Accounting for Survey Design in Multilevel Models

James Brown, University of Southampton
Nyovani Madise, University of Southampton
David Steel, University of Wollongong

Secondary data analysts are becoming more aware of the impact a complex survey design can have on model estimation of parameters and their associated standard errors in linear models. Using STATA allows the analyst to account for weighting, clustering, and stratification in model estimation with correctly estimated standard errors. The problem with this approach occurs when the impact of clustering is of substantive interest to the data analyst and the preferred method of analysis is multilevel modeling. Pfeffermann et al. (1998) propose a weighting approach that has been implemented in MLwiN. We investigate this approach through a small simulation study. Comparisons of the results with simple unweighted analysis, and a model-based approach to account for the design, demonstrate the importance of not ignoring the sample design when it is informative. Data from the Tanzania DHS and Ghana DHS are then analyzed to illustrate the above approaches.