Prices and poverty in India Angus Deaton Alessandro Tarozzi Research Program in Development Studies Princeton University This version, December 13,1999 (First version, October 1999) Contact address: [email protected]. We wish to acknowledge the assistance of India’s National Sample Survey Organization, its Director, Professor Pravin Vasaria, and Joint Director Dr S. S. Shukla for providing the data as well as documentation and help on their use. We would also like to thank Anne Case, Gaurav Datt, Jean Dreze, Valerie Kozel, and Salman Zaidi for their assistance and comments on an earlier draft. 0. Introduction In India, as in other countries, indexes of consumer prices perform many important functions. Millions of workers have their wages indexed to some measure of the price level. Just as important is the issue that is our main focus here, which is in the estimation of poverty. Indian poverty rates are defined as the fractions of people living in households whose real per capita total expenditure falls below the poverty line. Data on total expenditures are collected by the National Sample Survey (NSS) in money terms so that, for each new round of data, the real poverty line must be converted to current rupees by multiplying by an index of prices. Inaccuracy in the estimation of the index, for example overestimation of the price increase relative to the base, will result in corresponding inaccuracy of the poverty estimates, for example an underestimation in the rate of poverty reduction. At a time when the data show historically high rates of GDP growth without much reduction in poverty, especially rural poverty, it is important to establish the accuracy of the price and poverty calculations. The measurement of inflation is not the only role of price indexes in measuring poverty in India. Price indexes are required, not only to establish the rates of inflation in the urban and rural sector of each state, but also to compare price levels between them. Finally, in a country where many states are larger than most nations in the world, price indexes are needed to make comparisons between states. Differences in poverty rates between Indian states affect the amounts of transfers from the center to the states, and influence discussions about poverty reduction strategies among international lenders such as the World Bank. For purposes of tracking poverty over time, the two most important price indexes in India are the Consumer Price Index for Industrial Workers (CPIIW) and the Consumer Price Index for 1 Agricultural Labourers (CPIAL). In a simple average with a third index, the Consumer Price Index for urban non-manual employees, (a version of) the CPIIW is used for updating urban poverty lines, and (a version of) the CPIAL for updating rural poverty lines. Until it was revised in November 1995, the CPIAL was based on prices regularly collected from a sample of 422 villages, weighted using an expenditure pattern that dates back to 1960–61. By the time of the revision in 1995, which was later than any of the survey data used in this paper, the weights for the CPIAL were more than three decades out of date, an unusually long time period for any major price index anywhere in the world. It should also be noted that agricultural laborers, whose expenditure weights were collected in the 1960–61 survey, are only a fraction of the rural population. Even if the weights were not so old, their coverage would give cause doubts about their representativeness. The CPIIW was revised in October 1988. Prior to that date, the base year was 1960, with weights derived from a survey of workers from 1958–59. Prices were collected from shops in a number of markets in 50 industrial centres throughout India. After October 1988, the weights were updated using a 1981–82 survey, with a new basis in 1982, some old centres were dropped and some new ones added, and the total rose to 70. For both indexes, and in order to maximize comparability over time, the specifications of items priced, the units, the shops, the markets, and the day and the time of the visits were held fixed throughout the life of the series. To the extent that there are problems with these price series, they are likely to come from the unusually long periods between revisions. Not only are the weights of these Laspeyres indexes long out of date by the time of transition, but there must also be concern about the continued representativeness of the villages, centres, and markets over such long periods. Whether or not the price indexes are 2 seriously affected is a ultimately an empirical question, though it is often supposed (for example in the comparable debate over the CPI in the United States) that Laspeyres indexes will increasingly overstate inflation as the base period recedes into the past, a tendency that will be exacerbated by the failure to pick up new goods (whose prices are often falling rapidly) and discard old (whose prices may be stagnant or even rising.) The purpose of this paper is to provide an independent set of calculations of Indian price indexes using data for two year-long periods, 1987–88, and 1993–94. We provide estimates of the rate of inflation over the six year period for All India and for 17 of the largest states, by sector, plus Delhi. We also provide separate price indexes for the rural and urban sectors for each of the 17 states and for All India in both periods, as well as estimates of price levels across states by sector in each of the two periods. The sources of data for both prices and expenditure patterns are the 43rd and 50th rounds of the NSS, both of which collected extensive data on consumption of individual items. The use of expenditure surveys to calculate weights for consumer price indexes is standard practice throughout the world. The innovation here is the use of information from the surveys on the prices themselves. For most of the commodities in the NSS surveys, respondents are asked to provide information on how much they spent on the item and on the physical quantity purchased, for example 8 rupees on 2 kilos of rice. The ratio of expenditure to volume provides a measure of price, or more precisely, a measure of unit value. Compared with the use of the CPIAL and CPIIW, this approach has both advantages and disadvantages. One strength is the size of the samples involved and their representativeness across states and sectors. More than 3.5 million pairs of expenditures and quantities are sampled in each of the two rounds, and the NSS samples are designed to be representative at the state and 3 sector level (and indeed beyond). In consequence, it is possible to construct, not only price indexes that track inflation over time, but also price indexes that compare price levels across states and sectors. A second advantage of unit values is that they relate to actual transactions, not to prices listed or reported by shops. Third, because the transactions are linked to the people who made them, it is possible to stratify prices and price indexes by socioeconomic characteristics, such as level of living, or occupation, or demographic structure. There are two main disadvantages to the use of unit values. Not all goods and services have readily defined quantities. In particular, while unit values are available for most foods, for alcohol and tobacco, and for fuels, they are not collected for such items as transportation or housing. In India, the covered goods comprise between two-thirds and three-quarters of the budget, which makes the exercise worth doing, but which would obviously not be the case in a country such as the United States. Even so, when using price indexes constructed from unit values, it is always important to keep in mind the likely effects of the excluded categories, and in particular the effects on comparisons between urban and rural sectors of omitting the prices of housing and of transportation. The second disadvantage is that unit values are not prices. Even when goods are defined at the maximum feasible level of disaggregation, many goods are not perfectly homogeneous, so that any given unit value will reflect, not only price, but the mix of varieties within the category. As a result, unit values differ from one purchaser to another in a way that is not caused by differences in prices. In particular, richer households have higher unit values than poorer households, a fact that has been used to study the choice of quality since Prais and Houtkakker (1955), see Deaton (1988, 1997, Chapter 5) for modern treatments. The quality problem can be 4 dealt with in part by disaggregating to the maximum extent permitted by the data. In the analysis below, we work with more than two hundred items of expenditure. Even so, the literature shows that the total expenditure elasticity of unit values is small, even for fairly broad aggregates of goods—such as “cereals” or “pulses.” Beyond that, it is important to inspect the data on unit values and to document their price-like characteristics, for example that in a given round and state that a large number of people report the same unit value, and that the unit values have the appropriate patterns of variation over regions and seasons of the year. The rest of this paper is organized as follows. In Section 1 below, we explain how the unit values and expenditure weights are calculated from the detailed survey data, and we provide summaries of the results and of the methods used to obtain them. Because there are so many observations, more than 7 million unit values, and because each must be examined before being incorporated into the price indexes, the data processing stage of this work has been both long and complex. Nevertheless, we have tried to provide enough detail to permit replication of our results, and our STATA code is available on request. Section 2 is also methodological, and presents the index number formulas used in the calculations, as well as the strengths and weaknesses of each. It also explains why some of the indexes are much more difficult to calculate than others, in particular, why the interstate comparisons of prices are likely to be much less reliable than comparisons over time, with comparisons between urban and rural prices somewhere in between. Section 3 presents the main results and compares the price index numbers calculated here with the two official price indexes. Section 4 considers the implications for the calculations of poverty rates between 1987–88 and 1993–94, as well as for the 1993–94 rates themselves under alternative assumptions about interstate and intersector price variation. 5 Section 5 offers some tentative conclusions, as well as an outline of the work that remains to be done. Section 1: Using the NSS data to calculate unit values and expenditure patterns The NSS samples from the 43rd and 50th Rounds are described in Table 1 which shows the distribution of sample households over states and sectors, as well as the total number of purchases recorded for the food, alcohol and tobacco, and fuel categories, the 228 detailed components of which are listed in Table A1. Table 1 lists information for All India, which is the complete sample, and for the 17 largest states plus Delhi; clearly it is only for urban Delhi that there are a significant number of sample households. There are 128,101 sample households in the 43rd Round and 115,354 in the 50th Round. By dividing the “purchases” column by the “households” column, we see that, on average, urban households, who have access to a wider range of goods and are typically better off than rural households, record expenditure on around 40 of the items used here, while rural households reported purchases of about 30 items. In total, over the two surveys, there are 8.3 million quantity/expenditure pairs available for analysis; as we shall see below, this number will be reduced somewhat as we proceed. For each item of expenditure, household respondents are asked to report both the quantity and value of purchases over the last 30 days. The NSS records expenditures in considerable detail which is shown (for the goods used here and for the 50th Round) in Table A1. The comparable list of goods for the 43rd Round is almost identical apart from the important difference that goods bought from the public distribution system (PDS in the Table) are recorded separately from goods from other sources in the 50th Round but not in the 43rd. We have made no attempt to work with the data on clothing and footwear, where there is also some information on quantities 6 purchased, e.g. dhotis and sarees in meters, or shoes in pairs. For a few of the commodities listed in Table A1, it is effectively impossible to measure quantities, and the questionnaire does not attempt to do so. These commodities (or commodity groups) are therefore dropped from the analysis. They are as follows: egg products, other fresh fruits, other beverages (Horlicks, etc.), biscuits and confectionery, salted refreshments, prepared sweets, other processed food, other drugs and intoxicants, dung cakes, gobar gas, and other fuel and light. Several of these fall into the “other” or residual category within a larger group; for example, other drugs and intoxicants is the residual category in a group that contains toddy, beer, liquor, ganja, and opium. As we shall see later, not only these but several other residual categories do not have well-defined units. The are also few cases where the units change between the two rounds. For example, lemons (guavas) were measured in kilograms (units) in the 43rd Round and in units (kilograms) in the 50th Round. We retain such items for comparisons between states or sectors within each round but, since we have no way of knowing how many lemons or guavas are in a kilo, we drop them when making comparisons between the two rounds. For each consumption record, a unit value was calculated by dividing expenditure by quantity. The NSS collects data separately for commodities purchased in the market, for commodities produced or grown at home, and for commodities obtained as gifts or loans. Unit values were calculated by dividing the sum of the three kinds of expenditure by the sum of the three kinds of quantity. This procedure effectively weights each of three possible unit values by the shares of expenditure devoted to each. Working commodity by commodity, the unit values were then checked for plausibility as indicators of price. There is no foolproof way of doing this. Nevertheless, there are a number of obvious problems to guard against, and procedures can be 7 developed to detect them. One such is the difficulty of defining units for physical quantities. Expenditures are always measured in rupees, and the concept and its units are clear. For quantities, there is sometimes a choice of units; for example, items can be bought one by one, or by weight. There are also local variations in units, so that what works in one place may not work as well somewhere else. Some customary units are not well-defined in terms of weight; goods bought by the bunch, box, bag, or packet will be converted by the respondent or the enumerator, but the conversion may be less than accurate. To the extent that errors are made—eggs measured in units for some households, and in dozens for others—the unit values will have multi-modal distributions, with peaks corresponding to each distinct unit. A second problem, which can also be detected by looking for multiple modes, is when two or more distinct goods are included within a single commodity. For example, if milk (liquid) and milk products (expensive sweets) are lumped together, the unit values will cluster around the milk price and the sweet price. While gross contamination is avoided by using the maximal detail, inspection of Table A1 shows that, even with so many items, there is still room for heterogeneity within many of the categories. If the compounded goods are sufficiently similar, the unit value may still give a useful indication of prices. The problem arises when there are spatial or temporal differences in the mixture, so that a compound of (cheap) A and (expensive) B is primarily A in state 1, but is primarily B in state 2. Our procedures are part graphical, and part automatic. For each commodity, we draw histograms and one-way plots of the logarithms of the unit values, using each to detect the presence of gross outliers for further investigation. In some cases, outliers are isolated cases that result from errors of misreporting, miscoding, or misinterpretation of units, and are deleted. In 8 other cases, a problem with units or with contamination can be identified and corrected. An automatic method for outlier detection was also used, and unit values eliminated whose logarithm lie more than 2.5 standard deviations from the mean of logarithms. Note that this does not remove the need for the graphical inspection, since gross bimodality would not necessarily be detected by the standard deviation rule. Log unit values were also inspected for plausibility after deletion of outliers. The resulting distributions of unit values were also examined to assess how many purchases clustered at the median—if the unit values are close to being prices, we would expect substantial such clustering—and tested using analysis of variance for cluster (PSU), district, and subround (seasonal effects)—which should be present if variation in unit values is dominated by price variation rather than quality effects or product heterogeneity within the commodity group. The data examination led to the deletion of a number of goods where unit values appeared not to be reliable, or where there were other unsolved problems of interpretation. Tables A2 through to A5 list the goods involved. Table A2 lists the goods omitted because the NSS instructions do not call for quantities to be collected; for reasons that we do not understand, quantity data exist for these commodities, but the associated unit values would have been deleted by our inspection procedures in any case. Tables A3 and A4 list the additional commodities that were excluded from the 43rd and 50th Rounds respectively. Most of these cases are “other” residual categories within larger subgroups, where the unit values showed dispersion to match the obvious heterogeneity in the definition of the group. Some are goods where there is clear evidence of bimodality (e.g. Liquid Petroleum Gas in the 43rd Round, where there is apparently more than one unit of measurement, but where we were unable to make a suitable correction. 9