1. Introduction
2. Comparability, consistency and accuracy
3. Impact of china and India on regional/international levels and comparability of poverty statistics
   1. Table 1. Population, poverty and per capita income in selected ESCAP developing countries, (1999)
4. Data collection methods
5. Concepts and methods used in deriving poverty statistics from survey data
6. Income or expenditure?
   1. Table 2. Energy thresholds and Engel's coefficients used in selected ESCAP countries
7. International comparability
8. Intra-country comparability or consistency
9. Combining reporting systems, censuses and surveys to produce small area statistics

References

I. INTRODUCTION

1. Poverty statistics in this note will be confined to those that measure absolute poverty, as exemplified by the indicators in the International Development Targets (IDTs), a subset of UN Conference resolutions in the 1990s. Statistics based on relative poverty lines, such as half the median per capita income, will not be dealt with here.

2. Some clarification of terms is helpful even within the very few indicators in the IDTs, more so since countries and writers tend to use some of the terms interchangeably. The first IDT is a reduction by one half in the proportion of people living in extreme poverty by 2015. Extreme poverty is associated with not having enough food to eat, more specifically with an energy intake below a specified threshold, such as 2100 kilocalories/capita/day. Food insecure or food poor are other terms used interchangeably with extremely poor.  If to the cost of 2100 kcal (also called food poverty line) the cost of acquiring specified non-food basic needs is added, the result is a total poverty line. Absolute poverty is used to describe the condition where a person earns less than the total poverty line. Since the two poverty lines are in money terms, extreme poverty and absolute poverty are sometimes referred to as income poverty statistics or indicators.

3. The IDTs include also non-income poverty indicators. These revolve around improving access to primary education and basic health services and eliminating gender disparity. The 2015 IDTs are universal access to primary education, eliminating gender inequality in primary and secondary education, reduction by two-thirds in the infant and child mortality rates, reduction by three-fourths in maternal mortality, and universal access to reproductive health services for all individuals of appropriate ages.

4. This report deals largely, though not exclusively, with income poverty statistics. It is to be noted, however, that many of the recommendations for improving income poverty statistics also impact positively on the non-income or human poverty statistics.

II. COMPARABILITY, CONSISTENCY AND ACCURACY

5. Comparability of poverty indicators, or statistics for that matter, is many-sided.  There is comparability or lack of it originating from the primary data -- different ways of collecting the basic data lead to different recorded values of the same item or variable. And then there is non-comparability emanating from differences in the treatment of the primary data, up to the point that the statistics are calculated.  There is spatial or geographic comparability, of statistics between countries and between areas and domains within a country.  And there is temporal comparability, such as from one measurement period to the next.

6. Comparability has not been dealt with extensively in theoretical statistics. Consistency, however, is a well-worked concept with formal methods of defining and measuring it in certain circumstances. In official statistics, consistency is a related but not equivalent concept that is nevertheless sometimes used interchangeably with comparability. In practice analysts often use consistency to indicate whether the statistics remain comparable over time or, less frequently, whether the sub-national statistics in a country estimate the same target parameter or distribution, as in whether the poverty lines of different regions measure the same welfare level (that defines absolute poverty).  Consistency has not been used to describe inter-country or international comparability of statistics.

7. There is congruence between theoretical and official statistics on the concept and measurement of accuracy.  It is the distance between the estimate and the target parameter. A measure is provided by the mean square error (MSE), the average of the squared distance over all possible samples, which simplifies into sampling variance + (bias)¹.  Sampling theory offers universally valid estimators of the sampling variance for a vast array of situations; in general, however, majority of the official statistics of developing countries are still not accompanied by sampling variance estimates.  The situation regarding biases (or non-sampling errors) is gloomier, in large part because these are very location- and situation-specific, even survey-specific; hence there is limited scope for developing universally applicable methods for measuring them quantitatively. Official statisticians often resort to describing qualitatively the procedures they incorporate in a survey that are intended to control/reduce non-sampling error.

8. Their specificity to situations requires that most non-sampling error studies be integrated with the actual survey or production of the statistics; hence it is not surprising that survey practitioners and official statisticians are more into them than theoretical statisticians.^[1] It is also not surprising that only a small proportion of applied non-sampling error studies get published.  It is worth exploring ways to increase support for more non-sampling error studies and to improve the distribution/sharing of the results among countries

III. IMPACT OF CHINA AND INDIA ON REGIONAL/INTERNATIONAL LEVELS AND COMPARABILITY OF POVERTY STATISTICS

9. Of the world's 6 billion people in 2000, 3.3 billion were in developing Asia, and 2.3 billion were living in just two countries - China and India (Table 1).  Clearly, these two countries play pivotal roles in the perception of the level and the degree of international comparability of poverty statistics.  Their influence permeates from the primary data collection methods used, to the treatment of the primary data. For example, by changing nothing but the recall period for food expenses from 30 days to 7 days in its 1999-2000 consumer expenditure survey, India found out that the estimated poverty head count ratio declined from 26.1% to 23.3% respectively (GOI Press Information Bureau, 2001).  The difference is equivalent to 28 million (out of India's 1 billion population at the end of 2000), which equals the total number of poor in Vietnam, or of the sum in the list of nine countries from Kazakhstan to Mongolia in Table 1.  It is a sobering thought that efforts towards more uniformity in data collection methods (taking into consideration the location-situation specificity limitations) between the medium and small sized countries would be worthwhile only if the big two move in the same direction.

10.The effect of differences in methods of deriving the statistics once the primary data have been collected has been the subject of comparatively more attention. The results are no less sobering than the effect of changes in data collection methods. For example, some studies suggest that China's official poverty statistics correspond roughly with a $0.66/day/per capita poverty line.  Raising the poverty line to $1/day/capita increased the number of absolutely poor in 1998 from 42 million to 106 million, or from 4.6% to 12% in terms of head count ratios (LGPR/WB/UNDP Report, 2000).  The 64 million increase is more than the total population of Thailand, and is surpassed only by the number of poor in India in all the countries represented in Table 1 

Table 1.  Population, poverty and per capita income in selected^a ESCAP developing countries, (1999)

Country	Population (million)	Population in povertyb		Per capita
		(%)	(million)	(GNP, US$)c	(GDP, PPP $)d
China	1254	3.2	42	780	3620
India	991	26.1	258	440	2250
Indonesia	207	23.4	48	600	2860
Pakistan	134	32.2	43	470	1830
Bangladesh	128	44.7	57	370	1480
Philippines	77	40.0	31	1050	3800
Vietnam	77	37.0	28	370	1860
Thailand	62	12.9	8	2010	6130
Uzbekistan	24	22.0	5	720	2250
Malaysia	23	8.1	2	3390	8210
Nepal	22	42.0	9	220	1240
Sri Lanka	19	26.7	5	820	3280
Kazakhstan	15	31.8	5	1250	4950
Cambodia	12	36.1	4	280	1360
Azerbaijan	8	68.1	5	460	2850
Tajikistan	6	83.0	5	280	1030
Turkmenistan	5	48.0	2	270	3350
Lao PDR	5	38.6	2	290	1470
Kyrgyz Republic	5	53.3	3	300	2570
Papua New Guinea	4	21.7	1	810	2370
Mongolia	2	35.6	1	390	1710
Total	3080	18.3	564

^a Excluding countries/territories with < 1 million population; developed and with small proportion of poor people, viz, population of Republic of Korea; Singapore; Taiwan Province of China; and Hong Kong, China; and where poverty estimates are not available, e.g. Afghanistan, Myanmar.
^b  The percentages are the latest available estimates, mostly between 1998-2000.  These are multiplied by the 1999 populations to provide approximations for the number of poor people during the 1998-2000 period.
^c  Estimate from World Bank Atlas 3-year-based method.
^d  From UNDP Human Develoment Report 2001, rounded to the nearest 10.

Sources:

• ADB, Key Indicators 2001 Sources:  ADB, Key Indicators 2001
• UNDP, Human Development Report 2001
• LPGR/WB/UNDP, China: Overcoming Rural Poverty, 2000
• NSCB, Philippines
• Government of India Press Information Bureau, Poverty Estimates for 1999-2000
• New Delhi, 22 February 2001

IV. DATA COLLECTION METHODS

11. The countries' surveys that are the main sources of poverty statistics may agree or differ on a number of features and practices. Three of these - sampling procedure including sample size, method of data capture, and frequency or periodicity - exert major influences on (the differences in) quality of the basic data, demonstrate the dissonance or lack of comparability, and point to key survey design and poverty monitoring issues confronting official statisticians in most developing countries. In some countries, such as Bangladesh, China, and Vietnam, the NSOs conduct more than one poverty monitoring survey - for different reasons, but often essentially for the same purpose

12. In China, the Rural Survey Organization of the National Bureau of Statistics (RSO-NBS) conducts an annual Rural Household Survey (RHS) that is the source of the official poverty statistics for the rural areas. (Tang, et. al., 2001). RSO also conducts an annual National Poverty Survey (NPS) that is likewise a household survey covering specifically all the 592 nationally designated poor counties that are all rural. The likelihood of significant overlaps between the 592 NPS counties and the 857 RHS sample counties is high (there are 2,400+ counties in China), but the actual situation is hard to confirm from the published literature.  What seems clear are that the two surveys were designed and are being conducted independently, and that the NPS results are exclusively for use of the Office of the Leading Group for Poverty Reduction (OLGPR) [David and Sangui, 2001]. The Urban Survey Organization (USO-NBS) of the same bureau conducts urban household surveys as well; however, it is hard to discern (from what is published) how urban poverty statistics are derived and how these are combined with the rural poverty estimates to arrive at national estimates.

13. The Bangladesh Bureau of Statistics (BBS) used to run an annual Poverty Monitoring Survey of modest objective and sample size to match; this was expanded through external financial/technical assistance by increasing the sample size and using another method for computing the poverty statistics. Another donor provided assistance also for BBS to do a Household Expenditure Survey in 1995/1996 and in 1999/2000, still using another method.  As a result, Bangladesh has poverty statistics from three surveys and three methods. It is still unclear whether the Government/BBS has made a decision on which survey-method to adopt officially, and whether public funds will be made available to enable BBS to sustain the activity (David, 2000).

14. In Vietnam, the General Statistics Office (GSO) started a Multipurpose Household Survey (MPHS) in 1993 that has been repeated five times, the last one in 1999.  In 1992/93 and 1997/98, GSO with UNDP, SIDA and World Bank assistance conducted Vietnamese Living Standards Surveys (VLSS).  The MPHS and VLSS were very different in many aspects, from sample size (25,000 versus 4,800-6,000 households respectively), to data capture (interview with one year recall versus monthly interviews on 1/12 of the sample for a period of one year), to the method of analysis and, of course, the results. Reconciling MPHS and VLSS into one poverty monitoring and analysis survey system is a priority concern of GSO. 

15. Most of the surveys have been designed to be capable of producing statistically valid estimates for regions (States in India), provinces, and rural-urban classifications, by using these same geographic subdivisions as domains or strata, i.e. with independent sampling inside each one. The sampling is typically in two or three stages employing districts, enumeration blocks, villages and households as sampling units (up to four in China where the county is used as first stage sampling unit), with proportional-to-size selection probabilities at some of the stages, and with fully or partially self-weighting ultimate sampling units (households).  In most countries, however, the ex-ante survey design objectives did not fit ex-post results: what are being released are national, rural-urban and at best regionally disaggregated statistics; provincial and lower level breakdowns have unacceptably high sampling errors. Most of the 10 countries represented at a Workshop on Strengthening Poverty Data Collection and Analysis^[2] placed the redesign of their surveys to improve the precision of small area statistics among their top priorities.

16. Some of the surveys had panel samples, with partial rotation at the ultimate sampling unit level (e.g. China annually rotates one-fourth of the sample households in RHS); and some draw a fresh sample of households at every round (e.g. Thailand). Some countries have reported increasing non-response rates, e.g. 25% in Metropolitan Bangkok and 10% in the rest of Thailand. In the same Workshop mentioned above, the participants ranked high among their priority needs for technical assistance the designing of poverty monitoring surveys that strike a balance between practicable sample rotation schemes, controlling non-response, and computational simplicity.

17. The data capture methods used in the surveys range widely, from daily recording for one year (called diary book + visiting) in China; monthly interviews carried over for a year on 12 sub-samples, as in Thailand and in the more recent WB-supported surveys, e.g. Bangladesh; semestral, as in the Philippines where the January-June data are obtained by interview in July, and for the July-December data the same households are re-interviewed using the same questionnaire the following January; to one time interview for data covering an entire year, e.g. Indonesia.  In all cases, excepting China, the recall period used varies depending on the items, e.g. one week or one month for food, one month or six months for other consumable and semi-durable goods, and annual for durable equipment. In all cases, the data are annualized or expressed in average per capita monthly values. Income data in the majority of cases use the same recall and reference period of one year.

18. The sample sizes of these surveys vary widely, from the 6,000 or sometimes fewer than 4,000 households in the World Bank supported Living Standards Surveys, all the way up to 200,000 in Indonesia's Socioeconomic Survey (SUSENAS). In between there are surveys with 14,000-16,000 households (e.g. Pakistan's HIES and PIHS, Bangladesh's PMS), Thailand's SES with 32,000 households, Malaysia's HIS and Philippines' FIES with 40,000-41,000, and China's RHS with 67,000 sample households. The periodicities also vary, from annual (e.g. China, Indonesia), biennial (Thailand), triennial (Philippines), quinquennial (India, Malaysia's HES), to irregular, dependent by and large on availability of donor support (e.g. Bangladesh, Cambodia, Lao PDR).

19. The World Bank-designed Living Standards Surveys involve relatively smaller sample sizes because their questionnaires are very long - to support aims that go beyond poverty assessment and monitoring, to a research agenda on the nature and causes of poverty, including non-income poverty, and on standards of living. Also, the data collection employs splitting the sample into twelve sub-samples and stretching the field data collection over as many months, which is much more resource-intensive and time consuming, that the NSO could not possibly do it on a bigger sample without making some sacrifices to the agency's equally important responsibilities, including surveys, during the year.  Thus, these surveys cannot be expected to produce precise statistics for small areas, but only for national and rural-urban domains, or sometimes large regions.

20. The diversity in sizes and periodicities of the surveys provide indications of some of the NSOs' efforts to satisfy the various stakeholders' requirements for (a) statistics for smaller areas than normally produced from nationwide inter-census sample surveys - for more directed targeting of poverty alleviation interventions and (b) more frequent updating of the statistics - for more frequent monitoring of the impact of poverty reduction policies and programs. There is also a third demand, namely (c) for more statistics - to cover the collective requirements of donors (e.g. common country assessment or CCA indicators of the UN system) and of the non-income dimensions of poverty - that add to the breadth and length of the survey questionnaires.

21. The NSOs' efforts towards satisfying (a) have not been that successful for a basic but evidently frequently ignored reason. Sample surveys are the wrong instruments to rely on for producing the kind of small area statistics required, namely poverty indicators at village, district or geographic levels below regions (e.g. counties for China, sub-states for India, and provinces for the smaller countries).  That role is intended for censuses (of agriculture, population and housing in particular). However, censuses are so massive and costly that their questionnaires are necessarily thin and cannot support the estimation of poverty statistics like per capita consumption expenditure, poverty line and the like; and they are conducted generally once in ten years only.  Hence, by themselves, censuses fail in providing small area poverty statistics and fail miserably in satisfying (b), i.e. in updating statistics frequently. Combining surveys and censuses through small area estimation techniques, which is discussed at more length later, offers potential solutions.

22. How often should poverty statistics be updated? The object is to monitor change in, say the poverty headcount ratio between two succeeding surveys, P_t - P_t-1. The pivotal quantity is the t-value or ratio between the estimate p_t- p_t-1 and the standard error s(p_t-p_t-1). The power of the t-test, or the ability of the survey to differentiate between signal and noise (for a given level of significance), is a function of P_t - P_t-1 and the precision or standard error from the survey. This means that, unless there is reason to believe that the poverty situation is either improving fast or is deteriorating fast, it is not likely that a change can be detected and found significant from surveys that are only a year or two apart. Hence, there may be justification for more frequent monitoring during an economic crisis, as indeed there was in the aftermath of the 1987 Asian crisis. By the same token, less frequent monitoring would be justified in periods of economic calm, as well as when the poverty incidence has been reduced to such low levels that the range of P_t - P_t-1 gets severely limited.  There is even less logic in trying to monitor annually or even biennially the change in small areas where the standard errors are larger on account of the smaller sample sizes (David, 2000). Unfortunately, these elementary statistical truths also seem to be overlooked when some countries design their poverty monitoring systems.

23. Regarding (c), the time required to complete the questionnaire has an inverse adverse impact on quality of the basic data.  Accumulating empirical research tends to show that after one hour, the quality of the basic data obtained from face-to-face interview deteriorates quickly (for opinion polls 30 minutes is often the advocated maximum interview length).  Unfortunately, questionnaires that take up to 2 hours to complete, sometimes more, are not unheard of in socioeconomic or income and expenditure surveys. One possible solution, which will be described in more detail later, is to adopt a system of integrated surveys in which individual surveys are modules consisting of shorter questionnaires on a more limited range of subjects.

V. CONCEPTS AND METHODS USED IN DERIVING POVERTY STATISTICS FROM SURVEY DATA

24. This section will focus on the concepts and methods that impact significantly on the level and comparability of the poverty statistics: calculation of the food poverty line (fpl); calculation of the total poverty line (tpl); and metric used (e.g. consumption expenditure or income) to determine poverty incidences and number of poor persons.

25. How the two biggest countries derive their poverty statistics illustrate some of the sources of deviations in national practices.  China's method, described in Tang, et. al. (2001), is as follows: First, decide on an energy threshold that would distinguish between the food poor from the non-food poor. Based on studies by the Ministry of Health's Institute of Nutrition and Food Hygiene, the threshold was set at 2100 kilocalories/capita/day. (This is an average based on varying minimum calorie requirements for different age-by-sex groups of the population.)  Second, select a reference population (viz. households in the lowest quartile of the per capita income distribution) and, guided by the actual composition of the food consumed by this group, decide on a food bundle that would satisfy the energy threshold. This is where the Rural Household Survey is needed: for example, from the 1995 RHS the reference population was determined to be households with annual per capita income below RMB 800; and the selected food bundle comprised 100 items which were classified into 15 groups (e.g. grain, beans) and 27 sub-groups (e.g. under grain there are individual items like rice, wheat and a catch all item, other grains). This means that the estimated kcalorie content/unit weight of each item is available, which when multiplied by the average per capita consumption gives the kcalorie intake from that item; and the sum across items comes close to 2100 kcals.^[3]Third, the price paid by the reference population for each item in the bundle is multiplied by the quantity consumed (adjusted, see footnote); and the sum of the products across the items in the food bundle is the estimated fpl.^[4]In 1995,  fpl = RMB 477.  Fourth, an expense for 'non-food essentials' like clothing, shelter, and primary medical and basic educational services is estimated, which when added to fpl gives tpl. From 1995, the NBS has adopted a method proposed from the World Bank (Ravallion, 1994) which is premised on a particular definition of what constitute essential non-food basic needs: a household whose total expenditure = fpl still has to spend for items other than food; and whatever non-food goods and services he chooses to buy can be regarded as essential. Thus, the average non-food expenditure of households satisfying the above equality can be added to fpl to arrive at an estimate of tpl.  The problem, of course, is that no households will have total expenditure exactly equal to fpl. The proposed World Bank solution, adopted by NBS, is to run a linear regression of the share of food expenditure to total expenditure (S_i, which is also called Engel's coefficient) on log (X_i/fpl), where X_i is total expenditure, and _i runs through the sample households that belong to the reference population; i.e.

S_i  =  ?  +  ? log(X_i/fpl) + error.

Since log 1 = 0, it follows that the estimate of  ? (say a) may be used to estimate the food share of households whose total expenditure = fpl, or 1-a may be used to estimate the share of essential non-food expenditures. In 1995, 1-a = 0.17, hence,

 tpl = 1.17*fpl =  RMB 557. 

{It is to be noted that before 1995 tpl = 1.40* fpl, where the 40% adjustment was based on what was then the experts' opinion of a 'reasonable food share' of 60%. The big reduction in the adjustment factor from 40% to 17% has led to speculations that the pre-1995 estimates are not comparable to those from and including 1995, and that the latter may underestimate of the true magnitude of poverty (Park and Wang, 2000). Majority of the Asian developing countries use an adjustment factor in the neighborhood of 30%.}

26. And fifth, the household per capita net income distribution is estimated from the RHS, taking care to use the same prices as those used in the calculation of fpl and tpl.  The number or proportion of households in the income distribution whose incomes fall below tpl are absolutely poor, also taking care that the calculations reflect the statistical nuances of the RHS sampling design.  The number of persons in the absolutely poor households, or their proportion in the estimated total population, is the estimate of absolute poverty incidence or head count ratio.

27. the poverty lines fpl and tpl are recomputed from RHS data roughly every five years, viz. 1985, 1990, 1995 and 1998. In between, the lines are updated using consumer price indexes. However, calculation of the poverty incidence and other functions of the head count ratio require updating of the household per capita income distribution from new data supplied by the (annual) RHS.  There seems to be no quick and cheaper way to do this for the time being. 

28. NBS calculates or publishes poverty statistics at the national level only. It seems that corresponding estimates using fpl in place of tpl, i.e. statistics on the food poor or extremely poor, are not published. Given the size of the country, the absence of regularly published poverty statistics for the provinces, cities and counties is extraordinary. There is, in fact, an emerging need for poverty statistics down to the village level (LGPR/UNDP/WB, 2000; David and Sangui, 2001).

29. The official statistics are supposed to be a combination of estimates from RHS and the Urban Household Survey (UHS) done by another organization in NBS, except that there is even less published documentation on the latter survey and on the procedure for the merging of the estimates.  Tang, et.al. (2001) report that (a) both surveys miss residents in the semi-urban areas such as towns; (b) both surveys also miss the floating migrants, especially those who move to the cities part of the year to work; (c) income in the RHS is defined more expansively to include transfers on the one hand, but does not include imputed rent of owned house on the other hand; and (d) there is under-representation and possibly outright exclusion of the more remote villages and illiterate families (in which the daily diary method or record keeping would be a problem).

30. In India, the idea of quantifying poverty lines started as early as the 1950s. The present methodology has its roots in the report of a 1979 Planning Commission Task Force that worked out 1973-74 poverty lines based on average requirements of 2400 kcal for rural areas and 2100 kcal for urban areas and a set of minimum non-food requirements (Sharma, 2001) . The 1973-74 poverty lines were Rs. 49.09 and Rs.56.64 per month for the rural and urban areas respectively. Later, these poverty lines were disaggregated into state-specific poverty lines using inter-state price differentials by Fisher's index. The rural state poverty lines were continuously updated by using state-specific CPIs for Agricultural Laborers, and urban poverty lines are updated similarly using the CPIs for Industrial Workers. The state poverty line estimates are weighted averages of the rural and urban estimates; and in turn the national estimate is a weighted average of the states estimates. It is to be noted that, unlike in China and most of the countries, the India poverty lines have not been recomputed from new consumption expenditure data since 1973-74, but were merely updated or adjusted using price indexes.

31. The Planning Commission using the National Sample Survey Organization's Household Consumer Expenditure Survey (CES) estimates the headcount ratios and number of poor every five years. State-specific distributions of per capita monthly consumption expenditure are constructed from the CES, which together with the poverty lines give off the poverty incidence estimates. After 1973-74, estimates were compiled in 1977-78, 1983, 1987-88, 1993-94, and 1999-2000. Thus, while China uses income to produce poverty incidence statistics every year, India uses expenditure and updates her statistics every five years. China regularly publishes national estimates only, while India releases national and states estimates disaggregated into urban and rural domains. India's poverty lines are based on higher energy thresholds than China's. 

32. Majority of the countries use a food bundle or food basket approach along with an energy threshold to calculate fpl. An exception is the Philippines where, instead of just an energy threshold (2000 kcal), 100% of the minimum protein (50 grams/capita/day) and of the other nutrients and vitamins are specified as well. Moreover, instead of building these nutrient requirements into a food bundle, these are used as guide to construct one-day "poor person" menus separately for each rural and urban domain in each region. The fpl is computed as the sum of the (weight x unit price) of the ingredients that comprise the menu in a domain. Estimates for bigger areas are built up as weighted (by population) averages.  It has been hypothesized that the use of one-day  menus and of more than just an energy threshold are key reasons for the unusually high official poverty statistics of the country (David and Maligalig, 2001).  Other authors have proposed alternative methods with results that were significantly much lower than the official statistics (Balisacan, 2001; Kakwani, 2001). 

33. Instead of the World Bank proposed regression method of estimating the adjustment for essential non-food expenditures, the Philippines takes the non-food share to total expenditure (net of expenses for tobacco, alcohol, etc) among households whose per capita incomes lie within ± 10% of fpl. This is a much simpler method (that Laos used also, but using per capita household expenditure in lieu of income; see Kakwani, et.al., 2001). Like China, the Philippines also uses per capita income distribution in calculating poverty incidence and the number of poor.

VI. INCOME OR EXPENDITURE?

34. The estimation of fpl is generally based on expenditure for food that satisfies certain pre-specified nutritional requirements. The food items are those consumed by a reference (lower income) population and are commonly available in the area or country. The range of the energy thresholds adopted by the countries is surprisingly small, with a midpoint in the neighborhood of 2100 kcal/capita/day. (See Table 2). Beyond that, it would be unreasonable - say in the interest of strict comparability - to expect countries to agree on the same food bundle.

Table 2.  Energy thresholds and Engel's coefficients used in selected ESCAP countries.

Country	Energy thresholds kcal/capita/day)	Engel's coefficient (fe/te or fpl/tpl)
China	2100	0.83b
India	2400 rural; 2100 urban	.
Indonesia	2100	0.70 - 0.75c
Pakistana	2000-2100	.
Bangladesh	2122	0.76
Philippines	2000	0.65 - 0.71
Vietnam	2100	0.72
Thailand	2100	0.60
Uzbekistan	.	.
Malaysia	1982d	.
Nepal	2124	.
Sri Lanka	.	.
Kazakhstan	.	.
Cambodia	2100	0.75 - 0.79
Azerbaijan	.	.
Tajikistan	.	.
Turkmenistan	.	.
Lao PDR	1983	0.80
Kyrgyz Republic	2100	0.83
Papua New Guinea	.	.
Mongolia	2100	0.67 - 0.74
... = not available a Pakistan does not have official poverty statistics. b Refers to 1995 onwards; this figure is 0.60 prior to 1995. c Figure used since the early 1990's; earlier, this was 0.90. d Based on 9910 kilocalories/day for a family of five.

5. The upward adjustment of fpl to a tpl is generally based on expenditure also, of a bundle of essential non-food requirements. There is diversity in country practices here: Some adopt Ravallion's proposed definition of essential non-food requirements as those procured by households whose total expenditure = fpl. Other countries use ratio of food expenditure (fe) to total expenditure (te) of a reference group of households, e.g. those with per capita incomes within a narrow band around fpl. Sometimes the expenditure for tobacco, alcohol and durable equipment like motorcycles are excluded in the computation of fe/te. The modal value of fe/te appears to be in the neighborhood of 2/3 (Table  2).

36. In determining the incidence, number, depth and severity of poverty, the countries are split between using per capita income or per capita expenditure. China uses income, while India uses expenditure, for instance. The two metrics can lead to very different estimates even in the same country and from the same survey.  Using the 1997 Family Income and Expenditure Survey (FIES) data, Balisacan (2001) reported a 25% poverty incidence for the Philippines using expenditure (although he also changed a few other steps in the official methodology), compared with the 37% official estimate based on income. In rural China, they sometimes refer to the poor as those in households whose per capita net income is below tpl and whose per capita expenditure is below 1.5 x tpl (Tang, et.al., 2001). Although no definite numbers are given, it is possible that this definition has been prompted by an observation that the proportion of households with per capita expenditure < tpl is significantly less than the proportion with per capita net income < tpl.

37. conceptual theory favoring expenditure abound in the published literature; e.g. borrowings and transfers which offer added opportunities for consumption (expenditure) are not part of the usual or UN-recommended definition of income, so that expenditure is a broader measure of welfare and is more able to reflect consumption smoothing. Thus, expenditure could lead to more meaningful and stable poverty statistics, whereas income tends to measure transitory, hence less stable, poverty. From a practical standpoint, it has been argued also that expenditure can be more accurately measured and cheaper to obtain from surveys; e.g. a year of income data obtained over a number of visits would normally be needed, while a week for certain consumables like food, a month for semi-expendable items, and a year for durable items may suffice for expenditure.

38. On the other hand, the difficulty in recording the income of the very rich has much to do with the perception that income survey data is less accurate than expenditure data. But for purposes of estimating poverty incidence, it is adequate to estimate household income distribution truncated at a point just above the anticipated value of tpl. Majority of those in this truncated distribution are salaried and other fixed income workers who can more accurately tally their incomes than their expenditures. It is not at all certain also that expenditure is cheaper to obtain than income. The Philippines 1997 Family Income and Expenditure Survey questionnaire is 70 pages long, 42 of which are on expenditure and only 12 are on income (Virola, et. al., 2000). Jinho Hur (2001) reports that in the Korean Household Income and Expenditure Survey questionnaire, items of income and expenditure are classified by commodity in accordance with ILO classifications, and the income portion has 23 subordinate items, while the expenditure portion has 516 subordinate items.

39. More applied research is needed to study the comparative advantages of using income or expenditure in certain specific situations and countries.  In these studies, and for purposes of analyzing and assessing poverty, it is neither necessary nor advisable to adhere strictly to the UN recommended definitions; and very few countries do.  Countries that use income - China and Philippines, for example - include some transfers in the computation of income. Likewise, a critical review of the operational definitions of income and expenditure that are used by the countries in their poverty statistics compilation would be arduous but useful.

VII. INTERNATIONAL COMPARABILITY

40. In Figures 1 and 2, the head count ratios in Table 1 are plotted against US$ GNP per capita (World Bank Atlas method) and PPP$ GDP per capita respectively. Both scatter plots show "outlier" behavior by the same subset of countries. China's head count ratio is comparatively low relative to her income; some of the probable reasons have  been  discussed above, including a low tpl, under-representation and exclusion in the sample of some areas (e.g. of high illiteracy, inaccessibility), and underestimation of urban poverty incidence.  On the other hand, the Philippines' head count ratio looks inordinately high relative to her income, for reasons analyzed earlier (e.g. Kakwani, 2000, Balisacan, 2001,  David and Maligalig, 2001).  Why some of the Central Asian Republics also exhibit very high poverty incidences relative to their per capita incomes need closer scrutiny; here, however, the accuracy of the national income data needs to be studied also.

41. It is not difficult to envisage that between harmonizing the countries' data collection methods or their computational methods towards improving the comparability of poverty statistics, the former will definitely be much more difficult to do. All one has to do is try imagining what and how long it would take to convince India to revise its quinquennial Consumer Expenditure Survey and ask China to do the same with her RHS and NPS. The World Bank's  LSMS or LSS package seems to have found more acceptance in the smaller transition economies, countries with relatively undeveloped statistical systems, and/or those that badly need technical/financial assistance. It may be too early to tell whether the initial acceptance will lead to countries' repeating LSS on their own once the external assistance stops. Aside from China and India, the LSS package also has had less success of adoption in countries with relatively more developed statistical systems, e.g. Malaysia, the Philippines.

42. The quest for international comparability in poverty statistics should be guided by practicality and pragmatism. On the one hand it had been said that truly comparable statistics are either difficult or impossible to develop; and the more comparable they are made, the less relevant and useful they would be to individual country situations and needs. If I recall correctly, this line of  thinking led to a recommendation in one ESCAP meeting, to let individual countries go about their own way in defining and deriving their poverty statistics. On the other hand, 150 countries are signatories to the Millenium Summit in September 2000, that included a consensus on achieving International Development Targets (IDTs) that focus on reducing abject poverty by 2015. Monitoring of the IDTs is through a set of mostly (income and non-income) poverty indicators that need to be updated and aggregated globally during milestone years leading to and including 2015. There is, therefore, a clear need for indicators that are more or less comparable to an extent that combining them regionally and globally would make sense.

43. The International Comparison Programme (ICP) provides many lessons on the pitfalls of developing and getting countries to cooperate in the production of comparable statistics, i.e. GDP.  After more than 30 years, most of the analytical, conceptual and methodological difficulties may have been solved. However, the actual data production and updating continue to be beset with delays stemming from issues that appear to remain unresolved to the satisfaction of some countries; e.g.:  some countries are satisfied with their national currency GDP and have not been sufficiently convinced that PPP GDP's provide enough additional value to them to justify the added cost of participating in ICP price surveys; the role of the countries is reduced to that of data collector, with the actual computation of the PPP GDPs done elsewhere (World Bank, Univ. of Pennsylvania), and the way that is done is a "black box" to many; and consequently, developing country statisticians are not fully prepared to explain to their clients how the PPP GDPs came about, or whether or not there is systematic bias in the results in which developing countries all end up looking richer and the developed countries proportionately poorer.

44. The computation of the World Bank's $1/day/capita poverty lines starts with converting the national currency poverty lines of countries to a common numeraire, i.e. PPP $ (World Bank, 1990). Thus the methodology is subject to exactly the same unresolved ICP issues mentioned above. There is even a parallel in the earlier World Bank assurances that use of the PPP GDPs' will be confined to analytical purposes and not on operational policies and decisions regarding the borrowing countries. Similarly, it has been said that $1/day poverty estimates are not intended for use in or for individual countries, but are simply meant to be inputs to calculating regional and global estimates; however, their publication virtually guaranteed their propagation and use to compare, rank and analyze individual countries, such as in UNDP's Human Development Reports.

45. There is much to recommend in exploring simpler means to an incremental approach to improving the inter-country comparability of poverty statistics. For example, we could exploit the observation that energy thresholds and Engel coefficients in current use are concentrated around 2000 -2100 kcal and 2/3 (see Table 2).  Convincing countries to use the same or approximately the same energy threshold, or else produce additional tabulations in addition to their official threshold, should be easier than asking them to adopt the same food bundle. The individual countries' estimates of head count index may then be said to be comparable in the sense that they are in the same units and that they are based on the same energy threshold (David and Maligalig, 2001). The same goes with the estimates of the number of extremely poor people.^[5]  The procedure circumvents the use of fpl, which is in different national currencies.  Similarly, use of a constant Engel's coefficient (viz. 2/3) to compute tpl = (2/3)fpl will mean freedom from having to recompute the coefficient every so often, at the same time that it ensures a more temporally consistent tpl. Furthermore, the resulting headcount indexes and number of absolutely poor people will be comparable in the sense that they are expressed in common units of measure, and are based on the same energy threshold as well as on the same food to total expenditure ratio.

VIII. INTRA-COUNTRY COMPARABILITY OR CONSISTENCY

6. There are significant advantages to be gained from having comparable sub-national poverty statistics. The poor are spatially distributed, e.g. rural poverty incidence is generally higher than urban poverty incidence, but not always (in Mongolia it is the reverse); the probability is higher for female-headed households to be poor, but not always (Philippines, Thailand and Vietnam are exceptions); and areas characterized by poor soil, inhospitable terrain, extreme climate (very arid or flood-prone), and far from markets tend to have more of the poor in the population. Poverty incidence tends to be linked with ethnicity also. Comparable statistics provide more accurate information on the specific location and count of the poor, which lead to more appropriate and efficient targeting of poverty alleviation programs.

47. Making sub-national poverty statistics comparable is easier because all prices are in a single currency. Two things remain to be ensured, however. One is to define being poor nationally, so that the same welfare standard (of extreme or absolute poverty) is being applied uniformly.  Suppose the reference population is chosen to be the lowest quartile in the per capita income distribution of all households, and let x be the cut-off point for the quartile, i.e. Pr[per capita income < x] = 0.25. Then x also determines the reference sub-populations in all the sub-national domains of interest, with the probabilities allowed to vary accordingly. For example, in the usual case where per capita incomes are higher in urban than in rural areas, x defines a smaller proportion of households below it in the urban domain than in the rural domain. The other important thing is to apply in the poverty statistics calculations the prices that these reference sub-populations actually pay for the goods and services they consume. 

48. Without downgrading the importance of having comparable sub-national poverty statistics, it is worth noting that there are situations when the need for non-comparable indicators arise. When poverty intervention programs are run by local governments, the latter find it essential to make the definition of who is poor and the resulting count commensurate with the available resources. This is the case in China where the local governments are required to supplement from their own budgets the poverty alleviation allocations from the central government.

49. In many countries, the tasks of implementing and/or allocating resources for poverty alleviation programs fall on central ministries and specialized agencies. Examples include the Office of the Leading Group for Poverty Reduction (OLGPR) that is directly under the State Council in China; National Family Planning Board and State Ministry for People's Welfare and Poverty Alleviation in Indonesia; National Anti-Poverty Commission in the Philippines; Community Development Department of the Ministry of Interior in Thailand; and Ministry of Labor, Invalids and Social Affairs in Vietnam. These agencies engage in poverty targeting for which they need indicators to identify/count the beneficiaries or rank villages and districts according to indicators of the incidence, depth and severity of poverty. Unfortunately, and as mentioned previously,  the national statistical offices (NSOs) in general have managed to provide useful poverty statistics for large areas such as regions and whole country only. Thus, there is either an unfilled need for small area poverty statistics, or the agencies produce their own set of numbers that, when aggregated at higher levels, are found to be at odds with the official statistics from the NSOs. NSO-subject matter agency collaboration is urgently needed to ease these problems.

IX. COMBINING REPORTING SYSTEMS, CENSUSES AND SURVEYS TO PRODUCE SMALL AREA STATISTICS

50. In market economies, the official statistics come mainly though not exclusively from censuses and sample surveys. Some, e.g. education and health statistics, come partly from administrative reports, but the latter are not designed specifically as sources of statistics. Thus, censuses are the primary sources of small area statistics. What has not been explored fully is combining surveys and censuses, with the latter functioning as  auxiliary information sources, to produce small area statistics. In transition countries, many of the statistical reporting systems from central planning days are still intact, and sample surveys and censuses have also been initiated; consequently, the official statistical databases of the latter countries are mixtures from these three different sources. Thus, in these latter countries, there are two potential sources of auxiliary information, namely statistical reporting systems and censuses.

51. Consider China, for example, where the provinces are divided into counties, then to towns/townships, administrative villages, and natural villages.  The (NBS) has a reporting system that reaches all these levels; the reporting forms are summarized progressively upward and what reaches NBS are county summary sheets containing 128 indicators. This NBS county indicators database is updated annually; however, only 20 of the indicators are published regularly (David and Sangui, 2001). The village summaries are left with the townships, the township summaries sheets stay with the counties, (and in the case of the other ministries' reporting systems, the county summaries are submitted to the provinces, and only provincial summaries reach the central ministries). NBS conducted a full enumeration agriculture census in 1996, from which an Abstract (1999) has been published containing provincial and national summaries mainly, and has put in the Internet a one percent public use sample. As mentioned previously, NBS also runs an annual Rural Household Survey (RHS), from which only national level poverty statistics are published regularly.

52. Thus, anybody will be hard put to assemble a county level set of crosscutting indicators (such as a set of income and non-income poverty indicators) if the indicators are not all covered in the NBS county database. From Beijing, it will not be possible to construct a table of township indicators, much less village indicators. There is no complete set of published poverty incidence estimates for the 34 provinces and cities.

53. The Statistics discipline has for well over two decades now developed and improved methods for estimating small area statistics.  For example, the RHS can provide valid design-based estimates for each of the provinces; that it cannot do so for all the counties is obvious from the fact that only one third of them are represented in the sample (857 out of 2,412 total counties).  Theoretically, it might be possible to produce estimates for each of the RHS sample counties; however, the errors in these estimates will in all likelihood be unacceptably high. Suppose the objective is to estimate the proportion of the energy deficient or extremely poor (e.g. per capita daily consumption < 2100 kcal) in each of the 2,412 counties. One way to attempt improving the estimates for the sample counties and constructing estimates for all the other non-sample counties is to find predictor variables present in the RHS sample counties and in all the counties from an auxiliary source, such as the most recent census or the NBS county indicators database; compute the proportions of extremely poor people directly in each of the 857 sample counties and regress these against the predictor variables, using either ordinary or weighted least squares for example; input the predictor variables from the auxiliary source into the fitted regression equation to get the estimated proportion of extremely poor people for each of the 2412 counties. The quality of the estimates will be a function of the goodness of fit of the regression model as well as the accuracy of the indicators in the auxiliary source.  The latter takes on added importance when the auxiliary source is from the reporting system that uses a different method of data collection than the RHS.^[6]  The quality can and should be estimated also, e.g. by comparing the pairs of direct estimates and regression estimates from the 857 RHS sample counties.

54. It is to be noted that the county level predictor variables must be available both in the RHS and the auxiliary source. Thus, if the 1996 agriculture census is the chosen auxiliary source, these variables are actually county statistics that have to be estimated/computed from RHS and the census. When the auxiliary source is an existing county indicators set from the reporting system (of NBS), the choice of indicators as predictors must be guided by the condition that these can be estimated from the RHS as well. 

55. In the example above, township or village can replace county.  Also, there are many more small area estimation techniques developed by survey sampling statisticians..  These have proven themselves very useful that a short course, "Introduction to Small Area Estimation" has become a regular satellite activity in conferences of the International Statistical Institute, International Association for Survey Statistics and International Association for Official Statistics. The reference list to the last short course run by J.N.K. Rao (August 2001) is a good guide for those who wish to read more about this very important area.

56. Based on their recent published works, analysts in the poverty research area tend to favor methods that are similar to the above example, except that the model used is brought down to the level of the household (e.g. Hentschel et. al., 1998; Bigman and Deichmann, 2000; Fofack, 2000;  Henninger, 1998; Schady; Elbers, et. al. 2001).  For example, the total food expenditure per capita of the household (or its logarithm), which is available in RHS but not in the agriculture census, is expressed as a linear function of a predictor variables set, and a regression analysis is run on the RHS sample households.  The choice of  predictor variables is limited to those available from the household records of the auxiliary source, such as the 1996 agriculture census or the 2000 census of population and housing. The fitted equation and the census data can be used to estimate the total food expenditure per capita of every household. However, researchers in general do not recommend that the results be used to classify or rank individual households. Misclassification errors of both kinds (i.e. poor but actually not, and not poor but in fact poor) can be high at the household level.  Instead, the household results are aggregated to small areas of interest. For instance, all persons in a village that belong to households whose estimated per capita food expenditures fall below a pre-specified threshold may be used to estimate the number (or proportion) of undernourished persons in that village. Alternatively, the average predicted per capita consumption in the village is computed, which can be used to rank villages.  Similarly, it is possible to estimate error rates e.g. by comparison of the regression estimates from the survey estimates.

57. Unlike the first example that requires small area aggregates only, the latter is very data intensive, as it requires the household data of both the survey and the census.

References

Asian Development Bank. (2001). Key Indicators of Developing Asian and Pacific Countries, Vol. XXXII, Oxford University Press, Hongkong.

Balisacan, A. (2001) Is What We Know about the Poor Robust? Presented at the ADB Asia and Pacific Forum on Poverty, Manila, 5-9 February 2001.

Bigman, D. and Uwe Deichmann (2000). Geographical Targeting: A Review of Different Methods and Approaches, Ch. 1 in Bigman, et. al. (Editors),. Geographical Targeting for Poverty Alleviation, 2000.

Boonperm, Jirawan and Sopon Tatiyanunphong (2001). Data Collection for Poverty Monitoring and Analysis. Presented at the Regional Workshop on Strengthening Poverty Data Collection and Analysis, 30 April - 3 May 2001, Manila.

China: Overcoming Rural Poverty. (2000). Joint Report of the Leading Group on Poverty Reduction/ UNDP/World Bank, Beijing, May 2000.

David, I. P. (2000).  Poverty Statistics and Indicators: How Often Should They Be Measured?  Presented at the Ninth Meeting of the Heads of National Statistics Offices of East Asian Countries, 23-25 August 2000, Gotenba, Japan.

________. and D. S. Maligalig  (2001).  Issues in Estimating the Poverty Line. Presented at the Regional Workshop on Strengthening Poverty Data Collection and Analysis, 30 April - 3 May 2001, Manila.

________. and Sangui Wang. (2001).  Establishing a National FIVIMS in China. Presented at a National Seminar on Food Insecurity and Vulnerability Information and Mapping System, 21-22 November 2001, Beijing.

Elbers, C., J. Lanjouw and P. Lanjouw (2001). Welfare in Villages and Towns:Micro-Level Estimation of Poverty and Inequality. WIDER Development Conference on Growth and Poverty, 25-26 May 2001, Helsinki.

FAO. (2000). Guidelines to National FIVIMS: Background and Principles. IAWG Guidelines, Series No. 1, Rome.

____. (2000). The State of Food Insecurity in the World 2000. Rome.

Fofack, H. (2000). Applying Household Expenditure Survey Data to Improve Poverty Targeting: The Case of Ghana. Ch. 5 in Bigman,, et. al. (Editors).

Government of India, Press Information Bureau. (2001). Poverty Statistics for 1999-2000. Press Release, 22 February, 2001, New Delhi.

Henninger, Norbert (1998). Mapping and Geographic Analysis of  Human Welfare and Poverty - Review and Assessment. World Resources Institute, Washington, D.C., April 1998.

Hentschel, Jenjo, J. A. Lanjouw, P. Lanjouw and J. Poggi (1998). Combining Census and Survey Data to Study Spatial Dimensions of Poverty: A Case Study of Ecuador. Draft, 8 April 1998, World Bank, Washington, D.C.

Hur, Jinho (2001). Introduction to Household Income and Expenditure Statistics in Korea. Presented at the Regional Workshop on Strengthening Poverty Data Collection and Analysis, 30 April - 3 May 2001, Manila.

Kakwani, N. (2001) On Specifying Poverty Lines. Presented at the ADB Asia and Pacific Forum on Poverty, Manila, 5-9 February 2001.

_________, Sisouphanthong, B., Souksavath, P. and B. Dark. (2001).  Poverty in Lao PDR. Presented at the ADB Asia and Pacific Forum on Poverty, Manila, 5-9 February 2001.

National Bureau of Statistics (1999) Abstract of First National Agricultural Census in China, Beijing, 1999.

Park, Albert and Sangui Wang (2000). China's Poverty Statistics. Forthcoming.

Rao, J.N.K. (2001). Introduction to Small Area Estimation. International Association of Survey Statisticians Short Course, Seoul, Korea, 20-22 August 2001.

Ravallion, M. (1994). Poverty Comparisons. Harwood Academic Publishers.

________, G. Datt and D. van de Walle (1991). Quantifying Absolute Poverty in the Developing World. Review of Income and Wealth, 37(4):345-361.

Schady, Norbert (?). Picking the Poor: Indicators for Geographic Targeting in Peru. World Bank, Washington, D.C.

Sharma, Savita (2001). Country Paper: India. Presented at an ADB Inception Workshop on Building a Poverty Database, 30 July-1 August 2001, Manila.

Surbakti, S., P. Surbakti, La Ode Syaffudin and W. Imawan. (2001). Indonesia: Poverty Data Policy, Management, and Implementation. Presented at the Regional Workshop on Strengthening Poverty Data Collection and Analysis, 30 April - 3 May 2001, Manila

Tang Ping, Sheng Laiyun and Wang Pingping (2001). Poverty Monitoring in Rural China.  Presented at the Regional Workshop on Strengthening Poverty Data Collection and Analysis, 30April - 3 May 2001, Manila.

UNDP.  Human Development Report 2001. Oxford University Press, New York.

Van Tien, Nguyen. (2001). Poverty Data Collection Activities in Vietnam. Presented at the Regional Workshop on Strengthening Poverty Data Collection and Analysis, 30 April - 3 May 2001, Manila.

Virola, G.A., V. N. Ganac and I. S. Bacani (2000). Poverty Assessment in the Philippines. Presented at a series of Users' For a on Poverty Statistics, sponsored by the National Statistical Coordination Board, March -August 2000.

World Bank (1990). World Development Report 1990: Poverty. New York, Oxford University Press.