help
help > RE: second-level analysis in Conn
Oct 8, 2015 01:10 PM | Alfonso Nieto-Castanon - Boston University
RE: second-level analysis in Conn
Hi Diana,
Some thoughts on your questions below
Best
Alfonso
Originally posted by Diana Parvinchi:
No, the "analysis results" table multiple-comparison correction only corrects at the analysis-level (i.e. for that particular analysis -with a particular seed ROI and particular symptom score- it corrects across multiple target ROIs). Since you are performing multiple analyses (e.g. you are repeating this procedure for multiple seed ROIs) you would typically need to apply an additional correction to take that into account (unless each repeated analysis can be reasonably framed as a separate individual experimental hypothesis). Typically, in addition to simply using a more conservative Bonferroni-corrected set of individual thresholds, there are a couple of more-senstive ways to correct across multiple seed ROIs:
a) if the set of seed ROIs is not too large, you may simply perform an F-test across all of your seeds: i.e. repeat your analyses, but instead of selecting a single seed ROI in the "sources" list select now all of the seed ROIs that you want to test simultaneously and leave the default eye(N) between-sources contrast. That will test for connectivity associations with clinical scores across any of the selected seed ROIs
or b) alternatively click on "results explorer", then select in the "sources" list all of your seed ROIs. You will be now looking at the matrix of effects between all selected seed ROIs and all selected target ROIs, and you have multiple ways to threshold the results here while properly correcting for multiple comparisons (e.g. using seed-level thresholds as well as connection-level thresholds, using network-based-statistics, etc.)
2) Given that the symptoms severity scores are not standardized, meaning a score of "0" indicates the absence of that symptom, not the average, is our interpretation (e.g. functional connectivity between right Intracalcarine and right Frontal Orbital Cortex is negatively correlated with severity of Repetitive Behaviour in our third group) correct in the attached graph?
Yes, the centering of the symptom score variables does not affect the interpretation of the sign of the association (between connectivity values and symptom scores). What changes with centering is the interpretation of the y-intercepts (the first three bars in your plot). In your case it means that the connectivity between those regions is typically (in the absence of symptoms) positive, and that this positive connectivity decreases with severity of symptoms.
3) When we see a significant effect in one of our three age-groups, how can I run pairwise comparisons (following an interaction effect) to report the p-value associated with that particular effect? For example, in the attached graph, there's significant correlation between connectivity and repetitive behaviour only in our third group, how can I determine the p-value for this particular effect?
To look at the association between connectivity and symptom scores only in your third group, simply select the 'Group3' and 'Group3_scores' effects in the "subject-effects" list and enter a [0 1] between-subjects contrast.
Hope this helps
Alfonso
Some thoughts on your questions below
Best
Alfonso
Originally posted by Diana Parvinchi:
Hi Alfonso,
Thank you very much for all your help. I have three questions and deeply appreciate your help in confirming these issues before we submit the results for publication purposes. Just to refresh your mind, we examined the correlation between functional connectivity and core symptoms of autism spectrum disorder in three age-groups of participants. This is how I analyzed the data - followed by my three questions:
I selected 6 regressors (three age-groups and their corresponding scores on a given clinical measure) from the "subject effects" list in the 2nd-level results (ROI to ROI), entered the contrast [0 0 0 1 0 0;0 0 0 0 1 0; 0 0 0 0 0 1] in the "between-groups contrasts", then selected a single ROI from our list of ROIs and ran the contrast. I repeated this process for each ROI in our list (for each symptom - total of 5 symptoms).
1) Are the results shown in the "analysis results" table corrected for multiple comparisons, given the route I took explained above?
Thank you very much for all your help. I have three questions and deeply appreciate your help in confirming these issues before we submit the results for publication purposes. Just to refresh your mind, we examined the correlation between functional connectivity and core symptoms of autism spectrum disorder in three age-groups of participants. This is how I analyzed the data - followed by my three questions:
I selected 6 regressors (three age-groups and their corresponding scores on a given clinical measure) from the "subject effects" list in the 2nd-level results (ROI to ROI), entered the contrast [0 0 0 1 0 0;0 0 0 0 1 0; 0 0 0 0 0 1] in the "between-groups contrasts", then selected a single ROI from our list of ROIs and ran the contrast. I repeated this process for each ROI in our list (for each symptom - total of 5 symptoms).
1) Are the results shown in the "analysis results" table corrected for multiple comparisons, given the route I took explained above?
No, the "analysis results" table multiple-comparison correction only corrects at the analysis-level (i.e. for that particular analysis -with a particular seed ROI and particular symptom score- it corrects across multiple target ROIs). Since you are performing multiple analyses (e.g. you are repeating this procedure for multiple seed ROIs) you would typically need to apply an additional correction to take that into account (unless each repeated analysis can be reasonably framed as a separate individual experimental hypothesis). Typically, in addition to simply using a more conservative Bonferroni-corrected set of individual thresholds, there are a couple of more-senstive ways to correct across multiple seed ROIs:
a) if the set of seed ROIs is not too large, you may simply perform an F-test across all of your seeds: i.e. repeat your analyses, but instead of selecting a single seed ROI in the "sources" list select now all of the seed ROIs that you want to test simultaneously and leave the default eye(N) between-sources contrast. That will test for connectivity associations with clinical scores across any of the selected seed ROIs
or b) alternatively click on "results explorer", then select in the "sources" list all of your seed ROIs. You will be now looking at the matrix of effects between all selected seed ROIs and all selected target ROIs, and you have multiple ways to threshold the results here while properly correcting for multiple comparisons (e.g. using seed-level thresholds as well as connection-level thresholds, using network-based-statistics, etc.)
2) Given that the symptoms severity scores are not standardized, meaning a score of "0" indicates the absence of that symptom, not the average, is our interpretation (e.g. functional connectivity between right Intracalcarine and right Frontal Orbital Cortex is negatively correlated with severity of Repetitive Behaviour in our third group) correct in the attached graph?
Yes, the centering of the symptom score variables does not affect the interpretation of the sign of the association (between connectivity values and symptom scores). What changes with centering is the interpretation of the y-intercepts (the first three bars in your plot). In your case it means that the connectivity between those regions is typically (in the absence of symptoms) positive, and that this positive connectivity decreases with severity of symptoms.
3) When we see a significant effect in one of our three age-groups, how can I run pairwise comparisons (following an interaction effect) to report the p-value associated with that particular effect? For example, in the attached graph, there's significant correlation between connectivity and repetitive behaviour only in our third group, how can I determine the p-value for this particular effect?
To look at the association between connectivity and symptom scores only in your third group, simply select the 'Group3' and 'Group3_scores' effects in the "subject-effects" list and enter a [0 1] between-subjects contrast.
Hope this helps
Alfonso