Causal inference tools for pharmacovigilance: using causal graphs to identify and address biases in disproportionality analysis.

Introduction: Disproportionality analysis, finding associations in the co-reporting of drugs and events, is used in pharmacovigilance to detect signals of potential adverse drug reactions. Due to its susceptibility to biases, disproportionality analysis is generally relegated to hypothesis generation, and its integration into the broader evidence landscape is controversial. Objective: We showcase how Directed Acyclic Graphs (DAGs) can enhance disproportionality analysis for causal inference and simplify its inclusion in evidence synthesis. Methods: We introduce a DAG-based causal framework to systematically address biases in disproportionality analyses (e.g. confounding, collider, measurement errors, and reporting biases). We chose case studies from the FDA Adverse Event Reporting System to illustrate its application, using the Information Component as a disproportionality metric and restriction as conditioning. Results: DAGs enable the formalization of existing knowledge and causal assumptions, optimize the design of disproportionality analysis to mitigate biases – enhancing sensitivity against negative biases, and specificity against positive biases –, improve communication, and guide follow-up studies to address residual confounding. Under-characterization of reports makes so that restricting to reports mentioning the confounder is generally better than the opposite.Conclusion: Evaluating and mitigating biases with DAGs leads to more reliable and knowledge-based safety signals, reducing and mapping the gap between what we find (association) and what we look for (causation). Further research is needed to tailor DAGs to pharmacovigilance challenges, map factors underlying differential reporting bias, and develop workflows for better-integrating disproportionality analysis results into evidence synthesis.