There is one (most) important test: the test for the interaction. This test obviousely needs all the data anyway.

There might be other interesting hypotheses to test, like a gender-effect in normal-weight people, or an overweight-effect in females etc. These tests *should* be done using post-hoc tests, that is, these tests should use all the data to get the best and most robust estimates of the standard errors. Since the hypotheses of these tests are not related (it's not a screening to identify candidates from a set of possibilities), I don't think that any correction for multiple testing is required here.

Hello Mr. Wilhelm,

thank you for the quick reply.

In attachment an R script with the HSD test on the interaction + letter grouping.

Marius Paulus, if you had only one dependent variable, I think it would indeed by OK if you merely looked for a main effect for weight, a main effect for gender, and an interaction effect.

However, there is an issue concerning how you have scored the DEBQ as your dependent variable. As I understand it, the DEBQ has three, or perhaps four, domains. If you conduct separate analyses for each of those domains you don't need post hoc tests in the traditional sense, but I suspect you should make some adjustment to your p value to acknowledge multiple comparisons having taken place.

In offering the above advice, I hope I've understood the nature of your research sufficiently - and all the best with it.

Dear Marius Paulus

There is a complete chapter / document about “ Two-Way ANOVA Post hoc analysis” via the following link which might useful regarding your quries:

https://webstat.une.edu.au/unit_materials/c7_anova/twoway_post_hoc.htm