 |
 |
HI George, Yes, this would be the same as the 2nd stage of the dual regression. All the best, Anderson On 28 September 2015 at 15:40, G Ch <[log in to unmask]> wrote: > Hi Anderson, > > 'Do you mean in Y=X*b+e, Y is the original timepoints by voxels and X the > components' spatial maps? That'd be the first stage of the dual regression. > Or do you mean regressing the components' spatial maps on the component's > timecourses? That wouldn't be meaningful. Or do you mean something else?' > > I meant regressing the subject-specific component time courses on the > original filtered_func_data image to obtain beta maps for each component, > and then multiplying beta values of the component of interest by the > original time course of each voxel to obtain a time course 'weighted' by > the component. Does it make any sense? I am guessing those beta values are > the equivalent of the parameter estimates from the second dual regression > stage. > > Thanks a lot again, > > George > > 2015-09-28 10:33 GMT+02:00 Anderson M. Winkler <[log in to unmask]>: > >> Hi George, >> >> Please see below: >> >> On 27 September 2015 at 21:16, G Ch <[log in to unmask]> wrote: >> >>> Thank you Anderson. One thing I also thought about is to obtain beta >>> values for each voxel and each component (by doing a multiple regression >>> using the subject specific timeseries), >>> >> >> Do you mean in Y=X*b+e, Y is the original timepoints by voxels and X the >> components' spatial maps? That'd be the first stage of the dual regression. >> Or do you mean regressing the components' spatial maps on the component's >> timecourses? That wouldn't be meaningful. Or do you mean something else? >> >> >>> and multiplying the beta value for the component of interest in each >>> voxel by the original time series. This should get us an estimate of the >>> 'contribution' of the component to the original time course, before doing >>> further analyses on said time series. What do you think? I know I am >>> playing well out of my field, but I am just wondering if I can get anywhere >>> in trying to answer my original question. >>> >> >> In my previous email (quoted below) I left some questions. They are >> answerable. Perhaps you could try splitting the DMN as in your original >> suggestion, split the other ICs with the same partitioning as the DMN >> (consistent), mean-center within region, and run the regression with all in >> the model. The results up to this stage seem interpretable. Then these can >> be passed to DCM. >> >> Cheers, >> >> Anderson >> >> >> >> >>> >>> Best regards, >>> >>> George >>> >>> 2015-09-24 8:33 GMT+02:00 Anderson M. Winkler <[log in to unmask]>: >>> >>>> Hi George, >>>> >>>> All regions in the map of a given component share the same timecourse >>>> for that component. I suspect that, all regions having the same timecourse, >>>> there won't be much information to infer relationships among them with DCM. >>>> >>>> About splitting the component, one of the nicest features of the dual >>>> regression is that the maps are continuous, and they are so in two senses: >>>> they span the whole brain with no interruption, and they take continuous >>>> values. You can split, sure, but what about the other components in the >>>> first stage of the dual regression? Should them be included? If yes, should >>>> them be divided too? If yes, should the division follow the DMN to be >>>> consistent, even for ICs related to entirely different processes, or should >>>> there be one split pattern for each individual IC? In the first stage, >>>> should the divided components be mean-centered (within region) or not? Each >>>> of these questions need careful consideration and I don't know the answers >>>> for all or how these choices would impact results. You may want to >>>> investigate. >>>> >>>> All the best, >>>> >>>> Anderson >>>> >>>> >>>> >>>> On 23 September 2015 at 11:09, G Ch <[log in to unmask]> wrote: >>>> >>>>> Thanks Anderson. To tell you more about the analysis, I ran dual >>>>> regression and I found an area with a significant difference between the >>>>> two groups for a certain component. Now I want to investigate connections >>>>> between subregions of the component with the significant area through DCM. >>>>> Extracting time series from the non-decomposed resting state data won't do >>>>> the job cause the initial finding is specific for the component; that's why >>>>> I would like to have voxel or region-wise time-course for the component to >>>>> run the analysis between those subregions. Is multiplying the component map >>>>> with the component time course the way to get what I want? >>>>> >>>>> Someone on this forum also suggested to run dual regression again but >>>>> after dividing the component of interest into subregions and feeding the >>>>> analysis the new melodic_IC 4D image, in order to get component time series >>>>> for the images. I appreciate your help. >>>>> >>>>> Best, >>>>> >>>>> George >>>>> >>>>> 2015-09-23 9:41 GMT+02:00 Anderson M. Winkler <[log in to unmask]> >>>>> : >>>>> >>>>>> Hi George, >>>>>> >>>>>> Not impossible: the dual regression are two regressions, and in >>>>>> Y=X*b+e, once you have an estimated b, the fitted response, without the >>>>>> noise, is X*b. It isn't something interesting, though: it will be just the >>>>>> spatial map for the DMN multiplied by its estimated timecourse from the >>>>>> previous step. I think the best is just leave the dual regression as it is. >>>>>> >>>>>> All the best, >>>>>> >>>>>> Anderson >>>>>> >>>>>> >>>>>> >>>>>> On 22 September 2015 at 14:10, G Ch <[log in to unmask]> wrote: >>>>>> >>>>>>> Thank you Anderson. It would be impossible then to regress out all >>>>>>> components of non-interest + noise and leave DMN signal? I'm asking because >>>>>>> I'm interested in voxel-wise DMN residuals for further DCM analyses. >>>>>>> >>>>>>> Best, >>>>>>> >>>>>>> George >>>>>>> >>>>>>> 2015-09-22 10:36 GMT+02:00 Anderson M. Winkler < >>>>>>> [log in to unmask]>: >>>>>>> >>>>>>>> Hi George, >>>>>>>> >>>>>>>> I'm a bit worried with this suggestion: regressing out all other >>>>>>>> timecourses will leave DMN + noise. Perhaps better than is to simply take >>>>>>>> the DMN fit (or the fit for whatever other component that is of interest), >>>>>>>> which is what the dual regression does. >>>>>>>> >>>>>>>> All the best, >>>>>>>> >>>>>>>> Anderson >>>>>>>> >>>>>>>> >>>>>>>> >>>>>>>> On 21 September 2015 at 15:29, G Ch <[log in to unmask]> wrote: >>>>>>>> >>>>>>>>> Hi Andrew, >>>>>>>>> >>>>>>>>> I would like to suggest another method though perhaps someone else >>>>>>>>> can confirm if this is valid as I also thought about this. >>>>>>>>> >>>>>>>>> You can use the filtered_func_data and regress out from time >>>>>>>>> courses all components except the DMN so you would end up with the >>>>>>>>> contribution of the DMN to each voxel. Then from the residuals you can use >>>>>>>>> a mask for each region and measure the functional connectivity between >>>>>>>>> regions of interest in a seed analysis. >>>>>>>>> >>>>>>>>> Best, >>>>>>>>> >>>>>>>>> George >>>>>>>>> >>>>>>>>> 2015-09-21 15:49 GMT+02:00 Andrew Song <[log in to unmask]>: >>>>>>>>> >>>>>>>>>> Dear FSL experts, >>>>>>>>>> >>>>>>>>>> I am interested in figuring out functional connectivity between >>>>>>>>>> sub-components of independent components (IC) by MELODIC. >>>>>>>>>> For instance, I am interested in functional connectivity between >>>>>>>>>> pcc and IPL/IPS within DMN. >>>>>>>>>> >>>>>>>>>> The first method I tried was to simply increase the number of IC >>>>>>>>>> to be identified in MELODIC (from 25 to 60/70), in hopes of replicating the >>>>>>>>>> Smith PNAS 2009 paper. However, DMN did not seem to break down - It still >>>>>>>>>> emerged as a single IC. >>>>>>>>>> >>>>>>>>>> The second method I am thinking of is 'manually' breaking down >>>>>>>>>> the ICs. I would break down DMN into several sub-regions manually using >>>>>>>>>> masks and re-integrate them as 'separate components' into the dataset to be >>>>>>>>>> fed into dual regression. This was suggested a while ago in this forum. >>>>>>>>>> >>>>>>>>>> Following this, I have two questions. >>>>>>>>>> >>>>>>>>>> 1. In the first method, is simply increasing the number of ICs >>>>>>>>>> not the correct answer to get the sub-regions? >>>>>>>>>> 2. Perhaps more important, is the second method valid? I am >>>>>>>>>> worried that manually breaking down the components will violate the spatial >>>>>>>>>> independence between components, which is the very premise of ICA/MELODIC. >>>>>>>>>> >>>>>>>>>> Thank you very much, >>>>>>>>>> >>>>>>>>> >>>>>>>>> >>>>>>>> >>>>>>> >>>>>> >>>>> >>>> >>> >> >