help
help > RE: Cluster-forming threshold for 1st l. analysis
Feb 12, 2016 09:02 PM | Alfonso Nieto-Castanon - Boston University
RE: Cluster-forming threshold for 1st l. analysis
Dear Johann,
Not exactly. The first-level pFDR images provided by CONN apply a voxel-wise False Discovery Rate correction, not a cluster-wise FDR correction (e.g. if thresholding using pFDR<.05 the implication is that among the resulting suprathreshold voxels less than 5% of these voxels will be false positives; compare that to cluster-wise pFDR where the implication is that among the resulting suprathreshold clusters less than 5% of these clusters will be false positives).
The voxel-wise pFDR values are computed by using the uncorrected p-values associated with the correlation values and then applying an FDR correction across the entire brain. This is not the same as first thresholding the correlation values based on uncorrected voxel-level p-values (cluster-forming threshold) and then applying a family-wise or FDR correction across the resulting clusters (e.g. cluster-level p-FDR or p-FWE thresholds). The latter would require computing cluster-level p-values (estimating the distribution of cluster sizes under the null hypothesis), which require either SPM smoothness estimation in combination with Random Field Theory assumptions, or randomization/permutation tests (in CONN both of these options are only available for second-level analyses, but not for first-level correlation maps). If you want to be able to use a combination of both voxel- and cluster- level thresholds on your first-level correlation maps I would probably suggest to replicate these first-level analyses within SPM (unfortunately you cannot simply use CONN's output files to infer cluster-level stats). Another alternative would be to use for example alphaSim to estimate the maximum cluster size expected under the null hypothesis for these first-level analyses and then manually thresholding your resulting clusters with the resulting cluster-size threshold (i.e. threshold CONN's p_corr values below your chosen cluster-forming threshold and then disregard any resulting cluster with size below the alphaSim-estimated threshold)
Hope this helps
Alfonso
Originally posted by Johann Philipp Zöllner:
Not exactly. The first-level pFDR images provided by CONN apply a voxel-wise False Discovery Rate correction, not a cluster-wise FDR correction (e.g. if thresholding using pFDR<.05 the implication is that among the resulting suprathreshold voxels less than 5% of these voxels will be false positives; compare that to cluster-wise pFDR where the implication is that among the resulting suprathreshold clusters less than 5% of these clusters will be false positives).
The voxel-wise pFDR values are computed by using the uncorrected p-values associated with the correlation values and then applying an FDR correction across the entire brain. This is not the same as first thresholding the correlation values based on uncorrected voxel-level p-values (cluster-forming threshold) and then applying a family-wise or FDR correction across the resulting clusters (e.g. cluster-level p-FDR or p-FWE thresholds). The latter would require computing cluster-level p-values (estimating the distribution of cluster sizes under the null hypothesis), which require either SPM smoothness estimation in combination with Random Field Theory assumptions, or randomization/permutation tests (in CONN both of these options are only available for second-level analyses, but not for first-level correlation maps). If you want to be able to use a combination of both voxel- and cluster- level thresholds on your first-level correlation maps I would probably suggest to replicate these first-level analyses within SPM (unfortunately you cannot simply use CONN's output files to infer cluster-level stats). Another alternative would be to use for example alphaSim to estimate the maximum cluster size expected under the null hypothesis for these first-level analyses and then manually thresholding your resulting clusters with the resulting cluster-size threshold (i.e. threshold CONN's p_corr values below your chosen cluster-forming threshold and then disregard any resulting cluster with size below the alphaSim-estimated threshold)
Hope this helps
Alfonso
Originally posted by Johann Philipp Zöllner:
Dear Alfonso,
thanks as always for your succinct reply! I implemented your advice but I'm afraid still fail to grasp the statistical implications here...
If I use the pFDR images to mask my corr-maps using the imcalc method described above, would this be equivalent to masking the corr-maps with the p_corr images and *then* performing an FDR correction on the resulting masked files using conn_fdr? Is the latter even a statistically correct way to do this? I tought this up because I want to control the cluster-forming threshold a bit like it's possible in SPM.
I know this might be very specific and a bit of an awry method with regards to CONN's working structure, but maybe you (or someone statistically inclined) can answer this!
Thanks a lot!
Johann Philipp
thanks as always for your succinct reply! I implemented your advice but I'm afraid still fail to grasp the statistical implications here...
If I use the pFDR images to mask my corr-maps using the imcalc method described above, would this be equivalent to masking the corr-maps with the p_corr images and *then* performing an FDR correction on the resulting masked files using conn_fdr? Is the latter even a statistically correct way to do this? I tought this up because I want to control the cluster-forming threshold a bit like it's possible in SPM.
I know this might be very specific and a bit of an awry method with regards to CONN's working structure, but maybe you (or someone statistically inclined) can answer this!
Thanks a lot!
Johann Philipp
Threaded View
| Title | Author | Date |
|---|---|---|
| Johann Philipp Zöllner | Feb 3, 2016 | |
| Alfonso Nieto-Castanon | Feb 4, 2016 | |
| Johann Philipp Zöllner | Feb 11, 2016 | |
| Alfonso Nieto-Castanon | Feb 12, 2016 | |