The h volume is a little nasty (blame Anders:). If you want the stddev of the residual error, there should be a rstd.mgh file there. If you want to know about the h volume, read on.

doug

selxavg-sess creates a directory called bold/analysis in which several volumes will be stored. One of these is the "h volume", ie, h_000.bfloat, h_001.bfloat, ..., corresponding h_000.hdr. There are also single h.bhdr and h.dat files.

The h volume will have many frames, and the interpretation of these time points is quite obscure and hard to understand. But that is the purpose of this document.

The number of frames will be equal to: Nframes = 2*(Nc+1)*Na

Nc is the number of conditions (argument to -nconditions in mkanalysis-sess).

Na is the number of regressors per condition. This number can change depending upon how you are doing the analysis.

For an FIR model (ie, not using the -gammafit or -spmhrf flags), then Na = TimeWindow/TER, where TimeWindow is the argument of the -timewindow flag. TER is the argument of the -TER flag. If you have not specified a TER, then it defaults to the TR.

When assuming a gamma function shape to the HRF (ie, -gammafit), there will be one regressor per condition (Na=1).

When assuming an SPM HRF function shape to the HRF (ie, -spmhrf Nd), then Na = Nd+1 (Nd is the number of temporal derivatives).

Here's what each of the frames means: 1 to Na : always 0 Na+1 to 2*Na : all frames are the same and equal to the residual stddev 2Na+1 to 3*Na : regression coefficients for condition 1 4Na+1 to 5*Na : stddev of coefficients for condition 1 5Na+1 to 6*Na : regression coefficients for condition 2 6Na+1 to 7*Na : stddev of coefficients for condition 2 ...

The stddev of a coefficient is computed first by computing the matrix M = inv(X'*X) * resvar, where X is the design matrix and resvar is the residual variance. M is the covariance of the regression coefficients. d = diag(M) is a vector of the variances of the regression coefficient. The stddev is the sqrare root of this.

Sebastian Moeller wrote:

Hi Doug,

a related question, why do the h volumes show differing values for voxels at the odd "time points", supposedly the residual variance per condition, while fast_ldsxabfile only returns one value per voxel for the residual error variance? My expectation is to just have a single eresvar per whole model (or for each real timepoint of the model, but I digress), but what is really inside of the odd timepoints of the h-volumes then?

ahoi Sebastian

On 11. Jul 2006, at 12:07 Uhr, Sebastian Moeller wrote:

...

Hi Doug, hi List,

I got a question regarding the h volumes inside the analysis sub directories. The content of those file are the regression coefficients for each voxel (XY planes) for each condition (Z stack). Supposedly those estimates are corrected for the noise and would make really nice descriptors of the overall activation of voxels/rois for the given conditions. Now, here is my question, what is the unit of the values in the maps? And is there a way to transform those values into something ala percent change? Would dividing a XY voxel of a slice though the matching voxel value from the h-offset volume result in percent of total?

ahoi Sebastian -- Sebastian Moeller

Tel.: 04 21 - 2 18 - 78 38 oder 96 91 Fax.: 04 21 - 2 18 - 90 04 GSM: 01 62 - 3 25 45 59 moeller@brain.uni-bremen.de

AG Kreiter / FB 2 Institut fuer Hirnforschung III Abteilung Theoretische Neurobiologie Universitaet Bremen Biogarten Hochschulring 16a Postfach 33 04 40 28359 Bremen

Hi Doug,

On 11. Jul 2006, at 19:56 Uhr, Doug Greve wrote:

Hi Sebastian,

those are scaled to a global mean of 1000 (so you can think of them as tenths of percents) -- but it's global, not local.

So (I am a bit dense, sorry for that), I could just use those values as basis to directly (multiply by 10 ;)) compute the percent change induced by the different conditions. Without having to resort to the h-offset files, and ugly edge effects.

Yes, you can divide the h volume by the h-offset to give you local percent signal change.

What is the difference between local and global in this matter.

This makes me a little uncomfortable because you can get some funny business at the edge of the brain.

A, how does the global scaling avoid that, or is the division the problem (say values close to zero)?

ahoi & thanks for the information & the fine tool Sebastian

doug

Sebastian Moeller wrote:

...

Hi Doug, hi List,

I got a question regarding the h volumes inside the analysis sub directories. The content of those file are the regression coefficients for each voxel (XY planes) for each condition (Z stack). Supposedly those estimates are corrected for the noise and would make really nice descriptors of the overall activation of voxels/rois for the given conditions. Now, here is my question, what is the unit of the values in the maps? And is there a way to transform those values into something ala percent change? Would dividing a XY voxel of a slice though the matching voxel value from the h-offset volume result in percent of total?

ahoi Sebastian

-- Douglas N. Greve, Ph.D. MGH-NMR Center greve@nmr.mgh.harvard.edu Phone Number: 617-724-2358 Fax: 617-726-7422

In order to help us help you, please follow the steps in: surfer.nmr.mgh.harvard.edu/fswiki/BugReporting