I Introduction

II Current Status

In India, public expenditure contributes a significantly small percentage in healthcare expenditure. Table 1 provides comparison of public expenditures on health. The comparison of health expenditure with some Asian countries suggests that India’s public health expenditure is only 17.9 per cent of the total.

Public healthcare expenditure in India is composed of states and central government allocations. The centre provides direct and partial (matching grant) support to the states for meeting recurring and non-recurring expenditure of programmes under this policy initiative. The states’ share in the total revenue expenditure has been declining due to their fiscal problems and central support in their budgetary allocations is increasing. This is illustrated in Figures 1 and 2, though the change is significant in percentage terms. Trend of public healthcare expenditure at state level: State government revenue expenditure for medical and healthcare increased more than three times from Rs 50 billion in 1992 to over 150 billion in 2001. Around 1996 there was a sharp dip in public healthcare expenditures across all states (as shown by the dotted line in Figure 3), after which it again increased. Keeping in mind the sharp dip in 1996, if we divide the period being studied into two parts, from 1990 to 1996 and from 1996 to 2002, we can see (Table 2) the variations in public healthcare expenditures for these two as well as for the entire time period.

Public health expenditure of all states except Assam went down in the period 1990-96 but increased during the period 1996-2002 for all states except Uttar Pradesh and Assam. Overall, in this period it increased for most of the states except Assam, Gujarat, Orissa and Uttar Pradesh. But, if we observe per capita health expenditure as per cent of per capita gross state domestic product (GSDP), in real terms for the same period, a different picture emerges. The percentage spending of state governments shows a declining trend (Figure 4).

Figure 1: State Government Revenue Expenditure on Health Figure 2: Central Government Revenue Expenditure on Health

l l Per cent 94 93 92 91 90 89 88 87 86 18000 16000 14000 12000 10000 8000 6000 4000 2000 0 Percentage of State Governmentin Total Expenditure for Health 1987 20031994 Per cent

l l i 10 12 14 16 18 l l Percentage of Central Governmentin Total Expenditure for Health 14 13 12 11 10 9 8 7 6 18000 16000 14000 12000 10000 8000 6000 4000 2000 01987 2003 1994

19871988198919901991199219931994199519961997199819992000200120022003

19871988198919901991199219931994199519961997199819992000200120022003

Total Revenue Expenditure on Health

State Government Revenue Expenditure on Health

An analysis of public healthcare expenditure as a percentage of GSDP shows that Bihar and Uttar Pradesh have not fared badly. Comparatively, major states like Maharashtra, Madhya Pradesh and Gujarat, have not done well. States like Bihar, Assam, Andhra Pradesh and Punjab show high fluctuation in public health expenditure to GSDP ratio while for other states like Maharashtra and Gujarat this ratio has remained constant. The analysis clearly shows that in almost all the states, public health expenditures as per cent of GSDP has not increased much during the past decade. In fact, during the period 1994 to 2002 healthcare expenditure as a percentage of GSDP shows a declining trend (Table 3).

From the Table 3 we can see that for all the states public health expenditure as a percentage of GSDP went down significantly in the period 1990-96; for the period 1996-2002 it again went down except for Andhra Pradesh, Madhya Pradesh, Maharashtra, Orissa, Punjab and West Bengal. This shows that government priority for healthcare expenditure is decreasing over the years in all the states as indicated below.

Between 30 to 40 per cent Andhra Pradesh, Karnataka, Kerala, Uttar Pradesh, Tamil Nadu

Selvaraju (2000) reasons that when economic liberalisation was initiated in 1991, healthcare, as other sectors, faced expenditure contraction. The expenditure subsequently went up in 1996 after the situation improved.

Private expenditure on health (PHE) as a per cent of per capita income has almost doubled since 1961. Table 4 shows the average per capita private health expenditure as per cent of per capita income (PCI) in different periods since 1961. The PHE as per cent of PCI has increased from 2.71 during 1961-70 to 5.53 during 2001-03:

Total Revenue Expenditure on Health

Central Government Revenue Expenditure on Health

This implies that PHE has grown at a much higher rate than the per capita income over the years. Table 5 provides information about the growth rates of PHE, PCI and private final consumption expenditure in different periods. During the period 1991-2003, PHE grew at 10.88 per cent per annum in real terms whereas per capita income grew at 3.76 per cent during the same period. The growth in PHE has been much higher than income growth or private final consumption expenditure.

III Literature Review

Studies related to relationship between health expenditure and income: The relationship between health expenditure and income has received significant analytical attention. One important finding in earlier studies has been that the ratio of healthcare expenditure to GDP increased as countries developed economically and industrially. Abel-Smith (1963, 1967) found that after adjusting for inflation, exchange rates and population, GDP is a major determinant of health expenditure. In a seminal paper, Newhouse (1977) raised the question about what determines the quantity of resources a country devotes to medical care and suggests that per capita GDP of the country is the single-most important factor. This study found a positive linear relationship between the percentage of health care expenditure to GDP and GDP.

Gerdtham, et al (1992) used a single cross-section of 19 OECD countries. They found PCI, urbanisation, and the share of public financing in total health expenditure as positive and significant

Table 1: Public Expenditure on Health as Percentageof Total Expenditure, 2001

Country Percentage

Bhutan 90.6 Maldives 83.5 Democratic People’s Republic of Korea 73.4 Thailand 57.1 Sri Lanka 48.9 Bangladesh 44.2 Nepal 29.7 Indonesia 25.1 India 17.9 Myanmar 17.8

2002 r i j r t r l r l r r i j t i l t r r t 2002 Ratio 0.014 0.012 0.01 0.008 0.006 0.004 0.002 0 Bihar Kerala Rajasthan Tamil Nadu Gujarat Madhya Pradesh Maharashtra West Bengal Andhra Pradesh Orissa Assam Kerala Punjab Uttar Pradesh

r r j r t r tr l r r tr i j j t il t r r t lValue (Rs crore) 160 140 120 100 80 60 40 20 0 Punjab Kerala Bihar Tamil Nadu Rajasthan Maharashtra West Bengal Uttar PradeshOrissa Assam Madhya Pradesh Andhra Pradesh Gujarat Karnataka

Economic and Political Weekly January 7, 2006

variables. Hitris and Posnett (1992) used 560 pooled time series and cross section observations from 20 OECD countries over the period 1960-87 and found a strong and positive correlation between per capita health spending and GDP. Other related studies, such as of Hansen and King (1996), McKoskey and Selden (1998), Gerdtham and Lothgren (2000), Karatzas (2000) agree that healthcare expenditure is dependent on the GDP of the country. Studies related to elasticity of health expenditure and income: Whether healthcare is a luxury or necessity is important from the viewpoint of estimating future expenditure on it. This is because if healthcare is a luxury product it will consume an everincreasing share of national income. It also has implications for the link between healthcare expenditure and economic wellbeing. Generally, normal measures of health like infant mortality rate, death rate, morbidity, etc, are more or less similar for almost equally developed countries (for example in OECD countries). But healthcare spending may differ more than these normal measures. There may be a situation in which the utility of healthcare expenditure can be very low as we can see from Engel’s curve and Engel’s law. Newhouse (1977) argues that since income elasticity of healthcare expenditure is greater than one, it could be treated as a “luxury” good. Some studies [like Newhouse 1977; Gerdtham et al 1992] found the elasticity greater than one while many other studies [Manning et al 1987; McLaughlin 1987; Di Matteo and Di Matteo 1998] found it much less than one. Getzen (2000) in his paper analyses the literature and concludes

1961 to 1970 2.71 1971 to 1980 3.27 1981 to 1990 3.72 1991 to 2000 3.26 2001 to 2003 5.53

Studies related to use of panel data: Panel data sets for economic research possess several major advantages over conventional cross-sectional or time series data sets [e g, Hsiao 1985a, 1995, 2000]. The main advantages of panel data are that they usually give a large number of data points, increasing the degree of freedom and reducing the collinearity among explanatory variables – hence improving the efficiency of econometric estimates. The use of panel data also provides a means of resolving or reducing the magnitude of the problem of presence of omitted (unmeasured or unobserved) variables that are correlated with explanatory variables. Although panel data has many advantages, a problem can arise if several trending variables are present in panel data regressions, for example, health expenditure and GDP. Phillips (1986) shows that regressions involving nonstationary variables might lead to spurious results showing apparently significant relationships even if variables are generated independently.

Notes: PHE: private health expenditure, PCI: per capita income, and PCE: private final consumption expenditure. Subscripts n and r denote variables expressed in nominal and real terms respectively.

Im et al# ADF-Fisher Chi-square## ADF - Choi Z-stat## t-statistics probability t-statistics probability t-statistics probability

IV Analysis of Public Expenditure on Health

To estimate the relationship between income and public healthcare expenditure, we use real per capita GSDP to represent income and real per capita state public health expenditure (PHCE). In this analysis, healthcare expenditure refers to states’ expenditure and does not include the central government’s allocation for family welfare programmes. The expenditure on health also does not include budgetary allocations to water supply and sanitation. For the purpose of this study we have included those 14 states that account for more than 90 per cent of the total population of the country. The states included in the study are: Andhra Pradesh, Assam, Bihar, Gujarat, Karnataka, Kerala, Madhya Pradesh, Maharashtra, Orissa, Punjab, Rajasthan, Tamil Nadu, Uttar Pradesh and West Bengal. The time period of the study is from 1990 to 2002 and the main sources for the real expenditure were the database maintained by Centre for Monitoring Indian Economy (CMIE) and different government publications.

Most of the time series data generally have trend, cycle, and/ or seasonality. By removing these deterministic patterns, it should be possible to make the remaining series stationary. In case the time series variables are not stationary, they can produce invalid inferences. Granger and Newbold (1974) have shown that in a non-stationary series, the estimation can lead to a problem of spurious regression with a high R-square. The Durbin-Watson statistic near to zero is mainly due to the use of a non-stationary data series. This means that the estimates of the model may turn out to be statistically significant but the relationship may have no meaning. Hence, we should first study the stationarity property of the time series variables. Testing stationarity: The first step in checking the stationarity of data is plotting it and observing how it behaves. The GSDP and PHCE plots of the chosen 14 states give a rough idea about how these variables are behaving across them. From these we observe that GSDP shows an upward trend for all the states, as expected with rising incomes. But there is no trend as such in PHCE.

For using time series data, the test of unit root is very important for testing stationarity. Here we are using panel data which is time series data for a number of states so we use panel unit root tests. Testing unit root in panel data is a recent phenomenon. Levin and Li (1993) gave a test for unit root in panel data but this test (and others) involves too restrictive assumptions on the slope parameters of pooled regressions. Now more flexible tests are available like IPS test [Im et al 1997] and MW test [Maddala and Wu 1999] which are based on the Augmented Dickey-Fuller (ADF) equation.

The standard approach to test for non-stationarity of each observed time series y observed over T time periods in a panel is to estimate an ADF regression here including a time trend:

pΔy =αi +δ +βiy −+ ∑ t ρij Δy − +ε , i = 1,…,N; t = 1,…,T

where Δyit = yit − yi, t − 1 . The number of lags p should be large enough to make the residuals serially uncorrelated. The null hypothesis H0: βi = 0, that the data generating process for the series for panel group i can be characterised as a difference stationary I (1) process, are tested against the trend stationary alternative H1: βi < 0 based on the t-statistic of the βi estimate [Campbell and Perron 1991 and Hamilton 1994 for thorough treatments of the univariate unit root tests].

Since unit root tests are known to have low power in distinguishing between the non-stationary null and a stationary but persistent alternative, in testing the individual state series with only nine years of annual data, the parameters of ADF equation will not give precise measures. Using the cross-section dimension of data one can increase the power of the unit root test. Im et al treated data as N independent, perhaps homogenous processes, that either contain unit roots or not. Thus, as the time and the cross-section dimension increase, unit root test statistics can be derived that converge to normally distributed variables. Im et al propose an approach to perform unit roots tests for panel data. Another test, which was performed for unit root is MW [Maddala and Wu 1999] based on the ADF equation. If the two series have unit roots and it is of the same order then the next step is to test for cointegration. However, if both the series do not show unit roots then there is no need to test for cointegration.

Unit root tests were done on both ln PHCE and ln GSDP separately (Table 6). While estimating presence of unit roots in PHCE we used constant and it was found that PHCE does not show presence of unit roots. In both the tests, i e, Im et al and ADF, unit root was not found. This means that real per capita public healthcare expenditure is stationary.

We observe the presence of trends in GSDP data, therefore while estimating unit root in GSDP we estimate it in the presence of constant and trend. Here also we find no evidence of unit root by both Im et al and ADF tests. Therefore we can say that real per capita GSDP is stationary. The state level results are presented in Tables 7 and 8. From Tables 7 and 8, we can see that for Karnataka, Maharashtra and Uttar Pradesh, PHCE shows the presence of unit roots while for Bihar and West Bengal, GSDP shows presence of unit roots. Since both these series are stationary, there is no need to test for cointegration.

From the literature per capita public healthcare expenditure by state governments is assumed to be a function of per capita GSDP. The model is specified in log-log form so that the coefficient estimates are elasticity and therefore enable us to interpret the relationship of healthcare expenditures and income. We use the following model to estimate this relationship: ln PHCEit = α + β1 *ln GSDPit + ε where β1 will give the elasticity of PHCE with GSDP and εis the residual. Method of estimation: Another important consideration is whether panel data is a fixed effect (FE) model or a random effect (RE) model. The FE model is a reasonable approach when we are confident that the differences between units can be viewed as parametric shifts of the regression function. In other settings, it might be better to view individual specific constant terms as randomly distributed across cross-sectional units. This would be appropriate if we believed that sampled cross-sectional units were drawn from a large population. The generally accepted way of

Notes: # Exogenous variables: Individual effects. ++ Exogenous variables: Individual effects, individual linear trends. Figures in bracket shows number of lags

choosing between FE and RE is running a Hausman test. The Hausman test checks a more efficient model against a less efficient but consistent model to make sure that the more efficient model also gives consistent results. Given a model and data in which FE estimation would be appropriate, a Hausman test checks whether RE estimation would be almost as good. In a FE kind of case, the Hausman test is a test of H0: that RE would be consistent and efficient, versus H1: that RE would be inconsistent. The result of the test is distributed as chi-square. So if the Hausman test statistic is large, one must use FE. If the statistic is small, one uses the RE model. Elasticity computation: The Hausman test was used to determine whether it is a fixed or random model. Three different models were used to calculate elasticity. From the Hausman test (H-value 0.14) it was found that RE model is appropriate for this case. But we see here that there is not much difference between the coefficients of GSDP in all the three models. Elasticity calculated by this method comes to between 0.65 and

0.71 (Table 9). The state level elasticity results are presented in Table 10.

The analysis presented in the previous section suggests that for every 1 per cent increase in state PCI, state level health expenditure has gone up by 0.684 per cent. To estimate the target PHCE/GSDP ratio states follow, we used the methodology of adaptive expectation model and estimated the ratio of health expenditure as per cent of GDP which governments incur on healthcare. Expectations are important in economic models of dynamic processes, particularly in macroeconomic models, and finding ways to model them is often a difficult task for the applied economist using time series data. The adaptive expectations model has been one of the earliest approaches developed for this purpose.

Suppose one postulates that target public health expenditure (PHCE*) at time t is related to GSDP as follows: PHCE* = λ0 + λ1 GSDPt + μt where λ1 is target ratio of health spending as per cent of GSDP. One assumes that states are not spending exactly as per this ratio. It aims to achieve this target over a period of time with some speed of adjustment. This can be modelled as follows: (PHCEt – PHCEt-1) = δ (PHCEt* – PHCEt-1)

Simplifying this equation and substituting the value of PHCE* in above equation gives us the following equation:*

From the above we can estimate the target ratio follows: λ1 = α1/(1–α2) where λ1 is target of GSDP which should be spent on PHCE.

The estimation of the above model using panel data poses methodological problems. Though the panel estimation through the FE and RE models controls for unobservable heterogeneity, the specification of the model is dynamic as it has a lagged dependent variable as an independent variable. The presence of a lagged dependent variable on the right hand side causes considerable difficulty in estimation, as the error term may be auto-correlated; but more seriously, the lagged dependent variable is correlated with the error term when we use FE or RE models [Greene 2003]. Related literature suggests the use of instrumental variables (IV) estimators and panel generalised method of moments (GMM) estimators in the estimation of dynamic models. Here to avoid the problems of heterogeneity and biases caused by the lagged dependent variable, we use the panel GMM procedure based on Arellano and Bond (1991) and Arellano and Bover (1995). The panel GMM estimator uses instrument variables. In our estimations, we use instruments for lagged dependent right hand side variables. The results of the GMM estimator based on Arellano and Bover (1995) are given in Table 11 and the estimation based on it uses orthogonal deviations and removes the individual effects.

spending, this goal looks very ambitious and to achieve it the state governments need to reform significantly and prioritise spending in the health sector. This will require significant political will and change of mind-set. This is all the more so when with every 1 per cent increase in GSDP the health expenditure increases by a mere 0.684 per cent.

IV Analysis of Private Healthcare Expenditure

We plotted the behaviour of per capita private health expenditure (PHE), per capita income (PCI) and private final consumption expenditure (PCE) in terms of their levels and first differences of levels. Examination of these plots suggests that the variables are not stationary and contain a linear trend. This characteristic has been incorporated while specifying the model and analysing the data.

The first step in statistical testing of the non-stationarity of time series data is to test for the random walk using unit root test of levels and first dfference of series. As discussed earlier using the non-stationery series in estimating relations may give spurious results. In case the first difference is stationary (has no unit root) then the series is described as having integration of order 1 and is denoted by I(1). If two time series are integrated of order or I(1), the correlation coefficient between them will tend towards plus or minus unity, whether an economic relationship between them exists or not. In case we do find unit root presence in first differences, we carry out the process of taking the further difference till the unit root problem persists. The stage at which we find the absence of unit root, we are able to identify the order of the integrated process for the series.

In order to test the unit root of a series it is useful to formulate its behaviour as simple auto-regressive process. For example, if we consider a simple AR(1) process:

yt =ρyt − 1 + xt 'δ+εt (1) where xt’ are optional exogenous regressors which may consist of constant, or a constant and trend, p and δ are parameters to be estimated, and the εt are assumed to be white noise. If I p l> = 1, y is a non-stationary series and the variance of y increases with time and approaches infinity. If IρI <1, y is to test the unit roots of these residuals. The EG results suggest

a stationary series. Thus, the hypothesis of stationarity can be evaluated by testing whether the absolute value of ρ is strictly less than one. Three tests which are standard in literature, Augmented Dickey-Fuller (ADF), Phillips Perron (PP) and Ng and Perron (NP), were carried out to find that whether unit root is present in the data or not (see Appendix 1 for the details of these tests). Table 12 presents the results of unit root tests of PHE and PCI. All the three tests indicated that there is a unit root in the data.

The results indicate that the PCI and PHE are not stationary in their levels. All three tests indicated in Table 12 suggest that the first difference of log values of PCI and PHE (both expressed in real terms) are stationary. Hence both ln(PCI) and ln(PHE) are integrated to the order 1 or I(1). It is well documented in the literature that unit root tests have low power to reject the null hypothesis. Hence, if the null hypothesis of a unit root is rejected, there is no need to proceed further. One important property of variables having I(1) property is that their linear combination can be I(0). This means the linear combination of

Per capita income (PCI)

PCI -2.3374 0.1659 2.6669 1.0000 1.8152 -1.1501 d(PCI) -0.9214 0.7713 -2.3228 0.4128 -0.5711 -1.9786 PCIr 0.1737 0.9970 1.2100 0.9999 4.5663 0.1593 d(PCI) -7.0050* 0.0000 -7.5871* 0.0000 -3.1407* -3.2097**

rln(PCI) -1.9987 0.5847 -2.3943 0.3771 0.2403 -1.2062 d[ln(PCI)] -5.7721* 0.0001 -5.7721* 0.0001 -2.7506** -3.0038** ln(PCI) -1.6610 0.7508 -1.4862 0.8186 2.9624 -1.2096

rd[ln(PCI)] -7.5441* 0.0000 -11.4397* 0.0000 -3.2815* -3.2022**

Per capita private health expenditure (PHE)

PHE 1.1943 0.9999 4.3162 1.0000 -5.7449 -12.6312 d(PHE) -3.1522 0.1115 -2.1994 0.4773 -1.0495 -1.9840 PHEr -3.8501 0.0251 1.6278 1.0000 0.5394 -3.1391 d(PHE) -3.0428 0.1336 -2.9633 0.1546 -2.1213 -2.5357

rln(PHE) -1.5835 0.7823 -1.0436 0.9265 0.6218 -1.9859 d[ln(PHE)] -4.2213* 0.0094 -4.1744* 0.0106 -2.7998 -2.9910** ln(PHE) -4.3543* 0.0076 -0.7027 0.9663 1.2753 -1.9255

rd[ln(PHE)]-4.1385* 0.0116 -4.0853* 0.0133 -2.9374 -3.0006**

Notes: All estimations are with constant and trend. d(..) is first difference. r at the end of each variable is indicating variable at constant prices. * and ** indicate significance levels at 1 per cent and 5 per cent respectively. ADF test and PP test statistics have been estimated with constant and trend. Ng-Perron (MZt GLS) is based on HAC corrected variance (Spectral GLS-detrended AR) and asymptotic critical values are as follows:

Note: Standard errors and t-statistics are based on Newey-West HAC Standard Errors and Covariance (lag truncation = 3).

non-stationary series of I(1) can be stationary. These variables are described as cointegrated variables. Cointegration analysis also helps us to perform a analysis of long-run relationships in a set of variables. For a multivariate time series, after testing unit roots for each variable, a cointegration test should be carried out to ensure that the regression model is not spurious.

The concept of cointegration was first introduced by Granger (1981), and has since then become a standard tool in the time series analysis of economic data. An economic relationship exists, however, when two I(1) series are cointegrated, such that a linear combination of the series is stationary and two series share a common stochastic trend. Several studies have estimated the relationship between health expenditures and income using this approach. The lack of cointegration, on the other hand, would imply that the series could wander apart without having any fundamental relationship. To test whether PHE has a long-term and equilibrium relationship with PCI, we estimate the relationship between these two variables and test for their cointegration. The Granger representation theorem also shows that any cointegrating relationship can be expressed as an equilibrium correction model (ECM). Cointegration tests: The concept of cointegration is used to determine whether there is a long run equilibrium relationship between private expenditure and health and income. Engle and Granger (1987), hereafter referred as EG, have developed a simple method whether two variables integrated of the same order are cointegrated. As per this method we first determine whether the two variables have integration of the same order. The cointegration test is to be applied only for the same order integrated series. Given that both PHE and PCI series are integrated series of order one, the long-run relationship: ln( PHE t ) =β0 +β1ln( PCI t ) +εt will be meaningful only if the error εt is free of unit root. The error εt represents the deviations from the long-term relationship. If these deviations are stationary then the two series have a cointegrated relationship and estimation is not spurious. Alternatively, one can also obtain regression residuals for unit root tests obtained from a cointegrating equation which includes a trend variable. By rejecting the null hypothesis of unit root on the residuals, the variables in the regression equation are said to be cointegrated. Table 13 presents the regression results of the cointegrated model.

The regression estimates suggest that the income elasticity of private expenditures in India is 1.43. This implies that for every 1 per cent increase in PCI the per capita PHE increases by 1.43 per cent. These results are acceptable if the error term of this regression does not have unit root. We use an EG residual based test to examine this. Table 14 presents these results.

Table 14 presents the values of the t-statistics that we obtain from applying ADF tests to the fitted residuals of the above equation. We also present the PP and NP (MZt GLS) test statistics to test the unit roots of these residuals. The EG results suggest that PCI and PHE are not cointegrated. The results cannot be used as they may be spurious. Since the EG residual based test has low power, it is possible that this test may fail to detect cointegration, when it is actually present. This may happen because it is difficult to reject a unit root in the residual due to the low power of the unit root test. Thus, we also used the Johansen rank based test to find cointegration. In this test we examine the null rank of zero for no cointegration against the alternative rank greater than zero for the presence of a cointegrating vector. Table 15 presents these results.

when the stationary alternative hypothesis is true is indeed compromised. In fact, the power of these tests reduces. There have been some attempts to provide alternative unit root tests in the presence of structural breaks. Perron suggests a modified version of the Dickey-Fuller unit root test by including dummy variables to deal with one exogenous break point. This break point is provided exogenously in Perron (augmented type) Test. Lee and Amsler (1995) have also developed Lagrange Multiplier (LM) based test assuming a given break point. Later on, research on this issue evolved towards the development of test modifications allowing for endogenously determinate break points. The Zivot and Andrews (1992) minimum test is the endogenous procedure most widely used to select the break point when the t-statistic testing the null of a unit root is at its minimum value. Recently, researchers have raised the possibility of the existence of more than one break point in economic time series [Lumsdaine and Papell 1999]. It is possible to test for two structural breaks in a series [Lee and Strazicich 1999c]. In this paper we focus on the analysis of unit root with one structural break.

There are three structural break models developed in Perron (1989). These are: (i) the model allowing for a one-time change in level, termed as crash model (CM); (ii) the changing model which considers a sudden change in slope of the trend function (TM); and (iii) a third model that allows for changes in level and trend, called break-trend (BT) model. Since the third modelincorporates the changing model, only the CM and BT models are taken into account in this paper. Based on Perron (1989), the framework of these two models can be constructed as follows:

PCIt =γ+θDUt +βT +δDUM t +αPCIt−1 +∑λi ΔPCIt−i +εt

In the above specification DUt is dummy variable assuming value 1 for all t>Tb and DUMt is taking the value equal to 1 for t = Tb+1. Tb is the endogenously determined time of the break. The methodology searches over all possible break points and chooses the break point at the minimum value of the t-statistic. The above model allows a change in intercept only. The unit root test is performed using the t-statistic for null hypothesis that α = 1 (a unit root) in the regression. The t-statistics α is used for testing α = 1, with a break date Tb and truncation lag parameter k. Tb and k are treated as unknown and are determined endogenously.

Under the BT Model both a change in the intercept and the slope are allowed and is constructed as follows:

PCIt =γ+ θDUt + βT + ϕDTt+δDUMt +αPCIt-1 + Σλi Δ PCIt-i + εt

We apply two tests based on Zivot and Andrews (1992) and Lee and Strazicich (1999b) to calculate break point. For each of them, we admit two possibilities for the model set up: CM and BT models. As discussed earlier, standard unit root tests do not take into account the existence of break points in the time series while these two tests consider them. The programmes automatically take into account the appropriate lag length. The results of applying these procedures are presented in Table 16. The results for CM with a change in the intercept only show an interesting pattern.

We are not able to reject the null in case of PCI in the Lee and Strazicich method whereas under the Zivot and Andrews method, PCI is significant only at 10 per cent. For the BT model we are able to reject the unit root null under both the methods.

We view these results as generally consistent with the hypothesis that most of the series are best characterised as stationary around a breaking mean and/or trend function. The econometric implications of this mis-specification are relevant in that, following the structural break analysis of PCI and PHE, we can deduce that the acceptance of the no-cointegration null hypothesis may be caused by ignoring the presence of changes in the long-run relationship. Also, it was shown that the power of cointegration tests reduces if there is any structural break in the data [Gregory and Hansen 1996]. Since it is difficult to know such break points a priori, Gregory and Hansen (1996) propose a statistic that attempts to test the null hypothesis of no-cointegration against the alternative cointegration with a structural break at an unknown point of time. It can lead us to draw appropriate inferences on cointegration when the parameters of the cointegrating vector are not constant. Adopting the original notation to the case of PHE and PCI, these statistics are based on the estimation of the OLS residuals of the following models:

ln(PHEt ) =β0 +β1ln(PCIt ) +β2 DUt +εt

ln(PHEt ) =β0 +β1ln(PCIt ) +β2 DUt +β3(DUt *ln(PCIt )) +νt

where PHEt and PCIt have been previously defined, and where DUt is a dummy variable that takes the value 1 whenever t > Time of Break (TB) and 0 otherwise.

Three different Gregory-Hansen test statistics are shown. Model

(A) allows a level shift in the cointegrating relation, Model (B) augments model (A) with a trend in the cointegrating relation while model (C) allows for a regime shift (i e, for the value of the cointegrating parameter to have changed). In all these cases we get values of ADF*, Zt* and Z*.

aThe distribution of these statistics is derived in Gregory and Hansen (1996), where the asymptotic critical values are also tabulated. Thus, these statistics allow us to test for the noncointegration null hypothesis when the parameters of the cointegration relationship may change across the sample. All these aspects will play a crucial role in the following section, where we analyse the relationship between private healthcare expenditure and the GDP of India. Results of Hansen test: By adopting the test in Gregory and Hansen (1996) we get the ADF*, Z* and Zt* values with the

abreak points. Here we see that ADF*, Z* and Zt* are significant

aat 1 per cent. So we can say that with the normal cointegration tests which take the null of no cointegration against the cointegration, we do not find any cointegration but when we consider structural break we find evidence of cointegration. This has important implications. It means that the two series after taking into account structural breaks are cointegrated and there is a long-term relationship in the PCI and private healthcare expenditure.

Crash Model Break-trend Model Method and variables Tb lag (k) t-statistics Tb lag (k) t-statistics

Zivot and Andrews Model PHE 1987 7 -5.47* 1998 7 -4.76* PCI 1992 4 -3.68** 1982 0 -4.05** Lee and Strazicich Model PHE 1998 6 -3.92** 1983 6 -5.70* PCI 1997 8 -2.75 1981 8 -4.33*

Note: Significance level: * 1 per cent, ** 5 per cent and *** 10 per cent.

Based on the results obtained after incorporating the structural breaks we estimate the relationship between PHE and PCI and estimate elasticity using the following models:

ln(PHEt) =β0 +β1ln(PCIt) +β2DUt +β3(DUt ln(PCIt)) +νt

In the above equation,β1 will give the elasticity. Another important point here is that the dummy variable is chosen according to the break point suggested by Zt* statistic as recommended by Hansen. Table 17 presents these results. These results are based on fully modified OLS estimates. When a traditional OLS is implemented with non-stationary variables, test statistics, being biased, cannot be interpreted in the usual way. Generally the asymptotic distributions of the OLS estimator involves the unit root distribution and it is also non-standard; because of which inferences on β using the usual t-tests in the OLS regressions will be invalid. The Phillips-Hansen methodology corrects these test statistics using a semi-parametric procedure by suggesting fully modified least squares (FM-OLS) regression method. This particular method is appropriate in situations of cointegrating regressions. The method modifies least squares to account for serial correlation effects and for the endogeneity in the regressors that result from the existence of a cointegrating relationship. The model also provides estimates when there is drift in independent variables.

The regressions results after introducing the structural breaks dummy and regime shifts suggest that dummy variables in Model 1 (with level shift only) and both dummy and interaction variables in Model 2 are significant. The statistical significance of dummy variable in both models also provides evidence in favour of a structural break. The evidence also suggests that the long-term relationship between income and PHE exhibits as structural break and therefore there is no stable relationship between income and PHE across the sample period 1961 to 2003. The results, without recognising the structural break, would be distorted and raise questions about the validity of the conclusions. Table 18 provides estimates of elasticity coefficients which are 2.19 in Model 1 and 1.95 in Model 2 and both these coefficients are significant at the 1 per cent level. The results indicate an increase in elasticity from 1.39 estimates based on fully modified OLS after the introduction of dummy variables for a structural break. Because of the significance of both dummy variables, we select Model 2 for the purpose of our estimation according to which the income elasticity of PHE is 1.95. This elasticity is also statistically different from 1.

We have presented results which suggest that ln(PHE) and ln(PCI) are best characterised as stationary processes around a breaking trend function. We also find that these series are consistent with cointegrated representation and after introducing the structural breaks the two series are cointegrated.

V Discussion

This study suggests that PHE has grown substantially faster than real incomes. For each 1 per cent increase in real PCI, the real per capita expenditure on health has gone up by 1.95 per cent. During the last decade PHE has grown by 18 per cent per annum in nominal terms and about 11 per cent in real terms. The main concerns emerging from these findings are as follows: Financing mechanism and provider payment system: The way healthcare expenditures are financed has important implications for the health delivery system. For example, insurance coverage for healthcare expenditure is very limited in India. About 4 to 5 per cent of total health expenditure is reimbursable under any insurance or reimbursement schemes. Studies have shown that in the absence of reimbursement mechanisms, people borrow substantially to finance healthcare. In some cases, borrowing has been as high as their annual incomes. With relatively large amount out-of-pocket costs incurred by households, it is

Table 18: Fully Modified OLS Estimates of Relationship Between PHE and PCI with Structural Breaks

Variables Model 0 Model 1 Model 2 Co-t-Statistic Co-t-Statistic Co-t-Statistic efficient efficient efficient

Constant -6.82* -8.32 -13.69* 11.82 -11.61* -8.61 ln(PHE) 1.39* 14.90 2.19* 16.30 1.95* 12.47 Dummy -0.55* -6.83 -6.88* -2.96 ln(PHE)*Dummy 0.70* 2.72 Wald statistic (χ2) 55.84 128.18* 88.61* Wald statistic (χ2)@ 17.27 78.38* 36.80*

Notes: Estimates are based on Fully Modified Phillips-Hansen Estimates using Parzen weights and zero truncation lag. Wald statistic without any restrictions and tests whether all estimated coefficients together are significantly different from zero.@ Wald statistic is with one restriction: Coefficient of ln(PHE) = 1.

* statistically significant at 1 per cent level.

Figure 5: Medical Equipment Imports

(in Rs million)

180000 160000 140000 120000 100000 80000 60000 40000 20000

perceived better quality care in the private healthcare system is making people increasinly receptive towards the profit oriented, “fee-for-service” private sector. Private household expenditure is predominant in curative primary care, which is about 46 per cent of total health expenditure. Secondary and tertiary (hospital) care accounts for 27 per cent of the total. Although direct treatment costs in most public hospitals are largely subsidised, households have to bear substantial costs for purchase of medicines so that illnesses impose a heavy burden on the poor. Therefore, there is an urgent need for increasing government funding in providing health services.

Email: rbhat@iimahd.ernet.in

References

Abel-Smith, B (1963): ‘Paying for Health Services’, World Health

debatable whether people are getting value for their money, especially in catastrophic illnesses where the financial burden is high. Production function of private heath services delivery system: Various concerns about the growth of the private sector pertain to quality and cost of care, equity and efficiency. Cost concerns arise because, with the growth of the private sector, one is not informed about the scale on which private healthcare services are being produced. Data on the private sector suggests that many health facilities are small in size [Bhat 1994].

Given the morbidity and mortality conditions, India will certainly need more resources to meet the health needs of the population. However, without any regulation and monitoring of the performance of private sector health spending, it is possible that additional income buys costlier treatments at the margin, which produces very little impact on health outcomes. This to some extent gets reflected by the high income elasticity of PHE. Newhouse suggests the high elasticity may imply that people do not buy “cure” but buy “care”. It also diverts resources from more important health needs. The high expenditure may be also driven by the higher investments in technology. The data on medical equipment imports during the last 13 years shows an increase of about 25 per cent per annum (Figure 5). Demographic trends and epidemiological transition: Due to the major demographic and epidemiological transition in the second half of the 20th century, India’s middle income group is a vast base of around 250 million. The proportion of households in the low income group has declined from 58.8 per cent in 1990, to 49 per cent in 1996 and the middle and higher income-group has increased from 14 per cent to 20 per cent. The greater population in the age group of 15-60 years, altered life expectancy, increased burden of non-communicable diseases along with periodic recurrence of various communicable diseases have further burdened the ailing health system of India. Dwindling financial support to public health system: India has created a huge system of public health service delivery but more than 60 per cent of the health budget is spent in the recurring costs of staff salary. Social sector allocations are almost all absorbed by staffing costs. Little remains for capital investment and maintenance of essential infrastructure. Governmental resource constraint and the compelling need for upgrading infrastructure, together pave the way for private sector growth. Dwindling financing support to the public health system and the

– (1967): ‘An International Study of Health Expenditure’, World Health Organisation, Geneva, Public Health Papers No 32.

Arellano, M, S Bond (1991): ‘Some Tests of Specification for Panel Data: Monte Carlo Evidence and an Application to Employment Equations’, Review of Economic Studies 58, pp 277-97.

Arellano, M, O Bover (1995): ‘Another Look at the Instrumental-Variable Estimation of Error-Components Models’, Journal of Econometrics 68, pp 29-51.

Bhat, Ramesh (1993): ‘The Private/Public Mix of Healthcare in India’, Health Policy and Planning, 8(1), pp 43-56.

– (1996): ‘Regulation of the Private Health Sector in India’, International Journal of Health Planning and Management, Vol 11, pp 253-74.

Di Matteo L, R Di Matteo (1998): ‘Evidence on the Determinants of Canadian Provincial Government Health Expenditures: 1965-1991’, Journal of Health Economics, 17, pp 211-28.

Duggal, Ravi and Suchetha Amin (1989): ‘Cost of Healthcare: A Household Survey in an Indian District’, Foundation for Research in Community Health, Mumbai.

Duggal, Ravi (1996): ‘Health Sector Financing in the Context of Women’s Health’, March, ISST Occasional Paper No 4/96.

Engle, R F, C W J Granger (1987): ‘Cointegration and Error Correction Representation, Estimation and Testing’, Econometrica, 55, pp 251-76.

Gerdtham, U G, M Löthgren (2000): ‘On Stationarity and Cointegration of International Health Expenditure and GNP’, Journal of Health Economics, 19, pp 461-75.

Getzen, T (2000): ‘Healthcare Is an Individual Necessity and a National Luxury: Applying Multilevel Decision Models to the Analysis of Healthcare Expenditure, Journal of Health Economics, 19, pp 259-70.

GoI (1946): Health Survey and Development Committee Report (4 vols), Bhore committee, Government of India, New Delhi.

Granger, C W J and P Newbold (1974): ‘Spurious Regressions in Econometrics’, Journal of Econometrics, 2, pp 111-20.

Granger, C W J (1981): ‘Some Properties of Time Series Data and Their Use in Econometric Model Specification’, Journal of Econometrics, 28, pp 121-30.

Greene, William (2003): Econometric Analysis, Prentice Hall, New York.

Gregory, A W and B E Hansen (1996): ‘Residual-based Tests for Cointegration in Models with Regime Shifts’, Journal of Econometrics 70, pp 99-126.

Hansen, Paul, and Alan King (1996): ‘The Determinants of Healthcare Expenditure: A Cointegration Approach’, Journal of Health Economics, 15, pp 127-37.

Hitris, T and J Posnett (1992): ‘The Determinants and Effects of Health Expenditures in Developed Countries’, Journal of Health Economics, 11, pp 173-81.

Hsiao, C (1985a): ‘Benefits and Limitations of Panel Data’, Econometric Review, 4, pp 121-74.

– (1995): ‘Panel Analysis for Metric Data’, Handbook of Statistical Modelling in the Social and Behavioural Sciences, G Arminger, C C Clogg, M Z Sobel (eds), Plenum, pp 361-400.

– (2000): ‘To Pool or Not to Pool Panel Data’, Panel Data Econometrics: Future Directions, J Krishnakumar, E Ronchetti (eds), Elsevier, Amsterdam, pp 181-98.

Im, K S, M H Pesaran, Y Shin (1997): Testing for Unit Root in Heterogeneous Panels, Department of Applied Economics, University of Cambridge, Mimeo.

IIM (1987): ‘Study of Healthcare Financing in India’, Indian Institute of Management, Ahmedabad. Karatzas, G (2000): ‘On the Determination of the USA Aggregate Healthcare Expenditure’, Applied Economics 32, pp 1085-99. Lee J, Amsler C (1995): ‘An LM Test for a Unit Root in the Presence of a Structural Change’, Econometric Theory, June. Lee J, Strazicich M (1999b): ‘Minimum LM Unit Root Tests’, Working Paper, University of Central Florida.

– (1999c): ‘Minimum LM Unit Root Test with Two Structural Breaks’, Working Paper, University of Central Florida.

Lumsdaine, R L, D H Papell (1997): ‘Multiple Trend Breaks and the Unit Root Hypothesis’, The Review of Economics and Statistics, 79, pp 212-18.

Maddala, G S, S Wu (1999): ‘A Comparative Study of Unit Root Tests with Panel Data and a New Simple Test’, Oxford Bulletin of Economics and Statistics, 61, pp 631-52.

Mahal, Ajay (2000a): ‘Diet-linked Chronic Illness in India, 1995-96: Estimates and Economic Consequences’, Draft, National Council for Applied Economic Research, New Delhi.

– (2000b): ‘Equity Implications of Private Health Insurance in India: A Model’, National Council of Applied Economic Research, New Delhi.

Manning, W et al (1987): ‘Health Insurance and the Demand for Medical Care Evidence from a Randomised Experiment’, American Economic Review, 77(3), pp 251-77.

McCoskey, S K, T M Selden (1998): ‘Healthcare Expenditures and GDP: Panel Data Unit Root Test Results’, Journal of Health Economics, 17, pp 369-76.