Weighting Adjustments for Panel Nonresponse

• Activity Number: 87
• Type: Contributed
• Date/Time: Sunday, August 4, 2013 : 4:00 PM to 5:50 PM
• Sponsor: Survey Research Methods Section
• Author(s): Qixuan Chen*+ and Andrew Gelman and Melissa Tracy and Fran Norris and Sandro Galea
• Companies: Columbia University and Columbia University and Department of Epidemiology, Columbia University and Department of Psychiatry, Dartmouth Medical School and Department of Epidemiology, Columbia University
• Keywords: adjustment cells ; design variables ; panel surveys ; response propensity

Abstract

Weighting adjustment for panel nonresponse needs to incorporate information about nonrespondents collected in the earlier waves of the panel. We propose a cross-classified method for panel surveys with complex sampling designs by first grouping respondents and nonrespondents with similar estimated response propensities to form response propensity strata and then cross-classifying the propensity strata with design variables. Simulation shows that when design variables are not related to nonresponse, the cross-classified method yields survey estimates that have bias and root mean squared error similar to the estimates weighted by reciprocals of the response propensities and the response propensity stratification method. When design variables are related to nonresponse, the cross-classified method yields estimates with smaller bias than the other two methods if design variables are not included as covariates in the response propensity regression, but is comparable in the bias if the response propensity model is correctly specified. We apply these methods to the Galveston Bay Recovery Study, a panel study of trajectories of wellness in a community following a disaster.