An epidemiologic critique of current microbial risk assessment practices: the importance of prevalence and test accuracy data.

Data deficiencies are impeding the development and validation of microbial risk assessment models. One such deficiency is the failure to adjust test-based (apparent) prevalence estimates to true prevalence estimates by correcting for the imperfect accuracy of tests that are used. Such adjustments will facilitate comparability of data from different populations and from the same population over time as tests change and the unbiased quantification of effects of mitigation strategies. True prevalence can be estimated from apparent prevalence using frequentist and Bayesian methods, but the latter are more flexible and can incorporate uncertainty in test accuracy and prior prevalence data. Both approaches can be used for single or multiple populations, but the Bayesian approach can better deal with clustered data, inferences for rare events, and uncertainty in multiple variables. Examples of prevalence inferences based on results of Salmonella culture are presented. The opportunity to adjust test-based prevalence estimates is predicated on the availability of sensitivity and specificity estimates. These estimates can be obtained from studies using archived gold standard (reference) samples, by screening with the new test and follow-up of test-positive and test-negative samples with a gold standard test, and by use of latent class methods, which make no assumptions about the true status of each sampling unit. Latent class analysis can be done with maximum likelihood and Bayesian methods, and an example of their use in the evaluation of tests for Toxoplasma gondii in pigs is presented. Guidelines are proposed for more transparent incorporation of test data into microbial risk assessments.