DescriptionUS Food and Drug Association (FDA) presented a draft guidance of E9(R1) Statistical Principles for Clinical Trials Addendum: Estimands and Sensitivity Analysis in Clinical Trials in June 2017. This draft guidance has been widely referenced during recent research discussions. The aim of the draft guidance was to clarify the concept of estimand and to connect estimand with the concept of trial objective. An emphasized discussion about the impact and handling methods of missing data was also addressed. The draft guidance introduced the concept of ‘intercurrent event’ to describe all events that would cause either potential missing data or discontinuation from initial randomized treatment assignment. Five intercurrent event handling strategies were proposed with each strategy targeting a specific estimand which eventually represents a trial objective. Some of these strategies would result in missing data problem that required additional assumptions regarding the missing mechanism. In this dissertation, I will propose an alternative procedure in terms of connecting the intercurrent event handling strategy with estimand specification. The proposed procedure can be considered as an event-type driven strategy that selects the desirable estimand based on not only primary trial objective but also potential intercurrent event types. In this dissertation, I will also discuss the importance of sensitivity analysis and the relationship between sensitivity analysis missing mechanism assumptions and primary analysis missing mechanism assumptions. Literature review will be focused on recent developments on the topic of sensitivity analysis methods, especially the reference based imputation (RBI) method and the δ-adjustment tipping point analysis method. The benefits and drawbacks of both methods will be discussed in detail. This dissertation will contain a proposal of a modified Mixed Model Repeated Measure (MMRM) model that targets the ‘De Facto’ estimand when rescue medication is offered in a randomized clinical trial. Primary estimator can be represented as a linear combination of this modified MMRM model parameters. Delta approximation method will be used to directly derive the inference of the estimator. The result and performance will be compared with the result using multiple imputation method. Secondly, I will propose an alternative sensitivity analysis method called ‘decay model tipping point analysis method’. The highlights of this method are as follows, 1) It is capable of covering all possible sensitivity scenarios, including but not limited to the ones studied using RBI method. 2) The adjusted missing data effect is associated with dropout time. Patients who dropped out early will be adjusted with a greater value comparing to those who dropped out later. The adjustment decreases for time points that are further away from the patient dropout time point. This is a more reasonable approach comparing to δ-adjustment tipping point method which adjusts the effect at each time point with a same constant. 3) The range of the adjustment can be set within the a clinical meaningful boundary. This will avoid the over-adjustment problem in δ-adjustment tipping point method. 4) The decay rate parameter serves as a unified sensitivity parameter. It can be compared between different studies as a measurement of robustness of primary analysis result in terms of the missing mechanism assumption. 5) The tipping point can be solved directly without iterative searching the total domain sensitivity parameter, therefore saving computing resource and power. Simulation studies will be conducted to verify the modified MMRM model. Inference derived based on delta approximation method will be verified using empirical inference result from simulation. A simulated study will be presented to verify the direct calculation for the tipping point and demonstrate the features of decay model tipping point method. In addition, a case study of using the decay model sensitivity analysis in a real world rare blood disease trial study will be presented. The possibility of extending the decay model beyond continuous endpoint will be briefly discussed and some technical issues occurred in the current research will be included in future research plan.