Abstract:

In a typical twoarm (test, reference) randomized clinical trial, the endpoint of interest (e.g., change from baseline in HAMD17 at 6 weeks) is not observed for dropouts. The resulting missing data problem is commonly tackled by invoking a missing at random (MAR) assumption and proceeding with a mixed model repeated measures (MMRM) analysis. If the MAR assumption is incorrect (it usually is), the estimated betweentreatment difference in endpoint means can be notably biased for the estimand of interest. We will discuss biasreducing methods in which the implicitly imputed mean for testarm dropouts in the MMRM analysis is explicitly replaced with the estimated mean for either all referencearm patients or referencearm dropouts only. The socalled jumptoreference (J2R) method involving patientlevel imputation will also discussed. All three referencebased imputation approaches will be contrasted with a "dropout=failure" approach in which an extreme "bad" outlier is imputed for all the dropouts followed by quantile regressionbased quantile averaging and a nonparametric bootstrap for inference. Two real datasets and simulations will be used to reinforce the key points.
