Relevant, accessible sensitivity analysis using multiple imputation
*James R. Carpenter, London School of Hygiene & Tropical Medicine 
Michael G. Kenward, London School of Hygiene & Tropical Medicine 
James R. Roger, Research Statistics Unit, GlaxoSmithKline 

Keywords: missing data, clinical trials, multiple imputation, sensitivity analysis

Missing data due to early withdrawal are unavoidable in clinical trials, and imply that assumptions are needed for the analysis that cannot be verified from the data at hand. To tackle this issue in a coherent way it is therefore necessary to examine carefully these assumptions in the light of the goal of the analysis.

To do this we define two questions:

De facto - 'What would be the effect seen in practice if this treatment was applied to the population defined by the trial inclusion criteria'

De jure - 'Does the treatment work under the best case scenario'

With reference to these two questions, we further define the concept of a deviation from the protocol relevant to the question, which we refer to as a 'deviation'.

We then describe a variety of approaches for framing post- deviation distributions of outcome, conditional on observed data. Once this has been done, we argue that multiple imputation provides a flexible practical tool for inference. We illustrate our approach with data from an asthma study.