Accuracy versus convenience: A simulation-based comparison of two continuous imputation models for incomplete ordinal longitudinal clinical trials data

Multiple imputation has become an increasingly utilized principled tool in dealing with incomplete data in recent years, and reasons for its popularity are well documented. In this work, we compare the performances of two continuous imputation models via simulated examples that mimic the characteristics of a real data set from psychiatric research. The two imputation approaches under consideration are based on multivariate normality and linear-mixed effects models. Our research goal is oriented towards identifying the relative performances of these methods in the context of continuous as well as ordinalized versions of a clinical trials data set in a longitudinal setting. Our results appear to be only marginally different across these two methods, which motivates our recommendation that practitioners who are not computationally sophisticated enough to utilize more appropriate imputation techniques, may resort to simpler normal imputation method under ignorability when the fraction of missing information is relatively small.