In analysis of variance the usual method of determining the sample size needed to detect a certain maximum deviation between population means (assuming common variances) with specified significance level and power, uses the operating characteristic curves of the non-central F distribution. These curves, however are quite restrictive with respect to significance level and the number of treatment groups. An alternative procedure for approximating the sample size is proposed which is adaptable to a much broader range of analysis of variance situations. Monte Carlo simulation of the new procedure indicates that it is a reasonable method of determining the sample size for the specified power.