SnPM for RFX
===========================================================================

This document describes how to use SnPM to perform second level random
effects analyses for fMRI.  If you don't already have SnPM, or to
learn more about it, see

	http://www.fil.ion.ucl.ac.uk/spm/snpm



  Very, very brief installation instructions
---------------------------------------------------------------------------

Briefly, the installation amounts to getting 

	ftp://ftp.fil.ion.ucl.ac.uk/spm/snpm99b.tar.gz

and all of the updates in

	ftp://ftp.fil.ion.ucl.ac.uk/spm/snpm99b_updates

and including the resulting directory in your Matlab path.  

Note, it is *very* important to get the updates, especially if you are
using SPM under windows.



  Preparing for the analysis
---------------------------------------------------------------------------

To perform a random effects analysis you need to have fit identical
(or near identical) models for each subject.  You will need a contrast
image from each subject that expresses the effect of interest.

It may be helpful to copy or link each of the relevant contrasts into
a single directory.



  Setting up the analysis configuration file
---------------------------------------------------------------------------

Start SnPM with the command
      snpm

Click 'Setup'.

Select the 'MultSubj: 1 conditions, 1 scan per subject' plug in.  This
is a one-sample t-test.


'# of confounding covariates'  

Select 0 confounding covariates.  (This would be useful if, say, you
were analyzing structural data and you had the age of each subject).


'<Num> Perms. Use approx test?'

It will tell you how many permutations there are, and if you want to
use an approximate test.  Unless this is a huge number (greater than,
say, 5000), say 'No' approximate test.  If it is larger and you say
'No', it may take a very long time to run.


'FWHM(mm) for Variance smooth'

To compare directly with parametric results, enter 0 for no variance
smoothing.  Normally, however, you will want some of
variance smoothing to (essentially) increase the degrees of freedom of
the t statistic and reduce artifactual roughness in the statistic
image.   5 mm smoothing is a good starting place.  


'Collect Supra-Threshold stats?'

You cannot have cluster-level p-values unless you say 'Yes' here.
However, recording this extra information can take an extraordinary
amount of disk space.

'Select global normalization'

Always select '<no global normalization>' for second level analyses.
The contrast images need no scaling.

'Select global calculation...'

Even if you select no global normalization, you need to select a
global calculation method (sorry).  Easiest is to select the default
(mean voxel value).

'Threshold masking'

Because contrast images are NaN masked, there is no need for explicit
masking.  Select 'none'.

'grand mean scaling...'

Again, the contrast images need no scaling.  Select '<no grand mean
scaling>'


The set up of the analysis is finished.  A SnPMcfg.mat file has been
written and you're ready to 'Compute'




  Estimating the model and computing permutations
---------------------------------------------------------------------------

To fit the model and build up the permutation distributions, press the
'Compute' button.


Select the SnPMcfg.mat file to start the process.

It will report on the number of permutations.

Finally it will produce a summary, like:

    Correct Perm has max t 7.62625 & rank 3 out of 64 completed permutations

From this you can determine the overall significance of the experiment
(the p-value for the maximal voxel).  It is the rank divided by the
number of permutations.


 Examining the results
---------------------------------------------------------------------------

Before anything, look at the permutation distribution

       load SnPM
       hist(MaxT(:),20)

Note where the correctly labeled max t value falls on this.  (The
correctly labeled max t is MaxT(1), and is printed at the end of
computation, as noted above).


Now examine the results... click 'Results' and select the SnPM.mat file


'Positive or negative effects?'

Choose your sign.

'Write out statistic img?'

If you want...

'Write full SS adj p-value img?'

Here SS means "single step", and is a subtle reference to step-down
p-values, something not implemented in SnPM.  This is simply an image
of corrected p-values.  Usually won't want it.

'Corrected p-value for filtering'

SnPM only provides a corrected p-value threshold.  Hence, sometimes
you might want to enter crazy-low thresholds (e.g. 0.5) just to see
what's going on.



Note, that if you've used variance smoothing, the statistic values
reported are *pseudo-t* values!  They cannot be directly compared to
any other t-values.  In particular, your pseudo-t values will surely
be smaller than your t values (from 0 variance smoothing).  This
doesn't mean the pseudo-t is less significant... they're just
different.  You have to use (corrected) p-values to compare.




 %W% %E%