Abstract
Consider a two-way, n x r, classification or a randomized block design with n treatments and r blocks. The usual model is modified so that error components have variances for the j‐th column or block. Let Qj: be Grubbs1 estimator of
An approximate test of the hypothesis
is proposed and based on the statistic, −½(n−1)(r−1)(n−2) log Tr, where
, which is taken to have the central chi-square distribution with (r−1) degrees of freedom. Simulation studies for r=3(1)6(2)12 and n=10, 20, 30, 50 confirm that the approximate distribution is satisfactory for applications.