Estimation of infection prevalence from correlated binomial samples

Area of Science:

• Veterinary epidemiology
• Biostatistics
• Infectious disease modeling

Background:

• Infection prevalence is often estimated using grouped binary data, assuming binomial distribution.
• Correlated observations within groups can lead to overdispersion, violating independence assumptions.
• Existing methods may not adequately account for complex correlation structures in prevalence data.

Purpose of the Study:

• To demonstrate random-effects models for estimating infection prevalence from correlated data.
• To illustrate a nonparametric random-effects model for prevalence estimation in veterinary epidemiology.
• To compare the nonparametric approach with traditional and alternative statistical models.

Main Methods:

• Developed and outlined assumptions for a logistic-regression model with a nonparametric random effect.
• Applied the nonparametric method to Salmonella infection data in Irish pig herds.
• Compared results with a standard logistic model, a generalized estimating equation (GEE) model, and models with normally distributed random effects (SAS GLIMMIX and NLMIXED).

Main Results:

• The nonparametric random-effects model classified pig herds into distinct prevalence groups (low, medium, high).
• It also estimated the relative frequency of each prevalence category within the population.
• Population-averaged prevalence estimates were derived using numerical integration and Monte Carlo simulation.

Conclusions:

• A nonparametric random-effects model provides a valuable tool for analyzing correlated prevalence data in veterinary epidemiology.
• This approach allows for the identification of distinct prevalence strata within a population.
• The method offers an alternative to traditional models, particularly when overdispersion is present.