Noise regressors for SPM

[Deborah]

Hi,

I'm interested in using GLMdenoise to create noise regressors to then use in an SPM analysis pipeline. However, I believe the standard SPM design matrix differs from that of GLMdenoise in that there are separate columns for each condition for each run rather than one column per condition. Are the noise regressors valid in this case, or would I need to modify the design matrix in SPM to exactly match the output of results.parametric.designmatrix? Alternatively, is it possible to modify the design matrix in GLMdenoise to have one regressor per condition per run?

Thanks in advance for your help!

Best,
Deborah

[Kendrick Kay]

Hi Deborah,

This is a good question (and hard question).

The theory behind GLMdenoise dictates that you need to create a design matrix where a condition is repeated in multiple runs.  So the short answer is...  no, GLMdenoise is not compatible with an approach that requires separate regressors for each run...

You could try to just use the empirically derived noise regressors and hack them into the SPM-like scheme.  But I'm worried about the validity of that situation.  (Basically, the regime where you add noise regressors into the design matrix in GLMdenoise presupposes a coding of the experimental regressors (with repeats across runs), but you would be acting as if the noise regressors also applies equally well to the SPM-like scheme, which I'm not sure would be true...   since the different schemes are subject to overfitting in different ways.)

A different approach might be to use the "denoiseddata" output idea.  That might be okay if you aren't trying to estimate the noise level for a single session / subject and instead are just treating each session / subject as contributing a single number.

I wonder what the ultimate rationale for the SPM-like scheme is?

Kendrick

--

Kendrick Kay, PhD
Assistant Professor
Center for Magnetic Resonance Research (2-116)
University of Minnesota, Twin Cities
   Web: http://cvnlab.net
E-mail: k...@umn.edu
  Cell: 510-206-1059
 Skype: kendrickkay

--
You received this message because you are subscribed to the Google Groups "GLMdenoise" group.
To unsubscribe from this group and stop receiving emails from it, send an email to glmdenoise+...@googlegroups.com.
To post to this group, send email to glmde...@googlegroups.com.
Visit this group at https://groups.google.com/group/glmdenoise.
For more options, visit https://groups.google.com/d/optout.

[dbar...@gmail.com]

Hi Kendrick,

Thanks so much for your quick reply!

That makes sense that the noise regressors would not transfer to the standard SPM design. I suppose one option would be to use the concatenate runs option in SPM, which would get closer to the GLMdenoise matrix (although SPM adds an intercept column to the end).

Is the denoiseddata option appropriate for estimating beta weights as inputs for RSA? If staying within the GLMdenoise framework, how would I obtain beta estimates for each run separately?

Best,
Deborah

[Kendrick Kay]

Hi Deborah,

I'd be curious to see how the SPM experience turns out, as I'm generally curious about the details of what other software/code approaches do.

For an approach taken for RSA-style analyses, I guess it depends on how exactly you want to divide up the trials/runs/subjects etc. In the default approach taken by GLMdenoise, it doesn't try to deliver to you estimates that are derived independently for different runs (it tries to deliver you the best/most-accurate beta that uses all of the runs together). But maybe it might work for your purposes to do an even-runs / odd-runs separation of your dataset. Or even, a cross-scan-session separation, or a cross-subjects separation.

Or, a completely different route might be to code different instances of the same condition as separate conditions. This is something we've done recently (and is described in a pre-print: http://dx.doi.org/10.1101/337667)

Many things to consider...

Kendrick

[Deborah Barany]

Hi Kendrick,

Thanks for the preprint and the suggestions!  Yes, there are a lot of different options, just trying to get a handle on what makes logical and practical sense.

I will follow-up with anything interesting regarding implementing the GLMdenoise-derived regressors back into SPM.

Best,

Deborah