Statistical analyses of scatterplots to identify important factors in large-scale simulations, 2: robustness of techniques

The robustness of procedures for identifying patterns in scatterplots gener ated in Monte Carlo sensitivity analyses is investigated. These procedures are based on attempts to detect increasingly complex patterns in the scatte rplots under consideration and involve the identification of (i) linear rel ationships with correlation coefficients, (ii) monotonic relationships with rank correlation coefficients, (iii) trends in central tendency as defined by means, medians and the Kruskal-Wallis statistic, (iv) trends in variabi lity as defined by variances and interquartile ranges, and (v) deviations f rom randomness as defined by the chi-square statistic. The following two to pics related to the robustness of these procedures are considered for a seq uence of example analyses with a large model for two-phase fluid flow: the presence of Type I and Type II errors, and the stability of results obtaine d with independent Latin hypercube samples. Observations from analysis incl ude: (i) Type I errors are unavoidable, (ii) Type II errors can occur when inappropriate analysis procedures are used, (iii) physical explanations sho uld always be sought for why statistical procedures identify variables as b eing important, and (iv) the identification of important variables tends to be stable for independent Latin hypercube samples. (C) 1999 Published by E lsevier Science Ltd. All rights reserved.