Compare means for heterogeneous variances

Valerie_Nedbal

I found the Welch test to compare the means in the case where the variances are heterogeneous (a result of the comparison test of variances), but when you have more than one modality, you can not go further (as compared averages 2-2 in the case of heterogeneity of variance), I see these comparisons of means in the option in case the variances are equal, but I can not find the option in the case where the variances are not equal. Is it possible? Hope i'm clear.

Thanks,

Valerie

malcolm_moore1

Valerie,

Looks like the situation is one of wishing to compare the location of one Y vs one categorical X and each level of the X has multiple modes in Y and different variance in Y compared to other levels of X. Is that the situation? If so, I would take a modelling approach and try to find the other X's (in addition to the current categorical X) that is impacting the location and variance of the Y. Then look at the residuals of the "final" model to see if they are randomly distributed with a single mean of zero. If so then problem solved - we've found and modelled the X's responsible for location and variance effects in the Y and we don't need to worry about the distributional assumptions of oneway anova. In summary the data may be hinting that there is more than one X that is effecting Y and a more informative approach would be to model collectively the set of important X's. If customer is not happy with that advice then I'd try partitioning with the one X, at least that will tell us which levels of the X have differences in the mean of Y without the distributional assumptions of oneway anova. Better still use partitioning with many Xs.

Malcolm