Dear Karl and/or others,
Now that we're on the subject, how much smaller would the minimum
recommended n be in PET for doing a random effects analysis in order to
compare across groups or generalize to the population at large?
Thanks for your help,
Leann Kinnunen
University of Chicago
Department of Psychology
---------- Forwarded message ----------
Date: Tue, 20 Mar 2001 14:13:41 GMT
From: Karl Friston <[log in to unmask]>
To: [log in to unmask]
Subject: Re: group x task interactions in clinical fMRI
Dear Eraldo,
> When using fMRI for clinical purposes (eg; mapping hemispheric
> dominance for language function) one needs to compare the activations
> in a patient with the activations in a group of controls to test the
> reliability of any apparent difference in the patient
>
> This can be done as group x task interaction effect (one group has
> only one subject, the patient)
>
> My understanding is that fixed effects models would not be good here
> and a random effect analysis is needed
I think the usual qualifications apply here. If you want to infer that
the patient activates abnormally in relation to a specific group of
control subjects than a fixed-effects analysis is appropriate. If you
want to say that the abnormality is in relation to the population from
which the normals were selected, then a random-effects analysis is
called for. Both are valid.
The special caveat here is that the fixed effects analysis may be more
sensitive, which has implications in a clinical setting. The key
question here is which has the greatest predictive validity about
intervention outcomes.
> My understanding is also that a random effect analysis would come at
> the cost of the residual degrees of freedom with an issue on the
> magnitude of the controls sample size to achieve a sufficiently
> sensitive analysis
> Is there any estimate on how big should be the size of the control group?
Indeed - I would recommend at least 16 for a random effects analysis
and ideally 32.
With very best wishes,
Karl
|