Causal inference tools for pharmacovigilance: using causal graphs to systematize biases, plan disproportionality analyses, and reduce the risk of spin.

Introduction: Disproportionality analyses are used to detect signals of potential adverse drug reactions. Because they are prone to biases and the risk of overstating results, they are relegated to exploratory hypothesis generation. Directed Acyclic Graphs (DAGs) formalize mechanisms and assumptions, and identify biases; thus, identifying which kinds of inferences could be reliably made, which additional assumptions would be required, and which studies could better identify causal effects.Objective: To explore possible applications of DAGs in disproportionality analysis, so to facilitate identification of transparent assumptions, discussion of causal inference limitations, and design of more effective studies. The aim is to enhance the reliability of conclusions drawn from disproportionality analyses and mitigate the risk of misinterpretation. Methods: We introduce a DAG-based causal framework to coherently systematize the biases inherent in disproportionality analyses (e.g. confounding, selection, and reporting biases). Results: Exemplary case studies from the FDA Adverse Event Reporting System demonstrate the usefulness of DAGs in adjusting for confounders, avoiding the introduction of new biases, and assessing the direction of residual confounding in disproportionality analysis. Conclusion: DAGs can be used both to increase sensitivity in signal detection and specificity in signal validation. Systematic bias evaluation and mitigation using DAGs achieves more reliable and well-founded safety signals, reducing and mapping the gap between what we find (association) and what we look for (causation). Further research is necessary to develop DAGs to meet pharmacovigilance specific challenges, and good workflows for evidence synthesis better integrating epidemiological and pharmacovigilance information sources under a common causal inference framework.