Hi
just an additional comment about validity: estimating effect size etc
is not a problem. However, when feeding in data-derived regressors
into the GLM you should no longer use the standard null-hypothesis
tests on the output (i.e. even corrected p-values will be wrong). I
sugggest you use mixture-model based inference instead.
cheers
christian
On 27 Mar 2006, at 10:30, Jesper Andersson wrote:
> Hi there John,
>
>> I'm trying to implement a first-level GLM with EVs representing
>> condition (two levels,negative affect and neutral), activity within a
>> specific region over that time series (say, avg. amygdala
>> activity), and
>> their interaction. My main interest is in the interaction effect
>> as a
>> representation of connectivity between the specific region (amygdala)
>> and the rest of the brain, as a function of negative affect (what I
>> believe is often referred to as an analysis of "psychophysiological
>> interactions"). I haven't seen anything on this listserv or in
>> the FSL
>> documentation on how specifically to carry this out via FEAT (let me
>> know if I've missed anything). But I imagine I could carry it out by
>> extracting post-processed intensity values within the amygdala,
>> dumping
>> this to a text file and adding it as a column in the design matrix
>> to be
>> run on that same time series. I could then add an EV for my task
>> to the
>> same design matrix, and use the GUI option to multiply them
>> together for
>> the interaction. My questions are:
>>
>> 1) Does this approach seem valid?
>
> It seems valid indeed. One thing you should think about is to not
> chose
> too small a region to extract values from (i.e. not a single
> voxel). An
> assumption of the GLM is that there is no error in the independent
> variables, i.e. that the regressors are "completely known". As soon as
> you enter a measured time-series as a regressor you will of course
> violate this to some degree, and chosing a large enough region is a
> way
> to at least ensure that the error in the independent variable is a lot
> smaller than that in the dependent variable.
>
>>
>> 2) Do I need to worry about the different heights of the EVs
>> representing my task (e.g., 0 to 1) and the amygdala activity
>> (e.g., 0
>> to the max. intensity value, in the thousands)? Specifically, I'm
>> concerned with the interpretability of the EVs if I enter them into a
>> higher level analysis examining the interaction within and across
>> groups
>> (e.g. higher-level analyses), as well as contrasting interactions
>> within
>> and across subjects. If height is an issue here, can you suggest a
>> correction I could apply to the amygdala EV before entering it
>> into the
>> analysis that might take care of this?
>
> First of all your task should be mean-corrected, i.e. -.5 to .5 rather
> than 0 to 1. Multiplying the amygdala time-series with a 0--1 task
> regressor would simply ignore the correaltion in the 0 level of the
> task, rather than contrasting the correlations as is your intention.
>
> I hear that FSL will automatically mean correct your task regressor
> for
> you, so this may not be an issue. Just check the finished regressor to
> see all went to plan.
>
> As for scaling and interpretability at a 2nd level, I have never
> before
> thought of that issue in the specific context of PPI's. I guess
> what you
> want is to be certain you compare like with like. Let us think of the
> case of just entering a time-series from one region (A) and regressing
> the rest of the brains voxels on that series. The way to interpret the
> corresponding beta in that case is "rate of change of area X per unit
> change of area A". When we extend this to an interaction with some
> task
> the interpretation changes to "difference between condition 1 and 2 of
> the rate of change of area X per unit change of area A". As far as
> I can
> understand this is a reasonable thing to compare across subjects. So,
> the important thing is that you encode the task in the same way for
> each
> subject, but that would be equally true if you just wanted to
> compare a
> task effect across subjects.
>
> Needless to say you should not attempt a contrast to compare the PPI
> regressor with any other regressor in the model.
>
> I guess one additional thing to consider would be to give some thought
> to how you define the amygdala ROI. Let us say you have delineated the
> "real" part of amygdala that exhibit a certain effect. Let us then say
> that you double the size of the ROI, i.e. you "dilute" the effect so
> that the task induced variance is just half of what it was before.
> That
> would amount to doubling the beta at any target region. The same would
> of course be true for an interaction regressor. Having said that,
> it is
> not clear to me just how to achieve that objective. I think you will
> just have to use your own discretion, and avoid deep-sea fishing by
> testing different ROIs.
>
> Good luck Jesper
--
Christian F. Beckmann
Oxford University Centre for Functional
Magnetic Resonance Imaging of the Brain,
John Radcliffe Hospital, Headington, Oxford OX3 9DU, UK
Email: [log in to unmask] - http://www.fmrib.ox.ac.uk/~beckmann/
Phone: +44(0)1865 222551 Fax: +44(0)1865 222717
|