Temporal correlation and MVPA
60 views
Skip to first unread message
MS Al-Rawi
Nov 17, 2014, 7:51:17 AM11/17/14
to MPVA
Hi,
I am having an issue and I thought of sharing it with the group and ask for help. The problem is known to be caused by nonneuronal noise components (usually called temporal/serial correlations), one of the best papers describing the problem is: http://www.researchgate.net/publication/223562435_To_Smooth_or_Not_to_Smooth_Bias_and_Efficiency_in_fMRI_Time-Series_Analysis ).
In GLM, one approach that has frequently been used is based on the idea that the estimated temporal autocorrelation structure can be used to prewhiten the data, prior to fitting a general linear model with assumed identical and independently distributed error terms. Nonetheless, my hunch is that these temporal correlations won't affect the results obtained via MVPA, but, I would like to try this prewhiten approach. Will these temporal correlations affect leave one run out cross validation? I think no.
Now, I need to remove these temporal correlations and was wondering if MVPA-tool box has any function to do this?
I would appreciate any idea on this issue.
Cheers,
-Rawi
I am having an issue and I thought of sharing it with the group and ask for help. The problem is known to be caused by nonneuronal noise components (usually called temporal/serial correlations), one of the best papers describing the problem is: http://www.researchgate.net/publication/223562435_To_Smooth_or_Not_to_Smooth_Bias_and_Efficiency_in_fMRI_Time-Series_Analysis ).
In GLM, one approach that has frequently been used is based on the idea that the estimated temporal autocorrelation structure can be used to prewhiten the data, prior to fitting a general linear model with assumed identical and independently distributed error terms. Nonetheless, my hunch is that these temporal correlations won't affect the results obtained via MVPA, but, I would like to try this prewhiten approach. Will these temporal correlations affect leave one run out cross validation? I think no.
Now, I need to remove these temporal correlations and was wondering if MVPA-tool box has any function to do this?
I would appreciate any idea on this issue.
Cheers,
-Rawi
Greg Detre
Nov 18, 2014, 5:29:29 AM11/18/14
to mvpa-t...@googlegroups.com
I don't remember for sure, but my recollection is that the toolbox doesn't include anything for removing auto-correlation - please correct me if I'm wrong.
That said, I think AFNI does include some tools, and I'd bet that Matlab almost certainly does. So you could just wrap Matlab's native functions.
Once you figure out the right function, 'apply_to_runs.m' is designed to apply one or more functions to generate new versions of a pattern on a run-by-run basis
g
--
You received this message because you are subscribed to the Google Groups "Princeton MVPA Toolbox for Matlab" group.
To unsubscribe from this group and stop receiving emails from it, send an email to mvpa-toolbox...@googlegroups.com.
To post to this group, send email to mvpa-t...@googlegroups.com.
Visit this group at http://groups.google.com/group/mvpa-toolbox.
For more options, visit https://groups.google.com/d/optout.
MS Al-Rawi
Nov 18, 2014, 10:51:20 AM11/18/14
to mvpa-t...@googlegroups.com
Thank you Greg for your reply,
Unfortunately, the toolbox does not have a function to do that, and AFNI help to do the whitening and resolve the temporal correlations using a function called 3dREMLfit (AFNI program: 3dREMLfit) and it should be applied per each run, and not the whole series. One of its outputs is **** estimation of the temporal auto-correlation structure, but, I am not sure how to use it to whiten the data yet. In general, after estimating the correlations, one needs to subtract them for each voxel.
On the other hand, apply_to_runs or another one called filter_runs, could also be used, but in this case, it applies a temporal filter (could be 4 to 6 s Gaussian or a smoothing filter) on each voxel.
Although some literature praise both approaches, I am interested in the AFNI's approach.
-Rawi
From: Greg Detre <gr...@gregdetre.co.uk>
To: mvpa-t...@googlegroups.com
Sent: Tuesday, November 18, 2014 10:28 AM
Subject: Re: [mvpa-toolbox] Temporal correlation and MVPA
Reply all
Reply to author
Forward
0 new messages