Abstract:
|
In this work we discuss handling of intercurrent events (ICEs) and missing data for longitudinal binary outcomes in clinical trials. As missing data may occur due to various reasons, it is important to differentiate and appropriately account for their distinct sources. To minimize missing data and increase the statistical efficiency, we propose to follow the estimand framework. For binary endpoints, one may employ the composite variable strategy, to define response status considering clinically relevant ICEs. For example, rescue and premature discontinuation due to lack of efficacy may be treated as treatment failure, hence non-response imputation (NRI) can be used to handle these ICEs. Additional missing data may be considered as missing at random and can be handled by generalized linear mixed model (GLMM) or multiple imputation based generalized estimating equation (MI-GEE). The proposed methods are referred to as NRI-GLMM and NRI-MI-GEE, respectively. We compare the proposed methods to GLMM and MI-GEE under the hypothetical estimand where data collected after ICEs are considered missing. Extensive simulations are conducted to evaluate the performance of the proposed methods.
|