I have applied ANOVA on 3 independent groups of unequal sample sizes. I doubt that the ANOVA result are showing Type 1 error because of heteroscedasticity in data points (checked using F-test as well as Bartlett's test). I want to know is there any other good statistical method to solve this problem other than the t-test for unequal variances.
$\begingroup$
$\endgroup$
1
-
$\begingroup$ ANOVA is based on the homogeneity-of-variances (= homoscedasticity) assumption. Anova is fairly robust to violation of the assumption unless the group means correlate with the group variances (i.e. the wider is the variance the higher is the mean). If that correlation is present, power transform of the dependent variable (to attain homoscedasticity) before ANOVA is a good cure. More sophisticated Box-Cox transform is another option. A nonparametric analysis (suggested by @Peter in his answer) is yet another option. $\endgroup$– ttnphnsCommented Sep 26, 2013 at 19:59
Add a comment
|
1 Answer
$\begingroup$
$\endgroup$
Welcome to the site.
One solution is the Kruskal Wallis test.
If you are using R you can do this with kruskal.test in the stats package. If you are using SAS you can use PROC NPAR1WAY.
If you show us more about your data, we will be able to give more explicit answers. Also, your sentence starting "I doubt that...." is a little confusing. You can't tell when you are making a type I error (if you could tell, you would never make them!) but perhaps you mean that you think the violation of assumptions might be causing a type I error (it's hard for us to tell since you didn't show any results).
answered Sep 26, 2013 at 10:49