Missing Data in Clinical Trial: A Critical Look at the Proportionality of MNAR and MAR Assumptions for Multiple Imputation

Randomized control trial is a gold standard of research studies. Randomization helps reduce bias and infer causality. One constraint of these studies is that it depends on participants to obtain the desired data. Whatever the researcher can do, there is a possibility to end up with incomplete data. The problem is more relevant in clinical trials when missing data can be related to the condition under study. The benefits of randomization is compromised by missing data. Multiple imputation is a valid method of treating missing data under the assumption of MAR. Unfortunately this is an unverified assumptions. Current practice advise the use of sensitivity analysis to adjust for departure from the MAR missingness. Data collectors’ knowledge, researchers’ insight, and statisticians’ experience can improve assumptions of missing data mechanisms. In practice, a mixture of possible assumptions can be made about missingness. In an attempt to exploit supplemental knowledge for the amelioration of inference from data with missing values, this dissertation explores the possibility of combining various proportion of MAR and MNAR assumptions. This exploration will be done by simulating data having normal, chi-square, and t distributions with varying proportion of MAR and MNAR assumptions. We propose influential exponential tilting in which the model for the non-respondents correspond to an exponential tilting of the model for respondents, and the specified function in the tilted model is the influential function of the parameter to be estimated. The proposed method will be combined with MI to overcome the issue of MNAR