Using Bayesian Predictive Distributions to Identify Outbreaks of Influenza-Like Illness

• Activity Number: 160
• Type: Topic Contributed
• Date/Time: Monday, August 2, 2010 : 10:30 AM to 12:20 PM
• Sponsor: Section on Statistics in Defense and National Security
• Author(s): Owen Devine*+ and Howard Burkom
• Companies: CDC and The Johns Hopkins University
• Keywords: Bayesian Methods ; Zero Inflated Poisson ; Influenza ; Over Dispersion ; Random Effects Models

Abstract

ILINet is a surveillance network comprised of volunteer health care providers who supply CDC with weekly reports on the proportion of patient encounters meeting influenza-like illness (ILI) criteria. CDC uses this system both to monitor trends in ILI and to identify potential influenza outbreaks. Due to substantial variation across weeks in both the number and specialties of the participating providers, frequentist approaches for outbreak identification can be highly unstable. We use Bayesian methods to generate predictive distributions of the ILI percentage during non-influenza periods enabling estimation of the probability of observing a surveyed rate if the area/week is in a non-influenza state. The models account for heterogeneity in the specialties of reporting providers over time and reflect the substantial over dispersion and excess zero counts common in ILI surveillance data.