Re-Weighting the South African National Household Survey Data: A Cross Entropy Estimation Approach

Nicola Branson, University of Cape Town

In the absence of South African longitudinal data, national cross-sectional household survey data is frequently used to analyse change. When these data are stacked side-by-side however, inconsistencies both in time trends and between household and person level data are found. These inconsistencies can introduce bias into research. This study calculates a new set of person and household weights for ten years of South African data between 1995 and 2004. A cross entropy estimation approach is used. This approach is favoured because the calculated weights are similar to the initial sample weights (and hence retain the survey design benefits) and consistent with aggregate auxiliary data. The weights are benchmarked to a consistent series of aggregates from the ASSA model and the 1996 and 2001 South African Census data. The new weights result in consistent demographic and geographic trends over time and greater consistency between person and household level data.