Issues Involved in the Modeling of Complex Survey Data

• Activity Number: 146
• Type: Luncheons
• Date/Time: Monday, August 4, 2003 : 12:30 PM to 1:50 PM
• Sponsor: Section on Survey Research Methods
• Author(s): Daniel Pfeffermann*+
• Companies: The Hebrew University of Jerusalem

Abstract

The modeling of complex survey data is different and often much harder than in other areas of statistics. This is due to the complicated nature of the finite populations from which the samples are taken and the special characteristics of the sampling designs used to select the samples, typically involving multistage cluster sampling and unequal selection probabilities that might be related to the values of the target outcome variables. The following issues will be discussed: definition of parameters of interest: should they be defined in terms of the super-population model or in terms of the corresponding finite population quantities?; incorporation of sampling weights: should the sampling weights be used when making inference on models as opposed to the prediction of finite population totals? How can they best be used?; definition of variances: should the variance account for the variability induced by all possible sample selections, the variability of the outcome values given the selected sample, both sources of variance?; Model identification and diagnostics: should one use different diagnostics from the ones used in "classical statistics"? Which specific diagnostics are available?