individuals. To summarize: commune health centers appear to primarily serve the low-income groups, and government hospitals primarily serve the high income group. However, as noted below, those who are covered by Vietnam Health Insurance (VHI) program are serviced by government hospitals. And insurance coverage under VHI is more extensive for the relatively better off groups; hence the observed higher usage for these groups may be due to a combined income and a price effect, as will become clear in our econometric analysis. Finally, as others have also previously noted (Gertler and Litvack, 1999), self-medication through drugs purchased at pharmacies appears to be the first line of defence against sickness, irrespective of their income class. Cost of access: For each type of service contact, the questionnaire collected responses on the total cost of transportation, room and board, and other related costs. Of course, these data were only collected for those who actually had non-zero usage. The data are censored for those 9 who had zero usage. For those who had positive usage level, the data can be used to estimate average extraneous cost of health service. Although this is useful information, it is insufficient for modeling individual choice of the type of service. Standard economic theory suggests that in choosing between two types of providers, e.g. commune health center and government hospital, the relative extraneous cost of the services of the two providers is relevant. The survey data only pertain to the average extraneous cost of the service actually chosen by the patient. By itself it cannot be used to construct a relative price for each user that is needed for modeling purposes. Even a simple measure such as distance from different types of providers may be used to construct a more appropriate measure of the extraneous costs under the assumption that such costs are closely related to the distance. The average extraneous cost of access, for those who used a provider, may still be useful as a rough benchmark comparisons. Average household medical expenditures for different types of households are shown in Table 5. The medical expenditure for the average urban household is nearly 50% higher than the average rural household. 3.2 Health insurance A major new development in the health sector since the 1992-93 VLSS survey is the emergence of a national health insurance program, Vietnam Health Insurance (VHI), initiated in late 1992, began effective operation in 1993. Three health insurance programs in Vietnam are provided under government sponsorship, comprising a compulsory national health insurance program and the two voluntary programs. During its first phase, insurance coverage was provided to current and retired civil servants and to salaried employees of state-owned and large private enterprises. Benefits include the full cost 10 of pharmaceuticals, ambulatory and inpatient care at governmentalfacilities to which enrollees are referred; a district or provincial hospital acts as a primary care provider. The mandatory VHI coverage does not extend to dependents of employees. A second voluntary VHI plan provides for coverage on a group basis to VHI dependents and some other groups such as communes. That is, groups rather than individuals must enrol in the program. The benefit level varies. A third tier of national health insurance is the voluntary plan, called Comprehensive Student Insurance (CSI). The benefits of the compulsory VHI are less variable than those of the voluntary component. The CSI plans and premiums are locally designed and administered, and show substantial variation in premiums and benefits among localities. One estimate of the number of total (compulsory and voluntary) enrollees in the VHI program comes from VHSR (1999) that estimates this at 9.8 million in 1998, which includes about 38% voluntary enrollees. This covers roughly 12% of the population. The coverage of the target population for the compulsory component is around 77%, but it is much lower for the voluntary component and largely consists of students. Having health insurance is positively related to income class. In the lowest income quartile, insurance coverage rate is 9.2%. A high proportion of this group may be those enrolled in the voluntary scheme. In the top income quartile, 24.5% have health insurance, see Table 6. Tabular analysis of the impact of insurance is provided in the VHSR (1999, Table 65). Controlling for income (by quintiles) the insured have significantly higher rates of service utilization of public providers, especially for in inpatient services in government hospitals. Because the VHI premium for the compulsorily insured is a fixed percentage of the employee's base salary, the cost of insurance varies and income serves as a partial proxy for the cost of insurance. One of the objectives of the empirical investigations in this article is to estimate the impact of health insurance on utilization - - an issue that was not relevant in analyzing 1992-93 VLSS data. The foregoing account raises an important econometric issue concerning the treatment of health insurance. For the compulsorily enrolled in the program, i.e., the majority, we may treat insurance status as exogenous, but for those who are voluntarily enrolled, there may be an element of individual choice, which is an argument for treating the variable as endogenous. However, as was noted above, enrollment is on a group, and not individual, basis. This factor diminishes the role of individual preferences in the choice of health insurance. There is also a related data problem. The health component of the questionnaire asked only two questions about insurance; first whether the respondent had health insurance, and second, the cost of health insurance in the previous 12 months. Without additional data, one cannot distinguish between those who were enrolled in the compulsory insurance program and those who were not. Therefore, insurance status will be treated as exogenous in the health care utilization equation. That is, we postulate that causality runs from health insurance to health utilization. In theoretical models with unrestricted choice of insurance, the choice of insurance and health utilization will be interdependent, or jointly (rather than recursively) determined. It is of some interest to compare utilization patterns among the insured and the uninsured individuals, conditional on positive expenditure over the previous 12 months. Total expenditure of the insured sample is about 20 percent higher, and this difference is statistically significant. The average difference in government hospital utilization between the insured and the uninsured 12 is also statistically significant, that for the insured population being higher by a factor of about 2.5. The average difference in the use of private health facility and drug vendors is significantly higher for the uninsured sample than for the insured. This general pattern is also consistent with the results of regression analysis in which we control for many socio-demographic variables. 4 Statistical Issues in Analysis of Individual Data In this section we shall develop econometric models for health insurance and for four categories of health care services that jointly account for about 99 percent of the total health care expenditure. The largest component (92.7%) is due to drug vendors and pharmacy visits; government hospitals and private health facilities account for another 5.5 to 6.0 percent, with the former being slightly larger. The commune health centers account for close to 1 per cent. Traditional (Eastern) medicine providers and other smaller components will not be analyzed. 4.1 Problem of zeros Individual utilization data are available for 27,731 cases. However, this includes a high proportion of cases of zero utilization, in part because survey period is truncated at 4 weeks; see Table 7. For PHARVIS the zero proportion is about 76%, but for the other three major provider categories it is between 96 and 98 percent each. Typically the observed frequency distribution shows positive probability mass at only a few other integer values, such as 1, 2, and 3, and with very small mass at higher integer values. For example, for PHARVIS, the frequency of 1 visit is about 10%, and of 2 visits about it is less than 5%. In the case of GOVHOSP, PHF, and CHC, the corresponding percentages are even smaller. 13 The handling of the zero problem depends on whether one's objective is to model health care expenditures or frequency of contact with the provider. In modeling expenditures the zeros pose a problem because they introduce a discontinuity in the distribution of expenditures. But in modeling a discrete random variable such as contact frequency this is not an issue. For the 1992-93 data, Gertler and Litvack (1999) chose the "two-part model", in which the first part models the split between zero and nonzero expenditures through a binary outcome model. That is, the focus is on modeling the probability of contact with the provider, using the logit, or the probit, or the linear probability model. The second part of the model is a linear regression in which the outcome variable is health care expenditure, or the logarithm of it, for those who had at least one contact with the provider. If y denotes the measure of health care use, e.g. expenditure, and X denotes the explanatory variables, then according to the two-part model, E[y|X] = probability[y > OIX]x E[yly > O,XI An attraction of the TPM framework is that it provides a solution to the awkward statistical issues arising from the presence of significant probability mass at y = 0.3 Note also that the zeros have two possible interpretations. The first is that they correspond to "corner solutions" in the consumer choice problem. That is, they indicate non-consumption, given current income, price, and health status. A second interpretation is that zeros indicate that the good under consideration is not in the consumer's choice set, for a variety of possible reasons; see Cameron and Trivedi (1998, chapter 6). The first part of TPM may be interpreted as a model of the probability of an interior solution to the choice problem, while the second part models the level of consumption, conditional on an interior solution being realized. Therefore, both parts yield estimates of economically interesting parameters. 3 Continuous probability distributions cannot allow for nontrivial probability mass at zero frequency. 14 A disadvantage of this framework is that the number of observations available for the second part of the model can be proportionately quite small, leading to a loss of precision in estimation. In these cases it is attractive to use a count data model, which can naturally accommodate a significant probability mass at zero, making it unnecessary to separate the zero and nonzero observations as in the case of TPM.4 Further, count data models work well in those cases where the outcomes are concentrated on a relatively few values of the outcome variable, which is the case for three of the four utilization measures we wish to model; an important exception is the number of nights spent in public hospital. In all count models the following specification is used to model the number of visits: log(E[#visitsto provider|X|])= lXjj. In this paper we shall use the binary outcome framework, a la Gertler and Litvack, in those instances where the frequency distribution of contacts is strongly concentrated on just two (O and 1) or three values (0, 1, and 2). For other cases, most notably PHARVIS, we shall use a count data model in which the modeling focus is on the average number of contacts as a function of observed characteristics of individuals. Aggregate health care expenditures are modeled both at the level of an individual and at the level of the household, but we do not model expenditure on individual components of utilization. 4A qualification is that the proportion of zeros may be "excessive" relative to the count model specified, thereby requiring that we use as flexible a count data model as feasible (see Cameron and Trivedi, 1998). 15 4.2 Problem of clustering Another significant factor is the clustering of responses by the primary sampling unit, the commune, used in the complex (stratified) sample survey methodology. In the case of VLSS data, clustering is by commune, which is sampling unit. The sample covers fewer than 200 communes, with variable number of observations per commune. Clustering implies lack of independence of observations. Clustering affects both the discrete and continuous outcome variables studied in this article. If there is significant within- commune homogeneity, perhaps due to common unobserved fixed or random components that affect all individual behavior within a commune, then assuming independence of observations will produce estimates with a spuriously higher degree of precision than is warranted. The correct sampling variances are larger than estimated under the independence assumption. Two statistical approaches to deal with the effects of clustering are used. The first is based on adjusting all standard errors for clustering by using the so-called "cluster robust" variance estimator. Typically this adjustment inflates the estimated variance of the coefficients. This approach is analogous to the use of the so-called Eicker-White robust variance estimator. All estimates of standard errors reported in this study are "cluster robust" unless stated otherwise.5 A second approach to clustered observations uses a different statistical model. In this case it is assumed that each commune has its own intercept, denoted aj wherej is the commune subscript. Correlation between responses for a given commune reflect the presence of a common intercept, which is treated as a cluster (fixed) effect. In this formulation of the cluster effect, both the point estimates of parameters as well as their standard errors, are affected, whereas the 5For example, this is possible in the computer program STATA 6.0 for a variety of estimators. 16 "cluster robust" approach only adjusts standard errors. An example of a linear fixed effects model is Deaton (1997, 288-292). In this study we want to allow for commune fixed effects, but our main interest is not in estimating the fixed effects but in eliminating their impact on economically interesting parameters like income elasticity. Neglect of the fixed (cluster) effects can bias the coefficients of interest if the relevant regressors are correlated with fixed effects. In fixed effect models parameters of interest may be obtained after "sweeping out" commune effects. However, unlike linear regression models where the sweep step is always possible, for nonlinear regression models used in this paper, this approach is feasible for only for a class of models, such as the logit (but not the probit) and Poisson regression models (Hsiao (1986); Cameron and Trivedi (1998, Chapter 9.3). An interesting feature of our use of the fixed effects count model is that it indirectly reduces the impact of the "excess zeros" problem. This comes about because the fixed effects model will drop all observations from a commune if the responses within a commune are identical. In general, therefore, fixed effect models are based on fewer observations since (as is the case in our sample) all observations from a commune are thrown out if every survey respondent records the same response (e.g. zero). For example, if in any particular commune the sample shows no usage of the CHC, then all observations for that commune are dropped. This reduces the sample size but also reduces the impact of the "excess zeros" problem. The number of "lost" observations will vary with the type of provider being considered. The two approaches for handling clustered data involve non-overlapping statistical assumptions. Hence there is no guarantee that the results from fixed effects model will coincide with those from the standard models. Nor is there a simple of way of choosing between the two models, in case they yield different results. 17 4.3 Econometric models The preferred model in this article for counted data is some variant of the count regression model that can handle the aforementioned statistical problems of clustering and/or excess zeros. In those cases where the excess zeros problem is very severe, the binary outcome model is used to model the probability of nonzero outcome. The justification here is that our ability to distinguish between factors that lead to (say) n+l units of utilization rather than $n$ is seriously reduced when the cell counts for those outcomes are small. Hence, to gain precision, we only attempt to distinguish between zero usage and positive usage, ignoring the extent of positivity. In all cases we have used the fixed effects version of an estimator to check the robustness of the estimates. Fixed effects estimators are usually adopted in the analysis of longitudinal data to deal with individual specific heterogeneity under the assumption that this component is fixed for any individual and can be "swept out" by an appropriate data transformation that is feasible given repeated observations on the same individual. In the case of VLSS data we have assumed instead that the fixed effect is commune specific and can be handled in an analogous fashion given more than one observation per commune. That is, the longitudinal (panel) models can be adapted to handle the present case. The underlying general method for handling clustered data is the conditional maximum likelihood approach which can be applied to binary, counted or continuous data. Fixed effects variants of the Poisson, the logit and the linear regression available in the literature for panel data can be adapted to handle clustered observations. 18 Counted data are sometimes modeled using the ordered probit or the Tobit regression. These rely on the normality assumption. Violation of normality due to the excess zeros ad clustering is so pervasive in the VLSS sample that these estimators are inappropriate. To summarize: the following alternative models specifications were estimated. Individual data * Probability of having health insurance. * Probability of visit to: commune health center (CHC); private health facility (PHF); government hospital (GOVHOSP); pharmacy or drug vendor (PHARVIS). * Number of visits to: CHC; GOVHOSP; PHF; PHARVIS. * Probability of admission to government hospital (HOSPADM). * Number of nights spent in government hospital (HOSPNITE). * Aggregate expenditure on health care provided by all providers in the 4 weeks preceding the survey and in government hospitals in the previous 12 months (MEDEXP) Household data * Aggregate expenditure on health care provided by all providers in the 4 weeks preceding the survey and in government hospitals in the previous 12 months (HMEDEXP). 5 Results The main focus of the discussion of results will be on the role of income and health insurance, after conditioning on a set of relevant covariates. This conditioning applies to all models unless stated otherwise. As is customary, the income variable is proxied by logarithm of total household expenditure, denoted log(INC). The only price variable is HLTHINS. The conditioning 19 covariates are AGE, SEX, MARRIED, EDUC, ILL, INJ, ILLDAYS, ACTDAYS. These variables are defined in Table 8. 5.1 Determinants of health insurance status Initially we do not distinguish between voluntary and mandatory enrollees into the VHI program. A logit model for health insurance (HLTHINS) status, using the full sample, shows that age (AGE), educational level (EDUC), and income (INC) are all strongly positively related to having insurance. See the first two columns of Table 8 for detailed regression results. The coefficient on log of total household expenditure, our proxy for income, is precisely determined with a t-ratio exceeding 5 in most specifications. High income and high education both increase the probability of compulsory coverage, hence the observed result is quite plausible. Surprisingly being married (MARRIED) is negatively related to having health insurance. One possible interpretation of this result is that it reflects the higher rate of health insurance among students through the CSI program, but it could also reflect relatively higher enrollment into the VHI by unmarried males. Controlling for the aforementioned factors, there is a negative association between having insurance and being female. In the third and fourth column the results are for a specification with a more flexible functional formn for the income variable to allow for different response coefficients in the four income quartiles, denoted INC 1, INC2, INC3, and INC4. Essentially we split log(lNC) into four ranges and consider whether the income coefficient varies and whether this spline functional form improves the fit of the model. This specification fits the data better. However, the fixed effects logit version, which allows for commune fixed effects, of the same specification fits even better. Overall, this regression gives a similar picture to that form the 20 simple logit, but it shows that the insurance decision is insensitive to income in the lowest quartile, and most sensitive in the two middle quartiles. Several variants of the fixed effects logit model were estimated by three age categories: less than 22 years, between 22 and 60, and greater than 60. The motivation for this disaggregation comes from the aforementioned differences in types of health insurance. It is expected that the student enrollees are predominantly concentrated in the youngest group, and the retired individuals in the oldest group, leaving the middle group containing the largest number of those mandatorily enrolled. Interestingly, the average rate of enrollment in three age groups varies only between 16% and 19%. The regression for the middle group are shown in the last two columns of Table 8, and the main features of this regression conform to those mentioned earlier in this section. The results for the youngest group show the highest income coefficient (.72), and those of the oldest group show statistically zero income sensitivity. The separate results for the young and old groups are not reported in Table 8 to save space. Although this disaggregation is rough, and a more careful modeling of the insurance decision is desirable, our results do not suggest that there is serious distortion due to the absence of distinction between types of health insurance. It is, however, possible that the main impact of different types of insurance will be on utilization. This issue will be explored in later sections on health care utilization. An important issue is whether there is adverse selection into the insurance program. To infer that there is, one would need to show that controlling for other factors, the insurance program enrols a disproportionately larger number of "bad risks". Identifying high risk enrollees is not easy because of the lack of information about the long term health status of individuals in the health survey. Our inference has to be more indirect. The survey provides information on the 21 injury and/or illness in the 4 weeks preceding the survey, and information on the number of illness days and days of limited activity. Information is also available on the smoker/nonsmoker status of the respondent. This latter variable can serve as a proxy for future health problems. The logit regression results suggest that there is no significant association between insurance status and the number of illness days (ILLDAYS) and/or limited activity days (ACTDAYS), or smoking habit. There is, however, statistically significant, but weak, positive association between incidence of illness (ILL) and injury (INJ) and having health insurance. This is consistent with those who are insured having a greater proclivity to reporting illness or injury. The sector of employment plays an important role in the insurance status because, as was pointed out Section 3, insurance is mandatory in some government and private sectors. If, as seems reasonable, membership of the sector is independent of the level of utilization, then exogeneity of the insurance variable is justified. In such cases it seems valid to argue that causality runs from insurance to utilization. To put this argument on a sound footing it is desirable to enter sector of occupation as an additional factor in the insurance equation and to confirm its role as an important factor after conditioning on income and educational attainment. Implementing this step requires additional data that were not available when this study was done. We now turn to regression analysis of utilization, treating the insurance variable as exogenous. 5.2 Commune health centers Detailed regression results are given in Table 9. In the case of CHC, the relative infrequency of use, the "excess zeros" problem has already been noted. This makes the results from direct application of the Poisson regression unreliable. Hence the reported results use either 22 the fixed effects (conditional maximum likelihood) variant of the Poisson regression, or simply fixed effects logit model for the probability of CHC use. The three regressions in Table 9 suggest that CHC is treated by users as an inferior good. The marginal impact of rising household income on both the probability and level of usage is negative and significant; see the last four columns. The impact of rising educational level is also negative but is only marginally significant. That is, CHCs are typically not used by the higher income and better educated groups. In their analysis of 1992-92 VLSS Gertler and Litvack have noted the low quality of CHC services (also see VHSR, 1999). Our evidence is consistent with their observation. However, it would have been more satisfactory to have introduced additional variables in the regression reflecting various observed features of CHCs, in order to pin down precisely why CHCs are shunned by users. For example, it would be useful to know whether there are differences in their ability to supply drugs or provide higher levels of service. However, this requires more data than currently available. To throw more light on the relation between income (log(INC)) and the level of use of CHCs, we report an additional regression in Table 9, first two columns, which uses the spline functional form to allow for different response coefficients in the four income quartiles, denoted INC1, INC2, INC3, and INC4. The fit of the model for probability of using CHC improves only slightly. These results indicate that the income coefficient is not significantly different from zero for the three lowest income quartiles, and is significantly negative for the highest income quartile. Qualitatively the pattern of coefficients that we find is that which we expect for an "inferior good". The coefficient of HLTHINS is positive. This is an unexpected result. When additional regressions were estimated for three age groups, the results indicated positive coefficient in each 23 case but with a low degree of precision. One possible explanation is that the result reflects the role of the CHCs as drug providers to eligible insured individuals. The results also indicate that AGE and CHC use are negatively related, i.e., older individuals avoid using CHCs. The strongest positive relation of CHC use is with short term health status. Those who are ill, or injured, and have suffered limitation in physical activity do use CHCs. For young and sick/injured individuals from low-income households, CHCs may serve as a first step in seeking health care. 5.3 Government hospitals This subsection will first discuss the results from models of probability of outpatient visits. This is followed by discussion of results for inpatient hospital admissions and number of nights spent in the hospital. 5.3.1 Probability of outpatient use A variety of estimation methods have been used to model the probability of use of public hospital outpatient services. Detailed regression results are given in Table 10. The results from the regular logit and fixed effects logit are qualitatively similar and precise. The probability of use is strongly and positively related to household income. The elasticity of probable use with respect to income is estimated around 0.4. There is also a strong positive relationship wiih having health insurance (HLTHINS), which confirms the strong connection between the use of public hospital outpatient services and having health insurance. These two results confirm that income and price effects in the use of public hospital outpatient 24 services are strong. Additional robustness checks do not show that the fit of the model can be improved by using a spline specification for the income variable. The use of public hospitals is also strongly related to being ill or injured and to the length of illness (ILLDAYS). Usage is also higher for married persons and for females. Somewhat unexpectedly, AGE is negatively related to use. However, inpatient utilization is strongly and positively related to age, as discussed in the following section. 5.3.2 Hospital admissions and nights in hospital About 5% of the sample reported admission into hospital in the 12 months prior to the survey. About 26% of these are insured. Of these the overwhelming majority, more than 97%, spent at least one night in hospital. The average number of nights spent in hospital, conditional on admission, is between 13 and 14, but the median figure is about 7, indicating a skewed, fat-tailed distribution. The probability of hospital admission is higher for the insured than for the uninsured, and the average length of stay in hospital is also longer for the insured, about 18 days versus 12 for the uninsured. The distribution of hospital expenditure is also correspondingly skewed and fat-tailed with very high non-normal kurtosis. That is, relatively small number of individuals account for a high proportion of the total hospital expenditure. Detailed regression results for admission and hospital nights are given in Table 11. The results indicate a strong positive relation between being insured and the probability of admission into hospital. There is also a statistically significant positive relation between income and hospital admission, but this link is weaker than that with insurance. Other factors that indicate bad health status also increase the probability of admission. When HOSPADM regressions are run separately for the insured and uninsured subsamples, hardly a single explanatory variable has a statistically significant coefficient for the insured subsample, but for the uninsured, both health 25 status and income levels are important factors. When the HOSPNITE regressions are run separately for the insured and uninsured subsamples, most of the coefficients in the uninsured equation are absolutely larger indicating that their greater sensitivity to other factors. That is, having health insurance reduces the role of other factors but does not eliminate it. Developing a robust regression model for the number of nights in hospital is difficult because of the awkward frequency distribution of the data. Several linear and nonlinear (count data) models were tried. None fit the data really well; that is, the data are intrinsically rather noisy. Overall, these results indicate that AGE, income, and HLTHINS are the most important explanators of hospital usage. When different specifications of equations are considered income is shown to be not a robust explanatory variable, but HLTHINS and AGE remain consistently significant. 5.4 Private health facilities The contact rate for private health facilities are significantly higher among the uninsured than the insured population. They are also slightly higher among the younger age group. Because the excess zero problem is very evident (95.5% of the sample report no use), the estimated regressions model the probability of contact. Detailed regression results are given in Table 11. In the full-sample regressions, HLTHINS has a negative coefficient reflecting higher contact rate among the uninsured. Since insured individuals are eligible for outpatient care in public hospitals, the private facilities are relatively more frequented by the uninsured. Health insurance, therefore, diverts usage away from the private facilities. Once we control for insurance status there is no clear evidence that income and private health care are positively related. The size of the income coefficient is found to be sensitive to 26 changes in specification. In the fixed effects logit model the income coefficient is small and relatively imprecise. The use of the spline specification also does not completely resolve the ambiguity, but there is slight evidence that in the higher income quartiles there may be a significant positive relation between income and the use of PHF. In an attempt to examine the role of aggregation bias in estimating the income effect, the fixed effects Poisson model was reestimated by age groups using both the regular and for the spline specification of the income variable. For the middle income group (2260) groups. These results suggest that as yet there is only weak evidence of income- induced demand shift towards PHFs and quite clear evidence of the negative impact of VHI plan on PHFs. In other respects the pattern of use is qualitatively similar to that of public hospitals. As in that case, usage is positively related to being ill (ILL) or injured (INJ), being female (SEX) or married, and with the length of illness (ILLDAYS), and negatively related to AGE. The last result may simply indicate a greater willingness on the part of the young to try out the emergent private health facilities. 27 5.5 Pharmacy visits Detailed regression results for the number of visits and the probability of visit are given in Table 12. Because of the overwhelming importance of expenditure on purchased medicines, the results for the frequency of pharmacy visits are of special interest. Although many variants of the Poisson regression and the logit model were used, the reported results are based on the commune fixed effects formulation.6 The most interesting results shown in the first two columns of Table 12 are: overall income effect is significantly negative and the health insurance effect is also negative. As in the case of GOVHOSP, using the fixed effects model lowers the absolute size of the income coefficient. This result is plausible and consistent with lower income household relying overwhelmingly on self-medication in the event of illness, injury, and activity limitation. In earlier section we have already cited evidence that indicates an increasing reliance on self- medication as the supply of drugs has improved and the retailing of drugs has become deregulated. Drugs can also be dispensed at public hospitals, but the evidence presented above suggests that this particular channel is available to, and more likely to be used by, the high income insured individuals. To more closely investigate the connection between income and pharmacy visits separate regression models were fitted for probability and number of pharmacy visits, using a flexible spline specification for income variable. These results suggest that pharmacy use is a normal good, with a positive (but imprecisely determined) income elasticity, in the lower income quartile. But with high probability it is an inferior good, with a negative income elasticity, in the two highest quartile. Unfortunately, the relatively large standard errors on the coefficients 6See Cameron and Trivedi (1998, chapters 3 and 9) for a detailed discussion of the count data models used here. 28 preclude a stronger statement. That is, pharmacy visits appear to be an inferior good for the rich, but a normal good for the poor. This result is different from that in some previous analyses based on the 1992-93 VLSS data (Gertler et. al, 1996) which suggest that pharmacy visits is a normal good at all income levels. The impact of HLTHINS on self-medication is found to be negative and statistically significant and sign-wise robust across a range of alternative specifications. The size of the impact is larger in models that do not control for clustering and which are not reported in Table 12. The fixed effects Poisson model yields the lowest estimate but even this is unambiguously negative and significant. Consistent with the results is the interpretation that self-medication is a risky form of health care and it is avoided as income rises and as alternative higher quality health care becomes available through health insurance. In Vietnam the higher quality care is provided in public hospitals. Note that the more highly educated individuals, and hence presumably those better aware of the risks of self-medication, also avoid pharmacy visits. The role of other factors - - such as being female, being married, having illness or injury, and the length of illness - -is similar to that which has been found for other types of health care; that is, they all increase the frequency of pharmacy visits. One difference is that AGE does have positive effect. The picture presented above is broadly consistent with the predictions of a theoretical model which treats self-medication as a risky alternative to professional care (see Chang and Trivedi, 2001). The reported results are robust and do not qualitatively change if the econometric analysis is carried out by insurance status or by using disaggregated age categories, or using other econometric estimators based on more flexible assumptions. 29 6 Analysis of Health Care Expenditure 6.1 Individual data This section is devoted to a regression analysis of medical expenditure. The dependent variable is log(expenditure) for each member of the household, conditional on a positive level of expenditure for that individual. All types of health care expenditure in the 4 week period preceding the survey are included. The sample size is 8081. The main focus in this analysis is again on the role of household income and HLTHINS. As before, we control for the AGE, SEX, MARRIED, EDUC, and health status ILL, INJ, ILLDAYS, ACTDAYS). The detailed regression results are shown in Table 13. The results indicate that whereas household income continues to show a strong explanatory power, the insurance variable is much less significant. The point estimate of the elasticity of individual health expenditure with respect to income is of the order of 0.3 to 0.4. The estimates of income elasticity from the fixed effects model are slightly smaller than those without fixed effects. The response to HLTHINS, however, is positive but with a relatively large standard error. This can be interpreted as follows: HLTHINS acts to divert demand from care of lower quality, such as that provided by commune health centers, to care of higher quality, such as that provided in public hospitals. The aggregate response of expenditure to insurance then would be small or zero if most of the impact takes the form such substitution. However, substitution in terms of number of visits need not have zero impact on total expenditure. If the substitution is towards care of higher quality, then total medical expenditure may increase. Such a situation seems consistent with the estimates we have obtained. The insurance effect on health care expenditure reflects substitution towards higher quality care. 30 The reported estimates are robust. Point estimates of income elasticity similar to that from the fixed effects model are obtained from several other variants including the following: separate models for insured and uninsured samples; random effects version of the estimated model; model in which income coefficient is allowed to differ by income quartile. To save space the details of these results is not included in Table 13. 6.2 Household data Analysis of medical expenses aggregated across all household members serves as a useful check on the results from individual data. It also yields estimates of Engle curves for medical expenditure. The main limitation of this approach is that we cannot control for health status of individual members of the household, and we have seen that to do so is important. We can control for some of the other relevant variables such as location (urban or rural) and size of the household, SEX, AGE and educational attainment of the head of the household. Descriptive sample statistics show that on average households with insured head spend about 20% more on heath care than those without. Average household health care expenditure is higher for urban than rural household despite the fact that rural households are typically slightly larger. Detailed regression results are in Table 14. The most interesting result is that point estimates of the income elasticity is around 0.6- 0.7, varying somewhat with the exact definition of total expenditure that we use. Health care in total is thus found to be a normal good, but not a 'luxury good". If total health care expenditure is a small part of the household budget, endogeneity of the total expenditure may be ignored. However, this assumption is easy to relax. 31 Instrumental variable estimates are given in the last two columns. The income elasticity still remains around 0.6. Additional robustness checks do not indicate that this estimate varies significantly by urban or rural location, or by level of income. This point estimate of income elasticity is larger than the corresponding estimate for individual expenditures. One possible explanation is that the estimate from aggregate household data may be upward biased because of the failure to take into account the effect of HLTHINS which, as was seen earlier, is positively correlated with the household income level. That is, the role of insurance has been absorbed into the income elasticity, causing it to become somewhat inflated. A second possibility is simply that demand for health care at the level of the household is indeed more elastic than it is for a single individual. Age and sex of the head of the household are also significant factors. On average households with female heads spend more on health care, and households with older heads also spend more. Educational level of the head of the household is not found to be a significant explanator. 7 Discussion of Policy Issues In this section we discuss three health care policy issues: the implications of the pervasive phenomenon of self-medication; the future of the CHCs; and the expected future changes in the pattern of health care utilization. This article has documented the pervasiveness of self-medication and the factors that promote it. The phenomenon is a common one in most developing countries. The World Health Organization (WHO) explicitly recognizes that self-medication has an important role to play in 32 most health care systems, and has observed that with the continued improvement in people's education, general knowledge, and socio-economic status, self-medication has been successfully integrated into many health care systems around the world. In a less developed economy like Vietnam, where both public and private health care infrastructure is relatively basic, self- medication is an important form of health care. Freer availability of drugs contributes to household welfare. World-wide, the purchase of prescription-only drugs without a prescription is far more common than the sale of over-the-counter drugs.7 Opportunities for self-medication are enhanced and aided by the Internet and by the deregulation of over-the-counter sales of pharmaceutical products with active ingredients. These tendencies have harmful consequences, of which the growing ineffectiveness of antibiotic drugs is the most alarming and most visible. When small doses are used to treat bacterial infections, presumably because the user cannot afford the cost of the prescribed full course, instead of clearing the infection the practice promotes the growth of antibiotic resistant strains of bacteria. This reduces the future potency of antibiotics for all users. Previous World Bank reports reflect the concern expressed by the medical profession and public health organizations. Inadequately supervised and administered drugs, often in incorrect dosages, are a major contributing cause of the growth of antibiotic resistant bacteria. The problem is a serious concern because it involves a negative intertemporal externality - - current actions of an individual have a negative future impact on the society as a whole. The public health issue is how to combat this problem. This article has found econometric evidence that the practice of self-medication is negatively associated with educational attainment, income levels, and the relative price and 7 WHO Drug Information, 14(1), 2000. This paper cites a consumer interview study carried out in six Latin American countries that found that only 34% of the dispensed medicines were classified as OTC. 33 accessibility of alternative providers. There is also strong a priori reason to believe that supply side deregulation has lowered the price of, and improved access to, drugs, and thereby encouraged self-medication. On the demand side we can expect that economic growth will reduce the seriousness of the problem. But this view needs to be qualified. Although it appears that in the highest income quartile self-medication is an inferior good, any reduction from this source due to income growth will only make a small contribution to net reduction. The larger positive contribution from lower income groups may outweigh the negative tendencies. Moreover, economic growth has been more rapid in urban than in rural areas, so the reduction in harmful types of self-medication in rural areas is likely to be even slower. Moreover, if affordable alternatives to self-medication exist, then its use is more likely to decline as there is increasing awareness of its dangers. Again such alternatives are less accessible in rural areas. These arguments suggest that in the absence of regulatory constraints the practice of self- medication is likely to persist. A discussion of the appropriate form of regulation is outside the scope of this study. Previous World Bank analyses, in expressing concern about the extent and continued growth of (harmful) self-medication in Vietnam (see Gertler and Litvack (1998), VHSR (1999)), have suggested that this growth is in part a consequence of low quality of care available to the lower income groups. A variety of policy prescriptions have been put forward. Gertler and Litvack (1998, pp. 246-247) suggest that improving the quality of service at the CHCs, for example by improving the supply of low-priced generic drugs, would reduce self-medication. The VHSR (1999, pp. 120-21) also mentions the proliferation of counterfeit and sub-standard drugs available on the market. It goes on to describe the recently initiated Vietnam National Drug Policy also aimed at rational and safe use of drugs. Our results suggest that an expansion of 34 the voluntary health insurance program would also have a qualitatively similar effect. It appears that enrollment into the voluntary health insurance program has stalled, although the reasons for such stagnation are unclear. However, our results indicate that enrollment into the insurance program is responsive to income growth. Again this suggests that continued growth will reduce the problem. However, the poorer rural sections of the population will benefit from this development more slowly and to a lesser extent. This article finds evidence that strongly suggests that commune health centers provide an inferior service whose consumption declines with income and education. Our analysis is not sufficiently detailed to pinpoint the CHC characteristics that are responsible for their decline. But it is plausible that the readily available alternative of self-medication is partly responsible. Note that it has been suggested that CHCs fail to carry adequate stocks of cheap generics and are generally poorly staffed and equipped. Given the continuation of the present trends, the CHCs may become even less important in the future. The growth of private health facilities is a relatively recent phenomenon. Our results do not suggest that income growth has a strong impact on the growth of this sector. It seems likely that this sector provides services to those who desire a higher quality of care than available outside public hospitals, but also are either ineligible or unable to get treated at public hospitals. Evidence indicates that higher income individuals with insurance get treatment and drugs at public hospitals. Currently, public hospitals seem to be the main and perhaps the only source of quality care. With economic growth will come increased demand for health care of higher quality, especially in urban areas. As long as the "catchment area" of the private health facilities is restricted directly or indirectly, future economic growth will put greater pressure on public hospitals, unless other alternatives superior to those currently available can be found. If, on the 35 other hand, the insured individuals can be treated at Private Health Facilities, then there will be an alternative to public hospital care. This will reduce pressure on public hospitals that is bound to arise if the income growth continues to be robust. Whether this will raise the average quality of care, or reduce rampant self-medication, would seem to depend upon the regulatory constraints that apply to the Private Health Facilities. 8 Summary and Conclusions From previous analyses of 1992-93 VLSS data a stylized pattern of health care utilization in Vietnam has emerged. Broadly, according to this stylized view, the richer sections of the population get their health care at public hospitals, the poorer at the commune health centers, and all groups use self-medication heavily, causing the latter to dominate as the principle source of health care. In this picture, other providers such as private health facilities, traditional medicine practitioners, home providers and so forth play a relatively minor role. The stylized description has little to say about the impact of health insurance on the observed use pattern. This study provides some confirmations as well as modifications of the above stylized description. * Evidence supports the view that those who are either ill or injured have ready access to some form of health care. At all levels of income, the commonest response to the need for health care is some form of self-medication. * The private providers collectively are more important as a fraction of total health care spending than the commune health centers. * The results suggest that both self-medication and commune health centers are inferior goods by the usual definition that their demand declines with rising household incomes. 36 Self-medication appears to be an inferior good especially for high income households, but a normal good at low income levels. In the aggregate, however, the net effect of income on pharmacy visits is estimated to be close to zero. These results are consistent with the view that both self-medication and commune health centers are low-quality and risky forms of health care in Vietnam. * Within the existing income distribution estimates show a negative relation between the use probability of commune health center and income. This negative relation is less robust in the lowest income quartile. * Health care provided by private health facilities is weakly related to income overall, but may be positively related for the lower income groups. * There is a strong positive relation between income and the use of both inpatient and outpatient care provided at public hospitals. * The net impact of health insurance on self-medication and the use of private health facilities is negative. That is, under the current organization of health care delivery, having health insurance diverts patients away from private health care and self- medication mainly towards public hospitals, and to a lesser extent towards commune health centers. The growth of services in the private health facilities is therefore affected in opposite directions by rising income and rising proportion of insured population. * Income, insurance status and age are the three major determinants of in-patient care (hospital nights) in public hospitals. * There appears to be no evidence that health insurance has had a significant impact on the total household out-of-pocket health care expenditures (excluding care at in-patient care 37 at public hospitals) in either direction. That is, much of the impact seems to be in the form of redistribution of care between types of providers. * Previous analyses have expressed serious concern about the dominant role of unsupervised and unregulated self-medication in Vietnam. It has been suggested that this is made possible by easy availability of a very wide range of pharmaceutical drugs. Both rising incomes and growth of health insurance reduce the extent of self-medication, but it is not clear whether these deterrents are strong enough. * Aggregate household income elasticity for health care is higher than indicated by previous studies. There are some qualifications. We have treated, with some justification, both health insurance and household incomes as weakly exogenous variables. No allowance has been made for measurement error in income. Standard statistical arguments suggest that this may cause us to underestimate the impact of income on the demand for health care. Second, we have distinguished between different types of health insurance only indirectly, and this may have caused some aggregation bias of an indeterminate nature. Although we have provided a statistical model of the probability of enrollment in the health insurance program, we have not, for reasons of lack of data, provided any estimates of price sensitivity of insurance demand. This is an important qualification to our finding that enrollment in the program is strongly associated with income. Finally, it is useful to review the main policy implications of the findings, which were presented in Section 7. First, the results suggest several avenues for reducing the heavy (and often inappropriate) reliance of Vietnamese on over-the-counter antibiotics. Increases in household income and education levels should reduce this reliance over the long-term. In 38 addition, expansion of the new voluntary health insurance program would reduce reliance on self-medication of antibiotics and other drugs. Second, there is clear evidence of dissatisfaction with health services provided by commune health centers; people switch to other sources of medical care as their income rises. Further research is needed on the source of this dissatisfaction; among the possible sources are inadequate stocks of medicine and inadequate staff. If no changes are made commune health centers will become an even less important source of health care for the Vietnamese population. Finally, the role of private health care providers in Vietnam needs further development. In the long run, they can provide high quality hospital care, unless restrictions prevent them from doing so. This would serve as an additional source of such care for individuals in Vietnam's voluntary health insurance program, and as that program expands the capacity to meet the demand for such health care must be expanded as well. 39 References Behrman, J.R. and J.C. Knowles (2000), "The Demand for Student Health Insurance in Vietnam", unpublished paper. Behrman, J.R. and J.C. Knowles (2000), "The Demand for Health Insurance in Vietnam: An Application of Contingent Valuation", unpublished paper. Cameron, A.C., and P.K. Trivedi (1998), Regression Analysis of Count Data, Econometric Society Monograph 30. New York: Cambridge University Press. Cameron, C., P.K. Trivedi, F. Milne, and J. Piggot (1988), "A Microeconometric Model of the Demand for Health Care and Health Insurance in Australia", Review of Economic Studies, 55, 85-106. Chang, F-R., and P.K. Trivedi (2001), "Economics of Self-medication: Theory and Evidence". Working paper. Deaton, A. (1997), The Analysis of Household Surveys, Baltimore: Johns Hopkins Dollar, D., P. Glewwe, and J. Litvack (1998), Household Welfare and Vietnam's Transition. World Bank Regional and Sectoral Study. Washington, D.C.: The World Bank. Gertler, P. and J. Litvack (1998), "Access to Health Care during Transition: The Role of Private Sector in Vietnam", chapter 8 in Dollar, D., et al., Household Welfare and Vietnam 's Transition, 235-256. Gertler, P., J. Litvack and N. Prescott (1996), "Health Care During Transition: The Role of the Private Sector in Viet Nam", mimeo. Hsiao, C. (1986) Analysis of Panel Data, Cambridge: Cambridge University Press. World Bank (1999), Vietnam Health Sector Review. World Bank report. 40 World Bank (1995), Vietnam: Poverty Assessment and Strategy. Report of The World Bank Country Operations Division Country Department 1. East Asia and Pacific Region. 41 Table 1: Definitions and descriptive statistics Variable Definition Mean Std. Dev PHARVIS number of pharmacy visits .51 1.31 PHARDUM 0/1 dummy for pharmacy visit .26 .44 GOVHOSP number of hospital outpatient visits .049 0.41 PHF number of visits to private health facility .11 0.80 CHC number of visits to commune health centers .04 0.34 HOSPADM 0/1 dummy for hospital admissions .051 0.22 HOSPNITE number of hospital nights 13.53 21.82 MEDEXP Total medical expenditure 1,520 5,139 (4W+12m) (4 weeks plus 12 months) HHEXP Total norninal household expenditure 15,273 13,020 ln(INC) log(HHEXP) 2.60 0.62 lnmedexp ( >0) log(MEDEXP) for positive expenditures 2.14 1.08 Inmedexp (insured) Inmedexp for the insured 2.32 1.11 lmnedexp (uninsured) Inmedexp for the uninsured 2.12 1.08 HLTHINS 0/1 dummy for health insurance status 0.16 0.37 AGE age in years 29.7 9.67 SEX 0/1 dummy for gender .51 .49 MARRIED 0/1 dummny for marital status .40 .49 EDUC completed years of education 3.38 1.94 ILL 0/1 dummy for illness in4 weeks before survey 0.41 0.49 INJ 0/1 dummy for injury in 4 weeks before survey 0.009 0.098 ILLDAYS number of days of illness/injury in 4 weeks before survey 2.80 5.45 ACTDAYS number of days of lirnited activity in 4 weeks before survey 0.06 1.11 URBAN 0/1 dummy for urban household 0.29 0.45 HHSIZE household size 4.73 1.96 42 Table 2: Annualized health service contact rates by provider GOVHOSP CHC PHARVIS OTHER PHF TRAD Public All Providers Providers 1993 0.32 0.19 2.14 0.03 0.66 0.03 0.54 3.37 1998 0.60 0.57 6.78 0.25 1.76 0.36 1.43 10.33 Source: World Bank, Vietnam Health Sector Review, 37-38. Notes: CHC: commune health center GOVHOSP: govemment hospital PHF: private health facility PHARVIS: visits to pharnacy or drug vendor TRAD: traditional (Eastern) practitioner OTHER: other government facilities. 43 Table 3: Mean number of visits to different health care providersbigskip N CHC GOVHOSP PHF PHARVIS TRAD HOMEVIS Full sample: mean 27331 .040 .049 .113 .512 .022 .038 Sicksample: mean 11322 .096 .113 .274 1.214 .054 .091 Insured sample 4496 .041 .105 .084 .406 .014 .037 Sick/insured sample 1818 .100 .248 .198 .971 .033 .086 Percentofhealthexpend. 8081 .96 3.22 2.61 92.17 .33 .41 Notes: CHC: commune health center GOVHOSP: government hospital PHF: private health facility PHARVIS: visits to pharmacy or drug vendor TRAD: traditional (Eastern) practitioner HOMEVIS: home visits 44 Table 4: Average contact rates with providers by health status Income class Commune Govt. Hosp Private HF Pharmacy Other All Sick All Sick All Sick All Sick All Sick Lowest 10% .0659 .1271 .0299 .0532 .1002 .1967 .5660 1.0688 .0279 .0586 Lowest 25% .0596 .1219 .0336 .0672 .0995 .2089 .5595 1.1368 .0256 .0537 Highest 25% .0119 .0326 .0762 .2005 .1406 .3878 .4526 1.2323 .0221 .0086 Highest 10% .0075 .0211 .1081 .2920 .1398 .3989 3985 1.1407 .0096 Source: Author's calculations Notes: The samnple sizes are as follows: Full sample- Lowest 10% - 2773; Lowest 25%- 6922; Highest 25% - 6939; Highest 10% - 2775. Sick sample- Lowest 10% - 1408 ; Lowest 25%- 3273; Highest 25% - 2483; Highest 10% - 945; 45 Table 5: Medical expenditure by households Household type Expend. 1 Expend. 2 N Average 1520 768 5999 Average ly>O 1822 788 5006 Urban 2000 997 1730 Rural 1325 674 4269 Farm 1772 942 2561 Nonfarm 1331 637 3438 Notes: Expend. 1: Medical expenditure; Expend. 2: medical expenditure excluding insurance 46 Table 6: Health Insurance and income status Income class Lowest 10% Lowest 25% Highest 25% Highest 10% Percentage insured 8.7 9.2 24.5 27 47 Table 7: Frequency Distribution of health service contacts Number of contacts Pharmacy GOVHOSP PHF CHC Other HOSPNITE 0 20,639 26,796 26,481 27,041 27,158 42 1 3,827 736 540 486 254 71 2 1,716 133 316 111 124 100 3 776 30 180 52 66 122 4 359 17 99 23 36 74 5 174 8 36 6 24 92 6 64 4 12 3 12 54 7 43 1 18 4 16 274 8 16 3 4 1 9 32 9 4 2 3 3 2 20 10+ Notes: sample size: 27,731; for HOSPNITE sample size = 1,463 GOVHOSP: outpatient contacts only. HOSPADM: admissions in government hospitals. HOSPNITE: nights in government hospitals. 48 Table 8: Determinants probability of health insurance enrollment Robust Logit Robust Logit Fixed effects logit Fixed effects logit Coeff. Std.err Coeff. Std.err Coeff. Std.err Coeff. Std.err cons -4.53 .2985 -3.9217 .6039 - - - - ln(INC) .3645 .0611 - - - - .5868 .0452 AGE .3786 .0412 .0105 .0011 .0211 .0018 SEX -.2940 .0368 -.2966 .0366 -.3509 .0365 -.2456 .0388 MARRIED -.5099 .0726 -.5008 .0733 -.4266 .0464 -.8005 .0616 EDUC .2952 .0226 .2963 .0227 .2045 .0105 .2144 .0110 INCI - -.0169 .2330 .1171 .1509 - INC2 - 1.3532 .2871 1.0391 .2234 - - INC3 - .3967 .2343 .6203 .1684 - - INC4 - .0837 .1177 .3004 .0762 - - -In-lik 11268 11242 9422 8220 Sample Full Full Full 22