Evaluation of multivariate statistical methods for analysis and modeling of immunotoxicology data

Abstract

In immunotoxicology, the critical functions of the immune system (host resistance to infection and neoplasia) cannot be measured directly in humans. It is theoretically possible to predict changes in host resistance based on changes in immunological functions known to mediate host resistance. However, quantitative predictive models of this type have not yet been achieved in humans or in animal models. Multivariate statistical methods were developed for analysis and modeling of the effects of several explanatory variables on a dependent variable, and they seem well suited for attempts to predict host resistance changes caused by changes in immunological parameters. However, these methods were developed with the assumption that all variables can be measured for each experimental subject. For a number of reasons, this generally cannot be done in comprehensive immunotoxicology evaluations. In the present study, the suitability of multivariate methods for analysis of variables measured in different experiments was examined, using a limited data set consisting of immunological parameters that could all be measured for each mouse. Analysis was done on the original data set and test data sets produced by randomizing data within dosage groups. This was done to simulate the random pairing of data that would occur if measurements were obtained from different sets of mice in different experiments. Statistical theory indicates that randomization will disrupt the correlation matrices that are central in multivariate analyses. However, the present results demonstrate empirically that for at least one immunotoxicant (dexamethasone), remarkably similar multivariate models were obtained for the original and 109 randomized data sets. In contrast, the randomized data sets produced substantially different multivariate models when data obtained with a different immunotoxicant (cyclosporin A) were analyzed. The major difference between the two data sets was that dexamethasone strongly and dose-responsively suppressed many more parameters than did cyclosporin A. Additional work is needed to determine whether there are consistent criteria that could be used to identify immunotoxicology data sets, which would be amenable to multivariate analysis.