Information anchored sensitivity analysis for randomised controlled trials via Multiple Imputation
Suzie Cro (Imperial College London)
Abstract:
Controlled Multiple Imputation procedures have been proposed for contextually relevant accessible sensitivity analysis of clinical trials with missing data. The so called `δ- method’ enables trialists to explore the impact of unobserved patients having a poorer/better response than those observed. Reference based imputation procedures allow us to explore the impact of unobserved patients behaving like a specified reference group, typically control/placebo. Such sensitivity analysis should not inject information ‘by the back door.’ Neither do we want to lose value information collected in the trial. In this presentation we argue that sensitivity analysis of the type proposed should preserve the information loss due to missing data in the primary analysis. We refer to this as the information anchoring principle. We present theoretical and simulation results which show how an information anchored variance estimate is obtained for the treatment estimator when using the discussed methods. This provides a solid justification for their practical use, which we illustrate with the analysis of a chronic asthma trial.