Ongoing list so I don't forget/miss the comments in the code when I've chosen some other route than specified:
Selecting the best de-noising pipeline was not apparent from the different metrics (probably due to noise in the bad voxels). The choice (24HMP_aCompCor_SpikeReg_4GSR) was also informed by which pipeline is the most rigorous, appears to correlate the least with the null model, one and by looking at fmridenoised connectivity matrix in the report.
The procedure to choose the GMM for voxel thresholding was based on prediction in relation to other Gaussians rather than thresholding the largest Gaussian. This is a more sound method than specified in the preregistration as the specified method did not care about the context of the other Gaussians. Also we chose to keep all but the lowest GMM instead of only keeping the highest. This was based on the visual inspection of the GMMs to not enforce a Gaussian shape on the average voxel histogram (i.e. multiple Gaussians together sometimes captured the shape better).
Adjusted mutual information is used instead of normalized mutual information with regards to comparing the cortex communities with the Yeo template communities. Otherwise this produced one large community or very small communities as having the best NMI. AMI corrects for this.
Not explicitly stated in the preregistration. Only positive edges are considered in the community detection step and the participation coefficient.
Not explicitly stated (but logical). There were some runs where there were multiple stimulation sites during the es run. These are excluded.
Unclear how the community detection in analysis one was to be calculated. Given Thompson et al 2020 HBM, we now know that it is best to use the same community partition for comparing PC values (which would be important in analysis 2). This was a motivation to use the post-op data instead of the pre-op data to calculate the stim site properties on. Thus, instead of using the preop data where we know there are differences (and would reduce the number of participants), we decided to concatenate all the es-runs and use this to identify the stimulation sites community and categorize the topographic properties of the stimulation sites. We calculated the FC on the concatenated FC data and derived the community detection. To identify the stimulation site, we placed a sphere between the two channels and took identified the stimulation site's parcel as the parcel that had the most overlap with the sphere. The preregistration stated that this would be 3mm, but for some subjects no parcel lay in that radius. Subsequently, we expanded the radius of the center by 1mm increments until all subjects had a parcel associated with the stimulation sites. All subjects had a parcel that could be assigned at 6mm radius, so this was the choice. This choice was done to maximize the amount of data in the next step. Importantly the overlapping ROIs were calculated on the smörgåsbord parcellation, not the subject maps. However, there is still a possibility that the overlapping parcel is not included in the subject's mask due to noise. If the parcel identified as the stimsite was not in the subject mask, then the run was dropped.
The initial pre-registered plan was to calculate the es per stimulation sequence (i.e. multiple connectivity estimates per run) and then place this in a hierarhical statistical model with subject as a layer. When writing the preregistration, we did not realize that this would only be <10 data points per stimulation run. We deemed this too few to get a accurate and this analysis was never performed. Instead, we did away. However, because there was no longer multiple PC estimates for each stimulation site, we now needed to use the displacement (i.e. es-on - es-off) instead of just es-on. Without comparing the change to baseline, the values will be hard to compare, as the stimulation sites may belong to communities with different PC-properties. Thus, we had to use the displacement values here to achieve meaningful results.
Ongoing list so I don't forget/miss the comments in the code when I've chosen some other route than specified:
Selecting the best de-noising pipeline was not apparent from the different metrics (probably due to noise in the bad voxels). The choice (24HMP_aCompCor_SpikeReg_4GSR) was also informed by which pipeline is the most rigorous, appears to correlate the least with the null model, one and by looking at fmridenoised connectivity matrix in the report.
The procedure to choose the GMM for voxel thresholding was based on prediction in relation to other Gaussians rather than thresholding the largest Gaussian. This is a more sound method than specified in the preregistration as the specified method did not care about the context of the other Gaussians. Also we chose to keep all but the lowest GMM instead of only keeping the highest. This was based on the visual inspection of the GMMs to not enforce a Gaussian shape on the average voxel histogram (i.e. multiple Gaussians together sometimes captured the shape better).
Adjusted mutual information is used instead of normalized mutual information with regards to comparing the cortex communities with the Yeo template communities. Otherwise this produced one large community or very small communities as having the best NMI. AMI corrects for this.
Not explicitly stated in the preregistration. Only positive edges are considered in the community detection step and the participation coefficient.
Not explicitly stated (but logical). There were some runs where there were multiple stimulation sites during the es run. These are excluded.
Unclear how the community detection in analysis one was to be calculated. Given Thompson et al 2020 HBM, we now know that it is best to use the same community partition for comparing PC values (which would be important in analysis 2). This was a motivation to use the post-op data instead of the pre-op data to calculate the stim site properties on. Thus, instead of using the preop data where we know there are differences (and would reduce the number of participants), we decided to concatenate all the es-runs and use this to identify the stimulation sites community and categorize the topographic properties of the stimulation sites. We calculated the FC on the concatenated FC data and derived the community detection. To identify the stimulation site, we placed a sphere between the two channels and took identified the stimulation site's parcel as the parcel that had the most overlap with the sphere. The preregistration stated that this would be 3mm, but for some subjects no parcel lay in that radius. Subsequently, we expanded the radius of the center by 1mm increments until all subjects had a parcel associated with the stimulation sites. All subjects had a parcel that could be assigned at 6mm radius, so this was the choice. This choice was done to maximize the amount of data in the next step. Importantly the overlapping ROIs were calculated on the smörgåsbord parcellation, not the subject maps. However, there is still a possibility that the overlapping parcel is not included in the subject's mask due to noise. If the parcel identified as the stimsite was not in the subject mask, then the run was dropped.
The initial pre-registered plan was to calculate the es per stimulation sequence (i.e. multiple connectivity estimates per run) and then place this in a hierarhical statistical model with subject as a layer. When writing the preregistration, we did not realize that this would only be <10 data points per stimulation run. We deemed this too few to get a accurate and this analysis was never performed. Instead, we did away. However, because there was no longer multiple PC estimates for each stimulation site, we now needed to use the displacement (i.e. es-on - es-off) instead of just es-on. Without comparing the change to baseline, the values will be hard to compare, as the stimulation sites may belong to communities with different PC-properties. Thus, we had to use the displacement values here to achieve meaningful results.