Regression Tree Models for Analyzing Survey Response

Daniell Toth and Polly A. Phipps

Abstract

Modeling various conditional response propensities for a survey based on known unit characteristics or contact history is important when analyzing survey response. One increasingly important technique is to model these conditional response propensities using non-parametric regression tree models. Regression trees provide mutually exclusive cells with homogeneous response propensities that make it easier to identify interpretable associations between class membership characteristics and response propensity, compared to other regression models. We provide examples of how regression trees have been used to analyze survey response, including: gaining insight into how characteristics of sample members are associated with response, incorporation of auxiliary variables and para-data for use in adaptive designs and follow-up procedures, and the identification of auxiliary variables for nonresponse adjustment.