We investigated the sources of data used in recently published predictive models of skin permeability. It was found that skin permeability coefficients for 63 compounds are poorly documented. We hypothesized that these coefficients were calculated using the simple two variable, three parameter 'Potts and Guy' regression equation and hence were not derived from experimental measurements. We therefore examined the distribution of residuals of these reported coefficients compared with the Potts and Guy predictions. The residuals cannot be described by a normal distribution. A substantial (51%) number of residuals equaled 0.00. Further analysis demonstrated that 89% (56 out of 63) of the skin permeability coefficients can be explained as being calculated by the Potts and Guy equation using different documented octanol-water partition coefficients, and/or transcription errors. The results strongly suggest that these 63 skin permeability coefficients are calculated and not experimentally determined-a conclusion subsequently confirmed by one of the developers of the data set. Continued use of these data would lead to biased model selection, underestimation of experimental variability, and overestimation of model predictive ability.