The Evolving Landscape of External Controls and Digital Twins: Current State and Future Directions

Randomized controlled trials (RCTs) provide internally valid estimates of treatment effects but are often costly and underpowered. In contrast, real-world data (RWD) offer large-scale, passively collected information that can improve efficiency when integrated appropriately. Recent methods have explored using predictive models trained on RWD—so-called digital twins—to generate individualized outcome predictions and augment RCT analysis. However, when the RWD and RCT populations differ, naively applying external models can induce model shift bias and efficiency loss, undermining the validity of causal conclusions.
We propose a new framework that combines RWD-based digital twin modeling with a calibration step using auxiliary outcomes. This calibration adjusts for systematic discrepancies between the trial and real-world populations, enabling valid and efficient treatment effect estimation in hybrid trial designs. Our approach generalizes classical covariate-adjusted regression and prediction-powered inference to settings with distributional shift between data sources. Theoretically, we show that the proposed estimator is consistent and achieves asymptotic variance no larger than that of the unadjusted baseline under a directional alignment condition. Notably, even if this condition fails to hold, the estimator remains asymptotically unbiased. Empirically, we demonstrate substantial gains in statistical efficiency and robustness through simulations and an application to multi-center neuroimaging data, while maintaining the internal validity of the original RCT.
 

Utilization of data from historical control animals to form virtual control groups (VCGs) is an innovative approach to embody the 3Rs (reduce, refine, and replace use of control animals) principle in research. However, there is no available systematic evaluation of statistical performance using VCGs in preclinical safety assessment. The optimal selection criteria and combination of VCGs and concurrent control group (CCG) also remain unclear. The VICT3R consortium on VCG controls sponsored by the Innovative Health Initiative (IHI) is currently working to fill in these gaps and refine the implementation of VCGs. This study evaluated the statistical ability as measured by sensitivity and specificity to detect test article effects for body weight and clinical pathology endpoints retrospectively in Pfizer's large animal toxicity studies using VCGs. In addition, exploratory analyses using Bayesian statistics (meta-analytic priors and hierarchical models) were conducted for Developmental and Reproductive Toxicology (DART) and Safety Pharmacology endpoints. Our results show that both VCGs and Bayesian statistics can preserve or improve the sensitivity of detecting toxicologically relevant effects. 

In this talk we will discuss the assumptions required for external data to improve inferences in clinical trials, and how different methods produce different ranges and magnitudes of benefit and risk. We will illustrate these assumptions through completed clinical trials that have resulted in regulatory approvals.