Dear Boris,
There are a number of constraints with the current MFX implementation that we will be reviewing.
These are:
1. All subjects must have at least some data for every condition.
2. There is no facility to specify within-subject contrasts.
3. There is no facility to specify second-level contrasts/matrices.
Solution of problem 3 would provide the option of having multiple groups at the second level.
In the mean time I suggest you pursue an analysis based on the summary statistic approach, as follows:
1. Create [1 -1] contrasts testing for differences in condition within each session. This will produce 4 con images per subject.
2. Enter the con images into a second level Full Factorial design with two factors (1) group and (2) session. As you have differential contrasts at the first level there is no need to have within-subject effects (these remove the mean effect for each subject) at the second level.
3. You can then test for group differences etc using the usual contrasts.
As you know, you are only likely to get significant differences between summary statistic and MFX approaches if the first level design matrices are highly unbalanced (eg ten times as many trials for one subject as another) or the first level models vary greatly in fitting accuracies (again, we're talking factors of ten).
To test if this is likely to be the case for your data just have a look at the first level residuals at a few selected voxels and look up the number of trials for each condition/subject.
As it stands, the current MFX implementation should let you implement MFX modelling for a single group. So, you could always compare single group MFX to single group summary statistic results, with similar results giving you confidence that MFX might not be necessary.
Best,
Will.
> -----Original Message-----
> From: SPM (Statistical Parametric Mapping) [mailto:[log in to unmask]]
> On Behalf Of Boris Kleber
> Sent: 24 May 2014 12:30
> To: [log in to unmask]
> Subject: [SPM] SPM12b mixed-effects setup
>
> Dear SPMers,
>
> After the issues mentioned in my previous thread got spontaneously
> resolved after I switched to a machine with 16GB memory, I kindly ask
> for your advice on how to set-up the contrasts for a mixed-effects
> second-level analysis.
>
> I follwed the five steps described in the manual. However, after
> estimating the MFX SPM.mat file I ended up with a model that has two
> columns for second-level contrasts. I have a 2 x 8 design, i.e. 2
> groups รก 4 runs with 2 conditions per run. Each colomn of the MFX
> model now has 200 data-points, which equals the total ammount of 25
> subjects X 8 conditions. The first column seems to reflect activation
> for all participants and conditions whereas the second column brings up some kind of DMN activation.
>
> At 1st level I specified t-cons (i.e., within run motor > perception
> conditions) but didn't specify any factorial design. Thus I ended up
> with 4 contrasts of interests, assuming that I would be able to select
> and attribute them at a later stage as I do in a fullfactorial design.
> I intended to get a 2nd level matrix that distinguishes between Group
> (2) and Conditions (4) in the contrasts manager.
>
> Obviously that didn't work when I entered all 1st level spm.mat files
> into the FFX specification. It seems that I need to specify groups and
> conditions at a previous level but couldn't find any instructions on how to do this.
>
> Did I misunderstood the mixed-effects model or did I simply miss an
> important set-up step? Do you suggest to perform this analysis rather
> with GLM flex instead. My motivation is to get results that also take
> within-subject error variance into account.
>
> Any help is greatly appreciated.
>
> Best,
> Boris
|
