Propensity Score-Based Approaches in High Dimension for Pharmacovigilance Signal Detection: an Empirical Comparison on the French Spontaneous Reporting Database.

Abstract

Classical methods used for signal detection in pharmacovigilance rely on disproportionality analysis of counts aggregating spontaneous reports of a given adverse drug reaction. In recent years, alternative methods have been proposed to analyze individual spontaneous reports such as penalized multiple logistic regression approaches. These approaches address some well-known biases resulting from disproportionality methods. However, while penalization accounts for computational constraints due to high-dimensional data, it raises the issue of determining the regularization parameter and eventually that of an error-controlling decision rule. We present a new automated signal detection strategy for pharmacovigilance systems, based on propensity scores (PS) in high dimension. PSs are increasingly used to assess a given association with high-dimensional observational healthcare databases in accounting for confusion bias. Our main aim was to develop a method having the same advantages as multiple regression approaches in dealing with bias, while relying on the statistical multiple comparison framework as regards decision thresholds, by considering false discovery rate (FDR)-based decision rules. We investigate four PS estimation methods in high dimension: a gradient tree boosting (GTB) algorithm from machine-learning and three variable selection algorithms. For each (drug, adverse event) pair, the PS is then applied as adjustment covariate or by using two kinds of weighting: inverse proportional treatment weighting and matching weights. The different versions of the new approach were compared to a univariate approach, which is a disproportionality method, and to two penalized multiple logistic regression approaches, directly applied on spontaneous reporting data. Performance was assessed through an empirical comparative study conducted on a reference signal set in the French national pharmacovigilance database (2000–2016) that was recently proposed for drug-induced liver injury. Multiple regression approaches performed better in detecting true positives and false positives. Nonetheless, the performances of the PS-based methods using matching weights was very similar to that of multiple regression and better than with the univariate approach. In addition to being able to control FDR statistical errors, the proposed PS-based strategy is an interesting alternative to multiple regression approaches.