Drug Adverse Event Surveillance Using the Multiple-Item Gamma Poisson Shrinker (MGPS)

• Activity Number: 161
• Type: Invited
• Date/Time: Tuesday, August 13, 2002 : 8:30 AM to 10:20 AM
• Sponsor: Section on Bayesian Stat. Sciences*
• Author(s): William DuMouchel*+
• Affiliation(s): AT&T Labs - Research
• Keywords: Data Mining ; Empirical Bayes ; Market Basket Problem

Abstract

MGPS is an empirical Bayesian method for identifying reliably large counts in a large sparse frequency table. The method has been used by FDA and other safety researchers to screen data bases of spontaneous adverse event reports for unusually frequent combinations of items (drugs and/or symptoms). We will present extensions of the methodology in two directions. First, improved estimates of the frequencies of itemsets with three or more items are achieved by shrinking towards the all-2-factor loglinear model rather than the less realistic independence model. Second, an extension of the model allows us to focus on detecting differences between itemset frequencies in different subsets of the data, or from one time period to another.