External Comparator Studies: Performance of Four Missing Data-Handling Approaches, Stratified by Four Different Marginal Estimators.

BACKGROUND AND

objectiveMissing data and unmeasured confounding may bias results of external comparator (EC) studies. Previous research quantified these effects, but there were still knowledge gaps in terms of studying a broader set of missing data-handling approaches. This knowledge gap is addressed by investigating four different ways to handle missing data for a set of four distinct marginal estimators.

methodsAn extensive simulation study was conducted based on two real EC case studies. Four different variants of missing data-handling approaches were assessed in terms of bias and other performance characteristics. Specifically, multiple imputation (MI) for the trial and EC cohorts was conducted by applying within-cohort MI, across-cohort MI and a mixed within-across-cohort MI scheme. Dropping a covariate from the analysis model if missingness exceeded a certain threshold was also added as an analysis strategy. All simulation results were generated for a set of four marginal estimators: the average treatment effect of the untreated (ATU), the average treatment effect (ATE), the average treatment effect of the treated (ATT), and the average treatment effect in the overlap population (ATO). Missingness was simulated to occur only in the EC cohort, and propensity score weighting was applied as causal inference method.

resultsOverall, within-cohort MI and the ATU showed best performance in terms of mitigating bias, while the strategy of leaving out prognostic factors (covariates) due to a higher percentage of missingness performed worst.

conclusionsPerformances of four missing data-handling strategies were assessed for a set of four different marginal estimators. Our results add clarity with regard to potential residual bias for researchers conducting EC studies when using propensity score weighting in the case of missing data or unmeasured confounding. This enables researchers to select most appropriate statistical approaches to minimise bias, potentially by including an additional bias estimation and correction step.