ANOVA for skewed data with heteroscedasticity?

[Ciaran DK]

Hi, I am trying to determine whether the differences in means of several groups are statistically significant.

Initially I ran an ANOVA on Stata but then realised my data set was neither normally distributed and there was no homogeneity of variances between groups.

From reading online it seems that Welch's ANOVA is appropriate for data which lacks homogeneity of variances but is normally distributed whilst Kruskal Wallis is appropriate for data which is not normally distributed but has homogeneity of variances.

Are there any specific tests which would be valid for my data and if so, how would I perform these on Stata?

Many thanks,

Ciaran

[Nick Cox]

Much depends on what kind of non-normality you have and what is the pattern of homoscedasticity you have.

Top of my personal list of alternatives would be a generalised linear model with appropriate link and family, but I guess most people interested in the question would want to see example data or summaries of data to advise in detail.

Otherwise put why use 1940s or 1950s techniques when you could use 1970s techniques?

[Nick Cox]

In #2 I mean heteroscedasticity. Sorry to anyone puzzled. I do know the difference: https://stats.stackexchange.com/ques...eteroscedastic

[Carlo Lazzaro]

Ciaran:
welcome to this forum.
As an aside to Nick's helpful comments, why not switching to -regress-?

regress <regressand> i.group

Then you can investigate heteroskedasticity via -estat hettest- after -regress- (in addition to visual inspection -help rfvplot-).

[Mike Lacy]

To avoid the normality and homoscedasticity assumptions, I'd consider using -permute-, e.g.
permute group F = e(F), reps(10000): anova y i.group