Why have some Indian states done better than others at reducing rural poverty?

______s l;q94 POLICY RESEARCH WORKING PAPER 1594 Why Have Some Indian Experience in India suggests that reducing rural poverty States Done Better requires both economic Than Others growth {farm and nonfarm) and human resource at Reducing Rural Poverty? development. Gaurav Datt Martin Ravallion The World Bank Policy Research Department Poverty and Human Resources Division April 1996  POLICY RESEARCH WORKING PAPER 1594 Summary findings The unevenness of the rise in rural living standards in the infant mortality rates - had significantly greater long various states of India since the 1950s allowed Datt and term rates of consumption growth and poverty Ravallion to study the causes of poverty. reduction. They modeled the evolution of average consumption By and large, the same variables that promoted growth and various poverty measures using pooled state-level in average consumption also helped reduce poverty. The data for 1957-91. effects on poverty measures were partly redistributive in They found that poverty was reduced by higher nature. After controlling for inflation, Datt and Ravallion agricultural yields, above-trend growth in nonfarm found that some of the factors that helped reduce output, and lower inflation rates. But these factors only absolute poverty also improved distribution, and none of partly explain relative success and failure in reducing the factors that reduced absolute poverty had adverse poverty. impacts on distribution. Initial conditions also mattered. States that started the In other words, there was no sign of tradeoffs between period with better infrastructure and human resources - growth and pro-poor distribution. with more intense irrigation, greater literacy, and lower This paper - a product of the Poverty and Human Resources Division, Policy Research Department - is part of a larger effort in the department to understand the causes of poverty in developing countries and the implications for public policy. The study was funded by the Bank's Research Support Budget under research project "Poverty in India: 1951-92" (RPO 677-82). Copies of this paper are available free from the World Bank, 1818 H Street NW, Washington, DC 20433. Please contactPatricia Sader, room N8-040, telephone 202-473-3902, fax 202-522-1153, Internet address psader(@worldbank.org. April 1996. (41 pages) The Policy Research Working Paper Series disseminates the findings of work in progress to encourage the exchange of ideas about developrnent issues. An objective of the series is to get the findings out quickly, even if the presentations are less than fully polisbed. The papers carry the names of the authors and should be used and cited accordingly. The findings, interpretations, and conclusions are the authors' own and sbould not be attributed to the World Bank, its Executive Board of Directors, or any of its member countries. Produced by the Policy Research Disseminationi Center  Why have some Indian states done better than others at reducing rural poverty? Gaurav Datt and Martin Ravallion* Policy Research Department, World Bank 1818 H Street NW, Washington DC, 20043, USA Policy Research Department, World Bank, 1818 H Street NW, Washington DC. These are the views of the authors, and should not be attributed to the World Bank. The support of the Bank's Research Committee (under RPO 677-82) is gratefully acknowledged. The authors are grateful to Berk Ozler for help in setting up the data set used here. They also gratefully acknowledge the comments of Paul Glewwe, Joanne Salop, K. Subbarao, Dominique van de Walle, and seminar participants at the World Bank, the University of New South Wales in Sydney, the International Food Policy Research Institute, the University of Wisconsin in Madison, and participants at the I Ith World Congress of the International Economic Association held in Tunis.   1 Introduction A key to sound development policy-making may lie in understanding why some economies have performed so much better than others in escaping absolute poverty.' One can postulate factors which could explain why, including differences in technical progress, public spending, macroeconomic stability, and initial endowments of physical and human wealth.2 A large literature has emerged aiming to test such explanations for cross-country and inter-regional differences in the rate of economic growth.3 Though it has not, to our knowledge, been done yet, the same approach could also be applied to cross-country differences in (say) the rates of change in poverty relative to some agreed international poverty line. There are, however, problems in using cross-country data for this purpose, not least of which is the lack of comparable survey data for tracking progress in raising household living standards and reducing absolute poverty. Changes over time in survey methods and differences between countries in survey-data and sources for right-hand side variables have been a long-standing concern in applied work (Deaton, 1995). But for one large developing country one can assemble a long time series of reasonably comparable household surveys for its composite states (some of which are larger than most countries) as well as reasonably comparable explanatory variables. That country is India. The regional disparities in levels of living in India are well-known.4 For instance, the proportion of the northeastern state of Bihar's rural population living in poverty around 1990 was about 58%, more than three times higher than the proportion (18%) in rural northwestern Punjab and For evidence on differences across countries in rates of poverty reduction see World Bank (1990, Chapter 3). 2 Recent theories of economic growth have suggested a potentially rich menu of such factors. For a reviews of the theory of growth see Barro and Sala-i-Martin (1995) and Hammond and Rodriguez-Clare (1993). For a survey see Sala-i-Martin (1994). See Nayyar (1991), Choudhry (1993), and Datt and Ravallion (1993).  Haryana. (We describe how we have estimated these numbers later.) Some of these differences have persisted historically; for example, Punjab-Haryana also had the lowest incidence of rural poverty around 1960. However, looking back over time the more striking-though often ignored-feature of the Indian experience has been the markedly different rates of progress between states; indeed the ranking around 1990 looks very different to that 35 years ago, as can be seen in Figure 1. For example, the southern state of Kerala moved from having the second highest incidence of rural poverty around 1960 to having the fifth lowest around 1990. This paper tries to explain the relative successes and failures at poverty reduction evident in Figure 1. We focus on the rural sector because that is where three-quarters of India's poor live (Ravallion and Datt, 1996). Much discussion, and debate, has centered on a number of questions concerning the determinants of poverty in this setting,5 including the extent to which agricultural growth "trickles down" to the rural poor (many of whom hiave little or no land of their own), the poverty impact of growth in the non-farm sector, and the extent to which economy-wide variables (such as the rate of inflation and the level of public spending) matter to the rural poor. Questions have also been raised about the extent to whichi initial investments in infrastructure and human resources pay off in the longer term through higlher houselhold welfare, and what "handicap" regions with initially poor infrastructure face in catching up to other regions. We aim to throw new light on these and related questions. The following section outlines our methodology. After discussing our data and the measures of living standards used. section 3 describes how overall progress in raising rural living standards has varied across states of India. Section 4 then tries to explain the variation over time and across states. Some conclusions are offered in the final section. For a review of the literature on thcsc and rclatcd topics sce Lipton and Ravallion (1995). 2  2 Modelling progress in reducing poverty 2.1 Motivation We assume that each region of the economy has a deterministic trend rate of progress in reducing poverty but that there are also period-to-period deviations from the trend. The measured level of poverty is P, for state i at date t. The observed rate of change over time in P,, is simply the sum of the deviation from trend plus the trend. The expected value of the trend rate of progress for region i is given by y'X1 where X, is a vector of regional characteristics, comprising initial conditions (notably the endowments of physical and human infrastructure) and trends in other relevant time-dependent variables (capturing the effects of such factors as technological progress and public spending). The deviation from trend is 7l(AlnY. - r1 ) (in expectation), where Y,, is a vector of (positive) time-varying exogenous variables with trend (compound) rates of growth of r/ also included in the vector X,. The rate of progress in reducing poverty is then AInPP = ir'(A1nYjt - r/) + y'X1 + residual,, (1) Notice that some variables may be common to both the deviation from trend and the trend. For example, agricultural yields may matter in two ways: a higher trend rate of yield growth due to technical progress in agriculture will presumably raise the trend rate of poverty reduction and so the trend in yield will appear significant in the X1 vector, while fluctuations in yield due to the vagaries of the weather would appear in the first term on the right hand side of (1). And the two could have very different effects; if, for example, the poor are well insured against bad weather then the relevant r coefficient could be small, yet the corresponding y coefficient could be large. 3  "Growth regressions" such as (I) have been widely used in investigations of the determinants of cross-country and regional differences in growth rates of average consumption or output per worker. Economic theory offers some guidance on the specification of the right-hand-side variables in such a model (Hammond and Rodriguez-Clare, 1993; Barro and Sala-i-Martin, 1995). In principle, any variable which influences the consumption of someone at-and for some measures below-the poverty line will also influence the evolution of the poverty measure. If we were modelling growth rates of consumption for individuals or cohorts, the carry-over from endogenous growth models to the present setting would be straightforward. However, the relationship between determinants of the growth rate for a representative household and the growth rate of a poverty measure defined on the distribution of consumption will be more complex, involving both micro- behavioral factors (preferences, budget, time and borrowing constraints, and the properties of household production functions), as well as the properties of the distribution of endowments and the specific measure of poverty used. We do not attempt to derive an estimable parametric model for poverty measures from explicit functional forms for these factors. Instead, we estimate "poverty- growth regressions" analogous to standard regressions for growth rates in average income or consumption. Our models of average consumption and the poverty measures have the same functional form and explanatory variables, which we discuss in section 4. 2.2 The econometric model On allowing for latent regional effects in the levels and a serially correlated error term to reflect the likely persistence of the poverty measures, we estimate the following econometric model for measured poverty in region i at date t (Pt,) corresponding to the growth model in equation (1): 4  WnPfr = iC'VInYi, + yXit + li + El, (i=l,..,N; t=1,..,7) (2) where VInY. = InY, - rYt measures the deviations of the time-dependent variables from their trend levels, the vector X1 includes the initial conditions as well as the trend rates of growth of the time- dependent variables r/, % are time-invariant state-specific effects, and c. is an error term which we assume to follow an AR(l) process: e = pr,Cit, + u (3) in which u1, is a standard (white noise) innovation error and r, is the time interval between the successive household surveys. (Since the household surveys we use are unevenly spaced, the autocorrelation parameter p is raised to the power of the time-interval -r, so as to consistently define an AR(l) process.) We estimate the model in the levels form of (2), rather than the "growth regression" in (1), so as to allow direct estimation of the T 's and to avoid the complex ARMA error structure of a "growth regression" induced by our unevenly spaced data. The AR(l) specification imposes the common factor restriction on a more general dynamic model with lags on all variables (Sargan, 1980). However, we are unable to estimate the more general dynamic panel-data model given the form of our data set. The main problem has to do with the unevenly spaced NSS consumption surveys. Starting from an AD( , 1) type model in annual time units, as we re-express the model for the observed NSS survey time periods, we end up not only with a nonlinear dynamic panel data model but also one with a non-uniform dimension of the vector of right-hand-side (RHS) variables. For different time-observations, the RHS variables have lags 5  of different order depending upon the gap between the successive NSS rounds. We do not know of any appropriate estimator for such models. Our models can be consistently estimated using a nonlinear least squares dummy variable (LSDV) estimator. This is the standard covariance estimator for static panel data models, adapted to deal with the nonlinearity due to the autoregressive error term and the uneven spacing of our survey data. The estimator thus belongs to the class of nonlinear generalized least squares estimators (Hsiao, 1986; Matyas and Sevestre 1992). The estimator is consistent whether or not the state- specific effects are orthogonal to the other explanatory variables in the model, though under orthogonality there may be more efficient estimates. Note that model (2) can also be written as (2') below: InP, = n/IXnY, + y*/Xit + q; + e*, (i=l,..,N; t=l,..,7) (2') where y' = y - x . This is a convenient form for estimation and the significance of y' directly tests for the equality of the impact of the trend and deviation from trend components. 3 Trends in rural living standards by state 3.1 The consumption data and poverty measures We shall use a new and consistent set of measures of absolute poverty and mean consumption per person for the rural areas of India's 15 major states spanning the period 1957-58 to 1990-91. The measures are based on consumption distributions from 21 rounds of the National Sample Survey (NSS) spanning this period. However, not all 21 rounds of the survey can be covered for each of 6  the 15 states.6 Altogether, we use 310 distributions, forming a panel data set which is unbalanced in its temporal coverage for different states. The NSS rounds are also unevenly spaced; the time interval between the mid-points of the survey periods ranges from 0.9 to 5.5 years. The cost of living index is the state-level Consumer Price Index for Agricultural Laborers (CPIAL). Monthly CPIAL indices for the 15 states are collated for the entire period beginning August 1956.7 We have incorporated inter-state cost of living differentials, using the Fisher price indices estimated by Chatterjee and Bhattacharya (1974).8 ' The final indices we use are averages of monthly indices corresponding to the exact survey period of each of the NSS rounds. ' For 12 states (Andhra Pradesh, Assam, Bihar, Karnataka, Kerala, Madhya Pradesh, Orissa, Punjab- Haryana, Rajasthan, Tamil Nadu, Uttar Pradesh and West Bengal) all 21 rounds are covered. (Only from 1964- 65 does Haryana appear as a separate state in the NSS data. To maintain comparability, the poverty measures for this and subsequent rounds have thus been aggregated using rural population weights derived from the decennial censuses). For Gujarat and Maharashtra, 20 rounds are included, beginning with the 14th round for 1958-59 (prior to 1958-59, separate distributions are not available for Maharashtra and Gujarat, which were merged under the state of Bombay). For Jammu and Kashmir only 18 rounds can be included beginning with the 16th round for 1960-61. For Jammu and Kashmir, while the NSS consumption distributions are available prior to round 16, we are constrained by the availability of data on the rural cost of living index. The earliest available data on CPIAL indices for Jammu and Kashmir are for 1964-65. For the period 1960-61 to 1964-65, we have used the rate of inflation implied by the consumer price index (for industrial workers) in Srinagar as a proxy, which enabled us to make use of the NSS distributions for rounds 16, 17 and 18. However, for the period before 1960-61, even the Srinagar consumer price index is not available. 7 For some states, the published data from the Labour Bureau had to be supplemented with the CPIAL estimates reported in Jose (1974). The states (and years) for which we used this source were: Gujarat and Maharashtra (1956/57 to 1959/60), Jammu & Kashmir and Uttar Pradesh (1956/57 to 1963/64), and Tamil Nadu (1956/57 to 1966/67). I These estimates are based on the 18th round of the NSS, for the period February 1963 to January 1964. Minhas and Jain (1989) and Planning Commission (1993) assumed that these differentials for 1993-64 also apply to 1960-61, which is the base period for the CPIAL series. We do not make this unnecessary assumption, which implies the same rate of rural inflation in all states between 1960-61 and 1963-64. The inter-state cost of living differentials for 1960-61 are easily derived using the price relatives for 1963-64 from Chatterjee and Bhattacharya (1974), and the state and all-India CPIAL indices for 1960-61 and 1963-64. 9 We also adjusted the state CPIAL series to correct for the constant price of firewood used by the Labour Bureau in its published series since 1960-61. However, since we do not have data on actual firewood prices for individual states, we assume that the price of firewood increased at the all-India rate in all states. The necessary adjustment to the state indices was then worked out using the state-level weights for firewood in the state CPIALs (ranging from 4.99 % in Punjab and Haryana to 8.79 % in Madhya Pradesh). 7  For the poverty measures, we use the poverty line originally defined by the Planning Commission (1979), and recently endorsed by Planning Commission (1993). This is based on a nutritional norm of 2400 calories per person per day, and is defined as the level of average per capita total expenditure at which this norm is typically attained. The poverty line was thus determined at a per capita monthly expenditure of Rs. 49 at October 1973-June 1974 all-India rural prices. The three poverty measures we consider are the headcount index (H), the poverty-gap index (PG), and the squared poverty gap index (SPG) proposed by Foster, Greer and Thorbecke (1984). H is simply the proportion of the population living below the poverty line. PG is the average distance below the line expressed as a proportion of the poverty line, where the average is formed over the entire population (counting the non-poor as having zero distance below the line). SPG is defined the same way as PG, except that the proportionate distances below the poverty line are squared, so that the measure will penalize inequality amongst the poor.'