Dealing with missing data in randomized and cluster randomized trials

This thesis investigated statistical methods for dealing with missing data in randomized controlled trials (RCTs) and cluster randomized trials (CRTs) by comparing the performance of various missing data methods under various missingness scenarios. This was done via simulation and formal proofs (where necessary), in terms of estimation of the treatment effect and its standard error, which were converted into various performance criteria. The results showed that for continuous outcomes in RCTs with missingness in a single covariate or in multiple covariates, all methods considered produced unbiased treatment effect estimates when the covariates were measured before randomization, regardless of the missingness mechanisms. For handling missing data in both the covariates and outcome in RCTs or in CRTs, bias of the treatment effect estimates for all methods depended on missingness of the outcome but not on missingness of the covariates. Further, in each scenario, there was at least one simple method that performed similarly to advanced methods such as multiple imputation and/or linear mixed effects models.