It was observed previously a small effect of heterogeneous variances on the Bonferroni's test, Student "t" test and Tukey's test. In this paper, it was tested whether such observation was due to the fact that the error used for the tests was the treat. x block interaction mean square besides the existence of certain heterogeneous variances. The authors considered the existence of the following situations: a) two groups of variance in each experiment, and b) three kinds of errors for "t" test: treat. x block interaction mean square, specific contrast error for each comparison; and, pooled contrast error. It was simulated two hundred experiments, in groups of 50, for each of 56 combinations of four coefficients of variation, two treatment effects and seven groups of heterogeneous variances, performing a total of 11,200 simulations. In the analysis of variance, for each experiment, the following contrasts between means were considered: m1 - m3 and m2 - m3, where m3 represents the control treatments mean. The two treatment means m1 and m2 were considered "a priori" more productive than the control by 8% and 5% (effect 1) and by 17% and 12% (effect 2), respectively for m1 and m2. Thus, the one tail "t" test was applied. From the results observed, it was concluded that: a) non significant differences occurred concerning the percentages of rejections of the null hypothesis using any of the three errors considered; b) the relation among variances greater than 6 was statistically heterogeneous, when the error used was that of specific contrast error; and c) the percentages of rejections obtained were indirectly proportional to the coefficients of variation and directly proportional to the range values of contrasts between means, even in case of heterogeneous error conditions.
heterogeneity of variances; error for means comparisons; t test