|
Abstract:
|
We consider the role of multiple imputation (MI) when analyzing non-inferiority clinical trials with missing data. We focus on the situation in which the binary endpoint (Z) is measured only at the final analysis but other variables predicting Z are measured prior to final analysis. We present simulation results for various scenarios. Among methods that control the type I error, the preferred method will produce unbiased parameter estimates under both the null and alternative hypotheses, and high power under the alternative. As in superiority trials, MI will have most benefit when the important predictor variables are included in the imputation model and strongly associated with the unobserved outcome. Unbiased estimation of the variance of the MI estimator is also important as it affects both the power and type I error rate. In our simulations, MI estimation may be highly biased when data are missing not at random, and an imputation model that includes baseline predictors, early outcomes and treatment group produces the best analysis for data that are missing at random. Situations in which the type I error rate is controlled but parameter estimates are biased are discussed.
|
