How to monitor poverty for the Millennium Development Goals Angus Deaton Research Program in Development Studies Princeton University Revised, March, 2003 First version, December 2002. For comments on earlier drafts I am grateful to Tony Atkinson, Tim Besley, Nancy Birdsall, Tom Griffin, Ivo Havinga, Sakiko Fukuda-Parr, Haishan Fu, Robert Johnson, Richard Jolly, Nora Lustig, Nick Stern, Miguel Székely, and Sanjay Reddy. The views expressed here are my own. How to monitor poverty for the Millennium Development Goals Angus Deaton ABSTRACT I consider two issues concerning how to monitor global poverty for the Millennium Development Goals, the selection of poverty lines, and the data sources for monitoring poverty over time. I discuss the choice of a single international line, converted using purchasing power parity exchange rates, versus the use of country-specific poverty lines. I note the difficulties in constructing purchasing power parity exchange rates but argue in favor of a single international line, converted at PPP rates, but which would be regularly updated using domestic price indexes. Re-basing, using updated PPP rates, would be done infrequently. For example, if the global poverty numbers were estimated annually, the PPP rates might be updated once a decade. In any case, it is important that the poverty estimates be calculated much more frequently than the PPP rates are revised. I discuss whether monitoring should be done using national accounts data on income or consumption, supplemented by distributional data so as to make inferences about poverty, or from household survey data. I argue that data from the national accounts are not suitable for measuring poverty and that their use requires assumptions that are unlikely to hold. In particular, monitoring poverty through the national accounts runs the risk of prejudging important issues that are properly the subject of measurement, not assumption, such as the extent to which aggregate growth benefits the poor. I argue that poverty should be directly measured using household survey data, and discuss what needs to be done to enable such monitoring to be placed on a sounder basis. Contents 0. Introduction 1. How to draw the line: poverty lines for monitoring the MDG 1.1 National or international lines 1.2 Purchasing power parity indexes 1.3 General problems with PPP exchange rates 1.4 Specific problems of the PPP indexes used in the world poverty counts 1.5 What to do, a proposal 2. How to monitor poverty for the MDGs 2.1 Two methods 2.2 Are national accounts or survey data more reliable? 2.3 NIPA consumption/income is not the same as survey consumption/income 2.4 From NIPA means to poverty: using distributional information 2.5 Inequality and the discrepancies between NIPA and survey estimates 2.6 Conclusions and recommendations 3. What to do now 4. References 0. Introduction I consider two important issues in monitoring world poverty for the Millennium Development Goals. Section 1 is concerned with the selection of poverty lines, whether they should be national or international, and if the latter, how the line should be converted into local currencies. It concludes with a proposal for modifying current practice in which purchasing power parity exchange rates are used to set baselines, but not for frequent updating. I also propose that the local equivalents of the international line be validated within countries. Section 2 is concerned with data sources, and whether monitoring should be done using national accounts and distributional data to infer poverty, or from direct measurement using household surveys. I argue in favor of household surveys, and again end with a series of proposals for survey work. Section 3 summarizes some current priorities. 1. How to draw the line: poverty lines for monitoring the MDG 1.1 National or international lines? The World Bank’s world poverty counts use international poverty lines of approximately $1 and $2 -a-day at 1993 international purchasing power parity (PPP) prices. The notion behind such lines is a simple one, that for the purpose of the global counts, or indeed for monitoring poverty reduction, we need a common international standard. While a common money standard is certainly not the only possibility, the rhetoric of a $1-a-day line has been enormously successful, particularly for the first-world and international agency audience to which it is primarily directed. Yet, as we shall see, the construction of such lines is a complex and difficult task, and the attraction of even the simple concept is apt to diminish once the details of its implementation are exposed to close inspection. For most of its country work, the World Bank (like policymakers in each country) uses the national poverty lines that exist for many countries, although not all. The MDG target, which involves cutting poverty rates by half, could conceivably be applied to poverty measures based on either national or international poverty lines. So there is a choice. There are strong arguments for an international line, if only because of the proven appeal of something like the $1-a-day concept, even if the reality falls some way short of the rhetoric. To start with an area where there is a wide degree of agreement, no one would argue that the US poverty counts, as computed by the US Census Bureau, should be included in world counts nor in the MDGs. Of course, the counts as computed by the World Bank do not include the rich countries, but the obvious interpretation of the exclusion is that the US poverty line is too high to be used in international comparisons, and that if a suitable international line were applied to the US, then there would be little or no poverty. Martin Ravallion’s (1994) survey of international poverty lines shows that the US phenomenon is quite general. While national poverty lines do not vary with the level of GDP among the poorest countries, after a little less than $3 per person 1.1 per day, national poverty lines are higher in the richer is the country. Evidence from surveys within rich countries also shows that people’s own notions of the poverty line rise as they get better-off. Which is one reason why some countries, such as Canada and the European Council of Ministers, have formalized the process by defining a poverty line in relation to the mean or median of the income distribution. These relative poverty lines clearly make sense for each country considered in isolation. But they hardly correspond to a universal notion of a state of extreme poverty and deprivation that is recognizable irrespective of average living standards in each country. Anthony B. Atkinson and Francois Bourguignon (2001) have suggested a way of using information from both national and international lines within a common global framework. They propose a lexicographic treatment of absolute and relative poverty, focusing on the former—with a fixed common poverty line—among the poorest countries, and then switching to the latter— using national poverty lines—among better-off countries, including the richest. A relative view of poverty, focusing on participating in the life of the country in which one lives, is something that only becomes relevant once absolute deprivation has been dealt with. Atkinson and Bourguignon’s synthesis provides an elegant way of thinking about global poverty in a unified way. In the context of the MDGs, however, it is absolute poverty that is relevant and the emphasis is on the absolute level of living standards rather than the relative income level that permits participation. It might be thought that, for monitoring change, the choice of line is not very important in practice. The poverty MDG is defined in terms of poverty reduction, not in terms of the poverty level, so that even if national poverty lines are not comparable, they do not prevent us from looking at poverty reduction in each country using its own individual poverty line. But this argument will not take us very far. Suppose, to illustrate, that all countries have the same lognormal distribution of income, with the same degree of inequality, differing only in the mean of log income. Then the amount of growth needed to halve the poverty rate is larger in a country with a higher poverty rate. For example, in a country where the poverty line is equal to median income, the poverty rate is 50 percent. To cut that to 25 percent, mean income has to increase by two-thirds of the standard deviation of log income; if the standard deviation is 0.5 (the figure for consumption per head in India), we need income to increase by about a third. For a country that starts at 25 percent, cutting poverty by half requires that mean income increase by only a quarter, and if the country is fortunate enough to have a poverty rate of only one percent, cutting it by half needs only a 12.5 percent increase in mean income. Countries that are richer at the start, and have adjusted their poverty lines upward, have a stiffer task to meet the goal, but it is not a task that necessarily has much to do with eliminating the extreme poverty and deprivation with which we are concerned. Which leads me to the position that some degree of international comparability is desirable, and that some kind of international poverty line is required. So we must either defend the $1-a-day line, or come up with something better. 1.2 The $1-a-day line has been criticized because it unrelated to any clear conception of international poverty, Sanjay Reddy and Thomas Pogge (2002). Their alternative is to specify some minimum level of living in terms of “the resources necessary to achieve a salient set of elementary capabilities.” Many national poverty lines, in both rich and poor countries, are constructed in a way that is consistent with such an idea. For example, both the Indian and US poverty lines were originally set with reference to the ability to buy a minimal food bundle or number of calories, see Chapter 3 of my Analysis of Household Surveys, Deaton (1997) for comparative discussions of those histories. However, a closer examination of these and other cases suggests that the link with food is more rhetorical than real. In both India an the US, the poverty line has been held constant in real terms, updated in nominal terms by a price index, and no attempt has been made to preserve the original link with food. And in both countries, updating according to the original methodology would generate poverty lines that are quite different from those currently in use. A more accurate interpretation of the history would be that the Indian and American poverty lines were originally accepted because they fell within the range of what was generally acceptable as a poverty line, a range that included, but was far from determined by, the incomes of those currently buying some minimally acceptable food bundle. These lines survived, not because of their link to any prescribed level of living, but because they have continued to fulfil a useful purpose for policymakers and in the policy dialog more broadly. Many other lines could presumably have served the same function, at least for a time, and all lines are continually up for challenge by politicians, commentators, and expert panels. Proposals to reset the lines to a meaningful poverty standard are frequently advanced, but rarely carry the day, in part because the original justification of the lines tends not to be convincing on close inspection, but also because it is so difficult to come to agreement on what it costs to maintain a minimal standard of living. A committee of experts sitting in judgment on what poor people need to consume, or worse still, what they ought to consume, is a particularly objectionable mechanism for setting poverty lines. By these criteria, the $1-a-day line has a good deal to commend it. On the one hand, it corres- ponds, at least approximately, to the national poverty lines of a number of the poorest countries. And this is surely the right place to look. Indeed, it is hard to think of a more appropriate definition for international poverty than being poor in the poorest nations, and we can rely on the political processes in these countries to throw up appropriate poverty lines. But just as important is that such lines are useful for the purposes to which they are put, in this case for monitoring by the international community of the number of very poor people in the world, and the $1-a-day concept has proven itself by the extraordinarily extensive use to which it has been put in the decade since it was first constructed. Indeed, its use in the MDGs is itself a testimony to its value. If we accept the $1 or $2 a day lines as a reasonable starting point, then we need a method for converting them into the local currency equivalents that are required for measuring poverty in each country. 1.2 Purchasing power parity price indexes One thing that we cannot do is to convert a common international poverty line using market 1.3 exchange rates. In poor countries, because labor is relatively cheap (because poor people are poor) the prices of non-tradeable goods, especially those with a high labor content, are low relative to the prices of tradeable goods. In consequence, the market exchange rate between the poor country and the US (for example), which is determined by trade, is unlikely to be an accurate measure of the cost of living difference between the countries. A dollar buys 50 rupees in the foreign exchange market, but food that would cost $10 in the US can be purchased for only 100 rupees, so that the implicit exchange rate is not 50, but 10. An Indian consumption level of 500 rupees a month would convert to only $10 a month at the official exchange rate, but would be worth $50 at the “food” exchange rate. In consequence, using the official exchange rate to convert consumption or income would vastly exaggerate the difference in levels of living between the two countries, and more generally between poor and rich countries, systematically overstating poverty in the former. This account also suggests the remedy, which is to take a bundle of goods consumed by poor people, and “price it out” in each country. If the cost of the bundle of goods in India (say) is the poverty line, the cost of the bundle in the other countries gives the appropriate value of the common poverty line in those countries. Alternatively, we can select a base country and calculate Laspeyres price indexes for all other countries relative to that base, so that if we have something that is approximately a $1 per person per day poverty line for the base country, we can convert it into other currencies by multiplying by the price indexes. Such price indexes are examples of the purchasing power parity exchange rates that we need if we are to have a common international poverty standard. The construction of Laspeyres indexes by “pricing out” a poverty bundle can be formally linked to methods that rely on calculating the cost of a minimal standard of living, or indeed of achieving a minimal set of capabilities. Such cost-of-living price indexes can be approximated by pricing out the goods purchased by people at the relevant level of living, and such approxima- tions are routinely relied upon by statistical offices who see their consumer price indexes in cost- of-living terms. The accuracy of approximation can be improved by calculating various “super- lative” index numbers, W. Erwin Diewert (1976), of which Fisher’s ideal index is one, and these are often used in one form or another by the statisticians who construct purchasing power parity price indexes. 1.3 General problems with PPP exchange rates Because PPP exchange rates are price index numbers, we can use the extensive body of theory about price indexes to explore their strengths and weaknesses. In particular, it is important to separate the general problems that are common to all price indexes, and which are therefore inherent in any attempt to convert a common international poverty standard into local currencies, from problems that are specific to current PPP exchange rates that are used for poverty measurement, and which might be ameliorated by better practice. 1.4 If poor people the world over consumed a single commodity—rice, say—the construction of PPP exchange rates for poor people’s consumption would be a simple matter. All that we would need would be the price in local currency. This is the case where the price index works the best, which is when it is unnecessary, because we have a price, not a price index. In reality, poor people consume a range of goods, whose relative prices are different in different places and times. Patterns of consumption are also different in different parts of the world, and often even in different parts of the same country. There are several different conceptual bases for constructing price indexes. Perhaps the simplest is to price out a common bundle which contains the kinds of goods consumed by poor people around the world. Alternatively, we can price, not a bundle of goods, but the cost of attaining a common standard of living. For price indexes appropriate for international counts of extreme poverty, this would be the living standard of the poorest, and there is nothing that stops us from thinking of that living standard in terms of capabilities. But the exercise is clearly a difficult one. We are asked to consider the conceptual experiment of how many pesos it would cost a landless laborer from Bihar to live as well (or as poorly) in Mexico as she lives in India. Once stated thus, it is not at all clear that we should expect any satisfactory answer. The standard economic theory of cost of living indexes is cast in terms of utility theory, with price indexes defined by comparing the cost of attaining the same standard of living (utility) at different set of prices. When people consume only one good, such as rice, the standard of living maps directly into rice consumption, and cost of living indexes are just the relative costs of a unit of rice. When there are many goods, but the relative prices are the same in all countries, the cost of a bundle of goods in country A relative to its cost in country B will be the same no matter what bundle we choose, and in this case too, there is an straightforward way to measure price indexes. But the reality is that the relative prices of goods are quite different in different count- ries. Consumption patterns are also quite different, presumably in part because of international differences in incomes and relative prices, but also because of differences in tastes. When tastes differ, the theory offers no consistent yardstick with which to measure the cost of living, and there is no reason why the relative costs of living in India and Mexico should appear the same to Indians as they do to Mexicans. For an international subsistence poverty line, we can perhaps (although only perhaps) ignore taste differences, specifying a basic level of human functioning, and calculate the minimum cost of achieving it in different countries. The extent to which such an index would be feasible, and the properties it would possess in practice, are topics that would be worth a good deal more research. There are a wide variety of price indexes. The simple Laspeyres index outlined in section1.2 is not always feasible—goods consumed by the poor in one country may not be available in another country—and when it is, it is unclear why one country rather than another should serve as the base. We could consider using each country in turn as base, but we will generally get as many different sets of PPPs as there are countries. These multiple indexes can then be “reconciled” by various kinds of averaging, sometimes leading to “better” index numbers according to some criteria. The important point here is the one noted long ago by Irving Fisher, that there does not exist any index number that satisfies all the “reasonable” requirements that might be placed on it. 1.5 For example, there is no guarantee that the price in A relative to B is the reciprocal of the price of B relative to A, or that if we go from A to B, from B to C, and from C back to A that we will finish up where we started. One desirable property can be guaranteed by choosing an appropriate formula, but only at the price of losing another. Price indexes are not prices. The constructors of PPP exchange rates, like those who are responsible for national price indexes, know this, and make compromises. But they are always open to criticism that some “obvious” or “elementary” requirement has been violated. There is no way of constructing a PPP index number that is immune to all such criticism. These problems exist even in the (possibly) conceptually easier case of within-country comparisons. Until the early 1990s, India had two poverty lines, one for rural and one for urban households. Over the last decade however, there have been different poverty lines for each state, and for each sector within each state. The methodology under which these lines were constructed and are updated is explained in Government of India (1993) and appears to be sound in principle. Yet the results have brought official Indian poverty measurement into something close to disrepute. Urban poverty lines are so much higher than rural lines that, in many states, the urban poverty rate is higher than the rural rate, something that few independent observers would accept. In addition to these unsatisfactory intrastate price indexes, the pattern of prices across states also appears to make very little sense. There is an internal Indian debate on price indexes that parallels the international debate about unsatisfactory PPP exchange rates. And at least part of the reason lies in the inherent difficulty of constructing such indexes. Alessandro Tarozzi and I (2000) report the results of a project to recalculate price indexes for Indian states and sectors for 1987–88 and 1993–94. The conceptual problems are immediately relevant in practice. If, for example, we wish to construct a price index for Kerala relative to Uttar Pradesh, we can start by calculating the bundle of consumption at or around the poverty line in Kerala, and then price it out in UP. But this is immediately problematic, because some of the goods that are heavily consumed in Kerala (for example, fresh fish, coconuts, and coconut oil) are rarely purchased in UP. The few recorded purchases are at very high prices, because these goods are exotica in UP, and cater only to a small minority. Conversely, atta is a basic staple in UP, but rarely shows up in Kerala. The cost of the Keralan poverty bundle in UP is therefore not a very good measure of the relative costs of being poor in UP relative to Kerala. If we start from UP, we have the same problems in reverse, and we get an answer for the cost of living in Kerala relative to UP that is not the reciprocal of the cost of living in UP relative to Kerala. One answer to these difficulties is to average the two; indeed, the geometric average is the Fisher ideal price index. There is a lot to be said for the Fisher index in this context; it moderates the extremes that we get from either of the two Laspeyres indexes and, because it is a superlative index in the sense of Diewert (1976), it captures at least some of the change in consumption patterns that would take place if the Keralan actually moved to the UP, or vice versa. But the Fisher ideal index is subject to criticisms of its own, see for example Reddy and Pogge. (2002) who show that changes in quantities (driven by factors other than prices) can lead to counter- intuitive effects on the price index. Another partial solution that is sometimes proposed is 1.6 geographical chaining. We can compare Kerala with UP by passing through a sequence of neighboring states, from Kerala to Andhra, to Maharashtra, Rajasthan, and eventually to UP, calculating an index only for pairs of nearby states with consumption patterns that are more similar than those at the beginning and end of our journey, and then chaining the results. Such chained indexes also have many desirable properties. But they can also be criticized on the grounds that such an index has the odd feature that the price index of UP relative to Kerala depends on what happens to prices and consumption patterns in Maharashtra, a state that is adjacent to neither. There are no general solutions to the conceptual problems of constructing PPP price indexes. Those who construct them know this, and make compromises at least some of which are undesirable. There is no alternative. 1.4 Specific problems of the PPP indexes used in the world poverty counts The PPPs used in the world poverty counts originally came from the Penn World Tables, but more recent versions have been constructed under the aegis of the World Bank. Constructing PPP index numbers is a major undertaking, requiring each participating country to supply information that is not part of its usual activities and, not all countries have done so. The number of these “benchmark” countries has increased over time, from 60 in the 1985 numbers to 110 in the 1993 version, including all of the largest countries. Non-benchmark countries have PPP exchange rates imputed to them through a regression procedure that effectively adjusts their official exchange rate by an amount that is similar to the adjustment for countries at similar level of development for which benchmark data exist. This procedure, while clearly sensible, can also be subject to substantial error, and in some extreme cases in the past, the Bank has published PPPs and associated poverty rates that were clearly incorrect. There are also concerns that country statistical offices do not always take these tasks very seriously; the ICP data do not feed into domestic policymaking and they have no domestic constituency, so that they tend to be assigned low priority and funding. The latest World Bank poverty counts, detailed in Shaohua Chen and Ravallion (2001) show some very large changes compared with earlier counts, even when computed for the same country in the same year, and even on a broad regional basis, and the changes are mostly attributable to the switch to new PPP numbers. Some of these changes almost certainly reflect a move from worse to better data—it is surely better to have 115 countries than 65, and there have certainly been improvements within countries—but some is probably due to “noise,” essentially random measurement error. Certainly the new figures paint a different picture of global poverty than do the old ones, see Deaton (2001). And although it is hard to predict what might happen with a new round of PPP numbers—and it is possible that they will be much more stable from now on—it is surely undesirable to measure world poverty with methods that are so unstable and unreliable. 1.7