For many countries, in particular in sub-Saharan Africa, Demographic and Health Surveys (DHS) are the main national source of data (depending on the subject). Several DHS collect latitude and longitude of surveyed clusters but the sampling method is not appropriate to derive local estimates: sample size is not large enough for a direct spatial interpolation. We develop a methodological approach for estimating a proportion by using kernel density estimators with adaptive bandwidths of equal number of persons surveyed. The method was tested by creating a fictitious country from which survey datasets were produced. We compared the prevalence surface estimated from survey data with the model’s original prevalence surface. This method makes it possible to achieve a smoothing effect that adapts to the high irregularity of spatial distribution among the survey clusters. The surfaces thus generated are relatively accurate for densely populated areas and strongly smoothed in sparsely surveyed areas. Although local variations were filtered out, the regional component in the spatial variation of prevalence was reproduced, and the estimated prevalence surfaces could be interpreted as regional trend surfaces. Furthermore, this approach could be easily applied using prevR, a dedicated package for the statistical software R.
Larmarange Joseph (2013) “Mapping Demographic and Health Surveys (DHS): a method to estimate regional trends of a proportion (prevR)” (poster, session 176), presented at the XXVII IUSSP International Population Conference, Busan.