Abstract:
|
Rankings of cases based on estimated response propensities have been used to create inputs to adaptive survey designs. These inputs may be needed during data collection as triggers for design decisions. Cases above or below a certain threshold may receive a special recruitment protocol. However, Wagner and Hubbard (2014) showed that estimates of response propensity models can be biased when fit on a daily basis during data collection using the incoming data. These biases may lead to inaccurate ranking of cases, which, in turn, leads to inefficient or even counterproductive interventions. The use of informative priors in Bayesian logistic regression is explored. The goal is to identify a method for developing priors from other surveys and expert opinion that reduces or eliminates any potential biases in the rankings of cases.
|