When observations are taken from populations that are normally distributed with unknown and possibly unequal variances, Bishop and Dudewicz (1978) proposed a two-stage sampling procedure for the analysis of variance problem under heteroscedasticity. The procedure is a design-oriented process which requires an initial sample at the first stage and additional observations at the second stage. In this article we discuss the relationship between the initial sample size n_0 and the final total sample size. Two indicators T and R are provided to help selecting the initial sample size.