Parametric analysis of rank transformed data for statistical assessment of genotoxicity data with examples from...

The utility of rank transformation followed by parametric analysis of the ranks has been assessed for determination of the statistical significance of genotoxicity data. Both nonparametric and parametric analytical methods have defects when used to assess the significance of results from routine regulatory tests. Superficially, the rank transformation method followed by parametric analysis of ranks appears to be an ideal solution. However, we considered that such a test might suffer a substantial loss of power when used to analyse normally distributed data with very low sample replication. To test this hypothesis we took 22 data sets from five 'borderline' positive compounds in mouse lymphoma treat-and-plate assays where treatment-related increases were between 1.5- and 3-fold the control and analysed these results by Dunnett's t-test using rank transformed data and weighted, untransformed data. In theory these mouse lymphoma data should show the rank transformation system at its worst in comparison with parametric methodology using weighted data. Surprisingly, the rank transformation methodology showed no loss of power and, overall, performed more consistently than the weighted data methodology. Based on this limited number of data sets, rank transformation followed by parametric analysis of ranks seems to be an approach very suitable for genotoxicity assays in general, particularly where distributions are non-normal or of uncertain form. It combines the general applicability of non-parametric methods with the power of parametric analyses. However, the methodology still requires to be further validated in use and by computer simulation.