On the current and future potential of simulations based on directed acyclic graphs.

Real-world data are playing an increasingly important role in regulatory decision making. Adequately addressing bias is of paramount importance in this context. Structural representations of bias using directed acyclic graphs (DAGs) provide a unified approach to conceptualize bias, distinguish between different types of bias, and identify ways to address bias. DAG-based data simulation further enhances the scope of this approach. Recently, DAGs have been used to demonstrate how missing eligibility information can compromise emulated target trial analysis, a cutting edge approach to estimate treatment effects using real-world data. The importance of simulation for methodological research has received substantial recognition in the past few years, and others have argued that simulating data based on DAGs can be especially helpful for understanding various epidemiological concepts. In the present work, we present two concrete examples of how simulations based on DAGs can be used to gain insights into issues commonly encountered in real-world analytics, i.e., regression modelling to address confounding bias, and the potential extent of selection bias. Increasing accessibility and extending the simulation algorithms of existing software to include longitudinal and time-to-event data are identified as priorities for further development. With such extensions, simulations based on DAGs would be an even more powerful tool to advance our understanding of the rapidly growing toolbox of real-world analytics.