Dear Oliver,
> could anyone please comment on these questions posted last week on the
> list - I urgently need your advice because my girlfriend will not be
> with me this weekend...
I am pleased to hear that, at least for some, SPM is such a potent
surrogate for a girlfriend (although it was brought to my attention, at
a party last weekend, that SPM is a three-letter word beginning with
'S'). Or perhaps your girlfriend advises you on SPM?
> Referring to Andrews mail from Wed, 19 May, I am wondering how to
> properly perform a mixed effects analysis in a study with multiple
> conditions.
>
> Due to the large amount of data (11 subjects, about 1500 timesteps per
> subject) it seems impossible to do a group analysis using a fixed
> effects model (because of memory problems). Therefore, I firstly fitted
> the model for each subject, and now I want to feed appropriate
> contrast/condition images into a second level mixed effects analysis. I
> have 7 conditions (epochs):
>
> B(aseline) A(ctivation) C(ontrol) A1 C1 A2 C2
>
> Firstly, I am interested on the comparisons A versus C, A1 versus C1
> and A2 versus C2. Secondly, I want to determine specific (and perhaps
> also general) effects of the interventions 1 and 2 on 'A versus C', for
> example using the following interaction contrasts:
>
> 0 -1 1 1 -1 0 0
> 0 1 -1 -1 1 0 0
> 0 0 0 1 -1 -1 1
> 0 0 0 -1 1 1 -1.
>
> Finally, I would like to show that the task component 1 (tested for by
> the contrast C1 versus C) and task A per se (A verus C) produce similar
> activations (anatomically, not necessarily quantitatively). (The same
> applies to possible general effects of the interventions 1 and 2.)
>
> My first questions are: for the latter purpose, do I have to use a
> conjunction analysis or, perhaps, masking? And if a conjunction would
> be more appropriate (by the way, why?), how do I have to set the
> thresholds to obtain a significance level of p < 0.05, corrected
Conjunctions of contrasts defined at the first level, but assessed at a
second level, are not possible (this is because a conjunction is
defined in terms of joint signficance and the only thing that enters
into the second level are the effects themselves, not their
significance). You could however specify contrasts [and their
conjunctions at the second level by simply entering the
condition-specific parameter estimates for each subject and modelling
condition-specific effects at the second level. The above contrasts
[and any conjunctions] would then be specified at the SECOND level.
Here you would be assuming sphericity as dicussed previously. To
obtain a corrected p value of 0.05 for the conjunction simply enter
0.05 as the required threshold (If you enter an uncorrected p value
then this is applied individually to contrasts in the conjunction).
> Secondly, what about the additional caveats for multiple conditions and
> non-orthogonal contrasts mentioned in your mail from 19 May? And,
> generally, should I enter contrast images or specific condition images
> into the second level of analysis?
In your case the specific condition images (which are a special case of
contrast images where the contrast is ([0 ..0 1 0..0]). The caveats
are (i) that all these contrasts should be orthogonal and estimable and
(ii) you are assuming sphericity (i.e. knowing the condtion-specific
estimate for one subject does not predict the same estimate for another
or vice versa).
I hope this helps - Karl
%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
|