Dear CONN Experts,
I would like to evaluate a group*condition interaction (2 groups with 2 exams each) with a second-level covariate. I believe it would be an example of a mixed 2x2 ANCOVA design (subject effects [1 -1 0] with between-condition contrast [1 -1]). Actually, the covariate is not a nuisance covariate but rather a change in behavioral scores between the sessions. Hence, I am more interested in the interaction covariate*condition, i.e., with subject effects set to [0 0 1]. However, I am concerned that the covariate (change in a score) is not really meaningful if conditions are modelled separately.
One solution would be to somehow include a session-specific covariate (i.e., original scores per session). I could not find such design description in the documentation and I searched the forum to find a several related posts still waiting for reply:
https://www.nitrc.org/forum/message.php?msg_id=38860
https://www.nitrc.org/forum/message.php?msg_id=40691
https://www.nitrc.org/forum/message.php?msg_id=41198
A generative AI suggested to include the session-specific covariate as a first-level covariate set to a constant value per subject. Unfortunately, I found no such recommendation here.
Thanks in advance!
Kind regards
Pavel
Dear CONN Maintainers,
I will really appreciate any hint or comment on how to include a session-specific second-level covariate in CONN.
Kind regards
Pavel
Dear CONN Experts,
I have re-evaluated my original question and I am concerned it might have been based on wrong assumptions (actually based on a misleading AI summary). Hence, I would like to ask whether it is valid to use a design described in my original post (mixed 2x2 ANCOVA design with subject effects [1 -1 0] and between-condition contrast [1 -1], in which the additional covariate represents a behavioral change between session 1 and 2) as it is. Could you confirm that CONN evaluates both sessions simultaneously using multivariate statistics? If so, no session-specific covariate would be needed.
Kind regards
Pavel Hok
Dear Pavel,
Entering a model of the form
Subject effects: GroupA, GroupB, BehavScoresChange
Between-subjects contrast: [-1 1 0]
Conditions: Exam1, Exam2
Between-conditions contrast: [-1 1]
Will evaluate your original group*condition interaction while controlling for changes in behavioral scores (i.e. you are asking whether the change in functional connectivity between the two tests is actually different between the two groups, while potentially controlling for / disregarding differences that could be explained by differences between the two groups in their BehavioralScore changes)
You could also evalute the condition*covariate interaction using the same model but changing the between-subjects contrast to [0 0 1]. This will ask whether the change in functional connectivity between the two tests is actually related to the changes in behavioral scores between the two tests, while potentially controlling for baseline differences in functional connectivity between the two groups.
Last, if you would like to evaluate the group*condition*covariate interaction, you need to enter the group*covariate interaction terms into your model, which means that you would instead use a model of the form:
Subject effects: GroupA, GroupB, BehavScoresChange*GroupA, BehavScoresChange*GroupB
Between-subjects contrast: [0 0 -1 1]
Conditions: Exam1, Exam2
Between-conditions contrast: [-1 1]
(note: to create the "BehavScoresChanges*Group#" covariates simply select the three covariates -groupA,groupB,BehavScoresChage- in the Setup-covariates 2nd-level tab and then click on the menu "Covariate tools -> Create interaction of selected covariates")
This three-way interaction test will evaluate whether the association between changes in functional connectivity and changes in behavioral scores between the two exams are different when comparing the two groups.
Hope this helps
Alfonso
Originally posted by Pavel Hok:
Dear CONN Experts,
I have re-evaluated my original question and I am concerned it might have been based on wrong assumptions (actually based on a misleading AI summary). Hence, I would like to ask whether it is valid to use a design described in my original post (mixed 2x2 ANCOVA design with subject effects [1 -1 0] and between-condition contrast [1 -1], in which the additional covariate represents a behavioral change between session 1 and 2) as it is. Could you confirm that CONN evaluates both sessions simultaneously using multivariate statistics? If so, no session-specific covariate would be needed.
Kind regards
Pavel Hok
Dear Alfonso,
Many thanks for your very informative reply!
Best regards
Pavel