Finite-sample efficiency gain of A-TMLE when augmenting RCTs with real-world data

Ascertain the magnitude of finite-sample efficiency gains achievable by the adaptive targeted minimum loss-based estimation (A-TMLE) estimator for the average treatment effect when integrating randomized controlled trial data with real-world data, and develop methods to quantify or predict these gains under practical conditions, including dependence on the learned bias working model’s complexity and sample sizes.

Background

The paper proposes A-TMLE for estimating the average treatment effect by combining randomized controlled trial data with real-world data, without imposing mean exchangeability across studies. A-TMLE is shown to be asymptotically normal and can be super-efficient relative to estimators that ignore external data, with efficiency gains driven by the complexity of the data-adaptively learned bias working model.

While simulations demonstrate notable improvements in mean-squared-error and confidence interval width in some scenarios, the authors note that practical expectations for efficiency gains can vary across settings. They explicitly state uncertainty regarding how much efficiency gain should be expected in finite samples, motivating the need for research to characterize and predict such gains.