The effects of cluster sampling in an African urban setting.

Cluster sampling was popularised by the sampling procedure promoted by the WHO/UNICEF for the evaluation of the expanded programme of immunisation (EPI). Without a clear understanding of the limitations of the sampling strategy used, this sampling strategy has been extended to other types of surveys. This article shows how to approach the assessment of cluster sampling techniques scientifically by calculating design effects (DEFFs) and rates of homogeneity (roh) and illustrates this scientific assessment with three case studies from Alexandra in South Africa. We report on the DEFFs and rohs for variables studied in these surveys. The DEFF for all the variables relating to housing tended to exceed two and was as high as 6.99 for the variable new development. The variables relating to health service utilisation and health practices, namely immunisation status, nutrition status, presence of Road to Health Cards (RTDCs), breast-feeding and knowledge of diarrhoea and oral rehydration all had a DEFF close to one. The variables relating to contraception use, literacy and schooling had DEFFs close to one and a half. For a few variables the DEFFs were below one and the rates of homogeneity less than zero. The highest values of roh were for environment factors (all above 0.1433). Rohs for factors related to utilization of PHC services were mostly between 0.0200 and 0.0499. No single class of factors seemed to be related to very low values of roh. These results are then discussed.