Empirical Performance of LGPS and LEOPARD: Lessons for Developing a Risk Identification and Analysis System

Background The availability of large-scale observational healthcare data allows for the active monitoring of safety of drugs, but research is needed to determine which statistical methods are best suited for this task. Recently, the Longitudinal Gamma Poisson Shrinker (LGPS) and Longitudinal Evaluation of Observational Profiles of Adverse events Related to Drugs (LEOPARD) methods were developed specifically for this task. LGPS applies Bayesian shrinkage to an estimated incidence rate ratio, and LEOPARD aims to detect and discard associations due to protopathic bias. The operating characteristics of these methods still need to be determined. Objective Establish the operating characteristics of LGPS and LEOPARD for large scale observational analysis in drug safety. Research Design We empirically evaluated LGPS and LEOPARD in five real observational healthcare databases and six simulated datasets. We retrospectively studied the predictive accuracy of the methods when applied to a collection of 165 positive control and 234 negative control drug-outcome pairs across four outcomes: acute liver injury, acute myocardial infarction, acute kidney injury, and upper gastrointestinal bleeding. Results In contrast to earlier findings, we found that LGPS and LEOPARD provide weak discrimination between positive and negative controls, although the use of LEOPARD does lead to higher performance in this respect. Furthermore, the methods produce biased estimates and confidence intervals that have poor coverage properties. Conclusions For the four outcomes we examined, LGPS and LEOPARD may not be the designs of choice for risk identification.