Collaborative Research for Developing ATD: Bayesian Methods in Syndromic Surveillance: CAR Models and Computational Implementation

National Institute Of Statistical Sciences, Durham NC

Investigators

Abstract

The collaborative team develops and implements computationally of a Bayesian formulation of syndromic surveillance targeted at early detection of disease outbreaks. The research spans statistical theory and methodology, as well as issues of algorithms and computational science. Conditional autoregressive models, or CAR models, form the basis of the formulation, in order to accommodate spatial as well as temporal effects. The principal challenges are scalability (7,500 major U.S. hospitals are linked into the reporting network), multivariate reports, complex dependence structure in the data, low data quality, and incorporation of regional and other covariates. The unifying theme is principled calculation of uncertainties. The research is framed by an application of national significance. There is enormous social and economic benefit from early discovery of disease. The stakes are immense: the US Office of Technology Assessment estimates that a release of 100 KG of anthrax spores upwind of Washington, DC would, if not detected rapidly, lead to as many as three million deaths and a trillion dollars of economic losses. The research produces actionable, easily interpretable information that supports decision makers: a clear statement of the probability that a particular disease is present somewhere in the US, or present in a particular city, together with the associated uncertainty. False alarm rates are controlled, making them commensurate with resources available for post-signal investigation and taking into account consequences such as the psychological and financial impacts of a false alarm.