Bayesian analysis to validate a commercial ELISA to detect paratuberculosis in dairy herds of southern Chile.

In Chile, Mycobacterium avium subsp. paratuberculosis (Map) has been isolated on several occasions and clinical cases have been reported. Nevertheless, diagnostic tests have not yet been validated for this agent in the Chilean setting. The objective of the study was to validate a commercial ELISA to detect Map shedding dairy cows in management conditions, prevalence and stages of infection existing in Southern Chile, utilising different statistical approaches. Blood and faeces were collected from 1333 lactating cows in 27 dairy herds (both large commercial and smallholder dairy farms) between September 2003 and August 2004. Within the herds up to a maximum of 100 dairy cows were selected based on age (>or=3 years old) and, if present, clinical signs of a Map infection. In herds with less than 100 cows, all cows >or=3 years old were sampled. Blood samples were tested using a commercial ELISA kit (IDEXX Laboratories, Inc.). Faecal samples were cultured on Herrold's Egg Yolk Medium (HEYM). Latent class models (i.e. maximum likelihood (ML) methods and Bayesian inference) were used to determine the validity of the ELISA. Map was cultured from 54 (4.1%) cows and 10 (37.0%) herds, which were all large, commercial dairy herds. As a result of empty cells in the cross-tabulations, the ML model provided the same results as the validation with faecal culture as the gold-standard. In the Bayesian model, the Se and Sp of the ELISA were estimated to be 26% (95% CI: 18-35%) and 98.5% (95% CI: 97.4-99.4%), respectively. For faecal culture, the Se was 54% (95% CI: 46-62%) and the Sp was 100% (95% CI: 99.9-100%). Interestingly, the prevalence in the smallholder dairy farms was estimated to be 8% even though there were no faecal culture positive cows detected in those herds. There was no significant correlation between the two tests. The advantage of Bayesian inference is that the Se and Sp of both tests are obtained in one model relative to the (latent) true disease status, the model can handle small datasets and empty cells and the estimates can be corrected for the correlation between tests when the tests are not conditionally independent. Therefore, Bayesian analysis was the preferred method for Map that lacks a gold-standard and usually has low cow-level prevalence.