Complex survey data analysis: Estimation of standard errors using pseudostrata

A proper analysis of data from complex sample surveys requires special consideration for estimating standard errors. Special techniques and software packages are available, including Taylor series linearization (delta method), balanced repeated replication, and jackknife. Before their use, it is often necessary to make certain modifications in original data structure, to conform to computing method requirements. The most common modification is to form pseudostrata by collapsing substrata or partitioning a string of geographic clusters. This paper examines the performance of the delta method when it is applied to a complex community survey data set in which sequentially drawn clusters of households are partitioned to form pseudostrata. Standard errors of rates, regression coefficients, and odds ratios are compared with those computed from the variation of replicates built into the sample design. The results demonstrate that an analysis of complex survey data should use an appropriate method for estimating standard errors, arid that pseudostrata would produce reasonable estimates of standard errors for rates and regression coefficients, with mixed results for odds ratios.