Detection of an Outlier and Evaluation of its Influence in Chronic Toxicity Studies

Abstract

Targeting quantitative data in repeated-dose toxicity studies using rodents, a method for detecting an outlier and a selection problem between parametric and nonparametric approaches, based on actual toxicity data, was investigated. The consistency between the judgments of veteran toxicologists and several statistical methods (skewness, kurtosis, and the studentized residual) to detect an outlier was evaluated. The studentized residual had the highest consistency with the judgments of toxicologists and was the most effective method for detecting an outlier.

The performance of the parametric (regression) method and the nonparametric method (Jonckheere test) for detecting dose-dependency was evaluated. Parametric and nonparametric approaches had a different result when outliers existed. Parametric methods are sensitive to the presence of outliers and they lose statistical power. In contrast, nonparametric approaches are robust to outliers. The identification of extreme individual measures is, however, a different objective for safety studies versus the detection of dose-dependency, and it may require different methods. Detecting an outlier, and investigating its impact on statistical analysis and biological interpretation, is very important and requires the use of appropriate methods.