Adaptive Survey Design to Reduce Nonresponse Bias.

Abstract

This dissertation is concerned with methods for measuring and dealing with nonresponse bias. A new measure for the risk of nonresponse bias is proposed – the fraction of missing information – as an alternative to the response rate. This measure was developed as part of methods for handling missing data. It measures our uncertainty about the values we would impute for current nonresponders. Under this guiding indicator, the goal of data collection would be to maximize the information in the final dataset. There is a need for new research methods for identifying how to obtain response from cases that have the maximum impact on this measure. Methods developed for clinical trials under the rubric of dynamic treatment regimes are proposed and implemented using observational data. These methods tailor the treatment to the characteristics of the patient, including the history of previous treatments. In the survey context, this means tailoring the survey design to the characteristics of the case, including the history of previous attempts to obtain an interview. Finally, a new rule for stopping data collection is proposed that attempts to account for the uncertainty due to nonresponse.