|Home | About | Journals | Submit | Contact Us | Français|
This study uses the second National Family Health Survey (NFHS-2) of India to estimate the effect of state-level public health spending on mortality across all age groups, controlling for individual, household, and state-level covariates. We use a state’s gross fiscal deficit as an instrument for its health spending. Our study shows a 10 % increase in public spending on health in India decreases the average probability of death by about 2%, with effects mainly on the young, the elderly, and women. Other major factors affecting mortality are rural residence, household poverty, and access to toilet facilities.
One of the most fundamental yet unresolved issues in health policy is whether public spending on health care improves health outcomes. Political concerns (especially demands by the general public), economic considerations (public goods, externalities, catastrophic cost, and cost-effectiveness), and ethical reasons (horizontal and vertical poverty reduction) have been driving forces behind national and international agencies advocating for more public spending (Musgrove, 1999), while budget deficit makes many governments consider public expenditure cuts. Therefore, it is important to estimate the extent to which population health is affected by public spending.
International cross-country studies of the relationship between public health spending and health outcomes, relying on aggregate health indicators, usually find little effect of public health spending on health outcomes. The level of income is often found to be the major determinant of a population’s health status, while public health spending is a relatively poor predictor of cross-country differentials in health indicators (Kim and Moody, 1992, Filmer and Pritchett, 1999, Musgrove, 1996, McGuire et al., 1993, Gupta et al., 2002). In their influential study, Filmer and Pritchett (1999) find that 95% of cross-national variation in mortality can be explained by a country’s income per capita, inequality of income distribution, extent of female education, level of ethnic fragmentation, and predominant religion, while public health spending explained only one-seventh of 1% of the observed differences in mortality across countries. These results confirm the findings of previous studies that had concluded that poverty and income, rather than public health spending, are the crucial determinants of health status (Carrin and Politi, 1995, Demery and Walton, 1998).
Similar research limited to OECD countries also shows a very weak relationship between public spending on health care and premature mortality (Or, 2000). In contrast, a few cross-country studies, limited to poor countries, find that public health spending has a statistically significant effect on health status (Anand and Ravallion, 1993, Hojman, 1996, Bidani and Ravallion, 1997)
Bidani and Ravallion (1997) use the two-dollar-per-day poverty line to decompose health indicators into subgroup averages for 35 countries using a random coefficients model. They find that the poor have comparatively worse health status, and that they are affected more by public spending on health care, than the better off. Gupta et al. (2003) use King’s methodology1 (1997) and demographic and health survey (DHS) data from 70 developing and transition economies, to extend Bidani and Ravalli’s approach, finding similar results. These results suggest that public spending on health may have different effects on the health of people of different socio-economic status (SES), due to differing levels of need and ability to substitute private spending for public spending.
A major weakness in much of the existing literature is the use of aggregate data. Utilizing aggregate data on indicators often mask important variations in health status and health service use by individuals living in households with different socio-economic backgrounds. Using household level data allows a more detailed analysis of health outcomes and the effects of public health spending. Some studies, for example, report that mortality is higher for girls (Kishor, 1993) and for rural, low-income, agricultural households (Schultz, 1993). Urban areas may also provide better access to health facilities (Stanton, 1994). Therefore, to have reliable estimates of the effects of public health spending on mortality we need to consider factors affecting health at state-, household- and individual-levels.
A second methodological issue in estimating the link between spending and health outcomes using cross country is the lack of comparability of data on both mortality and government health expenditure across countries. There is a great heterogeneity in health care provision, and mortality rates, across countries at different stages of the epidemiological transition, which makes pooling these disparate countries in a single analysis problematic.
Some studies have avoided the difficulties imposed by data heterogeneity inherent in cross country international analyses by using sub-national data (Pierre-Yves Crémieux, 1999, Deolalikar, 2005, Bhalotra, 2007). The study by Cremieux et al. examines the relationship between expenditure and outcomes across ten Canadian provinces over the fifteen-year period 1978-1992. They find that low healthcare spending was associated with high infant mortality rate and low life expectancy.
Previous studies of the effect of public health spending on mortality in India are inconclusive. Deolalikar (2005) uses an Indian state panel for 1980-99 and finds that current health expenditure does not have a significant effect on mortality rates, once state fixed effects and a linear time trend are included in the model. Bhalotra (2007) on the other hand, restricts the sample to rural households and allows for lagged effects, and finds a significant effect of health expenditure on infant mortality rates, with the long run elasticity being about −0.24.
In this study we explore the effect of variation in public health spending across the states of India on individual mortality using household level data from the National Family Health Survey (NFHS II) conducted in 1998-99. There are numerous advantages to using micro-data. Pooling data from states within India raises less of a problem of heterogeneity than comparing outcomes across countries. We can control for a wider range of variables at different levels of aggregation; individual, household and state level characteristics can be included in the analysis. The data on individual mortality also allows us to explore heterogeneity in the effect of health expenditure in sub-populations such as the rural versus urban population, men versus women, and different socio-economic groups.
India is a large federated union of states, where state governments play a larger role than the central government in funding health service provision. Health is a Concurrent subject2 under the Indian Constitution, but State governments are dominantly responsible for health provision, both in terms of health care and public health measures. State governments account for about two-thirds and the Central Government about one-third of the total public spending on health (National Health Accounts India, 2005). This decentralization of funding means that there is a wide range of public health spending across states of India due to different policy choices by State governments. While there are some specific central interventions, especially in various “Missions” as well as high-end curative facilities, the bulk of the health provision that affects most of the citizenry is the result of spending by State governments (Das Gupta and Rani, 2004).
As in Bhalotra (2007) we find effects of health spending on mortality outcomes. The difference in our approach is that we look at mortality over the whole age range while Bhalotra focuses on infant mortality. A cost of our approach is we only have a single cross section of data, while she was able to use retrospective fertility histories to construct a time series for infant mortality. This means we cannot control for state level fixed effects, or dynamics, in our specification. We address the problem of unobservable state level variables that may affect both health spending and health outcomes by checking the robustness of our results using an instrumental variable. We instrument public health spending with the overall state budget deficit, the argument being that a higher state deficit will be associated with lower health spending, but will not otherwise directly affect health outcomes. We use public spending in the year before the survey to explain observed mortality. Due to data limitations we do not address the issue of longer lags in the response of health outcomes to health spending and the provision of health services, though these lags may be important (Farahani et al., 2009).
The main objective of this study is to evaluate the relationship between public spending on health at the state level and the probability of death at the individual level. The specific research questions are: Does public spending affect the probability of individual mortality? And if so, are the effects different by sex, urbanization, socio-economic level, or age group? We find a significant effect of public health spending on mortality rates. However this effect is predominately for women, the young (under 18 years old) and the elderly (over 40 years old). Surprisingly we find no differential effect of public spending by urban-rural status, or by asset quintile (a proxy for socio-economic status).
We also investigate the effect of public health spending on the prevalence of a number of illnesses reported in NFHS II, namely asthma, tuberculosis, malaria, and jaundice. We find no significant effect of public health spending on these illnesses, though as we argue below this finding may be due to reporting errors. Our analysis leaves open the issue of the mechanism through which public spending is affecting health.
We use the second National Family Health Survey (NFHS II) conducted in 1998-99 in all 26 states of India existing at that time. We use the NFHS II rather than NFHS III, which was released in 2007, since mortality data for all age groups were collected only in NFHS II.3 This is a rich dataset containing information on individual characteristics and mortality. The survey response rate ranged from 89% to almost 100%, with 24 of the 26 states having a rate of higher than 94%. This dataset is supplemented by state-level data on government health spending, average household spending on health care and state gross income.
The analyses are based on a representative cross-sectional of 519,502 individuals from 91,573 households in 26 Indian states. The survey collected extensive information on household members and their characteristics. The data on each household member, including those who had died in the previous two years, consist of information on age and gender for both living and dead household members. The household survey provides current information on caste, religion, and household assets, which was linked to members who were alive at the time of the survey as well as to the deceased members. This linkage assumes that the household members who died in the two years prior to the survey had a standard of living, caste, and religion similar to those of household members who were still alive (Subramanian et al., 2006).
The Household Questionnaire asks questions on current household members and deaths occurring to usual residents of the household during the two years preceding the survey. The outcome of interest is whether the person is reported as dead or alive at the time of survey. The explanatory variables will be at three levels: individual, household and state. At the individual level the only variables common to both the dead and alive were age and sex. For the household covariates, following the literature (Subramanian et al., 2006, Esrey and Habicht, 1986), we use asset index quintile, caste, religious affiliation, residence (urban, rural), and access to safe water and sanitation.
Age was divided into five age groups: namely, infant (0-1), child (1-5), youth (5-18), young adult (18-40), and above 40 years of age. Because probability of dying increases steadily after age 40 we also use a continuous variable for number of the years the person lived after age 40. Religious affiliation was categorized into Hindu, Muslim, Christian, and others.
The standard of living is proxied by an asset index, constructed from the household’s ownership of assets such as an iron, radio, television, or car. The econometric evidence suggests that the asset index is as reliable as conventionally measured consumption expenditures (Filmer and Pritchett, 2001). To construct the asset index, we use principal component analysis, following the method set out Filmer and Pritchett (2001). The asset index is a weighted sum of the assets owned and does not have a natural interpretation as an absolute value. For the purpose of our analysis, we think of it simply as an ordinal ranking and we divide the index into quintiles.
The variable “caste” has five categories: scheduled caste, scheduled tribe, other backward classes, other caste and no caste. These are the terms used in the Census of India. The ‘other castes’ are considered to have the highest social status among these designations and the ‘scheduled caste’ and ‘tribe categories’ the lowest. Scheduled castes and scheduled tribes are recognized by the Government of India as socially and economically disadvantaged and in need of special protection from injustice and exploitation. The government of India designates members of scheduled caste, scheduled tribe and other backward class as eligible for affirmative action. We classified members of groups for which caste may not always be applicable and participants who did not report any caste affiliation in the survey as “no caste.”
Access to water is divided into three categories, regardless of the ownership (private or public): access to piped-water, access to well water, and other sources. Similarly, access to toilet facilities was divided into households that have access to a flush toilet, a latrine, or other means for human waste disposal.
We add state level data on government spending on medical expenses and public health (hence forth called public health spending) obtained from the Reserve Bank of India (Reserve Bank of India, 2007). We include all spending in the state on health, whether financed by the central government or the state. We did not include the government spending on family welfare, for that budget line is mainly allocated to family planning purposes.4 Given the fact that out-of-pocket expenditure accounts for the major part of health care spending in India, it would be best if we had data on private household health spending. However, because this data was not collected in NFHS, we used the Household Consumer Expenditure in India for 1998 to calculate the average household out-of-pocket medical expenses in each state in that year and use this as a state level explanatory variable.5 We also control for the overall level of economic development of the state, by including 1998-1999 per capita net state domestic product, at current prices, obtained from India Reserve Bank (Reserve Bank of India, 2007).
The individual and household level covariates allow us to control for these variables in our analysis. There are wide disparities in health outcomes in India by individual and household characteristics. In spite of an increase in public funds in health care, improvements in health have not been shared equally. Infant mortality declined by 68 percent (from 37 to 12 per 1,000) in Kerala but by only 52 percent in Uttar Pradesh (from 150 to 72) between 1981 and 2004. Uttar Pradesh alone contributes one-quarter of all infant deaths in the country and, along with Madhya Pradesh, Bihar, and Rajasthan, accounts for slightly more than half of the total. During 1994-1999, just 10 percent of India’s 602 districts accounted for nearly 30 percent of all infant deaths, and 23 percent of the districts accounted for approximately half of the total. Furthermore, there are variations across social groups. Scheduled Tribes have an under-five mortality rate of 117 deaths per 1,000 live births, and Scheduled Castes have a rate of 108, compared with an all-India rate of 93 (Yip and Mahal, 2008).
A multilevel probit model is used to estimate the effects of health spending at the state-level, on probability of death at the individual level, controlling for covariates at individual-, household-, and state-level (Leyland and Goldstein, 2001). The advantage of this model is that it simultaneously considers the household-level and individual-level predictors while allowing for non-independence of observations within groups (Snijders, 1999). This analysis of public spending is at the state level, which is meaningful for policy decisions; we lack data on public health spending at lower levels of aggregation.
The outcome of interest is binary response (dead or not) for individual i in cluster j. Assuming the binary response, yij is Bernoulli distributed with probabilities πij : yij ~ Bernoulli(1, πij). If πij = P(yij = 1) denotes the probability the ith individual is dead, the probit model is represented as
Where Φ−1(πij) is the inverse cumulative distribution function of the standard normal and i and j refer to the individual and cluster, respectively:
A challenge in studies of this kind is that the observations are nested within units and super-unit structures making it likely that the individual observations are not independent. Failing to deal with this lack of independence leads to biased parameter estimates. We estimate equation (1) using the probit MLE estimator. All standard errors reported in this paper are robust and clustered by the primary sampling unit (PSU). We adjustment for conditional heteroskedasticity and for correlation within clusters (Cameron and Trivedi, 2005). This procedure corrects our standard errors for the fact that the sampling design randomly selects primary sampling units (and then purposively selects within these) rather than randomly selecting individuals.
A major concern is the possibility that public health spending is endogenous, in the sense that it is to some extent been influenced by the level of health, or is driven by some unobserved variable that also affect health outcomes. Our solution to this problem is to use an instrumental variable (IV) (Wooldridge, 2002). A good IV will be correlated with public health spending and simultaneously exogenous to probability of dying. By using instrumental variables estimation, we present an important piece of evidence that the relationship we observe is causal.
The use of the instrumental variable approach is, of course, dependent on the availability of a valid instrument. Since resource allocation to health sector in India is influenced by the general fiscal situation of the respective State Governments6, we use the state gross fiscal deficit as the instrumental variable. A two-stage probit regression is used to control for the endogeneity of public health spending. The first-stage regression uses ordinary least squares to predict the public health spending as a function of the state gross fiscal deficit, the instrument, and other explanatory variables. In the second stage we use the predicted value of the health spending plus other exogenous variables to estimate the probability of death. More specifically, the IV estimator uses the instrument to predict the value of the potentially endogenous regressor. The predicted values are then used as a regressor in the original model. Under the assumptions that the instrument is correlated with the endogenous explanatory variable but has no direct association with the outcome under study, the IV estimates of the effect of the endogenous variable are consistent (Bound et al., 1995).
There are large variations in both public and private health spending per capita across states in India. Figure 1 shows public and private health spending by state; the states being placed in descending order according to the government’s per capita expenditure on health as a share of total health spending. Figure.2 shows the relationship between total health spending and the percentage that is publically funded. There is little relationship between these variables indicating that that low public expenditure may play a role in generating higher levels of private expenditure, so there is not a one to one mapping from public health spending to total health spending. This “crowding out” of private health spending is a potential reason why public health spending may be ineffective.
We focus on the effect of state level public health spending. It is noteworthy that there are rural-urban differentials in public spending, with per capita spending in urban areas generally much higher than in rural areas for most states (Peters et al., 2002), which may produce differential results in terms of health outcomes in urban and rural areas.
Public health expenditures seems to have a direct impact on the infant mortality rate (Bhalotra, 2007). Figure 3 shows the relationship per capita health spending of the State government and infant mortality7. Richer states tend to spend more per capita, and generally have lower infant mortality rates (IMRs). However, this is not always the case. Some of the richer states, such as Maharashtra and Haryana, have comparatively low per capita public expenditure on health. On the contrary, Sikkim, with relatively low per capita income, has high per capita public spending on health.
Table 1 reports summary statistics, based on data for individuals, with household and state level variables being linked to each individual. Overall, the probability of dying in the two year period under study in this sample is about 2 percent. Average per capita public health spending is 125 Rupees, ranging from 49 (Bihar) to 406 Rupees (Goa). Rural population comprises 69% of the sample population. Only 45% of the population under study has access to a toilet (28% to a flushed toilet and 17% to a latrine). Less than half (42%) of the people have access to piped water. Note that the 20% of households are in each asset quintile, but the percentage of people in each quintile may differ due to different household sizes across quintiles.
Table 2 presents the results of the probit model of mortality displayed in the methods section. We use the log of public health spending in the state for 1997-98 as the main explanatory variable. This model shows the inverse association between public health spending and mortality. Public health spending has a significant effect and reduces mortality rates. We estimate that a 10 percent increase in per capita public health spending, i.e., about 13 Rupees, in an average state with all variables at their mean values, decreases the probability of death for an average individual by about 0.0004. Given the overall mortality rate of 2% in the sample, this is a reduction in mortality rates of about 2%, giving an elasticity of mortality with respect to public health spending of about −0.2.
The estimated effect of the log of private household spending on health has a similar impact on health outcomes. Most health spending comes from private sources8; Berman(1998) shows that 82 percent of people in rural areas and 79 percent in urban areas go to private providers for primary care. Our results suggest that the same percentage increase in either public or private spending has similar effects. Given that public spending is much less than private spending this means that each rupee of public health spending is about 4 times more effective at reducing mortality as private health spending.
In table 2 we find that the log of per capita net state domestic product is not a statistically significant predictor of mortality. It contrast to the cross country literature that stresses the importance of average income in determining mortality (Carrin and Politi, 1995, Demery and Walton, 1998, Filmer and Pritchett, 1999), our findings indicate a more important role of household economic status, as proxied by the asset index level. We find higher mortality rates among households in the lower quintiles of the asset index, with mortality being more than 1.1 percentage points higher in the bottom quintile than the top quintile (the baseline). Given an overall mortality rate of around 2%, this means mortality is about 50% higher among the poorest quintile.
We find higher mortality in rural areas as compared with urban areas. We find access to a flush toilet or a latrine significant in reducing mortality, though there are no significant effects of water sources on mortality which his is consistent with prior evidence in the literature (Lee et al., 1997). We find no significant effects of caste or religion on mortality.
We find significantly higher mortality among men as compared to women. We take as our reference group infants aged zero to one year. Children aged one to five years have significantly lower mortality than infants, and mortality tends to fall with age at first. The lowest mortality is among those aged 14-40. Mortality rises with age after age forty.
Table 3 reports effects of interaction terms between public health spending and gender, place of residence (urban, rural), and socio-economic group of the household as well as with the age group of the individual. This interaction effects allow us to investigate if the effect of public spending on health is uniform across groups or is concentrated on specific populations.
Model 1 in Table 3 investigates the effect of public health spending on mortality by sex. The default group is women and we find that after adding this interaction the direct effect of public spending becomes twice as great as the model with no interaction (in Table 2) while the interaction term between public health spending and being male is positive. This indicates that public spending on health has a significantly larger impact on the mortality of women than men.
In Model 2 we investigate the effect of interacting public health spending with rural residence. We do not find a significant effect of the interaction term, indicating similar impacts of public health spending across urban and rural areas. Navaneetham and Dharmalingam (2002) also show that in the case of antenatal care, there is no significant rural-urban gap, due to the role played by the multipurpose health workers posted in the rural areas to provide maternal health care services.
Similarly, in Model 3, we find no differential impact of public health spending by the economic level of the household, proxied by the asset index quintiles. While we find that lower socio-economic status households have higher mortality rates we find no evidence that they benefit disproportionally from public health spending. The results of Model 3 contrast with previous studies which used more aggregated data and showed that the poor benefit more than the rich from public spending on health (Bidani and Ravallion, 1997, Gupta et al., 2003). Public spending in India has shifted significantly toward primary care (Mahapatra and Berman, 1995), resulting in improved access to public facilities by all income groups. However, the results of the India’s National Sample Surveys (NSS) of 1986-87 and 1995-96 showed a considerable decline in the utilization of public health services by the poor, especially the rural poor (Selvaraju, 2001). They show that the rich consumed public services at three times the rate the poor uses and had a much higher chance than the poor of admission to public health facilities.
Model 4 examines the effect of public health spending by age group. We find that the interaction term between the young adult (18-40 years old) age group and public health spending is statistically significant. The coefficient on this interaction is positive and almost completely offsets the overall reduction in mortality caused by public health spending in Model 4. This means that the 18-40 age group appears to receive little or no benefit in mortality from public health spending, with larger than average effects accruing to younger and older people. There is little differential in health care utilization by age group in India (Berman, 1998) but generally better health, and lower mortality of the 18-40 year olds may make the effects of health care less noticeable in this group.
The results in Tables Tables22 and and33 treat public spending on health as exogenous. It may be that there is a feedback from health needs to health spending, or some hidden variable that affects both health spending and health outcomes. Either of these would mean that the relationship between health spending and mortality rates we observe in tables tables22 and and33 was not causal. To address this concern our estimation strategy in this section is to identify the causal effect of public health spending on mortality using an instrumental variable. The instrument needs to be a variable that is correlated with public health spending but exogenous with respect to health outcomes, and it also needs to be a variable that is not driven by any unobserved “third” variable we might suspect might be causing both changes in public health spending and health improvement. Evidence from other countries suggests that whenever there is a fiscal consolidation and stress, social sectors like health and education are targeted for reduced budget allocations (Tanzi, 2000). In India, since public health at the state level is financed through general tax and non-tax revenue resources (the cost recovery from the services delivered has been negligible, at less than 2%), resource allocation is influenced by the general fiscal situation of the respective state governments (Selvaraju, 2001). As a result, an increase in the fiscal deficit and a general resource crunch results in budget cut in the health sector (NCMH, 2005). We use the state’s gross fiscal deficit as an instrument for its public health spending. We find that the deficit is indeed negatively correlated with the public spending on health and it seems unlikely that the fiscal deficit will be associated with mortality rates through other mechanisms9.
Table 4 presents the results of IV estimation of Equation (1) using the state gross fiscal deficit in 1997-1998 as the instrument10. Compared to the probit results without instrumentation (Table 2), the coefficients we estimate (−0.114 versus −0.120) and the marginal effects (−0.0044 versus −0.0045) are very similar. The similarity of the results with and without instrumentation indicates that our results are reasonably robust, though our first stage F-test of 5.51 indicates with our instrument may not be very strong (Wooldridge, 2002, Cameron and Trivedi, 2005) 11
In this paper we focus on the effect of public health spending on mortality. We would also like to measure the effect of public health spending on morbidity as well as mortality. However the data on the morbidity in NFHS II that covers all age groups are limited to the prevalence of asthma, tuberculosis, malaria, and jaundice. In table 5 we report marginal effects of our explanatory variables in a probit model explaining the prevalence of each of these illnesses using the same explanatory variables as were used in our mortality study.
In no case do we find that public state spending on health is a significant predictor of illness, even at the 10% significance level. We find higher prevalence of these illnesses in people living in households in the lower quintiles of our asset index. The prevalence of asthma and tuberculosis appear to increase with age, while malaria is higher among those over 1 year of age rather than below, while jaundice appears mainly in the 1-5 year old age group. Men appear to have higher prevalence of tuberculosis and jaundice than women. The effects of other variables on disease prevalence are not consistent, though there is some evidence that access to clear water is protective.
A major problem with these morbidity data is measurement error. The data are based on reports from the head of the household, there is a high risk of inaccuracies with regard to morbidity data due to errors in reporting and recall (Boerma and Sommerfelt, 1993, Manesh et al., 2007). Recall error has been well described in studies of self and proxy health reporting (Coughlin, 1990, Beckett et al., 2001). In addition, we only have disease prevalence for those alive at the time of the study; those who died from these illnesses are excluded (we do not have cause of death, or illness reports, for the deceased) which creates a selection bias. In addition, public spending on health may affect the outcome of these illnesses on health status, rather than their prevalence. For these reasons we do not regard our results for morbidity to be very reliable.
This leaves open the mechanism through which increased public health spending is affecting mortality rates. It may be that public health spending is particularly important in countries that still face a substantial burden of infectious disease where very cost effective preventive measures are available. It would be very interesting to see how public health spending affects mortality by cause of death, but our dataset lacks this information. We leave this question to future research.
We have identified a significant impact of public spending on health at the state level given individual-, household-, and other state-level variables, on the probability of dying in India. The effects seem to be concentrated on woman, the young, and the elderly. Our results bear out recent evidence finding significant effects of public health spending in poor populations. The other major factors determining mortality in our study are an individual’s age and sex, household poverty, and access to hygienic toilet. Using a plausible instrumental variable gives some support to the idea that the relationship between mortality and public health spending is causal and not just an artifact of reverse causation or incidental association. We did not find significant results for the effect of public health spending on disease prevalence, perhaps due to measurement error.
A potential issue is the applicability of our results to current policy debates, given that the data that is almost a decade old and India has undergone significant economic reforms, unprecedented GDP growth, and marked improvement in health over the course of the last decade (World Bank, 2009). Income growth has created higher expectations and demands for higher-quality health services. A widening socioeconomic gap between high- and low-income households also poses increased challenges to achieving the societal goal of equal health status and access to health care.
Before the economic reforms in the mid-1980s, public spending on health in India peaked at about 1.6 percent of GDP and 4 percent of the government budget. During the 1990s, government health spending failed to keep up with the expanding economy, and by 2001 it constituted 0.9 percent of GDP and 2.7 percent of the government budget. These numbers fell to 0.8 percent and 2.4 percent, respectively, by 2005. The 2006-07 budget reversed this trend, increasing allocations to social sectors (including education, health, and woman and child development) (Deolalikar et al., 2008). However these changes have not changed the health system in India substantially from what it was a decade ago and it remains to be seen what effect recent increases in public funding will have (Yip and Mahal 2008).
K25 HL081275/HL/NHLBI NIH HHS/United States
1King uses Bayesian statistics to determine the probabilistic values of subgroups in populations within the permissible bounds. It assumes random coefficients, allowing the distribution of health status to vary from country to country. To explain the variation in subgroup averages across i’s as a function of measured explanatory variables, he recommends performing instead a second-stage analysis, such as regression analysis, with the estimated subgroup averages as dependent variables.
2The Concurrent subject is the joint domain of both the State Governments and the Union territories of India as well as the Central Government of India.
3Correspondence with the MEASURE DHS (Demographic and Health Surveys) indicated that while data in infant mortality are collected in each round, comprehensive mortality data is scheduled to be collected every 10 years. It is expected be collected in the next round of the NFHS.
4We experimented with family welfare spending as an explanatory variable but did not find any statistically significant effect on mortality rates.
5Medical expenses include expenditure on medicines and medical goods including family planning appliances, payments made for medical treatment, and expenses incurred for clinical tests.
6For a complete discussion see Financing and Delivery of Health Care Services in India, 2005, pp. 245-250.
7The individual level data was aggregated to the state level using sampling weights to obtain state level IMR.
8Public expenditure constitutes 20.3 percent, private sector expenditure 77.4 and external support 2.3 percent of total health expenditure in India (NHA INDIA (2005) National Health Accounts India 2001-02. New Delhi, Ministry of Health and Family Welfare, Government of India.
9The deficit may be associated with state average income levels since a recession will hurt both income levels and tax receipts, but we control for average state income per capita in our model.
10We experimented with the following alternative IV estimators: whether the dominant political party in the State is in the dominant national political party, the time interval remaining to the next parliamentary election, and the percentage of women members in the state parliament, but did not find them good predictors of state public health spending.
11The regression for the first stage F test is weighted and clustered at state-level.
No funding for the study
The authors have no conflict of interest
Publication Types, MeSH Terms
Research Support, N.I.H., Extramural
Delivery of Health Care/economics*
Delivery of Health Care/statistics & numerical data*
Financing, Government/statistics & numerical data*
Residence Characteristics/statistics & numerical data