ROBUST TESTS BASED ON THE SAMPLE CHARACTERISTIC FUNCTION

A new method is described for robust analysis of variance in the balanced fixed effects case. The method uses the empirical characteristic function of the treatment samples, and has an interpretation in terms of S‐estimators. The test statistic, under the null hypothesis, asymptotically follows a central chi‐square distribution, and under contiguous alternatives a noncentral chi‐square distribution. A Monte Carlo study suggests that, for finite samples, this is reasonably well approximated by the usual F distribution used in analysis of variance. The test statistic has a bounded influence function. The new procedure competes well with Huber's and a Wald‐type procedure except in very heavy‐tailed cases.