The Role of Multiple Imputation in Noninferiority Trials for Binary Outcomes

We consider the role of multiple imputation (MI) when analyzing noninferiority (NI) clinical trials with missing data. When the endpoint is measured longitudinally, direct-likelihood methods can be used. In this article, the focus is on the situation in which the endpoint is not measured longitudinally but other relevant data are measured at or after baseline prior to planned collection of the primary endpoint data. Simulation results are presented for various scenarios based on the missingness mechanism, the dropout rate, and the size of NI margin. When the endpoint is binary, the ratio of the amount of missing data to the noninferiority margin will affect the operating characteristics of any analysis strategy (whether imputation based or not), an issue that is unique to noninferiority trials. Biased estimates of treatment effect under missingness, not completely at random, may arise when using a misspecified imputation model lacking treatment effect, resulting in substantially inflated Type I error rates in noninferiority trials by making the two groups appear more similar, opposite the usual impact in superiority trials. As in superiority trials, MI will have most benefit when data are missing at random, and the important predictor variables are included in the imputation model.