Purpose: We used novel Bayesian modeling techniques to produce more precise estimates of foodborne illness source attribution trends over time.
Methods: Two Bayesian statistical models were used to evaluate changes in the numbers of outbreaks associated with a set of food categories (e.g. meat, poultry, fruit, and vegetables) reported to the CDC Foodborne Disease Outbreak Surveillance System (FDOSS) for the years 1998-2014. The simple shrinkage trend model naively pools information from all food categories to estimate individual category intercepts and slopes; the latent cluster shrinkage trend model pools information from categories only to the degree that sets of categories appear to share a common change over time. These were compared with a fully stratified model, which effectively fits a separate intercept and slope model to outbreak counts for each food category.
Results: Both Bayesian methods produced more precise trend estimates than the fully stratified estimates; the latent cluster shrinkage model produced estimates with comparable shrinkage but greater precision than the simple shrinkage model. The median reduction in the width of the 95% credibility intervals (CI) of the slope estimates across the 14 food categories analyzed was 13% for simple shrinkage and 24% for latent cluster shrinkage. The mean reductions in widths were 22% and 29% respectively.
Significance: Both Bayesian models offered improved precision over a fully stratified model. Bayesian latent cluster shrinkage is a novel, more precise, and potentially less biased statistical method than full stratification or simple shrinkage for evaluating foodborne outbreak trends over short time periods with limited data.