Propensity Score Methodology for Nonignorable Nonresponse

When data are not missing at random, approaches to reduce nonresponse bias include subsampling nonresponding units and modeling. The objective of this thesis is to develop unbiased and precise model-assisted estimators of the population total that are applicable to data from a complex survey design with nonignorable nonresponse. When information from a nonrespondent subsample is available, weighting methods for missing-at-random data may be modified to reduce bias from nonignorable missingness in estimates of population totals. Propensity score methodology for nonignorable missingness is developed for use with the weighting class adjustment and with the Horvitz-Thompson estimator to account for the dependence between the outcome of interest and the response mechanism. The novel propensity score techniques for nonignorable nonresponse are applied to a binary outcome subject to nonignorable missingness from a complex survey of elk hunters and are also examined with simulation.