UNCTAD. The Least Developed Countries Report 2010: Towards a New International Development Architecture for LDCs

Transcription

1 UNCTAD The Least Developed Countries Report 1: Towards a New International Development Architecture for LDCs Background Paper Global Poverty: New National Accounts Consistent Estimates based on 5 Purchasing Power Parity Exchange Rates, with extension to the Least Developed Countries poverty trends Massoud Karshenas Independent Expert Background Paper No. 8 This study was prepared for UNCTAD as a background paper for the Least Developed Countries Report 1: Towards a New International Development Architecture for LDCs. The views in this paper are those of the author and not necessarily those of UNCTAD or its member states. The designations, terminology and format employed are also those of the author.

2 Global Poverty: New National Accounts Consistent Estimates based on 5 Purchasing Power Parity Exchange Rates, with extension to the Least Developed Countries poverty trends. By Massoud Karshenas Department of Economics, SOAS, University of London Russell Square, London WC1H XG mk@soas.ac.uk Revised March 1 Abstract: This paper provides new global poverty estimates using the latest purchasing power parity exchange rate estimates and taking into account both the information in living standard surveys and the national accounts. It combines the information in the two sources by using the national accounts data to calibrate survey means. The properties of the new estimates are compared with the World Bank estimates which are based on survey results and are in general incompatible with national accounts data. The paper also introduces an empirical framework which is used to make poverty projections for the LDCs, and can be also useful in assessing the attainment of the Millennium income poverty goals in the long run. A comparison between the new poverty estimates based on the 5 PPP exchange rates and the previous estimates based on the 1993 PPP values is made for the LDCs. We also statistically test the validity of the new results as compared to the World Bank global poverty estimates and discuss the sensitivity of results to the possible errors in distribution data.

3 1. Introduction The latest update of the purchasing power parity (PPP) exchange rates by the World Bank has led to major revisions in the global poverty count in various regions of the world. At the same time global poverty lines have been redrawn on the basis of new data on national poverty lines and the new PPP exchange rates. In this paper we provide an update of our 7 round of national accounts consistent poverty measures in the light of these new developments. The appearance of the new and more accurate PPP exchange rate data (the 5 round results), combined with the updating of the global poverty line, imply considerable changes to the poverty measures in the case of the low income countries, particularly the LDCs, irrespective of the estimation methodologies used. In section 2 we begin with a comparison of the results of the new round of PPP exchange rate estimates with the previous rounds, and discuss the new poverty line introduced by the World Bank. In section 3 we highlight the lack of consistency between national accounts and survey consumption means and derive the calibrated survey means which utilize the information in both the national accounts and household surveys. Section 4 deals with some of the criticisms of the calibration method. In Section 5 we discuss the new national accounts consistent estimates of global poverty and compare the individual country results for the LDCs with the previous estimates as well as with the World Bank estimates. In section 6 we discuss the estimation of poverty curve and its use in estimating poverty trends for the LDCs. Aggregate LDC poverty estimates are also presented in this section, including a comparison with the previous estimates. Section 7 deals with the validation tests of the results against the World Bank estimates in a series of model selection exercises. In this section we also analyse the sensitivity of the global poverty trend estimates to the possible errors in distribution data. 2. Purchasing Power Parity Exchange rates and the Global Poverty Line The $1 a day poverty line originally proposed by the World Bank has become a popular yardstick for extreme poverty at the global level. This poverty line was originally derived as the average of official poverty lines for a number of low income countries measured in 1985 purchasing power parity exchange rates (see, Ravallion et al 1991). Purchasing power parity (PPP) exchange rates are used to take account of price differences across countries, thus making the global poverty line comparable across countries in purchasing power terms. The PPP exchange rates used in the original derivation of the $1 a day poverty line were based on 1

4 early estimates of PPP exchange rates by International Comparison Program (ICP), published as the Penn World Tables version 5.6. In this early version, due to lack of information on prices for many developing countries, including most LDCs, the PPP estimates for these countries were based on econometric extrapolations which involved sizable errors. In the World Bank produced new PPP estimates for developing countries, which supplemented the price data in the existing ICP estimates with additional price data for some of the developing countries. The World Bank adopted the new estimates as more reliable data compared to the Penn World Tables version 5.6 for its round of global poverty estimates. Measured in the 1993 PPP exchange rates, the new global poverty line in 1993 international prices became $1.8 a day. 1 According to Chen and Ravallion (1), this new poverty line was the median of national poverty lines in ten poorest countries in Ravallion et al (1991) dataset, measured in 1993 PPP. Between and 7, the global poverty estimates by the World Bank were measured using the new 1993 PPP exchange rates and the $1.8 a day poverty line. We adopted the same global poverty line and the 1993 PPP exchange rate estimates by the World Bank in our 7 update of national accounts consistent poverty estimates (see, Karshenas 8 and UNCTAD 8). 2 In 8 the World Bank conducted another major revision of its global poverty measures, based on the results of the ICP 5 round of PPP exchange rate estimates (see, Chen and Ravallion, 8). Following the recommendations of the Ryten Report (UN 1998), the 5 round of ICP adopted new methodologies, and according to the World Bank benefited from more rigorous action to improve data quality. It also incorporated price data from a much larger number of countries (see, World Bank 8). For example, while the 1993 PPP exchange rates were extrapolated in the case of sixteen Asian and African LDCs, the 5 estimates appear to have relied on actual price data for all the Asian and African LDCs. This in itself would imply considerable improvement in the 5 PPP exchange rate estimates for the LDCs. Both to be in conformity with the new global poverty estimates by the World Bank, and to benefit form the improved data for the LDCs, in this study we rely on the ICP 5 exchange rate estimates to measure the new round of the national accounts consistent estimates of global poverty. This of course would imply a revision, and in the case of some countries sizable revision, of the previous poverty estimates. 1 See, Karshenas (4) for a comparison of global poverty lines under Penn World Tables version 5.6, 6.1 and 1993 World Bank PPP exchange rates and different country ordering assumptions. 2

5 A comparison between the implied price levels between the 1993 and 5 ICP rounds is shown in Figure 1. The figure shows the distribution of the percentage change in national price levels between the two ICP rounds in 5 for developing countries. 3 The values of 5 price levels for the 1993 ICP round are based on extrapolations made by the World Bank (see, World Bank 8a). As can be seen there are large differences in price levels between the two rounds, with a positive bias, indicating an upward revision of price levels for most developing countries. For this reason the 8 global poverty estimates indicate higher absolute poverty than the earlier measures based on 1993 PPP exchange rates (see, Chen and Ravallion 8). The same price comparison in the case of the LDCs, shown in Figure 2, indicate a similar range of price revisions, but with a much higher share of countries showing relatively high positive price revisions. This indicates that survey means at 5 PPP are likely to be lower than measured Figure 1, Change in Price Level between ICP5 and previous at 1993 PPP and hence global estimates in Developing Countries poverty measures for many LDCs 35 3 are also likely to be relatively 25 higher than hitherto estimated. 15 The revision of the PPP exchange 1 5 rates would of course also have implications for absolute poverty % change in price through its impact on global Source: World Bank 8a poverty line. The final effect on Figure 2, Change in Price Level between ICP5 and previous absolute poverty depends on the estimates in the Least Developed Countries (LDCs) outcome of the two effects, 12 1 namely the survey mean effect 8 and the poverty line effect. number of countries number of countries For the 8 global poverty 2 estimates the World Bank has % change in price introduced a new poverty line Source: World Bank 8a measured at 5 PPPs and based on a new dataset of national poverty lines. The new poverty line is based on a dataset of More More 2 In addition to the $1.8 a day poverty line, Karshenas (8) also reports global poverty based on food calorie intake global poverty line suggested by Kakwani and Son (3). 3 Price level is defined as the PPP exchange rate divided by the official exchange rate (both in domestic currency per dollar terms), indicating the price level in the country relative to international prices. 3

6 developing countries over period (see, Ravallion et al, 8). This new data is plotted against per capita consumption expenditure (from national accounts) at 5 PPP rates in Figure 3. The solid line if Figure 3 is a simple exponential function fitted to the data. As expected, national poverty lines tend to increase with per capita consumption expenditure, but the relationship appears to be flat for low income countries. This behaviour is very similar to the patterns exhibited in old data set of national poverty lines analysed by Ravallion et al (1991). The average of the national poverty lines for the poorest 15 countries (in terms of per capita consumption expenditure) is $1.25 a day, which is proposed by Ravallion et al (8) as the new lower global poverty line. 4 Figure 3, National Poverty Lines vs Per capita Consumption Expenditures Poverty line (5 PPP $) Per capita expenditure (5 PPP $, log scale) Source: Ravallion, Chen and Sangraula 8 In terms methodology, this new global poverty line is commensurate with the original $1 a day in 1985 PPP values, and $1.8 a day in 1993 PPP values used by the World Bank in previous estimation rounds. However, given the higher price levels in most low income countries in the 5 ICP round discussed above, one would expect the value of the new global poverty line to be lower than the earlier estimates. Ravalloin et al (8) show that the $1.8 poverty line in 1993 prices, adjusted for inflation factor, is equivalent to $1.45 a day in 5 values. The lower value of $1.25 a day global poverty line in the new 5 PPP rates, therefore, to some extent neutralizes the lower values of survey means resulting from the revised PPP data discussed above. 4 The 15-country average of the expected values based on the exponential regression line fitted in Figure 3 4

7 The median of poverty lines for the entire data set shown in Figure 3 is $2 a day, which can be taken as the higher global poverty line. Chen and Ravallion (8) have reported global poverty estimates for a range of additional international poverty lines, e.g., $1, $1.45, and $2.5 a day. The estimates based on the $1.25 and $2 a day poverty lines appear to be the most commonly referred to in the literature, and are likely to be also more consistent with the earlier global poverty lines measured at various older PPP exchange rates. In this paper we shall be reporting poverty estimates based on these two poverty lines, but the results can be readily replicated for the other poverty lines used by Chen and Ravallion (8). 3. Lack of Consistency between the Survey and National Accounts Averages The large errors involved in the estimation of PPP exchange rates have been a major source of uncertainty in global poverty estimates. This problem has been alleviated over time by getting access to better price data and use of improved methodologies. Another important source of uncertainty arises from the discrepancy between survey data and national accounts information on per capita consumption. The lack of consistency between survey means and national accounts based measures of average income or consumption has long been recognized. 5 The nature and extent of this problem can be glanced from Figure 4 which shows the scatter plot of survey means against national accounts averages. Observations in Figure 4 depict individual country data for one or more survey years for each country. In order to use the same conversion factors in the two series, the national accounts expenditure means are first converted to 5 values in domestic currency and then valued at international prices using the 5 PPP exchange rates. The discrepancy between the two series in 5 PPP values for each country will be therefore the same as when expressed in domestic currency in the survey year, and hence PPP conversions do not play any role in explaining the discrepancies. There are important differences in coverage and definition of per capita consumption in the household surveys and national accounts and hence they are not expected to be equal in the sense of following the 45 degree line in Figure 4. However, large and non-systematic produces the same result as reported by Ravallion et al (8). 5 See e.g., Pyatt (3), Ravallion (, 1, 3), Deaton (, 2, 3), Bhalla, Karshenas (1, 3), Sundaram and Tendulkar 2, and Kulshreshtha and Kar 2 5

8 differences between the two variables in many instances can make comparability across countries or over time problematic. For example, as shown in Figure 4, country A has a survey mean four times higher than its national accounts per capita consumption expenditure. On the other hand, country B exhibits a national accounts expenditure average, which is four time higher than its survey mean. Similar discrepancies can be observed in the behaviour of the two variables over time, where in some countries growth in national accounts based consumption data coincides with decline in the survey mean, or vice versa. Given the level of income distribution, this can create large discrepancies in poverty measure depending which source is taken as the indicator of average standard of living in the country, thus rendering comparability of poverty measures across countries and over time problematic. Figure 4, Survey vs National Accounts Means Survey Mean (5 PPP $), log scale 1 A (2, 978) 1 B (1654, 412) N.A. mean consumption (5 PPP $), log scale consumption survey income survey 45 degree line Various authors who have acknowledged this problem have either opted using one or the other source in their poverty measurements. For example the World Bank continues to use the survey means in conjunction with survey based income distributions to measure poverty on grounds that the nature of error in survey means is likely to be such that it does not affect poverty measures (see, below). Others, e.g., Bhalla (2) and Salai Martin (2) have used the survey distribution data in conjunction with national accounts averages to estimate global poverty. Each of these two options discards the information available in the other source. As we shall argue below, estimating poverty on the basis of the national accounts average consumption alone is highly problematic, as it can give rise to systematic errors in addition to those arising from the idiosyncratic measurement errors in average consumption. On the 6

9 other hand, reliance on survey averages alone can suffer from lack of comparability between surveys conducted in different countries and different times with different methodologies, sampling frames, sample sizes, recall periods, etc. An improved method may be to use the information in both sources by calibrating survey using national accounts statistics a method which is adopted in the present paper. A minimum set of criteria need to be obeyed for the survey means to produce poverty measures which are consistent with national accounts information. Firstly, if two countries A and B have the same income distribution, but country B has a higher per capita income or consumption than country A according the national accounts, then poverty in country B should be lower than country A. Secondly, if income distribution remains constant in a country, but national accounts data show growth of per capita income and consumption over time, then poverty should be decreasing in such a country. These minimal criteria imply a positive association between survey mean and national accounts averages, which we shall utilize to calibrate the surveys averages to be consistent with national accounts. A simple method of calibration of survey means which obeys these criteria is to take the mean of household survey expenditure (S) conditional on national accounts mean (N), E(S / N) = f(n), with the constraint that f(.) is a positively sloped function, as the calibrated survey mean (see, Karshenas, 3 and 4). We estimate f(n) by fitting a smooth curve to the data in Figure 4, and calibrated survey means are read off the fitted curve corresponding to the national accounts means for each country. Since the pattern of observations from income surveys and consumption surveys were clearly different, separate curves were fitted to the two set of observations. We fitted various smooth non-parametric curves to the data, but in all cases these gave rise to a negative slope between the survey and national accounts means for low income countries, which contradicted the minimum criteria discussed in the above paragraph. We next fitted parametric curves and chose functional forms which best mimicked the non-parametric curves for higher income countries. For both the income and consumption surveys the following quadratic functions turned out to be the best fits 6 : Consumption surveys: s = c c 2 R 2 =.75 (1.4) (.39) (.28) 6 Each country in the regression equation has one or more observations related to different dates of their surveys. To detect any time effects we included various time dummies for different decades, as well as the date of the survey, but none of the time variables turned out to be significant, either individually or jointly. 7

10 Income surveys: y= c 2 R 2 =.73 (2.12) (.57) (.39) Where s, c an y are logarithms of consumption survey mean, national account consumption mean, and income survey mean respectively. Standard errors are in parentheses. In both regressions per capita consumption from the national accounts are used as calibrating variables. The fitted curve to the consumption survey data is shown in Figure 5. Since all the LDC countries with the exception of Haiti report consumption expenditure in their surveys, it is the fitted line in Figure 5 which is used to calibrate survey means for the LDCs. Figure 5, Survey vs National Accounts Means with fitted Regression Lines Survey Mean (5 PPP $), log scale N.A. mean consumption (5 PPP $), log scale survey mean fitted line 45 degree line It is clear that the fitted survey means in Figure 5 obey the minimum consistency requirements discussed above. The shape of the fitted curve in Figure 5 also highlights other important information regarding the systematic relationship between the national accounts and survey averages. As pointed out by Deaton 2, the definitional differences between the national accounts and survey consumption concepts imply that the national accounts per capita household consumption is likely to grow faster than average consumption from the surveys, particularly in the early stages of development. One reason being that the household surveys incorporate information on income and consumption derived from informal activities which may not be captured by national accounts statistics. National accounts private consumption expenditure data on the other hand incorporates consumption by semi public and charity institutions which are not included in household surveys. As along the 8

11 development path and share of the latter items is likely to increase and the share of the informal sector activities is likely to fall, national accounts survey means are likely to growth faster than survey averages. This is clearly indicated in the shape of the fitted curve in Figure 5. A comparison between the fitted curve and the 45 degree line also highlights the problem with using national accounts averages for poverty estimation. Such a practice, as in Salai Martin 2 and Bhalla, clearly leads to an overestimation of poverty reduction rates, particularly in the case of low income countries. The same criticism applies to the World Bank extrapolations of poverty trends where survey averages are extrapolated forward by applying national accounts growth rates (see, e.g., Chen and Ravallion 8, page 15). 4. Criticisms of the Calibration Approach The approach adopted in this paper is not free from criticism, and has its own shortcomings though we argue not as serious as the other approaches. The use of survey means in poverty measurement by the World Bank has been based on the contention that survey mean error is mainly due to non compliance of the rich. In that case, even though the survey mean may be biased the surveys nevertheless generate correct poverty estimates. 7 Under these circumstances, as argued by Ravallion 3, and Deaton 3, the correction of the survey mean bias can lead to underestimation of poverty by unduly increasing the income of the poor. This is a valid argument, to the extent that it can be shown that the apparent error in survey means are in fact dominated by non-compliance error. One indication of the problems associated with the non-compliance hypothesis is that it assumes survey means to be systematically underestimated. However, as is shown in Figure 4, in many instances, particularly in the case of low income countries which are of interest to us, surveys means are well above national accounts means. More rigorous tests also indicate that survey mean errors cannot be solely due to non-compliance. For example, as shown in Karshenas (4), if the non-compliance hypothesis is correct, one should observe a positive relationship between survey mean error (underestimation) and poverty as measured by noncalibrated survey means, across the sample countries. This result is based on the fact that under the non-compliance hypothesis, poverty as measured by survey results will be accurate, 7 This is only true if the survey covers the entire population. In sample surveys where poverty is calculated on the basis of the sample proportions poverty estimation will be biased even if the poor are all compliant (see, Karshenas, 4). 9

12 even though the mean and distribution of the surveys are wrong. To test this hypothesis, we followed Deaton 3 by depicting survey error as the log ratio of the survey mean over national accounts mean, and regressed this variable on the World Bank measures of $1 a day and $2 a day poverty estimates. We repeated these regressions for samples including and excluding countries where survey means were larger than national accounts averages. In none of the regressions this relationship turned out to be significant. We also added distributional variables such as the gini coefficient to the regressions but both the distribution and poverty variables turned out to be insignificant. These results, which are similar to others found in the literature (See, e.g., Deaton 3 and Karshenas 4), are also highlighted by the scatter plots in Figures 6 and 7 which show the lack of a significant relationship between log difference in the two means and the $1 a day and $2 a day poverty lines respectively. log mean ratio (survey / N.A.) Figure 6, Log mean ratio against $1.25 a day Headcount Poverty y =.13x R 2 Headcount poverty, % below $1 a day =.7 log mean ratio (survey / N.A.) Figure 7, Log mean ratio against $2 a day Headcount Poverty y =.9x R 2 Headcount poverty, =.39 1

13 The above results do not of course mean that non-compliance of the rich is not a source of error in survey means. What these tests indicate is that in our sample countries there are other more important sources of error that overshadow the non-compliance error, thus lending support to our treatment of errors as more akin to random numbers rather than systematic underestimates as maintained by the non-compliance hypothesis. One may therefore argue that the lack of a significant relationship between poverty and the mean deviations between the surveys and the national accounts data, supports the practice of using the calibrated survey means combined with the distribution indicators from the surveys in poverty measurement followed here. This, however, can be criticized as it assumes the survey means to be error ridden but no adjustments are made to the distribution data from the surveys. As pointed out by Ravallion 3, how can one assume that the shape of the distribution is correct but its mean is error ridden? It is likely that both the survey mean and its distribution are subject to large measurement errors. The question is how significant these errors are and what can be done about them. Measurement errors in surveys will always exist, but the question that one needs to address is how important the errors are and how significantly they can affect poverty estimates. As to the first question, as seen in the previous section, the measurement errors in means appear to be too large to be ignored. The coefficient of variation of the log ratio of survey to national accounts mean is about 1.4 for consumption surveys, as compared to a coefficient of variation of.19 for the Gini coefficient for sample countries. What is more important to note, however, is that while the errors in survey means have first order effects on poverty measurements, the effect of the distribution errors is only of second order, and hence likely to be much less significant. This is shown in Karshenas 4, where the addition of a value as large as one standard deviation to Gini coefficients in the sample countries changes poverty measures relatively much less than those arising from mean adjustment resulting from the calibration of survey means. Given the lack of significant observable relationship between survey mean errors and distribution indices, it is not clear how best to adjust the decile data without further research. Given the relative stability of the decile distribution the best strategy may be to leave them as they are. As we shall see below, variations in mean consumption appear to have a more significant impact on poverty than distributional changes in the income ranges relevant to the LDCs. Furthermore, considering that in our sample countries on average over 7 per cent of expenditure or income belongs to the top 4 per cent of income groups, much of the 11

14 adjustment in survey mean, keeping the decile distribution constant, will be allocated to the rich households. 5. National Accounts Consistent Poverty Estimates in the LDCs Using the calibrated survey means in 5 PPP values and the decile distribution from the surveys we estimate different poverty measures for individual countries for the $1.25 a day and $2 a day poverty lines. The new poverty measures along with calibrated survey means are reported in Appendix A, Tables A1, A2 and A3, for headcount poverty, poverty gap and square poverty gap measures respectively. It should be noted that we are using the same PPP exchange rates as the World Bank and the poverty lines are also the same ones as used by the Bank (namely, $1.25 a day and $2 a day). Hence the differences between the present estimates and the estimates by the World Bank are the result of the calibration of the survey means in this paper in order to make our estimates consistent with national accounts data. A comparison between the World Bank s survey based estimates and the new estimates is made in Figure Figure 8, Headcount poverty in LDCs (), National Accounts vs Survey based estimates Survey Based degree line National Acounts based Figure 8 highlights the fact that contrary to the predictions of the non-compliance hypothesis, the new poverty estimates based on calibrated survey means do not systematically underestimate poverty as compared to the World Bank estimates. In a large number of observations the new estimates are higher than the World Bank estimates and in some cases the new estimates are considerably higher, both at the low and high end of income scale. 12

15 In the majority of the LDCs, the new poverty estimates based on the 5 PPP exchange rates and the updated poverty lines show increased poverty levels compared to earlier estimates. This is shown in Figure 9 which compares the new poverty estimates with the previous Figure 9, Headcount Poverty in LDCs, 5 PPP vs 1993 PPP based Estimates (a) World Bank Estimates 1 (5 PPP) % below $1 a day (1993 PPP) (b) National Accounts Consistent Estimates (5 PPP) % below $1 a day (1993 PPP) estimates. Panel (a) in Figure 9 shows this for the case of the World Bank estimates, and Panel (b) shows data for the national accounts consistent estimates for the latest years for African and Asian LDCs where comparable data are available. Figures 1 and 11 show the same phenomenon with names of countries specified. The increase in poverty estimates in most cases is substantial. With one or two notable exceptions, the increase in poverty measures, particularly in the case of some of the most populous LDCs implies that the new estimates irrespective of methodology imply substantial upward revisions in aggregate LDC poverty (see, below). Another interesting phenomenon is that the new more accurate PPP estimates have lead to a convergence between the survey based and national accounts based poverty estimates. This can be detected by a comparison of Figures 12 and 13, which show the previous and current poverty estimates by the two methods side by side. Figure 13 suggests that at least in relative 13

16 Figure 1, World Bank poverty estimates for the LDCs, based on 1993 and 5 PPP exchange rates % below $1 / $1.25 a day Benin Burkina Faso Cent. Afr. Rep Ethiopia Gambia Lesotho Madagascar Malawi Mali Mauritania Mozambique Niger Rwanda Senegal Sierra Leone Tanzania Uganda Zambia Bangladesh Cambodia Lao PDR Nepal Yemen, Rep PPP 5 PPP Figure 11, National accounts consistent poverty estimates for the LDCs, based on 1993 and 5 PPP exchange rates % below $1 / $1.25 a day Benin Burkina Faso Cent. Afr. Rep Ethiopia Gambia Lesotho Madagascar Malawi Mali Mauritania Mozambique Niger Rwanda Senegal Sierra Leone Tanzania Uganda Zambia Bangladesh Nepal 1993 PPP 5 PPP Figure 12, National accounts consistent and survey based poverty estimates for the LDCs, based on 1993 PPP exchange rates Benin Burkina Faso Cent. Afr. Rep Ethiopia Gambia Lesotho Madagascar Malawi Mali Mauritania Mozambique Niger Rwanda Senegal Sierra Leone Tanzania Uganda Zambia Bangladesh Nepal % below $1 a day World Bank N.A. Based Figure 13, National accounts consistent and survey based poverty estimates for the LDCs, based on 5 PPP exchange rates World Bank N.A. Based Benin Burkina Faso Cent. Afr. Rep Ethiopia Gambia Lesotho Madagascar Malawi Mali Mauritania Mozambique Niger Rwanda Senegal Sierra Leone Tanzania Uganda Zambia Bangladesh Nepal terms the difference between the two methods has narrowed. Indeed the correlation coefficient of poverty estimates by the two methods increases from.4 using the 1993 PPP 14

17 exchange rates to.7 using the new 5 PPP estimates. Of course the improved 5 PPP estimates would have the same effect on national accounts and survey based consumption. The convergence between the two poverty estimates indicates that the revision of PPP exchange rates has brought some outlandish observations on average consumption estimates for the LDCs closer to the international patterns as depicted by the fitted line in Figure 5 above. 6. Poverty trends in the LDCs The World Bank provides estimates of global poverty trends between 1981 and 5 on a three yearly basis, by interpolating poverty estimates for the reference years where survey results do not exist. Survey means are interpolated by applying growth rates of per capita household consumption from national accounts to survey means. For countries with only one household survey, poverty is estimated by applying the resulting average consumption estimates to the available distribution data, assuming the Lorenz curve remains fixed. For countries with two or more surveys, poverty rates for reference years between the surveys are estimated by a weighted average of the resulting estimates. This procedure, particularly given the anomalies between the national accounts and survey averages discussed above, leads to various inconsistencies and problems which admittedly have been dealt with only in an ad hoc and unsatisfactory manner (see, Chen and Ravallion 8, p15). At the heart of these inconsistencies lies the practice of updating survey means by use of national accounts growth rates. Apart from data inconsistencies, as argued in Section 3 (see, Figure 5) this practice can lead to an over estimation of household consumption growth on average, and particularly serious overestimation of poverty reduction for countries with one or two surveys over a long period of time. Many of these problems and inconsistencies are avoided by the interpolation procedure followed here, which is based on using calibrated survey means to estimate poverty trends for reference years in which survey data are not available. For each year the actual per capita private consumption from the national accounts is used to estimate calibrated survey means by the regression method discussed in Section 3 above. In this way the interpolation of per capita consumption for all the countries and all years is done in a consistent and coherent manner. As to the distributional assumptions for the reference years when surveys are not available we use a similar procedure as the World Bank, but instead of using weighted 15

18 averages of poverty estimates with fixed distribution assumption at survey points, we try to use more systematic cross-country data for the averaging process by the help of poverty curves as discussed below. For a given income distribution and average per capita income m, poverty is uniquely determined as a function of the distance between m and poverty line z, say f(m/z). In the case of headcount poverty the function f(m/z) defines the proportion of the population living below the poverty line z. For a given poverty line, as the mean and distribution of income evolves over the development path, f(m/z) traces the poverty curve as a function of m/z (see Karshenas 1 and Karshenas and Pyatt 6). The shape of the poverty curve depends on the relative variations of income distribution and per capita income over the development path, but given the relatively larger variations of mean income referred to in section 4, one would expect a downward sloping curve (see Karshenas and Pyatt 6 for a more detailed discussion). As in Karshenas 1, we estimate the poverty curve as the expected value of f(m/z) on the basis of the empirical observations of headcount poverty across countries for the sample countries with available data. The scatter plot of the poverty estimates based on $1.25 and $2. a day poverty lines, in relation to normalized mean income (m/z) is shown in Figure 14. The figure also shows the fitted poverty curve, which depicts expected poverty as a logistic function of m/z. Since the survey data for most LDC countries cover per capita consumption, Figure 1 is based on observations with consumption poverty. Hereafter, any reference to income distribution thus refers to the distribution of consumption expenditure. The vertical distances between individual observations and the fitted poverty curve indicate the divergence of poverty in each country from the average pattern due to idiosyncratic income distribution effects. Of course the poverty curve itself also incorporates systematic changes in poverty along the growth path, and hence it will be a mistake to view the poverty curve as a device to decompose poverty changes into independent components such as income distribution and growth. The close fit of the poverty curve to the data at the upper ranges of poverty was utilized in LDC reports 1 and 4 to make projections of poverty for LDC countries with high poverty rates. This may be an appropriate method for the poverty ranges observed in LDCs using the $2 a day poverty line. At the lower ranges of headcount poverty, below 5 to 6 per cent, however, the deviations from poverty curve become more pronounced and at even lower 16

19 poverty levels the deviations rather than m/z appear to dominate the movement of poverty. For this reason in the last update of LDC poverty projections we used the gini coefficient as an indicator of the effect of income distribution for the projection of LDC poverty levels. This led to a substantial reduction in the variances of our projections in the case of $1 a day poverty line with relatively lower poverty levels (see, Karshenas 8). It Figure 14, Headcount poverty vs normalized consumption m/z and poverty curve I should be noted however, that the effect of income 1 9 distribution on poverty is 8 not uniform. It depends on 7 6 the position of the poverty 5 line relative to mean 4 3 income, and in low income countries where the 1 international poverty line average consumption / povety line (m/z) can be higher than mean income, distributional changes can have perverse Figue 15, Headcount poverty against normalized Sen's index and effects. This is to some poverty curve II extent reflected in the 1 dispersion of observations 9 8 around the poverty curve in 7 6 Figure 15, referred to as 5 poverty curve II. 4 Headcount poverty (%) Headcount poverty (%) 3 Figure 15 differs from 1 Figure 14 in that headcount poverty is plotted against (1-gini) * average consumption / povety line (1-g)m/z the Sen s index divided by the poverty line (1-g)m/z rather than m/z, where g is the gini coefficient. As expected, the introduction of the gini coefficient has a remarkable effect in reducing the dispersion observations around the poverty curve at poverty levels below 6 per cent. The effect on the observations with very high poverty levels (where mean income is equal or less than the 17

20 Table 1, Regression of headcount poverty on mean consumption and gini coefficient Dependent variable, % below $1.25 and $2 a day poverty lines u 2.5 *** u2 -.3 *** u3.25 *** u4 -.1 *** u(1-g) *** u(1-g) *** u(1-g) *** u(1-g)4.13 *** ln(u) -.59 *** g g *** _cons 1.44 *** no. of observations 48 Adj R-squared.994 Notes: u is normalized mean consumption (m/z). Gini coefficient is g. * p<.5; ** p<.1; *** p<.1 poverty line) is however the opposite. As noted above, at such high poverty levels, income distribution can have perverse effects on poverty, which in Figure 15 is captured by increased variations around the poverty curve. We take this into account in specifying the variance structure of the estimated poverty curve in updating the 7 LDC poverty estimates in the present paper. In order to make poverty projections for the LDCs by incorporating information both on income distribution and average consumption, we have regressed headcount poverty on polynomials of mean consumption, gini coefficient, and cross products of mean consumption and gini coefficient. To include poverty measures based on both the $1.25 and $2 a day poverty lines, the mean consumption is normalized by poverty line (m/z) and the cross products are entered as (m/z)(1-gini), or the Sen index normalized by the poverty line. In order to incorporate the differential distributional effects on variances in the case of low income countries waited least squares was used, correcting for heteroskedasticity. Table 1 reports the results with polynomial degrees that achieved the best fit. It is not surprising to find that the regression achieves almost perfect fit, with an adjusted R squared of over 99.4 per Predicted headcount poverty Figure 16, Actual vs Predicted Poverty Actual headcount poverty cent. The standard deviation of predictions from this regression for all the observations has a mean of.4 and variance of.24. The predictions of the regression are plotted against the actual poverty estimates in Figure 16. As can be seen, the scatter plot closely follows the 45 degree line with minimal prediction errors. We have used the estimated regression in Table 1 to make headcount poverty projections for LDCs. 18

21 The question may arise that once we have the information on both the Gini coefficient and average income, poverty can be measured directly and hence no matter how precise the above type of indirect prediction method may be it will be devoid of practical value. This is not however entirely correct. Even if we can make relatively accurate extrapolations of the Gini coefficient for some countries, this will not be sufficient for estimating headcount poverty, unless we make further assumptions about income distribution - e.g., log-normality assumption. However if income distributions in LDCs follow a log-normal distribution, the regression method adopted here can be considered as a short cut method which should produce the same results as individual country estimates, as the log-normal distribution is fully specified by its mean and gini coefficient (variance). Furthermore, the complicated interactions in the regression results in Table 1 seem to indicate that the log-normal assumption may not be an appropriate assumption. Given the nature of the income distribution data available, the indirect method also allows the examination of the sensitivity of the poverty estimates to various distributional assumptions in a practical manner. As discussed in Karshenas 8, given the nature of the available data in the case of the LDCs, this proves a valuable tool in producing estimates of poverty while providing us with some idea about the degree of reliability of such estimates (see, Section 7). The LDC countries fall into four categories in terms of availability of data: Group A countries are those countries for which numerous surveys exit and national accounts data on per capita consumption in 5 PPP is also available. Group B are countries where distribution information is available for only one year, but national accounts mean consumption data are available for all the years. Group C countries are countries where no surveys exist but national accounts mean consumption in 5 PPP values is available. Group D are countries where there are no survey data available, nor do they have PPP exchange rate estimates. In this group we can also include countries for which we may have scant survey information but national accounts averages are missing. In the case of countries in group A we interpolate the distribution of consumption, as linear interpolations of the gini coefficient between two survey years, on the basis of exiting survey distribution data. For the end years, before the first survey and after the last survey, we assume gini coefficient remains constant. Given the relatively slow changing characteristic of income distribution, and depending on the frequency and spacing of the surveys, we should be able to come up with relatively reasonable estimates of gini coefficient for this group of 19

22 countries. For countries in group B we assume a constant gini coefficient throughout the period. Using the gini coefficient and mean calibrated consumption for these countries, we produce estimates of headcount poverty on the basis of the regression equation in Table 1. The sensitivity of the results to the errors in gini coefficients is examined in the next section. The conclusion of the sensitivity analysis in the last round of our poverty estimates was that the variations in poverty projections arising from distributional assumptions were negligible in the case of the $2 a day poverty line for the LDC countries, where headcount poverty was above 6 per cent. In the case of the $1 a day poverty line it was found that distributional assumptions mattered much more, but errors involved were relatively small compared the errors arising from the discrepancy between the survey and national accounts means (see, Karshenas, 8). In other words, even in the case of Group C countries where no distributional data are available, but mean consumption data exist, it may make sense to assume a gini coefficient equal to the average LDC level as a whole for the purpose of estimation of poverty trends. Nevertheless in the present paper we have chosen not to produce poverty estimates for countries where no distribution data is available. As it happens, for most countries in this group data on mean consumption in 5 PPP exchange rates are not available either. The only country which falls in this category with available national accounts consumption data in 5 PPP rates is Sudan, where there is no survey in the World Bank data set, but unofficial estimates indicate an average gini coefficient of consumption expenditure of.5 (A. Ali, 3). We have included Sudan in this round of national accounts consistent poverty estimates, by assuming a constant gini coefficient of.5 for the entire period. Other low income LDCs such as Congo Dem.Rep., for which we have only one survey of possibly dubious quality (here categorized in Group B) also come close to the case of Sudan. We shall discuss the sensitivity of poverty trends to the assumptions regarding income distribution for such countries in the next section. Finally there is the case of group D countries which is rather straightforward. In the case of this group of countries we cannot estimate internationally comparable poverty estimates. The lack of appropriate time series data on per capita consumption in 5 PPP exchange rates has been the main reason for excluding a number of the LDCs from our estimation of poverty trends below (Afghanistan, Angola, Somalia, Bhutan, and Myanmar are amongst the notable African and Asian LDCs in this group). Excluding this group of countries we are left with 33 LDCs in Africa and Asia plus Haiti, which constitute about 9 per cent of the African and Asian LDC population.

23 A comparison between the new and previous estimates of aggregate headcount poverty for the African and Asian LDCs is made in Figure 17 for the period. The Figure provides headcount poverty estimates for the lower poverty line adopted by the World Bank, that is, $1.25 a day in 5 PPP values which is believed to be equivalent to the old poverty line of $1.8 a day in 1993 PPP rates. Equivalent headcount poverty estimates are also provided using the 1993 PPP rates based on our previous round of poverty estimates. In order to make the new aggregate estimates comparable to the previous estimates, the number of countries has been curtailed from 33 to 24. It can be seen that the new more accurate PPP rates have led to increased headcount poverty of between 16 to percentage points for both the Asian and African LDCs for all the years over the period. As noted above, this is the combined result of the new PPP exchange rates which indicate higher price levels in LDCs than previously believed, as well as the new poverty line in 5 values. Applying these ratios to the overall LDC population it will be revealed that for example in 5 the errors involved in 1993 PPP estimates led to an underestimation of headcount poverty in African LDCs of about 68 million and 56 million in Asian LDCs, adding up to 124 million underestimation for the LDCs as a whole. Firgure 17, Headcount poverty trends New and Previous Estimates compared, a- African and Asian LDCs c- Asian LDCs 1996 New estimates b- African LDCs 1997 Notes: New estimates refer to % living below 1.25 a day at 5 PPP exchange rates. Previous estimates refer to % below $1.8 a day at 1993 PPP rates. Based on curtailed sample of 24 countries Previous estimates The substantial upward revisions of poverty count in LDCs resulting from the 5 PPP exchange rate revisions are larger than the 9 percentage point increase reported by Chen and Ravallion (8) for developing countries as a whole. As discussed in the previous section, the 5 revised PPP estimates have led to a much higher upward price revision for the LDCs as compared to other developing countries. The next question is the way such upward revisions affect the trends in poverty reduction over time. As pointed out by Chen and

24 Ravallion (8), since the new PPP estimates do not affect the Lorenz curves or consumption growth rates over time, the trends in poverty reduction between the new and previous estimates should not be materially different. As shown in Figure 17, this appears to have held for the case of the Asian and African LDCs as a group over the period as well. However, as shown in Figures 14 and 15, the relationship between poverty and per capita consumption is highly non-linear, and hence drastic changes in the measured level of poverty in the case of individual countries can lead to significant changes in poverty trends for any given growth and distribution of consumption. For example at poverty rates of 4 per cent, responsiveness of poverty to growth is expected to be much higher than at per cent or 7 per cent levels (see, UNCTAD 2, and Karshenas 1). Trends in headcount poverty in African and Asian LDCs and Haiti, based on 33 countries for which data are available, are shown in Table 2 for the period It should be noted that these results are not comparable to aggregate estimates plotted in Figure 17, particularly in the case of African LDCs where a number of large countries such as Congo Dem. Rep. and Sudan were omitted from the data in Figure 17 for the sake of comparability with previous estimates. In addition to the data in Table 2, the aggregate trends of headcount poverty for the larger list of LDCs is also plotted in Figure 18. As can be seen from Table 2 and Figure 18, despite the declining trends in headcount poverty since the mid 199s, the share of the population living below $1.25 a day in 5 for the LDCs as a whole, (at 55.3%), was still higher than the 198 level. The number of the people living below $1.25 a day in the meantime has doubled increasing from about 186 million to about 363 million in the same period. Much of the increase of about 18 million persons in poverty has been due to the African LDCs. Between 198 and 5 the number of people in extreme poverty in African LDCs increased by about 14 million. The number of people living below $2 a day in LDCs as a group increased from 273 million in 198 to 518 million in 5 an increase of 245 million. About 61.5 million of this increase was accounted for by the Asian LDCs, but the major part was due to increase in the number of the poor in African LDCs. 8 African LDCs include Benin, Burkina Faso, Burundi, Central African Republic, Chad, Congo Dem. Rep., Djibouti, Ethiopia, Gambia, Guinea, Guinea-Bissau, Lesotho, Liberia, Madagascar, Malawi, Mali, Mauritania, Mozambique, Niger, Rwanda, Senegal, Sierra Leone, Sudan, Tanzania, Togo, Uganda, and Zambia. Asian LDCs include, Bangladesh, Cambodia, Lao PDR, Nepal, and Yemen, Rep. 22