When we compute parcellated connectivity, we first compute the average timeseries within the parcels, and then correlate those, as it vastly reduces the impact of noise. If we first computed the correlations, and then averaged them within parcels, we would be losing a huge amount of power.
The per-run correlating first and then averaging that you propose sounds like a similar situation, though because there are only 4 runs, and each run has lots of timepoints, and the averaging isn't spatial, it won't be nearly as dramatic a difference. Keep in mind that the phase encoding direction dictates where signal dropouts will be, which will show up in any analysis of non-concatenated data. Tim On Wed, Mar 7, 2018 at 4:02 PM, Harms, Michael <mha...@wustl.edu> wrote: > > > In the case of correlations or partial correlations, I would tend to > compute those separately for each run anyway, Fisher transform them, and > then average the r-to-z values across runs. In which case no across-run > concatentation is necessary in the first place. > > > > I don’t know if a per-run DCM approach, followed by averaging of the DCM > outputs is a possibility. If it is, you might just want to consider that > approach instead. > > > > Cheers, > > -MH > > > > -- > > Michael Harms, Ph.D. > > ----------------------------------------------------------- > > Associate Professor of Psychiatry > > Washington University School of Medicine > > Department of Psychiatry, Box 8134 > > 660 South Euclid Ave > <https://maps.google.com/?q=660+South+Euclid+Ave&entry=gmail&source=g>. > Tel: 314-747-6173 <(314)%20747-6173> > > St. Louis, MO 63110 Email: mha...@wustl.edu > > > > *From: *"Glasser, Matthew" <glass...@wustl.edu> > *Date: *Wednesday, March 7, 2018 at 3:39 PM > *To: *David Hofmann <davidhofma...@gmail.com> > > *Cc: *"Harms, Michael" <mha...@wustl.edu>, hcp-users < > hcp-users@humanconnectome.org> > *Subject: *Re: [HCP-Users] Concatenating resting state runs > > > > The basic idea for variance normalization is to equalize the variance of > the noise. It is very helpful for ICA and regression-based techniques. > I’m not sure we have explicitly tested the effect on correlation. > Correlation is a ratio and so it would not matter at all for a single run, > though there may be benefits to doing variance normalization prior to > concatenation for correlation. Not sure of how this will interact with DCM > either. > > > > Peace, > > > > Matt. > > > > *From: *David Hofmann <davidhofma...@gmail.com> > *Date: *Wednesday, March 7, 2018 at 3:29 PM > *To: *Matt Glasser <glass...@wustl.edu> > *Cc: *"Harms, Michael" <mha...@wustl.edu>, hcp-users < > hcp-users@humanconnectome.org> > *Subject: *Re: [HCP-Users] Concatenating resting state runs > > > > Hi all, > > > > that being said, why is this regression approach for variance > normalization superior to a z-standardization? That is, will it practically > matter e.g. for correlations or partial correlations? > > > > 2018-03-07 19:31 GMT+01:00 Glasser, Matthew <glass...@wustl.edu>: > > Hi Mike, > > > > I doubt that matters for this application of making an unstructured noise > timeseries for the purpose of variance normalization. > > > > Matt. > > > > *From: *"Harms, Michael" <mha...@wustl.edu> > *Date: *Wednesday, March 7, 2018 at 12:09 PM > > > *To: *Matt Glasser <glass...@wustl.edu>, David Hofmann < > davidhofma...@gmail.com> > *Cc: *hcp-users <hcp-users@humanconnectome.org> > *Subject: *Re: [HCP-Users] Concatenating resting state runs > > > > > > Hi Matt, > > Right, that recipe is straightforward, but for completeness there should > be two additional steps if one wants to match the FIX cleaning precisely: > > 1) the 24 motion parameters should be filtered with the same HP filter > applied to the data > > 2) those HP filtered 24 motion parameters should then be removed from the > (‘signal’) ICA time-series prior to regressing that (modified) ICA > time-series onto the cleaned data (i.e., that modified ICA time-series > becomes the basis for deriving ‘betaICA’). > > > > Cheers, > > -MH > > > > -- > > Michael Harms, Ph.D. > > ----------------------------------------------------------- > > Associate Professor of Psychiatry > > Washington University School of Medicine > > Department of Psychiatry, Box 8134 > > 660 South Euclid Ave > <https://maps.google.com/?q=660+South+Euclid+Ave&entry=gmail&source=g>. > Tel: 314-747-6173 <(314)%20747-6173> > > St. Louis, MO 63110 Email: mha...@wustl.edu > > > > *From: *"Glasser, Matthew" <glass...@wustl.edu> > *Date: *Wednesday, March 7, 2018 at 11:24 AM > *To: *"Harms, Michael" <mha...@wustl.edu>, David Hofmann < > davidhofma...@gmail.com> > *Cc: *hcp-users <hcp-users@humanconnectome.org> > *Subject: *Re: [HCP-Users] Concatenating resting state runs > > > > Hi Mike, > > > > Not for the volume data that he is asking about and not for the MSMAll > data either unfortunately. I thought it was better to explain this method > on the list so that it can be applied to arbitrary data whether or not we > precomputed it. > > > > Matt. > > > > *From: *"Harms, Michael" <mha...@wustl.edu> > *Date: *Wednesday, March 7, 2018 at 11:21 AM > *To: *Matt Glasser <glass...@wustl.edu>, David Hofmann < > davidhofma...@gmail.com> > *Cc: *hcp-users <hcp-users@humanconnectome.org> > *Subject: *Re: [HCP-Users] Concatenating resting state runs > > > > > > Matt, > > Don’t we compute an estimate of the unstructured noise variance as part of > RestingStateState, and then place that into one of the packages? > > > > > > -- > > Michael Harms, Ph.D. > > ----------------------------------------------------------- > > Associate Professor of Psychiatry > > Washington University School of Medicine > > Department of Psychiatry, Box 8134 > > 660 South Euclid Ave > <https://maps.google.com/?q=660+South+Euclid+Ave&entry=gmail&source=g>. > Tel: 314-747-6173 <(314)%20747-6173> > > St. Louis, MO 63110 Email: mha...@wustl.edu > > > > *From: *<hcp-users-boun...@humanconnectome.org> on behalf of "Glasser, > Matthew" <glass...@wustl.edu> > *Date: *Wednesday, March 7, 2018 at 11:01 AM > *To: *David Hofmann <davidhofma...@gmail.com> > *Cc: *hcp-users <hcp-users@humanconnectome.org> > *Subject: *Re: [HCP-Users] Concatenating resting state runs > > > > Yes they should be in that same package: > > > > ${StudyFolder}/${Subject}/MNINonLinear/Results/${ > fMRIName}/${fMRIName}_hp2000.ica/.fix — Tells you which are the noise > components (so you can use setdiff to find the signal components from a > list of all components) so that you can exclude the noise component from > the regression below. > > ${StudyFolder}/${Subject}/MNINonLinear/Results/${ > fMRIName}/${fMRIName}_hp2000.ica/filtered_func_data.ica/melodic_mix — ICA > component timeseries (you should remove the mean of each ICA component > timeseries before doing the regression). > > > > Probably the time to read in and write the file will be longer than the > time to do the regression if you do it in matlab. Here is some example > code: > > > > betaICA = pinv(ICA) * TCS; #TCS is timepoints x space, ICA is timepoints x > components and should include only the signal components (since the noise > components were already removed). > > UnstructNoiseTCS = TCS - (ICA * betaICA); > > > > You then compute the temporal standard deviation of the unstructured noise > timeseries and divide the data by it to get the variance normalized data. > > > > Peace, > > > > Matt. > > > > *From: *David Hofmann <davidhofma...@gmail.com> > *Date: *Wednesday, March 7, 2018 at 10:47 AM > *To: *Matt Glasser <glass...@wustl.edu> > *Cc: *hcp-users <hcp-users@humanconnectome.org> > *Subject: *Re: [HCP-Users] Concatenating resting state runs > > > > Ah I understand. However, I'm not sure how to do this practically for the > FIX extended data. I'd need all the signal component timeseries and run a > regression for each voxel which might take a while. I'm not sure if the > signals are supplied in the dataset, or are they? > > > > Thanks for the support! > > > > 2018-03-07 17:07 GMT+01:00 Glasser, Matthew <glass...@wustl.edu>: > > The unstructured noise variance is the standard deviation of the > timeseries after you regress out all of the signal component timeseries. > By doing this you make the unstructured noise equal in magnitude across the > brain. > > > > I wouldn’t do smoothing unless it is constrained to the greymatter. > Really you won’t get an obvious benefit if you will be averaging voxels in > an ROI anyway and that is a more accurate way to do things. > > > > I guess I don’t know enough about your study to know if the order > matters. If you are interested in effects that might be related to order > (e.g. drowsiness being higher in later scans, then order might matter). > > > > Peace, > > > > Matt. > > > > *From: *David Hofmann <davidhofma...@gmail.com> > *Date: *Wednesday, March 7, 2018 at 10:02 AM > > > *To: *Matt Glasser <glass...@wustl.edu> > *Cc: *hcp-users <hcp-users@humanconnectome.org> > *Subject: *Re: [HCP-Users] Concatenating resting state runs > > > > Hey Matthew, > > > > not sure I understood where to get the unstructured noise variance from, > i.e. is it even possible to apply this to the FIX extended datasets? > > > > I thought about using 4mm smoothing (maybe 2mm) before extracting the VOIs > / ROI timecourses for each subject. This is then fed into the DCMs for each > subject. I experimented with some HCP data before and it seems > smoothing increases the effect sizes a little bit. What is smoothing > between parcellations btw.? > > > > Also, any comments on the order of concatenation? I concatenate all of the > data RL and then LR. > > > > 2018-03-07 16:17 GMT+01:00 Glasser, Matthew <glass...@wustl.edu>: > > I typically variance normalize before concatenation, but do this based on > the unstructured noise variance. > > > > I would take the mean time course over an ROI that I thought to be > representative of a meaningful neuroanatomical subunit. > > > > My understanding of how SPM’s DCM is typically implemented is that there > are large amounts of spatial smoothing, cross-subject alignment is done in > the volume, and ROIs are spheres of some radius. All this would lead to a > lot of mixing of timecourses. My suggestion was to use parcel timecourses > from some kind of parcellation. If you have a good amygdala parcellation > that might be fine, though I would avoid smoothing the data between the > parcels. > > > > Peace, > > > > Matt. > > > > *From: *David Hofmann <davidhofma...@gmail.com> > *Date: *Wednesday, March 7, 2018 at 9:12 AM > *To: *Matt Glasser <glass...@wustl.edu> > *Cc: *hcp-users <hcp-users@humanconnectome.org> > *Subject: *Re: [HCP-Users] Concatenating resting state runs > > > > Hi Matthew, > > > > ok, so temporal filtering separately for each run. Any comments on > concatenation and z-standardization? > > > > I think there might be a work-around to supplying a custom ROI timecourse > to the DCM VOI-files somehow, but which values to input as alternative to > the eigenvariate? The mean over all voxels in the ROI would also be an > option but not sure what you had in mind. > > > > Can you elaborate on the issue of spatial localization you mention please, > not sure I understood? I'm using mask files to extract the time courses and > I am especially interested in amygdala subregions. > > > > Also, what do you mean by areal ROIs and that they give a purer signal? > > > > Thanks :) > > > > 2018-03-07 14:51 GMT+01:00 Glasser, Matthew <glass...@wustl.edu>: > > You would want to apply temporal filtering separately to each run. I > wonder if there is a way you could just provide the ROI timecourses to > SPM’s DCM model without using its tools for extracting the ROIs so that you > could avoid the issues spatial localization that SPM has. If you used > areal ROIs, you likely wouldn’t even need the eigenvariate approach as you > would be getting a much purer signal. > > > > Peace, > > > > Matt. > > > > *From: *<hcp-users-boun...@humanconnectome.org> on behalf of David > Hofmann <davidhofma...@gmail.com> > *Date: *Wednesday, March 7, 2018 at 2:32 AM > *To: *hcp-users <hcp-users@humanconnectome.org> > *Subject: *[HCP-Users] Concatenating resting state runs > > > > Hi all, > > > > for a later analysis where I extract ROIs with SPM, I need to concatenate > the resting state runs and want to make sure I'm doing it correctly. SPM > extracts the first eigenvariate of a ROI, i.e. the component that explains > the most variance. > > > > I'm using the* Resting State fMRI 1 FIX-Denoised (Extended)* and *Resting > State fMRI 2 FIX-Denoised (Extended)* datasets. That is, the > files: rfMRI_REST1_LR_hp2000_clean.nii, rfMRI_REST1_RL > _hp2000_clean.nii asf. > > > > I chose the following approach: > > > > 1. z-standardize each session (each voxel timecourse), i.e. RL, LR > separately > > 2. Then concatenate them > > 3. Run the SPM routines which will also apply a high-pass filter of about > 128s on the already concatenated data (it's for the processing of a DCM > rather than functional connectivity) > > > > I have the following questions: > > > > 1. Is this approach correct? > > 2. Does the order of concatenation matter? That is, (RL/LR or LR/RL) or is > it important to concatenate it in the order it was acquired in each > subject? I read that it sometimes changes between subjects such that LR > came first in one subject and RL first in another. > > 3. Since SPM will run a hp-filter on the concatenated data, would it be > better to hp filter each run *separately* before concatenation? > > 4. Is this approach also applicable to the task data (i.e. standardize and > filter separately before concatenation)? > > > > Thanks in advance > > > > David > > > > > > _______________________________________________ > HCP-Users mailing list > HCP-Users@humanconnectome.org > http://lists.humanconnectome.org/mailman/listinfo/hcp-users > > > > > > > > > > _______________________________________________ > HCP-Users mailing list > HCP-Users@humanconnectome.org > http://lists.humanconnectome.org/mailman/listinfo/hcp-users > > > > > ------------------------------ > > The materials in this message are private and may contain Protected > Healthcare Information or other information of a sensitive nature. If you > are not the intended recipient, be advised that any unauthorized use, > disclosure, copying or the taking of any action in reliance on the contents > of this information is strictly prohibited. If you have received this email > in error, please immediately notify the sender via telephone or return mail. > > _______________________________________________ > HCP-Users mailing list > HCP-Users@humanconnectome.org > http://lists.humanconnectome.org/mailman/listinfo/hcp-users > _______________________________________________ HCP-Users mailing list HCP-Users@humanconnectome.org http://lists.humanconnectome.org/mailman/listinfo/hcp-users
