OBJECTIVES

Tree-based scan statistics (TBSS) appropriately adjust p-values for multiple-testing of correlated hypotheses for signal detection studies. It is unclear what to include in a propensity score (PS)-matched cohort design to adjust for confounding across thousands of potential outcomes being screened. We developed and evaluated a general PS for drug safety signal detection with different active-comparator cohorts.

METHODS

We selected 3 drug pairs with established safety profiles where few, if any signals were expected. We evaluated 5 candidate PS that included different combinations of: predefined general covariates (demographics, comorbidity, frailty, utilization), empirically-identified covariates (via high-dimensional PS algorithm) and covariates tailored to the drug pair. We identified 1:1 matched cohorts in MarketScan data using each PS, ran TBSS, ranked potential adverse events by log likelihood ratio (LLR) and set a threshold for alerting of p ≤ 0.01.

RESULTS

For each drug pair (N >290,000 each), the top ranked outcomes were similar across PS-matched cohorts. There were ≤2 unique alerts observed in each example. Outcomes that met the threshold for alerting were expected or explainable. For example, when comparing macrolides versus fluoroquinolones, the top ranked outcome for each PS-matched cohort was hypertension complicating pregnancy. This was explainable by channeling of azithromycin (a macrolide) to pregnant women. Pregnancy was not a covariate in the predefined PS; the outcome “hypertension in pregnancy” had p ≤ 0.01. When pregnancy was included as empirical and/or tailored covariates, p-values were between 0.13-0.43. However, fewer pregnant patients were matched, achieving better balance but reduced power.

CONCLUSIONS

Including empirical covariates may provide better proxy coverage of confounders for numerous outcomes than predefined covariates alone, but could increase variance. Other than pregnancy, covariates tailored to exposure did not appreciably impact results in our examples. Potential signals should be followed up with pharmacoepidemiologic assessment where confounding control is tailored to the specific outcome(s) under investigation.