Hi Georg,
one possibility is that the betas in multiple regressions reflect the
bit of variance that is unique to a given EV. So it is possible that
when you add interaction terms some of the variance that was originally
allocated to some other EV (when you do not include interactions) is now
no longer unique to that EV. Which may explain why you see different
results. Note that this can be the case even if your interaction terms
do not show any significant activation.
just a thought,
cheers
martin
G. Homola wrote:
> Hi all!
>
> Let me ask you something basic about interactions. I did a lot of runs on some data here and noticed the results show great variation depending on how many interactions I add to the model. Interactions, whether sensible or not, always strongly change the outcome for the contrasts of the main EVs, too - by their mere existence! On group level, in comparison to the run without interactions, sometimes some clusters completely disappear and others show up in entirely different regions. Does that make sense? And if so, which runs are more valid, the ones with, or without interactions?
>
> Thanks!
> Georg
>
|