An Optimal Hybrid Approach to Calculate Conditional Power

Abstract

When used in clinical trial adaptive designs, conditional power (CP) is most commonly calculated assuming the current trend, but this approach has been criticized for substantially deviating from the actual CP due to “double-dipping” the interim data (i.e., using the interim data as part of the final analysis test statistic and as the assumption for the remainder of the trial). By contrast, CP assuming the design assumption preserves the shape of the CP curve by “single-dipping” the interim data (i.e., using the interim data once to construct the final test statistic) but is less robust against misspecification. To improve CP estimation, we propose a “1.w dipping” hybrid approach by using an assumption that optimally combines the current trend assumption with weight w and the design assumption with weight 1-w. The optimization of w accounts for the uncertainty of the true treatment effect size and can be linked with study priorities through customizable penalty functions. Through simulation using sample size re-estimation designs, we demonstrate that the proposed approach maximizes the precision of CP-based interim decision making on average and leads to other desirable operating characteristics. Given its flexibility and superior performance, we recommend this novel approach over the existing methods.