In this study, the effect of using trimmed, winsorized, and modified means instead of arithmetic mean on type-I error rate was investigated when the assumptions of the one-way ANOVA were not satisfied. Therefore, random numbers were generated by simulation technique from the populations distributed by Normal (0,1), Beta (5,2) and chi(2) (3) for 3 and 4 groups. The results of 30 000 simulation trials demonstrated that all the means displayed similar type-I error rates when the variances were homogenous regardless of the distribution shape, sample size and the number of groups. When homogeneity of variances assumption was not satisfied, the most reliable result was obtained by using trimmed mean in terms of keeping the type-I error rate at nominal alpha level and it was followed by modified and winsorized means. The most biased results were obtained when arithmetic mean was used.