Enhancing Clinical Trial Diversity and Representation Through Data-Adaptive Targeted Learning: A Tale of Lurking Berkson Bias

OBJECTIVES: There is a ubiquitous need for enhanced diversity in clinical trials for generalizability. Diverse trial populations allow for more accurate predictions of the safety and efficacy of potential medicines with implications for equitable treatments. Further, the real-world data (RWD) collection process may result in biased estimates because of structural differences in access to healthcare, socioeconomic status, and cultural environments. Additionally, many traditional analytic methods can restrict cohorts and disproportionately exclude patient groups. Applying causal structural reasoning and data-adaptive targeted learning (TL) provides an efficient opportunity to bolster diversity and inclusion in trial analysis.

METHODS: We generated synthetic data from clinically-informed directed-acyclic graphs (DAGs). The synthetic data included a potential collider (C), which, when conditioning on C, gives rise to Berkson (collider) bias. We compare estimates to' true' effects for varying sampling strategies and analytic methods, e.g., propensity weighting and targeted maximum likelihood estimates (tmle).

RESULTS: We observe the accuracy of effect, i.e., average treatment effect (ATE), estimation is remarkably sensitive to the sampling strategies based on the collider variable. Plots of sampling dependent on colliders show increasing bias introduced through selection strategy and remedied through correct analytic adjustment.

CONCLUSIONS: A patient's presentation to the healthcare system, and subsequent availability of healthcare records, might introduce bias when estimating relationships. To preserve a diverse and representative cohort, causal structural reasoning and appropriate targeted learning methods can provide efficient and accurate unbiased estimates while retaining representative patient composition. We note the relationship between causal considerations and the analytic approach and note bias is not automatically addressed through any analytic method and must reflect the context of health delivery and patient identification. These methods and strategies will help provide better insights into the potential impact of medicines on patients and drive more optimal treatment solutions.