Some Nonparametric Confidence Intervals Are Non-informative, Notably in Bioequivalence Studies

Abstract

Nonpararaetric confidence intervals from two-period cross-over studies may become smaller than the corresponding parametric interval by (a) deviations from normality or log-normality, or (b) the presence of what is defined as intrasubject “nonparametric outlier ratios or differences”. From a certain critical value onwards such outliers cause a nonparametric confidence interval to remain constant. They can be identified by their index occupancy of the intergroup-intersubject averages in Mann-Whitney/Wilcoxon's cumulative probability distribution. This is of particular interest to the European Union, since a European guide-line allows nonparametric procedures always to be used for the statistical analysis of bioequi-valence studies. Nonparametric procedures should not be used in order “to pick out the best”, which might lead to a consumer risk higher than the nominal 5%.

Nonparametric confidence intervals keep their place in cases of deviations from (log-)normality of the underlying probability distribution, provided the absence of nonparametric outliers is proven.