Sensitivity Analyses for Missing Not at Random Data in Body Donor Program Studies

Abstract

Missing data on socioeconomic variables, such as education and occupation, is a common issue in survey studies and can be Missing Not at Random (MNAR), where the likelihood of missingness depends on the unobserved value itself. Even small amounts of MNAR can introduce bias, compromising the validity of model outcomes. This study examines potential MNAR issues in the University of Wisconsin-Madison Body Donation Program’s Life Course Research, which predicted whole-body accep- tance rates using registrants’ demographic data, including incomplete education and occupation variables. To evaluate the robustness of the acceptance prediction model under the MNAR assumption, a three-stage sensitivity analysis framework was developed. The frame- work consists of (1) assigning ranks to categorical covariates based on model coeffi- cients, (2) imputing missing data under the Missing at Random (MAR) assumption, and (3) simulating MNAR scenarios by introducing deviations from MAR using a sen- sitivity parameter. Bootstrapping was applied to generate mean prediction accuracies and confidence intervals to assess MNAR’s impact on model robustness. The results indicated that the differences between the MNAR and MAR missingness mechanisms had minimal impact on the model, suggesting that the prediction model is robust under the MNAR assumption. This framework is practical and interpretable for as- sessing model robustness with categorical covariates and can be applied to enhance decision-making in other studies facing similar missing data issues.