Data Mining for Adverse Drug Events With a Propensity Score-matched Tree-based Scan Statistic

Area of Science:

• Pharmacovigilance
• Statistical Data Mining
• Epidemiology

Background:

• Tree-based scan statistics are used for signal detection in vaccine safety but are limited in drug safety studies.
• Self-controlled designs are not always suitable for addressing complex drug safety questions.
• Confounding is a major challenge in evaluating drug-induced adverse events.

Purpose of the Study:

• To propose and evaluate a novel method combining tree-based scan statistics with propensity score-matched analysis for drug safety signal detection.
• To assess the performance of this combined method in new initiator cohorts.
• To compare the effectiveness of adjusted versus unadjusted analyses in controlling for confounding.

Main Methods:

• Plasmode simulations were conducted to evaluate the proposed method.
• Tree-based scan statistics were applied to propensity score-matched new initiator cohorts.
• Performance was assessed based on type 1 error control and statistical power under various confounding scenarios.

Main Results:

• Propensity score-matched tree-based scan statistics outperformed unmatched analyses in realistic scenarios.
• Adjusted analyses recovered prespecified type 1 error and preserved power, unlike unadjusted analyses.
• Even moderately mis-specified propensity score matching significantly improved performance over no adjustment.

Conclusions:

• Combining tree-based scan statistics with propensity score matching offers a robust approach for drug safety signal screening.
• This method shows promise for prioritizing potential adverse events for further investigation.
• Clinical review and targeted safety studies are essential follow-ups to quantify effect magnitude.