I understand that post-hoc Tukey contrasts are inadequate on a 1-way ANOVA that violates the homoscedastiscity assumption, and that Games Howell or Dunnet contrasts are among the recommended alternatives. Does anyone know if this holds true even if I refit my ANOVA with gls using a varIdent variance structure to address the heteroscedasticity? In other words: Can I use Tukey contrasts after managing heteroscedasticity with gls and a variance structure, or do I still need to look for alternative Post-hoc tests?

Thanks in advance


1 Answer 1


I don't agree with your premise: If you wanted to do all pairwise comparisons, why would you change that objective just because some underlying assumption is violated?

Instead, you can compute the t statistics for each pairwise comparison using the estimated standard errors for each difference, and adjust the P values using a Bonferroni correction (conservative); or, for an approximate critical value for comparing the ith and jth means, use $\sqrt{1/2}q_\alpha(k,d_{ij})$ from the Studentized Range tables, where $k$ is the number of means and $d_{ij}$ is the df for that comparison.

The lsmeans package in R implements this procedure but you need to wait for the next update or contact its developer because there was a bug in the support for the varIdent structure.


Your Answer

By clicking “Post Your Answer”, you agree to our terms of service and acknowledge you have read our privacy policy.

Not the answer you're looking for? Browse other questions tagged or ask your own question.