help
help > RE: Average within / between network connectivity
Feb 12, 2018 10:02 AM | Helene Veenstra
RE: Average within / between network connectivity
Dear Stephen,
while I would prefer to use voxel-wise FDR, as I am very aware of the issues with using uncorrected or badly corrected p-values; for this specific CONN command line (conn_withinbetweenROItest) I am fairly sure uncorrected measures are used for the calculation between-network connectivity averaged over ROIs.
So I am where I am, doing a group study, having those values as output and in need to apply some sort of correction, or alternatively not use this command line for network investigation and comparison at all, and instead go for ROI-ROI connectivity values. However, since my hypothesis is in regard of whole networks and their between-network connectivity (DMN and Salience Networks), I would prefer to find a way to incorporate network analysis.
Helene
Originally posted by Stephen L.:
while I would prefer to use voxel-wise FDR, as I am very aware of the issues with using uncorrected or badly corrected p-values; for this specific CONN command line (conn_withinbetweenROItest) I am fairly sure uncorrected measures are used for the calculation between-network connectivity averaged over ROIs.
So I am where I am, doing a group study, having those values as output and in need to apply some sort of correction, or alternatively not use this command line for network investigation and comparison at all, and instead go for ROI-ROI connectivity values. However, since my hypothesis is in regard of whole networks and their between-network connectivity (DMN and Salience Networks), I would prefer to find a way to incorporate network analysis.
Helene
Originally posted by Stephen L.:
Dear Helene,
I cannot answer for the second question, but for the first unfortunately p-uncorrected can never be assimilated to Bonferroni correction, as p-uncorrected does not perform any multiple comparison correction. Even for small volumes, it will always be more optimistic than Bonferroni. For more infos and a demonstration, you can read this excellent wiki page, where they advise to rather show unthresholded maps (but I don't know how well they are received by journal editors): http://imaging.mrc-cbu.cam.ac.uk/imaging/UncorrectedThreshold
An alternative can be to either use voxel-wise FDR, or even cluster-size FDR, as topological (cluster-wise) correction is more liberal than FWE.
Another alternative, particularly if you are doing a single case study, is to use non-parametric correction with a higher cluster-size or cluster-mass threshold than 0.05, using Alfonso's patch which fixes a few edge cases with non-parametric correction:
https://www.nitrc.org/forum/message.php?...
Indeed, it is usually recommended not to raise the cluster-size threshold above 0.05 nor the voxel-wise p-uncorrected threshold of 0.001 when doing parametric correction since Eklund et al's "Cluster failure" paper, because higher thresholds would lead to unquantified and exponentially higher nominal false positive rates, but with non-parametric correction you can use higher thresholds (both at voxel-wise and cluster-wise levels) as they reliably correct to the nominal rate as demonstrated in the same paper (and in others).
Hope this helps,
Stephen Larroque
I cannot answer for the second question, but for the first unfortunately p-uncorrected can never be assimilated to Bonferroni correction, as p-uncorrected does not perform any multiple comparison correction. Even for small volumes, it will always be more optimistic than Bonferroni. For more infos and a demonstration, you can read this excellent wiki page, where they advise to rather show unthresholded maps (but I don't know how well they are received by journal editors): http://imaging.mrc-cbu.cam.ac.uk/imaging/UncorrectedThreshold
An alternative can be to either use voxel-wise FDR, or even cluster-size FDR, as topological (cluster-wise) correction is more liberal than FWE.
Another alternative, particularly if you are doing a single case study, is to use non-parametric correction with a higher cluster-size or cluster-mass threshold than 0.05, using Alfonso's patch which fixes a few edge cases with non-parametric correction:
https://www.nitrc.org/forum/message.php?...
Indeed, it is usually recommended not to raise the cluster-size threshold above 0.05 nor the voxel-wise p-uncorrected threshold of 0.001 when doing parametric correction since Eklund et al's "Cluster failure" paper, because higher thresholds would lead to unquantified and exponentially higher nominal false positive rates, but with non-parametric correction you can use higher thresholds (both at voxel-wise and cluster-wise levels) as they reliably correct to the nominal rate as demonstrated in the same paper (and in others).
Hope this helps,
Stephen Larroque
Threaded View
Title | Author | Date |
---|---|---|
Ruth Shaffer | Apr 27, 2016 | |
Alfonso Nieto-Castanon | May 4, 2016 | |
Chabran Eléna | Sep 20, 2018 | |
Maria Beser | Feb 11, 2022 | |
max345 | Feb 1, 2022 | |
Alfonso Nieto-Castanon | Feb 3, 2022 | |
max345 | Feb 4, 2022 | |
Alfonso Nieto-Castanon | Feb 4, 2022 | |
max345 | Feb 9, 2022 | |
max345 | Feb 5, 2022 | |
Himanshu Joshi | Feb 16, 2018 | |
Alfonso Nieto-Castanon | Apr 28, 2018 | |
Himanshu Joshi | Apr 9, 2018 | |
Himanshu Joshi | Apr 28, 2018 | |
Himanshu Joshi | Mar 13, 2018 | |
Himanshu Joshi | Feb 26, 2018 | |
Jon Dudley | May 4, 2016 | |
Alfonso Nieto-Castanon | May 5, 2016 | |
Helene Veenstra | Feb 8, 2018 | |
Alfonso Nieto-Castanon | Feb 14, 2018 | |
Helene Veenstra | Feb 15, 2018 | |
Stephen L. | Feb 8, 2018 | |
Helene Veenstra | Feb 12, 2018 | |
Mariah DeSerisy | Apr 30, 2020 | |
Elizabeth Olson | May 16, 2020 | |
Alfonso Nieto-Castanon | May 19, 2020 | |
Nadine Smith | Feb 23, 2021 | |
ryanw412 | Dec 16, 2021 | |
Jose Maximo | Apr 28, 2021 | |
Olivier Roy | Oct 15, 2021 | |
Karli Carston | Nov 9, 2021 | |