# model for Pre/Post analysis

Hi AFNI Team:

Apologies if this has been addressed previously. I checked, but didn’t find the exact question. I am hoping you all can provide some guidance on which AFNI analysis would be most appropriate for my data. I have the following Pre/Post design:

Within Person: 1 fMRI run before a manipulation (PRE) and 1 fMRI run after a manipulation (POST).
Between Person Variable: Risk (1) or No Risk (0)
Between Person Variable: Age

I’m interested in a whole brain analysis PRE > POST with (A) main effects of Risk/No Risk and Age and (B) interaction of risk with age.

In our data, Age is definitely correlated with Risk/NoRisk grouping variable (r = ~.35), so we want to make sure we select the appropriate model to examine the potential interaction effect.

It’s possible we will have additional main effects to include in the model with additional interactions. For example, we may include an additional mental health variable ininteraction with Risk/No Risk.

At first we considered using 3dMEMA but our understanding is that we’d have to run a within subject t-test (pre run vs post run) on each person before doing the regression & that just seemed like a cludge for doing a between/within ANOVA or a multi-level model.

Next, we were considering 3dMVM, but we want to make sure that it will model this factorial ANOVA appropriately. We see 3dMVM described in AFNI documentation as ANCOVA, but in our case our covariate is correlated with our grouping variable, which violates a core assumption of ANCOVA. Factorial ANOVA seems like the appropriate model, but we want to know if 3dMVM is capable of running a factorial ANOVA.

Thanks in advance, and happy to clarify anything.

Age is definitely correlated with Risk/NoRisk grouping variable (r = ~.35)

Do you mean one group is older than the other? How do you want to interpret your results: with both groups at some common age or with each group at a separate age?

we considered using 3dMEMA but our understanding is that we’d have to run a within subject t-test (pre run vs post run) on each person
before doing the regression & that just seemed like a cludge for doing a between/within ANOVA or a multi-level model.

I have recently removed that restriction. So you could provide each condition (Pre and Post) as a separate input.

in our case our covariate is correlated with our grouping variable, which violates a core assumption of ANCOVA.

3dMVM would be fine with such a case. It does not violate any assumption in this case; instead, it is more an issue as to how you want to center the age variable.

Thanks, Gang! To answer your first question, we would want to interpret group differences at some common age (so mean center across the sample).

If we did 3dMEMA, would we still be able to test the Main effect of group, main effect of age, and their interaction?

Following up to your last point, I just want to make sure we’re appropriately modeling variance. If I’m understanding the concern about a covariate sharing variance with the grouping variable (and why that violates the ANCOVA assumption), it is that you are partialing out variance in the outcome (here, post > pre run) due to age, but age also shares variability with the grouping variable. So instead, we’d want to use a factorial anova approach where we asses the main effects and interaction of Group X age to account for variance in our outcome appropriately. So rather than saying “what are the group differences irrespective of age?,” we would ask, “how do the group differences vary as a function age (Interaction)?”

In a non-neuroimaing world, I’d typically run a moderated regression here. So that’s what I’m trying to translate to our particular case.

Thanks so much for your thoughts here.

If we did 3dMEMA, would we still be able to test the Main effect of group, main effect of age, and their interaction?

Yes, you would, but it might be a little tedious to set it up properly. So, 3dMVM is a better choice in terms of user interface.

rather than saying “what are the group differences irrespective of age?,” we would ask, “how do the group differences vary as a function age (Interaction)?”

The first question does not even make sense to me. You may ask such an abstract question, but that is far from the reality.