Bayesian Multiple Testing for Suicide Risk in Pharmacoepidemiology: Leveraging Co-Prescription Patterns

2026-01-25

AI Summary

Current pharmacovigilance studies assessing associations between drugs and suicide risk are often limited to single-drug analyses and employ overly conservative multiple testing corrections, which may obscure true safety signals. This work proposes a unified Bayesian spike-and-slab framework that innovatively incorporates pharmacological relatedness into a covariance prior derived from a co-prescription network of 150 million patients. This approach enables network-informed variable selection and Bayesian false discovery rate control in high-dimensional settings with rare events. The method enhances statistical power while successfully replicating known risk associations—such as alprazolam—and protective effects, including mirtazapine. Moreover, it identifies novel high-risk opioid combinations and several folate-related agents with potential protective effects, generating actionable hypotheses for clinical decision-making and regulatory evaluation.

Technology Category

• Reasoning under Uncertainty: Causality
• Cognitive Modeling & Cognitive Systems: Neural Spike Coding
• Search and Optimization: Mixed Discrete/Continuous Search

Application Category

• Graph Algorithms and Modeling for the Web: Representation, reconstruction, and subgraph or motif discovery in Web-related graphs
• Search and Retrieval-Augmented AI: Web evaluation methodologies and metrics
• User Modeling, Personalization and Recommendation: Psychology-informed user models and recommender systems

Abstract

Suicide is the tenth leading cause of death in the United States, yet evidence on medication-related risk or protection remains limited. Most post-marketing studies examine one drug class at a time or rely on empirical-Bayes shrinkage with conservative multiplicity corrections, sacrificing power to detect clinically meaningful signals. We introduce a unified Bayesian spike-and-slab framework that advances both applied suicide research and statistical methodology. Substantively, we screen 922 prescription drugs across 150 million patients in U.S. commercial claims (2003 to 2014), leveraging real-world co-prescription patterns to inform a covariance prior that adaptively borrows strength across pharmacologically related agents. Statistically, the model couples this structured prior with Bayesian false-discovery-rate control, illustrating how network-guided variable selection can improve rare-event surveillance in high dimensions. Relative to the seminal empirical-Bayes analysis of Gibbons et al. (2019), our approach reconfirms the key harmful (e.g., alprazolam, hydrocodone) and protective (e.g., mirtazapine, folic acid) signals while revealing additional associations, such as a high-risk opioid combination and several folate-linked agents with potential preventive benefit that had been overlooked. A focused re-analysis of 18 antidepressants shows how alternative co-prescription metrics modulate effect estimates, shedding light on competitive versus complementary prescribing. These findings generate actionable hypotheses for clinicians and regulators and showcase the value of structured Bayesian modeling in pharmacovigilance.

Problem

• suicide risk
• pharmacoepidemiology
• multiple testing
• drug safety
• pharmacovigilance

Innovation

• Bayesian spike-and-slab
• co-prescription patterns
• false discovery rate control
• pharmacovigilance
• structured prior