An assessment of trends in poverty in Ghana 1988 - 1992

81 PSP Discussion Paper Series 31097 An Assessment of Trends in Poverty in Ghana 1988-1992 Harold Coulombe Andrew McKay November 1995 Poverty and Social Policy Department Human Capital Development and Operations Policy The World Bank  Abstracts This Booklet of Abstracts contains short summaries of recent PSP Discussion Papers; copies of specific papers may be requested from Patricia G. Sanchez via All-in-One. The views expressed in the papers are those of the authors and do not necessarily represent the official policy of the Bank. Rather, the papers reflect work in progress. They are intended to make lessons emerging from the current work program available to operational staff quickly and easily, as well as to stimulate discussion and comment. They also serve as the building blocks for subsequent policy and best practice papers.  Preface This paper is one of a series undertaken as part of the Extended Poverty Study in Ghana. A synthesis of this work is available in Ghana: Poverty Past, Present and Future (Report No. 140504-GH, Population and Human Resources Division, West Central African Department, World Bank, Washington D.C., June 29, 1995). Four background papers are appearing in the Poverty and Social Policy Discussion Paper Series: Harold Coulombe and Andrew McKay, 'An Assessment of Trends in Poverty in Ghana, 1988-92.' Poverty and Social Policy Discussion Paper No. 81, World Bank, Washington D.C. (November 1995) Lionel Demery, Shiyan Chao, Ren6 Bemier and Kalpana Mehra, 'The Incidence of Social Spending in Ghana.' Poverty and Social Policy Discussion Paper No. 82, World Bank, Washington D.C. (November 1995) Andy Norton, David Korboe, Ellen Bortei-Dorku and D.K. Tony Dogbe, 'Poverty Assessment in Ghana using Qualitative and Participatory Research Methods.' Poverty and Social Policy Discussion Paper No. 83, World Bank, Washington D.C. (November 1995) Christine Jones, and Ye Xiao, 'Accounting for the Reduction in Rural Poverty in Ghana, 1988-1992.' Poverty and Social Policy Discussion Paper No. 84, World Bank, Washington D.C. (November 1995) The papers draw upon the work of many. The contributions of the Ghana Statistical Service, Ministry of Health, and Ministry of Education in the Government of Ghana, and of UNICEF and the Canadian International Development Agency (CIDA), are gratefully acknowledged. The authors of this paper would also like to express their particular thanks to Lionel Demery, Christine Jones and Martin Ravallion for very helpful comments and suggestions on the issues covered by this paper. The views expressed are those of the authors. They should not be attributed to the World Bank, its Board of Directors, its mangement or any of its member countries 1  I  Abstract The Ghana Poverty Profile study was based on the three rounds of the Ghana Living Standards Survey conducted in 1987/88 (GLSS 1), 1988/89 (GLSS 2) and 1991/1992 (GLSS 3). This paper considers the validity of using the three surveys to assess trends in poverty in Ghana over the period for which they were conducted. Taking the survey data at face value suggests that poverty in Ghana increased between 1988 and 1989, but showed a more offsetting decline between 1989 and 1992. Thus if the poverty line is set at two-thirds of the mean value of the standard of living measure in the first round, the estimates suggest an evolution in the incidence of poverty from 36.8 percent of the population in 1987/88 to 41.8 percent in 1988/89 and to 27.9 percent in 1991/1992; this would suggest a significant decline in poverty over the period. There are, however, significant issues of comparability between the three rounds of the GLSS. Most significantly, the questionnaire used for the third round was different from that used for the first two rounds; in particular, it included a much more detailed expenditure section, in which households were visited repeatedly at two - three day intervals over a month, and in which there was a significant increase in the number of items. There is reason therefore to think that the estimates of expenditure in the third round may be higher as a result of the shorter recall periods and more frequent visits; given that expenditure is the basis of the standard of living measure, this must be corrected for if possible in order to make over time comparisons valid. This was done by using experimental results for Ghana obtained by Chris Scott and Ben Amenuvegbe on the rate at which people are likely to forget items of expenditure as the recall period is lengthened. This is used to adjust the expenditure data in an attempt to make the three rounds as comparable as possible. This correction obviously reduces the estimate of the decline in poverty over the period, but, using the same poverty line as above, poverty is still observed to fall from 36.9 percent of the population in 1987/88 to 31.4 percent in 1991/92. This more moderate estimate of the decline is probably more credible. The paper also examines the sensitivity of the observed trends in poverty at the national and regional level to changes in the definition of the standard of living measure. It is argued that measures based on household income and household expenditure on food are clearly inferior; changes in the way in which the cost of living index and the size of the household are calculated do not change the observed trends of poverty significantly. The conclusion is the pattem of change in poverty indicated by the total expenditure measure is broadly robust; poverty does indeed seem to fall over the period considered here. What we can not say, however, based on the GLSS data, is whether this trend continues beyond the period in which the GLSS surveys were conducted. 2  I  1. INTRODUCTION In a continent in which economic performance has been very mixed over the 1980s and early 1990s, Ghana stands out as an example of a country which has achieved relatively impressive economic growth. Since the Economic Recovery Programme was initiated in 1983, GDP has grown by an annual average rate of 5.0 percent over the period 1984 - 1992, and by 4.6 percent over the period 1987 - 1992 considered in this study (Demery and Squire, 1995). These rates of growth exceeded the population growth rates over the corresponding period, so that per capita GDP increased. A large part of the growth in per capita GDP reflected the larger growth in per capita private consumption, which increased by an average amount of 3 percent per annum over the period 1987-92. The national accounts statistics therefore suggest that average living standards in Ghana have been increasing over the period since 1984, including over the period 1987 - 1992 on which this study is focused. Increases in average per capita private consumption do not mean that everyone benefits, and it is theoretically possible for poverty to increase at the same time as average living standards increases, if these increases accrue only to those at the upper end of the income distribution. However, while a theoretical possibility, such an outcome is unlikely. Even still, we can pose the legitimate question of whether the poor have benefited to the same extent as the non-poor from economic growth, or whether growth has been accompanied by increasing inequality so that such that poverty reduction has been more modest than it would be under distribution-neutral growth. The study of poverty requires the suitable household survey data which can be used to characterize the living conditions of households; to monitor trends in living standards and poverty over time requires the availability of at least comparable surveys. In the case of Ghana, we are fortunate in having available three rounds of results from the Ghana Living Standards Survey (GLSS), a nationwide, multi-purpose household survey. The available data relate to the periods from September 1987 - August 1988 (GLSS 1), October 1988 - September 1989 (GLSS 2) and September 1991 - September 1992 (GLSS 3). To the extent that these surveys are comparable, which is one of the major issues on which we focus in this paper, these surveys may be used for the evaluation and monitoring of trends in poverty over time. The main issue covered in this paper is the pattern of changes in poverty in Ghana over the period covered by the GLSS surveys, and the extent to which the results indicated by the survey data are robust. There are different dimensions to this issue of robustness. One is the issue of the extent to which the surveys are legitimately comparable, the extent to which adjustments or corrections can be made for any problems of comparability, and the effect of such adjustments on the results. A second issue is the question of robustness to the definition of poverty. Many different approaches can be taken to the measurement of poverty, in particular many different choices can be justified with regard to the measure of the standard of living and the poverty line chosen. It is 3  important to know to what extent indicated trends in poverty over time are sensitive to the precise definitions used. This paper is organized as follows. We begin in section 2 by summarizing some of the issues arising in the measurement of poverty which have a direct bearing on the sensitivity analysis conducted in this paper. Section 3 then summarizes the issues which arise in measuring changes in poverty over time, discussing both general issues and issues specific to the data sets used in this study. Section 4 then presents the trends in poverty over time indicated by the survey data, and considers the sensitivity of these to the precise definition of the standard of living, including making attempts to correct for problems of comparability over time. Section 5 uses dominance analysis to assess the robustness of the results to the precise choice of poverty line. The concluding section 6 summarizes our best present understanding of changes in poverty in Ghana over the period 1987 - 1992. 2. MEASUREMENT OF POVERTY AT A POINT IN TIME An extensive literature has developed in recent years on the measurement of household welfare and poverty based on the types of data typically available from household surveys. The main issues are usefully summarized by Lipton and Ravallion (1993) in their recent survey article, and we will only repeat this discussion here to the extent that it relates to the issues considered in this paper. The measurement of poverty based on household survey data requires that we have available: (i) a suitable measure of the standard of living; (ii) a suitable poverty line or lines; and (iii) indices of poverty which conveniently summarize its main dimensions. In terms of the measure of the standard of living, money metric measures are generally used for this type of analysis, while at the same type recognizing that some aspects of household living standards might not be satisfactorily captured by these measures. Most commonly, the money metric measure used is based on total household consumption expenditure; while this measure does not take account of intra-household inequality, an issue which may be important in practice (Alderman et al, 1995), this is because household surveys commonly do not collect the necessary information to allow such issues to be taken into account. Expenditure-based measures are generally preferred to the alternative of using measures based on total household income, based both on theoretical arguments (that expenditure is a better approximation to permanent income of living standards than is current income) and practical arguments (that expenditure is typically measured more accurately than income, which is often significantly and systematically underestimated-a view for which there is plenty of empirical evidence). Notwithstanding these objections, it is still possible to consider income-based measures of the standard of living, which, according to the permanent income theory of consumption may be effective at identifying transitory poverty. 4  An alternative approach which has been adopted in some instances (Anand and Harris, 1991; Kyereme and Thorbecke, 1987) is to focus only on a subset of expenditure, that on food. 'Food poverty' may be of interest in its own right, but in food expenditure is to be used in place of total expenditure as a measure of a household's standard of living, the justification must be a practical one: that food, being a frequent and relatively regular item of consumption, is both more completely and more accurately measured than non- food expenditure, and that, perhaps appealing to Engel's Law, that non-food expenditure anyway increases as food expenditure increases. Here it is an issue of judging the quality of the estimates of non-food expenditure which are available; where the quality is poor, or indeed where non-food expenditure estimates are not available, measures of the standard of living based on food expenditure might be justified. We thus have available three variables which may be used as the basic of a money metric measure of welfare: total household expenditure, total household income, household expenditure on food, among which, a priori, total expenditure is likely to be preferred. Whichever is used in any particular case, the definition must be sufficiently broad to include appropriate imputations corresponding to relevant non-market transactions, such as consumption of own output, payment in kind and the imputed rental value of owner-occupied dwellings (Johnson, McKay and Round, 1990). Whichever of these three money metric measures is chosen, the measure of the standard of living must also take account in differences between the needs of households, reflecting differences in the prices faced by households as well as differences in the size and composition of households. Within a household survey, prices may differ across households according to the region of residence (spatial variation) or according to the precise time at which they are surveyed (temporal variation). The latter is important in situations of high inflation or seasonal variation in prices, given that surveys often take place over a number of months (often a full year to attempt to limit the effects of seasonality). Variations in prices can be allowed for by deflating the money metric measure by a cost of living index, which may be calculated as a Laspeyres or Paasche index; the Paasche form is convenient as deflating a money metric measure with a Paasche price index gives the easily interpreted Laspeyres quantity index. In the cases that total household expenditure or total household income are used as the basis of the money metric measure of welfare, the Paasche cost of living index for region r (r =1, ..., R) and time period t (t = 1, ..., T) is defined as follows: 1 W Pil i=1 Pirt 5  where C,t is the value of the cost of living index for region r and time period t, pi,, represents the price of commodity i (i = 1, ..., n) in region i and time period t, pill represents the price of commodity i in the base region (r = 1) and time period (t = 1), and win represents the share of actual and imputed expenditure on commodity i in total expenditure (wit = 1). Where the measure of the standard of living is based on food expenditure only, the same formula applies, but using food items only and with the shares (wir) relating to total food expenditure rather than total overall expenditure. Indeed, it is the difficulty in constructing a satisfactory non-food cost of living index, reflecting serious problems of quality variation in measuring the price of non-food commodities, which constitutes the strongest practical argument for using food expenditure rather than total expenditure as the measure of welfare. A further issue which arises in the calculation of the cost of living index is the computation of the weights wid in the specific case of poverty measurement, which is the question of for whom these shares should be computed. These will normally be computed based on the household survey. If we are simply interested in the estimation of a cost of living index we would compute the weights over the whole sample, which gives the largest sample size and in principle the most accurate weights. But when we are concerned with the measurement of poverty, we may wish to take into account the fact that the average consumption basket for the poor may be quite different from that of the population as a whole, such that it might be more appropriate to use weights computed from the lower end of the distribution of the money metric measure, say the bottom 50 percent. There is a significant element of circularity in this argument, but it may be worth taking into account if consumption patterns differ significantly between poor and rich (which is an empirical question). The remaining issue is that of measuring household size. As pointed out by Lipton and Ravallion (1993), the most common practice in poverty analysis is simply to divide by the number of household members in the household. However, as this fails to allow for differences in composition (which are likely to reflect differences in consumption needs) and for any household level economies of scale, it will tend to exaggerate the incidence of poverty in large households (Lanjouw and Ravallion, 1994). One way to avoid this problem is to measure household size in numbers of equivalent adults, computing using an appropriate (ideally country-specific) adult equivalence scale. Examples of commonly used adult equivalence scales used in practice include those developed by Deaton and Muellbauer (1986), which uses a relatively coarse classification and does not allow for household level economies of scale, and that developed by McClements, which uses a more detailed disaggregation as well as allowing for household level economies of scale. 6  There is also an extensive literature on the appropriate definition of a poverty line, which for present purposes we will assume to be absolute. There are many ways in which a poverty line may be set, and there is invariably a significant element of subjectivity. Perhaps the best relatively straightforward procedure is that set out by Ravallion and Bidani (1994), which is based on a combination of food energy requirements and appropriate assumptions about the share of non-food in the expenditure of poor households. In this section we have set out a number of alternative choices which will affect the results of any measurement of poverty. Many of these relate to the measurement of the standard of living. In this regard, while total household expenditure is perhaps the most suitable basis for a measure of the standard of living, reasonable justifications can also be made for instead using food expenditure or household income. Having determined the basic money metric measure, the issue of the construction of the weights for use in the cost of living index arises: should these reflect the consumption patterns of the whole sample, or simply those of a group in the lower end of the distribution? The next issue is the measure of household size: should this simply reflect number of persons, or number of adult equivalents, and if the latter, which adult equivalence scale should be used? Finally, having settled on a measure of the standard of living, the question of the appropriate level to set the poverty line arises; the precise line used will always be open to question, at least within a range. In the specific case of Ghana considered in this paper we have therefore used data supplied by the Ghana Living Standards Survey to estimate the constituent variables of the measure of the standard of living. Three basic money metric measures have been calculated using the GLSS household survey data: total household expenditure, total household income, total food expenditure, defined in each case to include relevant imputations. In the case of the cost of living indices, we have had to deviate slightly from the formula set out above. We judged that only in GLSS 3 where the available price data sufficient in quality and coverage to estimate the spatial cost of living variations. We therefore used the price data collected under GLSS 3 along with commodity specific expenditure weights from the same source to estimate the spatial variation in prices for food and non-food commodities separately at the level of five localities; centering this on the median survey month in each locality, this index was then extrapolated backwards and forwards using the consumer price index, using separate aggregate indices for food and non-food in each of urban and rural areas. Three separate cost of living indices were estimated on this basis: an overall cost of living indices using weights computed from the full sample; an overall cost of living index using weights from the bottom 50 percent of the sample in the locality in question; and a food cost of living index, using weights from the a full sample in each locality. Finally, we constructed three measures of household size: household size in number of persons; household size in number of equivalent adults calculated using the Deaton and Muellbauer scale; and household size in equivalent adults calculated using the McClements scale. 7  This gives a number of different combinations for the measurement of welfare. The poverty line in each case must be defined in the same units in which the standard of living in measured. The procedure we have adopted in this paper for determining poverty lines is a considerable short cut: for any welfare measure, we take two thirds of its mean value, defined over individuals, in the base year as the basic poverty line, and one half of the mean value as representing an ultra-poverty line. While this sounds arbitrary, which to some extent it is, in the case of expenditure-based measures of welfare the implied poverty line in fact corresponds quite closely to the World Bank poverty line of $375 per person per year in 1985 prices, one of the two lines proposed in the World Bank World Development Report, 1990. All of these give many different ways of measuring poverty in any particular situation. In practical terms, our concern is the pattern of poverty obtained from any particular choice is in fact highly sensitive to the choices made. In other words, how robust is the pattern of poverty? In some respects this has been considered, for the case of a poverty profile at a point in time (Ravallion and Bidani, 1994). Our interest here is in poverty comparisons over time, which raises a number of additional issues, to which we now turn. 3. ISSUES IN COMPARING POVERTY MEASUREMENTS OVER TIME There are a number of difficulties in comparing data on poverty over time, even where one can be assured that the data sets are consistent over time. Firstly, the magnitude of changes in poverty between years which are adjacent or close to each other can be expected to be small, such that it may be significantly affected by the presence of sampling and measurement error. Secondly, especially in a small developing economy, one-off shocks can have a significant affect on living standards from one year to the next; an example of this would be a harvest failure. Comparison of the two years will therefore not necessarily give a good indication of longer term trends. Third, correcting for inflation over time may cause problems; while in principle the consumer price index can be used, this may be based on collecting data in centers which are not typical of the entire sample, if, for example, data are collected mainly in urban areas. These difficulties are compounded in a situation in which the data collection procedures differed from one year to the next, as is the case in Ghana. The questionnaires for the first two rounds were identical, though there is some evidence of discrepancies between the results for the two years in variables which might be expected to remain relatively stable, raising concerns about comparability. Potentially more significant, however, is the fact that the questionnaire was significantly revised between the second and third rounds, in such a way that the collection of expenditure data was much more detailed. This in and of itself may raise estimated expenditure levels in the third round relative to the earlier rounds. Finally, the fact of households not being uniformly surveyed through the year can raise issues of comparability. 8  Changes in questionnaire structure in the GLSS As regards over time comparisons, the accuracy of the results will depend on the comparability of the expenditure estimates used in the welfare measure and on the accuracy of the estimate of inflation. Considering the former first, the main issue is consistency of questionnaire design. The first and second rounds of the Ghana Living Standards Survey used exactly the same questionnaire, whereas the questionnaire was modified for the third round, in particular with the addition of a much more detailed module on household expenditures. Important changes in the construction of the expenditure section included the following: (i) a large increase in the number of categories, in relation to both food and non-food categories; and (ii) the introduction of a much shorter recall period for food expenses and subsistence consumption, and multiple visits by enumerators. In urban areas a three day recall period was used, and eight visits were made to collect information of food expenditures; in rural areas a two day recall period was used and eleven visits were made. Both of these changes can be expected to improve significantly the accuracy of the expenditure data relative to the GLSS 1/2 questionnaire, in which a smaller number of items were identified and a 14-day recall period was used for food and other frequent purchases. This in itself might not be a problem if we could be confident that no bias would be introduced. But this is very unlikely to be the case. Much of the error in estimates of expenditure is recall error and this can be expected to increase with the length of the recall period. Scott and Amenuvegbe (1990) provide empirical evidence of this phenomenon in a study of Ghana. Recall error can also be expected to be reduced when a larger number of categories is used, although there is less empirical evidence for this. Thus it is likely that expenditure on food in GLSS 1/2 will be more underestimated due to recall error than will be the case for GLSS 3. This fact is likely to mean that the change in welfare between GLSS 1/2 and GLSS 3 is likely to be overestimated when the welfare measures are based on food expenditure or total expenditure. This problem does not affect estimates of non-food expenditure to the same extent, given that the changes in the data collection procedures were less significant in this area. But unfortunately there is no reason to think that non-food expenditure is a good measure of welfare. This suggests the desirability of also examining measures which are likely to be less affected by changes in the questionnaire structure between GLSS 1/2 and GLSS 3, and/or to attempt to make an approximate correction for the likely effect of the changes in the questionnaire on recall error, and the effect of non-uniform sampling within localities over time in GLSS 3. 9  The comparability of GLSS 1 and GLSS 2 The fact that the same questionnaires were used for GLSS 1 and GLSS 2 should imply that comparisons can be made between these two years with more confidence. But unfortunately this is not necessarily the case. There is evidence of significant changes between these two years which are hard to explain as a genuine phenomenon. Specifically, significant changes are observed in household size between the two years and in the composition of income. Detailed analysis has not been able to provide a satisfactory explanation of these phenomena, but they appear to reflect changes in data collection procedures rather than genuine changes. This raises questions about the comparability of these two surveys in other areas, specifically in the present context, expenditure. In fact household size and the composition of income are much more similar comparing GLSS 2 and GLSS 3, than comparing GLSS 1 and GLSS 2. This might be interpreted as suggesting that the GLSS 2 data were more satisfactorily collected than the GLSS 1 data, though this hypothesis must remain tentative at present. Non-uniform sampling and seasonality Another issue which potentially arises in assessing trends in poverty at the locality level is a consequence of non-uniform sampling combined with seasonality. As in each round of the GLSS survey activities take place over a twelve month period, seasonality would not cause a problem in comparing average values (or other summary statistics) of a variable over time, if households in each locality were uniformly surveyed throughout the twelve months. But this is not in fact the case in the GLSS samples as demonstrated by Table 1-in any locality the number of households surveyed in a month varies significantly over the year. This means that over time comparisons of average values and other summary statistics of any variable subject to seasonal variation will be sensitive to in which particular months survey activity was concentrated (unless the seasonal pattern was the same in all three years, which is not the case here). The problem identified here is a problem associated with seasonal variation in quantities, not prices, which are already taken into account in the cost of living index. As such, it is likely to be a more acute problem in rural areas than in urban areas. Within rural areas a number of different components of consumption expenditure (or whichever money metric variable is used as the basis of the standard of living measure) might be affected, but the expenditure component likely to be'most affected will be consumption of own production, which can be expected to be associated with the seasonal variation in production itself This raises another issue of comparability between the three years survey results though, as we demonstrate later, it can in some way be corrected for. 10  The estimate of inflation The reliability of the denominator in the calculation of the welfare measure is as important as the accuracy of the numerator, so that the question of the accuracy of the estimates of changes in prices over time arises. The estimates of the cost of living index over time were obtained by forward and backward projection of the regional cost of living index previously defined using the consumer price index. Separate indices were used for food and non-food products and for urban and rural areas, although lack of information meant that uniform inflation had to be assumed within urban and within rural areas. Unfortunately, however, there is no alternative source of information on inflation with which to verify the resulting index. These changes in questionnaire structure and data collection procedures raise questions about the comparisons of the results of the different surveys, and in particular highlight the importance of attempting to make a correction for recall error in making comparisons between the results of the GLSS 3 survey and the earlier years and for the effects of non-uniform sampling in each of the three years. These issues will be considered in the next section, when we examine the pattern of change in welfare and poverty over time according to different measures of the standard of living. 4. POVERTY CHANGE OVER TIME: THE EFFECT OF THE MEASURE OF THE STANDARD OF LIVING In the forgoing discussion, we have identified a number of issues which arise in defining a suitable measure of the standard of living for households based on the results of the GLSS surveys. We have identified a number of important issues, as follows: (i) The basic money metric variable to use: while total consumption expenditure is the most obvious choice, arguments can also be made for using total household income of total food expenditure; (ii) Any adjustments to make: as well as the unadjusted data obtained from the GLSS survey results, consideration must also be given to the possibility of correcting for the differential recall periods (and so recall errors) and for the effect of non-uniform sampling over the year; (iii) The cost of living index to use: this should be appropriate to the welfare measure being used (e.g. where seasonality and/or recall error corrections are made to expenditure data, the corresponding adjustments should be made in computing the cost of living index), the only issue which then arises is whether the weights used should be computed from the whole sample or from a subsample at the bottom of the distribution (e.g. the bottom 50 percent); 11  (iv) The measure of household size: besides the obvious measure in terms of number of persons, this may also be measured in terms of the number equivalents, using for example the Deaton and Muellbauer or the McClements adult equivalence scales. These different choices give up to 72 (3 x 4 x 2 x 3) different combinations depending on which choice is made at each stage. It is clearly infeasible, as well as being uninteresting, to consider all the possible combinations in this paper. Instead, we work through the four issues sequentially, deciding the most appropriate choice at each stage and then adopting this choice in all subsequent analysis. In this way we can reduce the number of tables considered to eight. In all cases we use the poverty line set at two third of the mean value across individuals of the welfare measure in question. The sensitivity of the results obtained to the choice of poverty line is an issue to which we return in the next section. The choice of money metric variable Conventional wisdom suggests that the most appropriate basis for measuring the standard of living is likely to be total household expenditure. For present purposes we simply take the estimates computed from the sample in each of the three years, making no adjustments for the problems of comparability between the different years previously referred to. We initially express this measure on a per capita basis, and deflate by a cost of living index based on weights computed from the full sample (and again without any adjustments applied in computing the weights). Using this measure of the standard of living and the poverty line previously referred to, Table 2 presents values of the Pa poverty indices for the country as a whole and for the standard five localities, for different values of a(a=O, 1,and 2). Focusing on the geographic pattern of poverty, in each of the three years and for each value of a considered, Accra represents the least poor locality, while the Rural Savannah is by some way the poorest locality. Poverty is more prevalent in the Other Urban locality than in Accra, but in general, levels of poverty are lower in urban areas than in any of the three rural localities. Poverty appears therefore to be disproportionately rural, snd within rural areas to be disproportionately prevalent in the northern Savannah area. There is, however, also significant urban poverty, especially outside of Accra, and a slightly higher proportion of national poverty is accounted for by urban areas in 1991/92 than previously. Indeed, in 1991/92, poverty in the other urban locality is slightly higher than in the rural coastal zone, the least poor of the rural areas. Turning to the over time changes implied in this table, the results show that in aggregate the incidence of poverty increased from 36.8 percent in 1987/88, to 41.8 percent in 1988/89, before declining to 27.9 percent in 1991/92. Thus the decrease between the second and third years more than compensates for the increase between the first two years. The depth of poverty also declines slightly between 1988/89 and 1991/92. 12  Some differences to the aggregate pattern emerge at the locality level. In broad terms, in Accra poverty is observed to increase between the first two years, and only to decline very slightly between 1988/89 and 1991/92. In all other localities, the patterns of change in the incidence of poverty are the same as those at the aggregate level (an increase followed by a more than offsetting decrease), though the magnitudes of changes vary from one locality to another. In rural areas, the fall in the incidence of poverty is most marked in the rural coastal area, though is also quite large in the other rural zones. The fall in the incidence of poverty in the urban centers outside Accra is less than the fall in the rural coastal area, so that these two localities converge. Note that the data have not been corrected for the effects of changes in the questionnaire design (nor indeed for the effects of non-uniform sampling), so that the magnitude of changes in poverty between 1988/89 and 1991/92 is liable to be exaggerated significantly by this table. The issue of how to correct for this exaggeration will be addressed shortly. First, however, we consider the effects of using measures of the standard of living based on food expenditure and on total household income, both computed to include relevant imputations and both deflated by household size measured in number of persons and by relevant cost of living indices (the same as for expenditure in the case of the income-based measure, and a subset relating to food only in the case of the food expenditure measure). The values of poverty indices computed using these variables, and a poverty line defined as two third of the mean value of the variable in question in 1987/88, are presented in Tables 3 and 4 respectively, following the same format as before. Using food expenditure as the welfare measure, the differences in poverty between the localities are much smaller than with the welfare measures based on total expenditure, suggesting that much of the difference occurs in non-food expenditure (though this would need to be considered further using other poverty lines). Over time, from 1987/88 to 1988/89, poverty as measured using this measure of welfare increases somewhat. However, over the whole period, there is a decline in poverty overall, so that the fall in poverty which occurs between 1988/89 and 1991/92 more than offsets the increase which occurred between 1987/88 and 1988/89. This pattern of change is not observed in all localities, in particular it is not observed in urban areas, and in Accra poverty increases between both pairs of years. In all of the rural areas, poverty falls significantly between 1988/89 and 1991/92, much more than offsetting the increases between 1987/88 and 1988/89. Indeed, using this measure of the standard of living, by 1991/92 rural areas are now significantly less poor in terms of food expenditure than urban areas! This raises significant questions about the use of food expenditure as a measure, although it might still provide valid information about food poverty. The pattern of geographic differences in poverty suggested by the income-based welfare measure (Table 4) is very unstable from one year to the next. The ranking of the localities changes sharply from one year to another. For example, the income measure suggests that Accra is the poorest area in 1987/88, the least poor in 1988/89 and the 13  second least poor in 1991/92. It suggests that the rural Savannah is among the least poor areas in 1987/88 and 1988/89. Over time, poverty is again observed to increase between 1987/88 and 1988/89, and to decline by a not quite offsetting amount between 1988/89 and 1991/92. The suggested pattern of change over time varies significantly from one locality to another, which in the absence of a good explanation does not encourage confidence in the results. We have serious reservations about the use of income as a measure of welfare. It is subject to significant measurement error. The estimate of total household income is significantly below the estimate of total expenditure, and there is every reason to believe (based on internal consistency and experience elsewhere) that the measurement error is predominantly in the former. Further, the underestimation of the income estimate is likely to be systematic, in that some components (self employment income and transfers) are likely to be more under-estimated than others. Finally, there were also changes in the structure of the income modules of the questionnaire between GLSS 2 and GLSS 3, though in this case it is much more difficult to predict the direction in which this will affect the estimates. In summary, there is no reason to believe that the changes in poverty implied by the income and food expenditure estimates are more reliable than those implied by the total expenditure measures; indeed the reverse is more likely. We therefore consider expenditure as the most appropriate basis for a money metric measure of welfare in all subsequent discussion. We now turn to the issue of attempting to correct for the problems of comparability between different years following from (i) differential recall error; (ii) non-uniform sampling combined with seasonality. Adjustments to the basic measure of expenditure We consider first the problem of the differential in recall periods between GLSS 1/2 and GLSS 3. Overall there is no completely satisfactory way of trying to allow for this. But some empirical evidence is available on the likely magnitudes of this effect, based on a study by Scott and Amenuvegbe (1990). They provide evidence, based on a study of Ghana, that in lengthening the recall period from one to seven days, each additional day added to the recall period leads to a reduction in the estimate of overall food expenditure of 2.9 percent, but that beyond seven days lengthening the recall period does not introduce significant further bias. We therefore attempt to correct for this possible bias in the estimate of food expenditure, based on switching from a 14 day bounded recall period in GLSS 1/2 to a two or three day bounded recall period in GLSS 3. For each household we modify its estimate of total food expenditure by adjusting the estimates in GLSS 1 and GLSS 2 by inflating it by 2.9 percent for each additional day of the recall beyond the period used in the GLSS 3 survey, up to the seventh day. Thus expenditure in GLSS 3 is unadjusted, while food expenditure estimates in GLSS 1/2 are inflated by 1.029 to the power of five in rural areas and 1.029 to the power of four in urban areas. This will in principle correct for the non-comparability problem. As non- 14  food expenditure is much less affected by this change in recall period, no adjustment is made to the estimates of non-food expenditure. Again defining the poverty line as two-thirds of the mean value of the distribution of living standards over individuals in 1987/88, Table 5 presents poverty indices using the adjusted total expenditure per capita estimates. This shows that, at the aggregate level, we continue to observe the increase in the incidence of poverty between 1987/88 and 1988/89, and the decline between 1988/89 and 1991/92, with again the decline more than compensating for the increase (the incidence of poverty now falling from 36.9 percent in 1988/89 to 31.4 percent in 1991/92). Thus, there is still a significant fall in poverty over the period, but the magnitude is now slightly more credible. The geographic pattern is similar to that displayed by Table 1; the largest decline in poverty between 1988/89 and 1991/92 is observed in the rural coastal area, which converges towards the other urban locality; significant falls are also observed in the forest and Savannah zones, thought the Savannah remains the poorest zone by some way in 1991/92. In Accra poverty in fact increases. Again there is evidence of a convergence between rural and urban areas, but with the rural Savannah still being left behind despite significant falls in the incidence and depth of poverty since 1988/89 and 1991/92. This table represents our best assessment so far of changes in household welfare and poverty over time in Ghana over the period covered by the study. We now consider additionally the effects of making an adjustment for the effects of non-uniform sampling combined with seasonality. The problem here results from the fact that, within any locality, the number of households surveyed in a month varies sharply from one month to another. In some months therefore there are an above average (for the locality) number of households surveyed whereas in others the number is significantly below average. The procedure adopted therefore is to apply a weight specific to each locality and each month computed as the ratio of the average number of households surveyed in a month in a locality divided by the actual number surveyed and apply this weight in calculating summary statistics such as mean values and poverty indices. Households surveyed in the month of below average survey activity for the locality receive a relatively high weight (reflecting the fact that there are not enough such households in the sample); the opposite applies to those surveyed in a month of above average survey activity. These weights are computed so as to average to one. Applying the weights to the measure of expenditure, and so using them in calculating the poverty indices, gives the results presented in Table 6. At the aggregate level the application of this adjustment factor has a very small effect; the incidence of poverty is slightly higher in 1987/88 than was obtained in Table 5, while aggregate poverty indices scarcely change at all in 1988/89 and 1991/92. At the locality level, the values of poverty indices in the rural Savannah is higher in this table in all three years than the corresponding figures in Table 5, but the overall trend is still the same, and the figures continue to indicate a sharp fall in poverty in this area between 1988/89 and 1991/92. Table 6 continues to indicate a dramatic fall in poverty in the rural coastal area between 1988/89 and 1991/92; it also indicates a modest decline in poverty between 1987/88 and 1988/89, in direct contradiction to the increase 15  suggested by Table 5. Clearly the seasonality adjustment is having an important impact in the rural coastal area. In the remaining regions, the trend in poverty over the three surveys is similar to that indicated by Table 5, although in the case of the rural forest the increase in poverty between 1987/88 and 1988/89 is much smaller once the seasonality adjustment is applied. The choice of the cost of living index Continuing to use the measure of expenditure corrected for recall error, but not for seasonality, we now consider the effect of using a cost of living index constructed using weights from the bottom 50 percent of the distribution only. Despite certain logical contradictions in using a procedure which have been referred to previously, it has the advantage that the basket of goods used in constructing the cost of living index more accurately reflects that of poor households, such that the resulting index more accurately reflects the prices such households face. Following the standard format, the results of using this index are presented in Table 8. The appropriate comparison here is with Table 5. Such a comparison indicates relatively little difference between the two tables. Levels and trends in poverty indices in urban areas hardly change at all. In rural areas, this index implies marginally higher levels of poverty in the rural Savannah than did the original index, but the trend remains the same, and the sharp fall is still observed between GLSS 2 and GLSS 3 (indeed, it is now marginally larger). Similarly, the trends in poverty indices change little in the other rural areas, although the magnitudes of the indices themselves change marginally. The definition of household size We now turn to consider alternative household size deflators, adopting the expenditure measure corrected for differential recall error but not for seasonality. All the measures of welfare considered so far have been per capita measures, in other words, divided by household size measured in number of persons. Such measures take no account of household composition, treating children as equivalent to adults in consumption terms and not allowing for household level economies of scale. An alternative approach is to measure household size in numbers of equivalent adults, using an appropriate adult equivalence scale. Although not always common practice, in some sense this is likely to be more realistic; the problem which always arises is the appropriate adult equivalence scale to use. In practice, the way in which household size is measured will have a significant impact on the demographic pattern of poverty; use of household size measured simply in numbers of persons is likely to exaggerate the number of large households which are poor, for example. Our concern here, however, is whether the measure of household size used affects our conclusions about trends in poverty over time. 16  Using total household expenditure corrected for recall error as described in the previous section, Tables 9 and 10 examine the sensitivity of the trends observed in Table 5 to alternative definitions of household size. Table 9 calculates household size using the adult equivalence scale estimated by Deaton and Muellbauer which has been used in previous studies of Ghana; this scale counts children as fractions of adults, though does not allow for household level economies of scale. In Table 10 household size is calculated using the McClements adult equivalence scale, a scale which is more sensitive to the precise age of children and which does allow for household level economies of scale. The poverty line is again defined as two thirds of the mean value (over individuals) of the corresponding welfare measure in the first year. The results of both these tables show that while the choice of measure of household size does not significantly the general trend (outside Accra) of poverty indices increasing between GLSS 1 and GLSS 2 and then falling by a more than offsetting amount between GLSS 2 and GLSS 3, the choice of equivalence scale has an impact on the ranking of localities in GLSS 3. Using the Deaton and Muellbauer scale, the rural coastal area comes out as having the lowest incidence of poverty (or at any rate a level indistinguishable from those in Accra), whereas when either the McClements scale or, more strikingly, whenever household size is measured in number of persons, the rural coastal area is ranked third in terms of poverty. The use of an adult equivalence sale significantly reduces the differential between the rural coastal zone and urban areas, presumably because average household size is larger here than in urban areas, and the use of a per capita measure is exaggerating poverty among large households. In summary the trend in poverty over time, for the country as a whole and for individual localities does not fundamentally change when different measures of household size are used, although the magnitudes of upward and downward movements do change, these effects being of sufficient magnitude to throw into question the relative ranking of some localities in GLSS 3. The rural Savannah remains always the locality in which poverty indices are highest, the relative ranking of the other region in GLSS 3 (though not so much in earlier rounds) does depend on the measure of household size adopted, though as the differences in poverty indices is relatively small in GLSS 3, perhaps this is not a major source of concern. There is no clear criterion for choosing a 'best' measure of household size from among those considered here. While conceptually adult equivalent scales are more appropriate than measuring household size in numbers of persons, neither of the adult equivalence scales considered here are specific, or even necessarily particularly appropriate, to Ghana. In the absence (at present) of a Ghana-specific adult equivalence scale, we follow standard practice of using per capita expenditure measures of living standards. 17  S. POVERTY CHANGE OVER TIME: TBE IMPORTANCE OF THE POVERTY LINE Having identified our preferred measure of household welfare, we now consider the possibility that the observed pattern of changes in poverty and in household welfare may be sensitive to the precise level at which the poverty line is drawn or to the particular poverty index used. This will be considered in terms of first order dominance analysis, which is based on cumulative distribution curves of welfare. The analysis is based on cumulative distribution curves of welfare. These are presented in Figure 1 for the country as a whole, in Figures 2 to 6 for the five standard geographic localities, and in Figures 7 to 13 for the seven different socio-economic groups, defined as before. In each of these figures the horizontal axis (XI) represents the cumulative proportion of individuals, where these individuals have been ranked from the poorest to the richest, while the vertical axis represents the cumulative welfare of these individuals (Y1). The ranking by welfare level means that the curves display a positive and non-decreasing slope. One curve (Cl for example) is said to display first order dominance relative to a second curve (C2) if Cl lies above C2 throughout the whole distribution. Where this arises it can be said that the level of poverty associated with the Cl curve is unambiguously lower than that associated with the C2 curve for any (sensible) measure of poverty and irrespective of the poverty line. When this property arises, comparisons of poverty can be considered robust, in the sense that they are not sensitive to the definition of poverty or to the poverty line. Note, however, that problems of measurement error still remain. The effect of such measurement errors is likely to be particularly acute near the bottom of the distribution where the position of the curve is determined based on a relatively small number of observations. The detailed results are summarized in the Appendix. It is often hard in the figures to see whether or not the curves cross towards the bottom of the distribution. This was examined by 'blowing up' the bottom 30 percent of the distribution to look for crossing. These latter figures are not presented here, but the conclusions which resulted from their examination are summarized in the Appendix. The curves constructed at the national level do display first order dominance, allowing us to conclude that (subject to any possible effects of measurement error, as always), poverty unambiguously rose between GLSS 1 and GLSS 2, and then unambiguously fell between GLSS 2 and GLSS 3. Poverty in GLSS 3 was unambiguously lower than that in GLSS 1 as well. This provides strong support for the previous results, which indicated an increase in poverty between GLSS 1 and GLSS 2 which was more than offset between GLSS 2 and GLSS 3; the fact that first order dominance is observed between each of the pairs of curves implies that this result is not sensitive either to the poverty line chosen or to the poverty index used. 18  Once the data are disaggregated by geographic locality, some of these unambiguous results no longer hold. The only locality which mirrors the national pattern is the rural coastal area, where an unambiguous rise in poverty between GLSS 1 and GLSS 2 was more than offset by an unambiguous fall between GLSS 2 and GLSS 3. In the rural Savannah, poverty in GLSS 3 was unambiguously lower than it was in both GLSS 1 and GLSS 2; however, what happened between these latter two years is not unambiguously revealed by this comparison. In the Rural Forest, poverty unambiguously rose between GLSS 1 and GLSS 2, and unambiguously fell between GLSS 2 and GLSS 3, but strictly one cannot unambiguously conclude that the fall necessarily offset the rise (although in practice the point of crossing occurs at the second percentile, above which point the GLSS 3 curve lies above the GLSS 2 curve). In urban areas, the pattern is somewhat different. In Accra, poverty in GLSS 2 and in GLSS 3 is unambiguously higher than in GLSS 1, but comparisons about the change between GLSS 2 and GLSS 3 will be sensitive to the choice of poverty line so that a robust conclusion cannot be drawn. In urban areas outside Accra, no robust conclusion can be drawn, as all curves cross each other. Thus the unambiguous patterns of change in poverty at the national level are not necessarily also present at the locality level. In summary, the evidence for poverty having fallen over the period from GLSS 1 to GLSS 2 is strongest in rural areas, especially in the coastal and Savannah zones. In urban areas, no such unambiguous evidence exists, and in Accra there appears to be evidence of the reverse, an increase in poverty (though whether this is genuine or reflects a data problem in Accra in GLSS I is not clear. 6. CHOICE OF PREFERRED WELFARE MEASURE On the basis of the evidence so far considered, in particular the results in this paper, our preferred welfare measure is that which was presented in Table 5 of this paper: total household expenditure, per capita, in constant prices, adjusted for the effects of differential recall error between GLSS 1/2 and GLSS 3. The significant changes in questionnaire structure which occurred between the GLSS 1/2 surveys and the GLSS 3 survey raise serious issues of comparability over time no matter which data are used as the basis of the welfare measure. This is very unfortunate as it is the trend between GLSS 1/2 and GLSS 3 that is of particular interest in assessing what has happened to changes in poverty in Ghana over the period of the late 1980s and early 1990s. Of course, it is necessary to work with the information which is available, and this necessitates that some attempt is made to correct for the changes in questionnaire structure. Such corrections can only be crude; what we have done in this study is to use previous research on recall error affecting expenditure data in Ghana to make a correction for the likely impact of this effect. Even with this correction, the data continue to suggest that a fall in poverty in Ghana over the period 1987/88 to 1991/92, and particularly over the period 1988/89 to 1991/92. 19  While it is certainly possible to raise legitimate issues about the magnitude of the corrections applied, and also about the accuracy of the estimates of inflation used, the data presented in Table 5 above represent our best present assessment about what has happened to poverty in Ghana over the period 1987/88 to 1991/92. [insert section on page 18] It appears therefore, especially if we make a comparison between GLSS 2 and GLSS 3, that there has been a significant fall in poverty over this period in Ghana for all values of the parameter a, a decline which is particularly marked in rural areas. Poverty appears to have fallen significantly in all rural areas, including in the Savannah, but the Savannah region remains a long way behind the remaining rural areas. In Accra, by contrast, poverty in fact increases marginally over the period. Overall there is evidence of convergence between urban and rural areas, though the differential between the least poor urban area (Accra) and the poorest rural area (the Savannah) remains large. 20  References Alderman, Harold, Pierre-Andre Chiappori, Lawrence Haddad, John Hoddinott and Ravi Kanbur (1995), 'Unitary versus Collective Models of the Household: Is it Time to Shift the Burden of Proof?', World Bank Research Observer, Vol. 10, No. 1, pp. 1-20. Anand, Sudhir and Christopher Harris (1991) 'Food and Standard of Living: an Analysis of Sri Lankan Data' in Jean Dreze and Amartya Sen (eds.) The Political Economy of Hunger, Volume 1, Entitlement and Well-Being, Oxford: Oxford University Press. Demery, Lionel and Lyn Squire (1995), 'Macroeconomic Adjustment and Poverty in Africa: An Emerging Picture', World Bank Research Observer (Forthcoming) Deaton, Angus and John Muellbauer (1986), 'On Measuring Child Costs: With Applications to Poor Countries', Journal of Political Economy, Vol. 94, pp. 720- 744. Johnson, Martin, Andrew McKay and Jeffery I. Round (1990), Income and Expenditure in a System of Household Accounts, Social Dimensions of Adjustment Working Paper No. 10, World Bank, Washington D.C. Kyereme, S.S. and Erik Thorbecke (1987), 'Food Poverty Profile and Decomposition Applied to Ghana' World Development, Vol. 15, No. 9 pp. 1189-1201. Lanjouw, Peter and Martin Ravallion (1994) 'Poverty and Household Size.' Policy Research Working Paper No. 1332 (August) World Bank, Washington, D.C. Lipton, Michael and Martin Ravallion (1993), Poverty and Policy, Policy Research Working Paper No. WPS 1130, World Bank, Washington D.C. Ravallion, Martin and Benu Bidani (1994), 'How Robust is a Poverty Profile?', World Bank Economic Review, Vol. 8, No. 1, pp.75-102. Scott, Chris and Ben Amenuvegbe (1990), Effect of Recall Duration on Reporting of Household Expenditures: An Experimental Study in Ghana, Social Dimensions of Adjustment Working Paper No. 6, World Bank, Washington D.C. 21  Table 1: Distribution of the GLSS samples by month of interview and locality GLSS 1 (1987/88) +-------+ I Accl OU I RC I RF I RS I ALL MONTH October 1987 . 31 48 111 80 270 November I 161 .1 161 1421 471 221 ---------------+-____+___-_+____-+___-_+___-_+___-_ December 1987 | 471 941 481 1121 151 316 ---------------_+-____+___-_+____-+___-_+___-_+___-_ January 1988 | .| 321 161 321 31! 111 ---------------_+-____+____-+____-+___-_+___-_+___-_ February | 32! 64! 961 961 321 320 ---------------_+-____+____-+____-+____-+____-+___- _ March | 641 142! 321 80! 321 350 ----------------+-____+__-__+___-_+___-_+___-_+___- _ April | 641 961 631 801 48! 351 ---------------_+____ +_____+_----+-----+-----+----- May I .| 96j 481 891 471 280 June | 321 941 481 64! 641 302 July X 481 751 481 321 961 299 August 1988 | 321 1281 161 801 961 352 ----------------+-----+-____+___-_+___-_+___-_+___-_ ALL | 3351 8521 4791 9181 588! 3172 GLSS 2 (1988/89) Accl OU I RC I RF I RS | ALL MONTH October 1988 48 111 48 111 80 398 November I .1 32! 961 239! 47! 414 December 1988 | .1 32! 32! 46! 48! 158 January 1989 | 48! 112! 48! 79! 48! 335 February | 32! 80! 64! 96! 48! 320 March I .1 80! 95! 127! 31! 333 April | 161 951 16! 96! 73! 296 May | 32! 111! 64! 96! 48! 351 June | 127! 48! .| .1 16! 191 July I 321 961 .1 1121 1121 352 August 1989 | 48! 63! 63! 64! 48! 286 ALL | 383! 860! 526! 1066! 599! 3434 22  Table 1 (continued) GLSS 3 (1991/92) +-_____--__--___--_--___--__----__--______-+ I Acc| OU I RC I RF I RS I ALL ---------------+-____+___-_+___-_+___-_+___-_+___-_ MONTH September 1991 14 20 19 17 20 90 October | 311 741 601 841 691 318 November I .| 1551 301 1101 1091 404 ----------------+-----+-____+___-_+___-_+___-_+___-_ December 1991 .1 1441 201 1431 1001 407 ---------------+-____+___-_+___---_+____+-----__+-__ January 1992 1 .1 1341 201 1871 701 411 ----------------+-____+___-_+____-+___-_+___-_+___-_ February I .1 1441 401 1121 1201 416 ----------------+-----+-----+-----+-----+-----+----- March | 151 1291 391 1481 791 410 ----------------+-----+-----+-----+-----+-----+----- April I 451 901 80| 1201 811 416 May I 901 451 1301 110| 401 415 ----------------+-----+-----+-----+-----+-----+----- June | 881 851 110| 100| 301 413 ----------------+-----+-----+-----+-----+-----+----- July I 891 551 90l 1201 601 414 ------__---_____+____-+____-+___-_+___-_+___-_+___-_ August | 871 441 691 881 481 336 ----------------+-----+-----+-----+-----+-----+----- September 1992 | .| .| 101 321 311 73 ----------------+-____+___-_+___-_+___-_+-____+____- ALL | 4591 1119! 7171 13711 8571 4523 + _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _2 23  Table 2: Poverty indices by locality, measure of welfare: total expenditure, unadjusted, in constant prices, per capita 1987/88 +-- Sample |Average | I share |welfare P0 P1 P2 Co C1 C2 _-- ---------- +__-- __+____-___+______-_+-_____-_+______-_+_ __---_+_____-_- +___- ___ Accra | 8.31 295.91 0.0921 0.0181 0.0051 2.11 1.21 0.8 ---- ------------+-__ ___ -+_____-__+______-_+______-_+ _____-__+______-_+-_____-_+_ ____-__ Other Urban | 25.91 195.21 0.3381 0.1021 0.0451 23.81 22.41 22.3 ___ --------------+-__----- +_--- ___ -+____ -__+_ ____-_+-____-_+____---+ __-_--_+______-_ Rural Coastal | 14.11 178.01 0.3741 0.1271 0.0571 14.31 15.21 15.3 ------------- +__--- -_+_ ___ -+_______-+___-__-+_____---+_____-__+______-_+-____- Rural Forest | 29.31 177.51 0.3841 0.1131 0.0461 30.61 28.21 26.0 ----------------+-- _ ____+_____-__+______-_+ __ ____-+______--+-____-_+_____-__+_____-__ Rural Savannah | 22.41 161.81 0.4801 0.1731 0.0831 29.21 32.91 35.6 -----------------+-__-__-_+ _-__ _+____ -__+_ ___-__+-____-+__----+_-__--_+_____-__ All I 100.0| 188.51 0.3681 0.1181 0.0521 100.01 100.01 100.0 +--------------------------------------------------------------------- - - - - - - ---+ 1988/89 +-+ Sample Average share welfare P0 P1 P2 Co Cl C2 --------- ------------+ ------------------+- ------ - - ------ Accra I 9.01 258.81 0.2251 0.0631 0.0271 4.91 4.21 4.0 --------------- +_- -__+_ ___-_+______-_+_ ____--+______-_+_-___-_+____-_- -+_____-__ Other Urban | 23.21 201.61 0.3521 0.1031 0.0431 19.5| 17.51 16.5 --------------+---- -- +___-_-_- +_____-__+______-_+______-_+_-___-_+___-__-+_______- Rural Coastal | 15.51 162.91 0.4471 0.1521 0.0671 16.51 17.21 17.1 ---- -------------_+-__ _ -__+- _______+______-_+________+____ ___-+______-_+______-_+______-_ Rural Forest | 30.21 170.41 0.4161 0.1291 0.0561 30.11 28.81 28.0 ---------------+--- _____ +____-__-+______-_+______-_+______-_+______-_+______-_+______-_ Rural Savannah | 22.11 139.41 0.5491 0.1981 0.0941 29.01 32.31 34.4 -----------------+-__ -____+___-____+__ ____-_+______-_+_______-+______-_+_______-+______-_ All I 100.01 177.61 0.4181 0.1361 0.0601 100.01 100.01 100.0 + ___------------------------------------------------------------_--___-- - - - - - - ---------_+ 24  Table 2 (continued) 1991/92 +-+ Sample |Average share |welfare P0 P1 P2 Co Cl C2 Accra 8.21 260.41 0.2041 0.0481 0.0161 6.01 5.71 5.1 -------------- + __-_____+ __ __-+____---+-_____-_+__ ____-_+____---_+-___ -__+______- Other Urban | 25.01 224.81 0.2451 0.0611 0.0231 21.91 22.01 23.2 _-------- _- --+-_ ______+_____-__+_______-+ -____ _+__ __-_-+___-- --+___ __-+____- Rural Coastal | 14.21 217.81 0.2481 0.0571 0.0191 12.61 11.71 11.0 __------ __--- -+-_-_____+_____-__+__ ___ -+____---+-__---+___---_+___---+_____-_ Rural Forest | 29.61 213.51 0.2901 0.0711 0.0251 30.71 30.31 30.3 -----------------+--_-_ -_+ _ - _- _ -+ _ _ ___+ -_--_--- + __---_-+_-_---_+-___---_-_-+__ -_-_-_ Rural Savannah | 23.11 188.51 0.3491 0.0911 0.0331 28.81 30.31 30.4 ------------- +-_- ____+_ __- _ _+______-_+_-__-_--+_ ___ -+ ___-_--+_ _ __-_+_____-__ All I 100.01 215.01 0.2791 0.0691 0.0251 100.01 100.01 100.0 + _--------------------------------------------------------------__---__-- - - - - - - ------+ 25  Table 3: Poverty indices by locality, measure of welfare: total food expenditure, unadjusted, in constant prices, per capita 1987/88 +-+ Sample |Average | l l l share [welfare l P0 P1 P2 Co C1 C2 -----------------+--------- ---- ---------+- ------ - - ------ Accra | 8.31 147.81 0.2941 0.0751 0.0281 6.31 4.81 3.8 --------- --------+-__-____+_____-__+ ______-_+____ -_-+_ ____-_+_ _____--+ ___ _ -_+___ ___- Other Urban | 25.91 107.51 0.4721 0.1651 0.0811 31.71 33.11 34.5 ________-_______ -_+-_ _____+_____-_-+______--+_ ____-__+ __ __--_+______-_+______-_+______- Rural Coastal | 14.11 118.21 0.4091 0.1431 0.0681 14.91 15.61 15.6 __-- _-------+__ - ____+_____ -+_____--+-_ __--+_____-_-+____-___+___ __-+_____- Rural Forest | 29.31 139.31 0.3091 0.0921 0.0401 23.51 21.01 19.2 -------- -----------+-_ -_-___+-_______+___ __-__+___ ___-_+______-_+ ______-_+ _____-__+_____-_ Rural Savannah | 22.41 132.11 0.4061 0.146! 0.0741 23.51 25.41 26.9 --------- -------+- _ ____-+_-____ -_+_- ______+ _____-_- +_ __-____+_ _____-_+_____-__+_ ______ All I 100.01 127.21 0.3861 0.1291 0.0611 100.01 100.01 100.0 +___------------------------------------------------------------__---__-- - - - - - - ----- _+ 1988/89 +-4+ Sample |Average | I share lwelfare P0 P1 P2 Co Cl C2 -----------------+------ +--------------- - --+-+--------- Accra | 9.01 119.91 0.4281 0.1451 0.0721 8.41 8.61 9.2 _____--_---------+_ -____-_+-_______+__- ___ _+______-+_ _____-_+_-____--+__ ___-__+_-___- Other Urban I 23.21 102.21 0.5581 0.1851 0.0861 28.01 28.11 28.2 ---------------+---_-_-_ +____-___+__ ___-_+_ ____-_+______-_+_-____-_+__ ___--_+______-_ Rural Coastal | 15.5! 102.61 0.5371 0.1791 0.0811 18.01 18.21 17.8 ________-______ __+_-___-__+__ ___-_+______-_+______-_+__ ____-_+_-____-_+__ ___-__+______-_ Rural Forest | 30.21 120.51 0.3711 0.1141 0.0501 24.31 22.61 21.4 _______- --------__ +_-______+________+___-4.--+______-_+___ ___-_+ ______-_+______-_+______-_ Rural Savannah | 22.11 109.3! 0.4441 0.1551 0.0741 21.31 22.51 23.3 _______- --------__ +-- _____+________+____-_-_ +______-_+______-_+______-_+___ __-__+________ All I 100.01 111.01 0.4611 0.1521 0.0701 100.01 100.01 100.0 +------------------------------------------------------------------------ - - - - - - ------------+ 26  LZ + --+ 0'00T 0001T 0* 001T |E00 ftsL80o O |ZEO IL-LET 0*001T I iv --------+--------+--------+--------+--------+--------+--------+--------+_----__----------- b-ZI 10-1ST |Z-81T 180o-0 19SO-0 SS-o jO-vST ITE