Dear CONN users/experts,

I have two groups (group_F, group_M; n = 35), however, only 30 subjects have a behavioral measure of interest. After comparing between group differences, I am interested in the interaction model. For simplicity, I am using 6 subjects as an example (subject 5 does not have behavioral measure) and want to check whether I am doing it correctly or not.

group_F = [1 1 1 0 0 0]
group_M= [0 0 0 1 1 1]
covariate1 =[-2 -1 -0.5 0.5 1 2] #Demeaned (a.k.a., orthogonalized for all subjects)
trait_F = [x1 x2 x3 0 0 0]
trait_M= [0 0 0      y1 0 y2] #Note the fifth subject being zero in both trait_*

This is the contrast that I should use for the model below: [0 0 0 1 -1]
y ~ group_F + group_M + covariate1 + trait_F + trait_M

Three questions:
1. Am I doing all things right?
2. Do I need to also orthogonalize trait_F & trait_M? In respect to all subjects? or in respect to group_F and group_M, respectively?
3. Any positive t-stat result for this contrast (trait_F > trait_M) means that there is a significant group*trait interaction in that voxels (regression lines of activity ~ trait for each group are crossing each other for that region), in which there is a positive beta for activity~trait in group_F and negative beta for activity~trait in group_M, right?


Thanks in advance!

Bests,
Amir

Dear Amir,

Almost exactly correct, but in GLM you would need to exclude subjects without the behavioral measure from the interaction analyses. In your example that could be done simply entering NaN for those subjects in their corresponding trait_* variables, i.e.:

group_F = [1 1 1 0 0 0]
group_M= [0 0 0 1 1 1]
covariate1 =[-2 -1 -0.5 0.5 1 2]
trait_F = [x1 x2 x3 0 0 0]
trait_M= [0 0 0 y1 NaN y2]

Regarding (2), there is no need to orthogonalize trait for this interaction analysis

And regarding (3), typically interaction effects are tested using two-tailed/sided analyses, as the sign of the interaction term only reflects whether the slopes (association between trait and connectivity within each group) are larger or smaller in one group compared to the other (i.e. the regression lines of connectivity vs. trait for each group will cross each other for both positive and negative interaction effects, in positive effects the regression line in the F group has a higher slope than in the M group, and the opposite for negative effects)

Hope this helps
Alfonso

Originally posted by AmirHussein Abdolalizadeh:

Dear Alfonso,

Thanks for your reply. Regarding the interaction, to identify the direction of slope for each variable I have to do further analysis separately for each trait_group, right? For example, I have to use the following model again, however with contrasts [0 0 0 1 0] to identify slope for trait_F and [0 0 0 0 1] for the trait_M.

y ~ group_F + group_M + covar1 + trait_F + trait_M


Bests,
Amir