Hi Jen

The SPM 2-stage 'summary statistic'procedure and the 1-stage procedure used
in Friston et al are both mixed effects procedures. The standard SPM
procedure takes into account the first level covariance indirectly as this
is reflected in the behaviour of the betas taken to the 2nd level.

The 1-stage procedure in the paper makes fewer assumptions but requires
more computation - the example data are included as a reference point for
the usual, 2-stage, procedure. Its robustness is assessed using these data
as a practical comparison, since they violate its assumptions.

A somewhat simplistic version of these assumptions: The 2-stage procedure
assumes that the within-session (or subject) error covariance needs to be
the same for each session (or subject), and so does the first-level design
matrix. If these conditions are fulfilled, the contribution of the within
session variability to the second level is exactly the same for each
session (or subject) and can be estimated together with the between subject
variability, which takes the same form.

The 1-stage procedure used in Friston et al extends the 2-stage procedure
to estimate directly the contribution the unbalanced 1st level designs and
variances makes to the 2nd level covariance in the example data. It is a
version of the SPM2 estimation procedure (see Friston et al esp. last para
which explains this - with no equations!)

There are of course likely to be multiple differences between SPM and FSL
and as always it is difficult to compare - we can't know with real data
which is really 'better' as we don't know where the true positives are. I
hate to say it but it often boils down to convenience and compatibility
issues, and of course personal preference, as both packages are thought to
be pretty good! ;)

Perhaps someone who knows how FSL deals with this would like to
contribute...?

I hope this helps

Alexa Morcom



On Aug 10 2006, J Labus wrote:

>I have combed through this paper however I won't claim perfect
>understanding. It does seem this paper informed us that now SPM passes
>both the parameter estimates and the variances at the subject level to
>the second level analysis which fixes the negative variance problem and
>allows us to use unbalanced designs. Can anyone translate this article
>to tell me why this is? Is this everyone else's take on this article?
>
>One of the key plugs for the FSL mixed model and a criticism of SPM has
>been that for the mixed model in SPM level-1 variance within subject
>variance did not get passed up in the SPM analysis so this seems to have
>been addressed. However, I know that the mixed model in FSL passes up
>the parameter estimates, the variances and also, the degrees of freedom
>in the hierarchical analysis. I wonder what is gained by passing up the
>degrees of freedom?(which i am not certain that SPM does).
>
>Any input would be useful as these are concepts I am trying to wrap my
>head around and the mathematicalese can sometimes impede this process.
>Jen
>
>Rik Henson wrote:
>>
>> This paper may help:
>>
>>
>> Friston KJ, Stephan KE, Lund TE, Morcom A, Kiebel S.
>>
>> Mixed-effects and fMRI studies.
>> Neuroimage. 2005 Jan 1;24(1):244-52.
>>
>> Rik
>>
>>
>>
>> ----- Original Message -----
>> *From:* Matthew L. Senjem <mailto:[log in to unmask]>
>> *To:* [log in to unmask] <mailto:[log in to unmask]>
>> *Sent:* Thursday, August 10, 2006 4:01 PM
>> *Subject:* Re: [SPM] mixed-effects analysis
>>
>> Dear Jen,
>>
>> I am not an expert on this topic, but I'll give it a shot. Others,
>> please feel free to correct me. It is my understanding that FSL
>> allows
>> one to do a mixed effects analysis automatically through the GUI,
>> whereas SPM requires one to do the analysis piecewise: First do the
>> Fixed Effects analysis on each subject, then enter all the con*
>> images
>> into a random effects analysis manually. FSL will do this all for
>> you
>> automatically. Another difference is that FSL models the variance
>> between subjects at the "2nd level", whereas SPM does not really
>> have a
>> good way to model the between subject variance.
>>
>> I hope this helps some.
>>
>> Best wishes,
>> Matt Senjem.
>>
>> On Wed, 2006-08-09 at 11:49, jen labus wrote:
>> > It has been suggested that FSL is advantageous because it
>> implements a
>> > mixed-effects analysis. However, I noticed SPM implements an
>> mixed-effects
>> > analysis also. Can anyone speak to the differences and
>> similarities of the
>> > implementation of these analysis procedures in in SPM verus FSL?
>> --
>> Matthew L. Senjem
>> Analyst Programmer - Imaging Systems
>> Mayo Clinic and Foundation
>> Mail Code: RO_CE_09_IMAGING
>> 200 First St. SW
>> Rochester, MN 55901
>> [log in to unmask] <mailto:[log in to unmask]>
>> Fax: 507-284-9778
>> Phone: 507-538-0764
>>
>
>

--

Dr. Alexa Morcom
Department of Psychiatry
Downing Site
Downing Street
Cambridge CB2 3EB
Tel: 01223 764674
Fax: 01223 764675 (please contact first)
http://www-bmu.psychiatry.cam.ac.uk/people/amm96/