help
help > RE: Compare within-subj effects and between-group effects
Dec 23, 2022 12:12 AM | Alfonso Nieto-Castanon - Boston University
RE: Compare within-subj effects and between-group effects
Dear Vlanglois
That is an unusual contrast. I can definitely confirm that comparing F(pre-post) - F(group) is not a valid test for the difference between the main effect of 'group' and the main effect of 'treatment'. If you go ahead and estimate the average "clinical - non-clinical" differences between the two groups (averaged across the pre- and post- treatment conditions, e.g. 1/2 (ClinicalPre - NonclinicalPre + ClinicalPost - NonclinicalPost) ), and then compare that to the average post - pre differences between the two conditions (averaged across the clinical and non-clinical groups, i.e. 1/2 (ClinicalPost - ClinicalPre + NonclinicalPost - NonclinicalPre) ), you will see that the resulting difference would be simply exactly the same as the difference between the post-treatment non-clinical group vs. the pre-treatment clinical group, i.e.:
1/2 (ClinicalPre - NonclinicalPre + ClinicalPost - NonclinicalPost) - 1/2 (ClinicalPost - ClinicalPre + NonclinicalPost - NonclinicalPre) = ClinicalPre - NonclinicalPost
So strictly speaking you could run a two-sample t-test comparing those two conditions (clinical group pre-intervention vs. non-clinical group post-intervention) and that will indicate whether the size of the 'group' effect (Clinical - Nonclinical) is larger/smaller than the size of the 'intervention' effect (Post - Pre). That said, this all depends on the direction of the comparison that you are trying to accomplish. I imagine perhaps one possible logic of this contrast may be to evaluate whether treatment effects "normalize" the clinical group by compensating for the original (pre-treatment) differences between the clinical and non-clinical groups. If that is the case then the desired comparison would instead be:
1/2 (NonclinicalPre - ClinicalPre + NonclinicalPost - ClinicalPost) - 1/2 (ClinicalPost - ClinicalPre + NonclinicalPost - NonclinicalPre) = NonclinicalPre - ClinicalPost
which, again, could be estimated using a two-sample t-test comparing those two conditions (pre-treatment connectivity in non-clinical group vs. post-treatment connectivity in the Clinical group).
Hope this helps
Alfonso
Originally posted by vlanglois:
That is an unusual contrast. I can definitely confirm that comparing F(pre-post) - F(group) is not a valid test for the difference between the main effect of 'group' and the main effect of 'treatment'. If you go ahead and estimate the average "clinical - non-clinical" differences between the two groups (averaged across the pre- and post- treatment conditions, e.g. 1/2 (ClinicalPre - NonclinicalPre + ClinicalPost - NonclinicalPost) ), and then compare that to the average post - pre differences between the two conditions (averaged across the clinical and non-clinical groups, i.e. 1/2 (ClinicalPost - ClinicalPre + NonclinicalPost - NonclinicalPre) ), you will see that the resulting difference would be simply exactly the same as the difference between the post-treatment non-clinical group vs. the pre-treatment clinical group, i.e.:
1/2 (ClinicalPre - NonclinicalPre + ClinicalPost - NonclinicalPost) - 1/2 (ClinicalPost - ClinicalPre + NonclinicalPost - NonclinicalPre) = ClinicalPre - NonclinicalPost
So strictly speaking you could run a two-sample t-test comparing those two conditions (clinical group pre-intervention vs. non-clinical group post-intervention) and that will indicate whether the size of the 'group' effect (Clinical - Nonclinical) is larger/smaller than the size of the 'intervention' effect (Post - Pre). That said, this all depends on the direction of the comparison that you are trying to accomplish. I imagine perhaps one possible logic of this contrast may be to evaluate whether treatment effects "normalize" the clinical group by compensating for the original (pre-treatment) differences between the clinical and non-clinical groups. If that is the case then the desired comparison would instead be:
1/2 (NonclinicalPre - ClinicalPre + NonclinicalPost - ClinicalPost) - 1/2 (ClinicalPost - ClinicalPre + NonclinicalPost - NonclinicalPre) = NonclinicalPre - ClinicalPost
which, again, could be estimated using a two-sample t-test comparing those two conditions (pre-treatment connectivity in non-clinical group vs. post-treatment connectivity in the Clinical group).
Hope this helps
Alfonso
Originally posted by vlanglois:
Dear CONN users,
I am a bit limited with my statistical knowledge, and I would really appreciate anyone's advice on how to choose and set a statistical test to answer the following question:
So, I have 2 resting-state scans per subject, one at pre- and one at post-treatment, as well as 2 groups (clinical vs non-clinical), and I would like to know whether the within subject effect (pre vs.post) is greater than the between group effect (clinical vs. non-clinical).
So far we decided to go for a repeated-measures ANOVA and compare the main effect of group vs the main effect of pre-post (simply put, F(pre-post) > F(group)) but no one around me was confident enough to say this is a valid approach or nor say it is not valid...
Also, please note that I do NOT want to know whether the treatment worked in one group compared to the other, for which, if I understand correctly, I would have to compute a pre-post X group interaction term.
I would really appreciate any input :)
Best,
I am a bit limited with my statistical knowledge, and I would really appreciate anyone's advice on how to choose and set a statistical test to answer the following question:
So, I have 2 resting-state scans per subject, one at pre- and one at post-treatment, as well as 2 groups (clinical vs non-clinical), and I would like to know whether the within subject effect (pre vs.post) is greater than the between group effect (clinical vs. non-clinical).
So far we decided to go for a repeated-measures ANOVA and compare the main effect of group vs the main effect of pre-post (simply put, F(pre-post) > F(group)) but no one around me was confident enough to say this is a valid approach or nor say it is not valid...
Also, please note that I do NOT want to know whether the treatment worked in one group compared to the other, for which, if I understand correctly, I would have to compute a pre-post X group interaction term.
I would really appreciate any input :)
Best,
Threaded View
| Title | Author | Date |
|---|---|---|
| vlanglois | Nov 30, 2022 | |
| Alfonso Nieto-Castanon | Dec 23, 2022 | |