A series of issues arise when constructing welfare measures, setting poverty lines, calculating poverty and inequality indicators, and when comparing poverty over time or across areas. These can be due, among others, to the choice of instruments, measurement errors, sampling issues. A recent paper by Angus Deaton (Princeton University) and Valerie Kozel (World Bank) "Data and Dogma: The Great Indian Poverty Debate" (September 2004), addresses the issues of discrepancies between surveys and national accounts, the effects of questionnaire design, reporting periods, survey nonresponse, repairing imperfect data, the choice of poverty lines, and the interplay between statistics and politics. Although the paper focuses on the Indian example, the issues have wide resonance elsewhere. This site is still under construction. A more developed site will systematically analyze issues and potential solutions. Comparing poverty with different surveys When comparing data estimates using different surveys – typically establishing poverty trends in a country on the basis of a series of household surveys – a series of issues arise. They relate to the construction of the consumption aggregate, the sampling of the surveys, and the setting up of poverty lines. The following articles address some of these issues in depth: Testing for the robustness of poverty comparisons Due to the many assumptions involved in poverty measurement, it is important to test for the robustness of poverty comparisons between groups or over time. There are three main ways of testing for robustness (this section is based mainly on Coudouel et al. please refer to the full document and its technical notes, as well as to its list of references, for a deeper treatment of the issues): Standard errors: Poverty calculations are based on a sample of households, or a subset of the population, rather than the population as a whole. Samples are designed to reproduce the whole population, but they can never be exact since the information does not cover all households in a country. Samples carry a margin of error, and so do the poverty measures calculated from household surveys. The standard errors, which most statistical packages will easily calculate, depend on the sample design—stratification and clustering, essentially—and the sample size in relationship to the size of the total population (see Deaton 1997 and Ravallion 1992 for a description of the standard errors of various poverty measures). When the standard errors of poverty measures are large, it may well be that small changes or differences in poverty, although observed, are not statistically significant, and thereby cannot be interpreted for policy purposes. Tstatistics: When carrying out multivariate regressions, it is also important to compute the Tstatistics or standard errors, which inform on the degree of significance of the various coefficients. It might be the case that the coefficient on a specific variable is large, but that it is not significantly different from zero. Attention should be paid to these significance levels when interpreting the results. Sensitivity analysis: Apart from taking standard errors into account when comparing poverty measures between groups or over time, it is important to establish the robustness of the poverty comparisons to the assumptions made by the analyst. This may call for repeating the analysis for alternative definitions of the income aggregate and alternative ways of setting the poverty line. The sensitivity analysis may for example focus on the impact of changes in the construction of the income or consumption aggregate when imputations for missing values or corrections for underreporting of income in the surveys are implemented. Alternatively, one can test results with various lines, say the base poverty line plus and minus 5 percent in monetary value. Tests can also be conducted for checking the sensitivity of poverty comparisons to the assumptions regarding economies of scale and equivalence scales within households.
Stochastic dominance: Profiles allow a ranking of various household groups (or various time periods) in terms of their level of poverty. However, it is important to test whether the ranking is robust to the choice of the poverty line. This leads to a special type of robustness test, referred to as stochastic dominance, which deals with the sensitivity of the ranking of poverty levels between groups or between periods of time to the use of different poverty lines. The simplest way to do this (for the robustness of poverty comparisons based on the headcount index of poverty) is to plot the cumulative distribution of income for two household groups or two periods of time. One can then see whether the curves intersect. If they do not intersect, then the group with the highest curve is poorer than the other group. If they do intersect, then for all poverty line below intersection, one group is poorer and for all poverty lines above the intersection, the other group is poorer. See also the The Global Poverty Numbers Debate for a discussion of the recent debate over global poverty estimates.
