This thesis is a study of approaches to leveraging historical data for increasing power in randomised clinical trials (RCTs) with a continuously measured efficacy outcome. Existing methods based on populating the control arm with a synthetic control arm (SCA) fail to strictly control the type I error rate. Therefore, we focus on the novel statistical method of using digital twins (artificially generated patients receiving control medication) in an analysis of covariance (ANCOVA) model.

We show analytically that under certain assumptions, by ad- justing for the predicted outcome of a digital twin in an AN- COVA model, we obtain asymptotic efficiency of the average treatment effect estimator among a large class of estimators. This efficiency gain can then be used to decrease the sample size needed in a trial, while maintaining the same power and strictly controlling the type I error rate.
In a simulation study, we compare the performance of an existing SCA approach with the performance of the novel dig- ital twins approach in terms of power gain and type I error rate control. Under several scenarios, we find that the SCA approach provides at best only modest gains in power and is unreliable in terms of controlling the type I error rate. Conversely, the digital twins approach provides strict type I error control and a substantial increase in power, even when assumptions on analytical results are violated.

Lastly, we evaluate the method of digital twins in real world data originating from RCTs previously conducted at Novo Nordisk A/S. We find that the method manages to decrease the required number of subjects in a trial from 83 to 72, with possible improvements by further fine-tuning the method.