Pitting corrosion:: Comparison of treatments with extreme-value-distributed responses

In this article we develop statistical extreme-value theory as a method to validate and improve experiments with extremal responses, and to extrapolate and compare results. Our main motivation is corrosion tests performed at Volvo Car Company. Localized, or pitting, corrosion can limit the usefulness of aluminum, magnesium, and other new lightweight materials and makes judicious choice of alloys and surface treatments necessary. Standard methods for evaluating corrosion tests are based on weight loss due to corrosion and ANOVA. These methods fail in two ways. The first is that it usually is not weight loss but the risk of perforation (i.e., the depth of the deepest pit) that is of interest. The second is that the standard ANOVA assumption of homogeneity of variances typically is not satisfied by pit depth measurements, and that normality does not give credible extrapolation into extreme tails.