help
help > 2nd Level Multi-Session Result Interpretation
Sep 16, 2017 11:09 PM | Alfonso Nieto-Castanon - Boston University
2nd Level Multi-Session Result Interpretation
Hi Jeff,
Regarding the first reviewer question, the approach is not unusual at all, it is in fact rather standard (you may cite Henson and Penny's "ANOVAs and SPM" as one of the reference first introducing this approach in the neuroimaging literature, but the actual math equivalences behind this and described in https://www.nitrc.org/forum/message.php?... date back literally centuries)
Regarding the interpretation and need for post-hoc analyses, if I am interpreting it correctly the reviewer is simply right and this applies to all ANOVA and/or factorial designs/analyses. This is already true for a simple one-way ANOVAs (e.g. if you find a significant difference between the groups in connectivity, how do you interpret it? is the connection present in one group and not in the other? or is it present in both but with different strengths? are these positive or anticorrelations that we are talking about? etc. all of these things affect your interpretation and all of them are consistent with the found significant main effect). In two-way, three-way, or higher-order ANOVAs the need for interpretation, and the amount of different interpretations consistent with a significant interaction, simply grows exponentially. For example, your three-way group x time x working-memory-load interaction, would be consistent with many different scenarios. E.g. it might be that in control subjects, the differences in connectivity between the 2-back and 1-back tasks are constant in time (baseline and follow-up) while in patients those differences in connectivity decrease/increase? or it might be that in control subjects there are no differences at all in connectivity between the 2-back and 1-back tasks at neither baseline nor follow-up, while in patients there is are abnormal connectivity values in 2-back at baseline only? or many/many other potential scenarios all consistent with a significant three-way interaction but with different interpreations or clinical consequences. So the recommendation in ANOVA is always to display the connectivity in each of the individual cells (in your case 8 cells = 2x2x2) and base your interpretation on those or additional post-hoc analyses.
Hope this helps
Alfonso
Originally posted by Jeff Browndyke:
Regarding the first reviewer question, the approach is not unusual at all, it is in fact rather standard (you may cite Henson and Penny's "ANOVAs and SPM" as one of the reference first introducing this approach in the neuroimaging literature, but the actual math equivalences behind this and described in https://www.nitrc.org/forum/message.php?... date back literally centuries)
Regarding the interpretation and need for post-hoc analyses, if I am interpreting it correctly the reviewer is simply right and this applies to all ANOVA and/or factorial designs/analyses. This is already true for a simple one-way ANOVAs (e.g. if you find a significant difference between the groups in connectivity, how do you interpret it? is the connection present in one group and not in the other? or is it present in both but with different strengths? are these positive or anticorrelations that we are talking about? etc. all of these things affect your interpretation and all of them are consistent with the found significant main effect). In two-way, three-way, or higher-order ANOVAs the need for interpretation, and the amount of different interpretations consistent with a significant interaction, simply grows exponentially. For example, your three-way group x time x working-memory-load interaction, would be consistent with many different scenarios. E.g. it might be that in control subjects, the differences in connectivity between the 2-back and 1-back tasks are constant in time (baseline and follow-up) while in patients those differences in connectivity decrease/increase? or it might be that in control subjects there are no differences at all in connectivity between the 2-back and 1-back tasks at neither baseline nor follow-up, while in patients there is are abnormal connectivity values in 2-back at baseline only? or many/many other potential scenarios all consistent with a significant three-way interaction but with different interpreations or clinical consequences. So the recommendation in ANOVA is always to display the connectivity in each of the individual cells (in your case 8 cells = 2x2x2) and base your interpretation on those or additional post-hoc analyses.
Hope this helps
Alfonso
Originally posted by Jeff Browndyke:
========================================================================
Added from other thread as it has direct applicability to Nicole's question and possible needs for post-hoc explanation
========================================================================
Hi, Alfonso.
Just to tag onto your response regarding the three-way Group X Condition x Condition interaction model and set-up in CONN.
As you know, we performed something similar for our patients and controls at two time points (baseline and follow-up) looking at a working memory load effect (2-back > 1-back). We found a region for this three-way group x time x working memory load contrast interaction, and we have compared that region with performance on a separate measure of cognition. The results were written up and sent off for review, and I received the following question from reviewer with respect to the CONN model and three-way interaction:
"Please provide justification for the (unusual) approach to test an interaction analysis on the first level. Which analyses were conducted on the second level? Two-sample t-test? How do you know what the results on the second level mean? There are many possibilities that can produce a significant interaction effect. Thus, you need to show post-hoc t-tests to disentangle interaction findings."
Any ideas on how to best address these questions? My understanding is that CONN is comparing the difference of time and condition factors between groups at the 2nd level. Are post-hoc t-tests even necessary to "disentangle interaction findings" when handled as t-tests of difference images?
Warm regards,
Jeff
Added from other thread as it has direct applicability to Nicole's question and possible needs for post-hoc explanation
========================================================================
Hi, Alfonso.
Just to tag onto your response regarding the three-way Group X Condition x Condition interaction model and set-up in CONN.
As you know, we performed something similar for our patients and controls at two time points (baseline and follow-up) looking at a working memory load effect (2-back > 1-back). We found a region for this three-way group x time x working memory load contrast interaction, and we have compared that region with performance on a separate measure of cognition. The results were written up and sent off for review, and I received the following question from reviewer with respect to the CONN model and three-way interaction:
"Please provide justification for the (unusual) approach to test an interaction analysis on the first level. Which analyses were conducted on the second level? Two-sample t-test? How do you know what the results on the second level mean? There are many possibilities that can produce a significant interaction effect. Thus, you need to show post-hoc t-tests to disentangle interaction findings."
Any ideas on how to best address these questions? My understanding is that CONN is comparing the difference of time and condition factors between groups at the 2nd level. Are post-hoc t-tests even necessary to "disentangle interaction findings" when handled as t-tests of difference images?
Warm regards,
Jeff
Threaded View
| Title | Author | Date |
|---|---|---|
| Nicole Nissim | Jul 14, 2017 | |
| Jeff Browndyke | Sep 8, 2017 | |
| Alfonso Nieto-Castanon | Sep 16, 2017 | |
| Jeff Browndyke | Sep 17, 2017 | |
| Alfonso Nieto-Castanon | Aug 30, 2017 | |
| Nicole Nissim | Sep 7, 2017 | |