Abstract

The paper uses a simulation study to examine the impact of different model selection strategies on the outcomes in responsive and adaptive survey designs. In responsive and adaptive survey design, the designs are altered for subgroups in the sample (either before or during data collection) with the goal of improving the quality of survey data. In this study, the models are predicting nonresponse. The model predictions are used to allocate effort. Four model selection strategies are included in the study. First, selecting predictors of nonresponse drawn from paradata (e.g. number of contact attempts). The predictions from models of this type are used to maximize response rates. Second, models are selected using predictors that are highly correlated with the key survey variables. Models of this type create predictions that are used to limit nonresponse bias. Third, a compromise involving predictions from both of the previous models. Fourth, an intercept-only model is used as a control. This model allocates equal effort across cases. Reported outcomes in the study include response rates, number of interviews, and bias of key estimates.