Hello!! 

I am trying to do an ANCOVA in CONN with 3 groups and 4 covariates. The overall model seems to be significant but I'm very confused on how to figure out which groups are different from each other. Is this done through the effect size bar graph and seeing if the bars overlap? If so, my model is significant but all the error bars seem to overlap in the bar graph so it doesn't seem to be showing which groups are different.  Can some please help me with this??

Best,
Jessica

Hi Jessica,

Typically that ANCOVA model would be defined in CONN as:

 subject-effects:                   GroupA GroupB GroupC Covar1 Covar2 Covar3 Covar4
 between-subjects contrast:  [1 -1 0 0 0 0 0; 0 1 -1 0 0 0 0]

That model would evaluate any differences between the three groups when compared at the same-level of each covariate. When displaying the effect-size barplot each bar will represent the average connectivity within each of the three groups (all estimated at the zero-level of all your four covariates, i.e. the 'adjusted means'), and differences between the groups would typically be observable visually as minimally- or non-overlapping errorbars

Hope this helps
Alfonso
Originally posted by Jessica Busler:

Hi Alfonso,

I really appreciate you responding!!! Thank you!

Thank you for the information. I have a few follow-up questions and am attaching some screenshot examples.

One screen shot is my second-level model setup. As you will see I'm trying to do an ANCOVA and with two seed ROIs.  4 clusters emerged as significant from this but in the other screenshot all the error bars are overlapping pretty substantially so I'm not sure where the group differences might be. 

I set up the subjects effects and the between-subjects contrast in the same manner you specified in your response and then did an "any effects" for the two seeds as I want to see if there is an effect from either one of those seeds.  

I also wanted to double check what the effect size bar graph error bars mean.  Are they confidence intervals? and that's why visually we would be able to tell group differences by whether or not the error bars overlap?

One more area of question... I would like to do post-hoc tests to get the T values for pairwise group differences. Is there a way to do this? I've played around a little with importing the values and the calculator but I'm not sure how to do set things up properly. Would it be the same model setup there? I am also curious about when to include the covariates (i.e. during second-level set up or posthoc in the calculator).  Could you touch on how setting up the model in the second-level results tab differs from instead waiting to set the model up in the calculator area?

Sorry for so many questions!!! I appreciate so much your help!!

Best,
Jessica