![IconGems-Print](images/IconGems-Print.png)
338 – Analyzing Survey Data with Regression Trees
Using Classification and Regression Trees to Model Survey Nonresponse
Sharon Lohr
Westat
Valerie Hsu
Promontory Financial Group, Washington, District of Columbia
Jill Montaquila
Westat
In the computation of survey weights to be used for the analysis of complex sample survey data, an adjustment for nonresponse is often an important step in reducing bias. These adjustments depend upon estimated response propensities, which are traditionally obtained through empirical response rates within weighting classes or through logistic regression modeling. In this paper, we discuss possible benefits of using regression trees and random forests for estimating response propensities in surveys, and describe how these models might be used to reduce nonresponse bias. We review issues for their use with complex surveys such as the effect of survey weights and clustering, pruning criteria, and loss functions, and we explore the sensitivity of results to these conditions.