Parametric and Nonparametric Empirical Bayes Techniques Applied to Post-Marketing Vaccine Safety Surveillance

• Activity Number: 139
• Type: Contributed
• Date/Time: Monday, July 30, 2012 : 8:30 AM to 10:20 AM
• Sponsor: Section on Statistics in Epidemiology
• Author(s): Brock Stewart*+ and Michael M McNeil and Frank DeStefano
• Companies: CDC and CDC and CDC
• Keywords: Vaccine Safety ; Empirical Bayesian ; Statistical Learning ; Data Mining ; Drug Safety

Abstract

Prior to licensure, the safety of new vaccines are evaluated in clinical trials, which are typically not powered to detect rare adverse events (AE). Monitoring after a vaccine is licensed and more people are exposed may be more likely to detect rare AE. The Vaccine Adverse Event Reporting System (VAERS), co-managed by CDC and FDA, is the frontline spontaneous reporting system for potential vaccine-AE associations in the post-marketing setting. Established in 1990, VAERS contains 337,000+ reports, receiving more than 30,000 reports annually in recent years. Each VAERS report may contain multiple vaccines and/or AE: 79%, 38%, and 31% contain multiple AE, vaccines, or both, respectively. Currently, FDA monitors VAERS reports with the aid of a gamma-Poisson empirical Bayesian model with a mixture of two gammas as the prior. A signal is generated when a vaccine-AE pair occurs in substantially more reports than would be expected under independence. We contrast various configurations of gamma-Poisson and beta-binomial empirical Bayesian models, mostly with regards to false discovery rate, which is an important consideration when conducting statistical learning on large databases.