Sensitivity analysis for modeling nonignorable missingness in randomized controlled trials
*Mulugeta Gebregziabher, MUSC 

Keywords: nonignorable missing data, pattern mixture model, randomized trail, sensitivity analysis, shared

Missing data in randomized controlled trials are likely to be missing not at random (MNAR) and researchers should be careful interpreting evidence based on a single MNAR model. Analysis should be performed under a sensitivity paradigm in order to assess how conclusions vary as a function of assumptions made about models for the missing data and measurement processes. However, methods for MNAR missing data analysis have not been presented under a unified sensitivity framework and therefore no clearly defined guidelines for choosing one method over another exist. This paper proposes a unified sensitivity analysis for modeling MNAR longitudinal outcome data based on comparisons of selection or shared parameter models, pattern mixture models, and maximum likelihood. The approach is demonstrated on data from a stroke randomized control trial, and results are compared with those obtained from likelihood methods that assume missing data are ignorable. Two modeling scenarios for the measurement process (a generalized linear mixed model with random intercept or random intercept and slope) and two scenarios for the missingness process (a generalized linear or time to event model) are considered. The proposed unified approach provides evidence for or against suggested MNAR models by allowing for the influence of different missing data assumptions in the presence of model perturbations in the measurement or outcome process.