Hi Helmut,

I was specifically to the question about using the ANOVA first. In the case that you have specific hypothesis comparing 2 groups, then I would conduct 2 sample t-tests for each comparison separately. The issue of having all three groups in the model without the initial main effect is that you are using the overall error term for three groups for your two group testing. Essentially, you're adding a group to modify the error variance -- which may be either good or bad depending on the third group. 

Best,

Donald

Best Regards, 

Donald McLaren, PhD

On Wed, Jun 7, 2017 at 7:45 AM, MRI More <[log in to unmask]> wrote:

Dear Donald, dear Jennifer,

When going with three groups, did you have hypotheses about specific group differences? If so, then you could just start with planned comparisons according to your hypotheses (for which you should consider multiple testing though). This way you avoid the issue of how to conduct proper post-hoc tests. In fact it might be misleading to look at the main effect first in case you had more specific comparisons, as you could well conduct pair-wise comparisons despite the main effect being non-significant (except if you rely on a correction procedure for multiple comparisons that explicitly requires testing for a sig. main effect).

For example, your factor group might consist of three levels healthy young, healthy elderly, demented elderly. You might want to test for healthy young vs. healthy elderly to find correlates of healthy ageing, and you might want to additionally test healthy elderly vs. demented elderly for alterations due to dementia. Go with e.g. .0005 uncorrected on voxel level (instead of .001 due to the two pair-wise comparisons) in combination with a cluster correction, and you should be on the safe side.

Best regards

Helmut