How Small Can You Go? Combining Census and Survey Data for Mapping Poverty in South Africa

1 Further information on this grant can be obtained from the South Africa local government website at - http://www.local.gov.za/DCD/dcdindex.html.

2 Hentschel et al. (1999) state that: “In fact, a poverty map would have to be constructed at quite a high degree of spatial disaggregation before the standard errors increase significantly due to small populations. … Only when the [local] population falls well below 500 households does the corresponding standard error rise to levels which could compromise comparisons.”

3 We focus here on the best means of measuring income or consumption poverty and abstract from the debate on what other measures of household welfare add to a multi-dimensional understanding of poverty. See, Ravallion (1992) for further discussion on the measurement of poverty.

4 The sample was stratified by province, urban and non-urban areas, and population group.

5 Recent studies have indicated that the poverty ranking of households is sensitive to assumptions regarding the degree that households have scale economies as well as whether adult equivalency scales are assumed for children (Lanjouw, Milanovic and Paternostro, 1999). However, we do not address this possibility in the current study.

6 Note that the first figure is household poverty, while the latter is individual poverty. I.e., 28.8% of the households in South Africa have a monthly household income of less than 800 Rands, whereas 48.4% of the individuals live in households with monthly per capita income of less than R250.

7 We discuss the last four columns of Table 3 as well as Table 4-6 after the methodology for imputing expenditures is presented.

8 We also explored specifications which included either the number of households in the district or the square root of this number to see if smaller MDs had measurably greater deviation between the census and the IES data. The coefficient of cluster size were generally significant at the 10% level or less and with a sign consistent with the expectation that precision increased with the size of the cluster. However, neither the regression r-square values or the magnitude of the coefficient of other variables were affect by the inclusion of the cluster size. Thus, the regressions reported in the table do not included the number of households.

9 If we look at the correlation of average income from the IES and average expenditures from that survey, we find that at the province the correlations is 0.99. At the MD level the correlation is 0.96. For both levels the rank correlations are above 0.93.

10 The methodology employed here of calculating headcount indices from the imputed expenditures in the census is based on Hentschel et al. (1999).. More details can be found in that paper.

11 Data for each magisterial district as well as for town and place names are available from Statistics South Africa.