A Comparison of Weighting Adjustment Methods for Nonresponse

• Activity Number: 242
• Type: Contributed
• Date/Time: Monday, August 4, 2014 : 2:00 PM to 3:50 PM
• Sponsor: Survey Research Methods Section
• Author(s): R. Lee Harding*+ and Ronaldo Iachan and Kurt Peters
• Companies: ICF International and ICF International and ICF International

Abstract

The adjustment of survey data for non-response is typically based on weighting class methods. Weighting classes are formed using a set of core variables that are correlated with response behavior and with survey outcomes. The underlying principle is that respondents are more alike within a class than across classes. The choice of variables may be made using response propensity models (i.e., logistic regression models for the response indicator), or with a range of recursive partitioning, tree-based classification methods (e.g., CHAID or CART). Propensity models may also be used more directly to generate propensity scores that are applied to adjust for response probabilities. This research compares these methods using real and simulated data with origins in two kinds of multistage stratified sample surveys: samples of students within schools, and samples of patients within facilities. Comparisons are made along bias, variance and mean squared error of key survey estimates.