Beta-Binomial Statistical Model for Validation Studies of Analytes with a Binary Response.

BACKGROUND: The probability of detection or 'POD' model has had widespread application for statistically analyzing single and multiple collaborator validations studies with binary outcome data for a wide range of analytes over the last decade.

OBJECTIVE: The POD model is placed on a firm theoretical foundation, and extended to a more generalized beta-binomial model.

METHODS: The POD model is revisited and embedded in the beta-binomial model. This generalization includes collaborator reproducibility as a specific parameter. The new model includes only two distributional parameters: LPOD as the overall across-collaborator probability of detection and ICC as the intraclass correlation of collaborators, measuring irreproducibility. Differences between methods are measured by the difference in LPOD values, denoted dLPOD.

RESULTS: Accurate statistical estimators and confidence intervals are provided with validation by simulation. This new beta-binomial model will be applicable to a full range of candidate methods giving binary qualitative results, including microbiological, toxin, allergen, biothreat and botanical analytes.

HIGHLIGHTS: The new beta-binomial model provides easy equivalence tests to show the study clearly demonstrates (with 95% confidence) that method differences and collaborator reproducibility are acceptable.