Statistical issues on the no-observed-adverse-effect level in categorical response

The determination of the value of the no-observed-adverse-effect level (NOAEL) when observed responses can be categorized by severity (categorical data) and sample sizes are small is discussed. The common situation of only two categories, where only the presence or absence of an effect is observed, is addressed first (dichotomous data). Three tests for dichotomous data are critically examined, including the Brown-La Vange test, a modified version of that test, and Dunnett's multiple comparison test. Although the modified test is an improvement, all three procedures have shortcomings in determining the value of the NOAEL, particularly when the sample size is small. An alternative method is suggested, based on the Akaike information criterion (AIC), which performs well. This method is extended to severity data with an arbitrary number of categories. Use of a dose-response curve for the NOAEL is discussed.