Application of Statistical Classification Methods to the Evaluation of In Vitro Tests

Abstract

The first Amden Report defines validation as “The process by which the reliability and relevance of a test are established”. The recommendations of the Amden conference mention the need for biometric methods, but, unfortunately, most publications on alternatives to animal testing procedures do not pay much attention to statistical problems in the validation process. The biometric repertoire consists mainly of scatterplots, correlation coefficients and linear regression.

In the evaluation of an interlaboratory study, we applied multivariate statistical methods to analyse the ability of two in vitro tests to identify severely irritating substances. In this study, conducted by the German Bundesgesundheitsamt, two alternatives to the well-known Draize eye irritancy test — the neutral red uptake test and the HET-CAM test — were investigated. The statistical methods applied were discriminant analysis and the Classification and Regression Trees (CART) method.

The results presented here show that the combination of the endpoints of the in vitro tests, analysed by using the CART method, can give a sensitivity of about 67% with 76% specificity. Discriminant analysis models only achieve a sensitivity of 55%. The results obtained with the data presented here show that these statistical methods are very suitable and useful for this type of problem. This was confirmed by further analyses with discriminant analysis models, which gave a sensitivity of about 80%.