A comparison of methods to approximate standard errors for complex survey data.

Abstract

Complex survey designs are characterized by multistage selections with stratification and clustering. The departure from simple random sampling assumptions requires special consideration with regard to variance estimation. Specially designed software packages exist to generate variance estimates for statistics from complex survey data. The variance estimation techniques used include balanced repeated replication, jackknife, and Taylor series linearization. Many complex surveys generate thousands of tables. The computational and publishing costs soar if estimates of standard error are required for all statistics. To decrease these costs, several alternative techniques to approximate the standard errors of estimates are available. These include the widely used relative variance curve, a method based on the average relative standard error, and the average design effect model. In this paper these three methods are compared with respect to accuracy, computational and publishing costs, and ease of implementation.