Dear Paul
> I have conducted a learning experiment in a blocked design using fMRI.
> Learning proceeds across 6 blocks alternating with baseline. RT data were
> collected.
>
> I want to explore brain regions changing as a function of time and am faced
> with a number of possibilities. A linear or exponential change can be
> modelled. Alternatively I could use the RT data as a behavioural measure of
> learning. The problem with the former approach is that my model is
> arbitrary. Who is to say what is the most appropriate model for learning in
> this study? The problem with the latter is that RT data with small numbers
> (an average RT per epoch (x6) per subject (x12)) is noisy and I want to
> model my BOLD responses with as little noise as possible.
>
> In summary, I'm faced with a choice between a clean but artificial (and
> possibly incorrect model) and a noisy but real model. Currently I plan to
> take up a suggestion made recently by Rik Henson to produce a combination
> of the two. This would hopefully mean that subjects' actual performance
> contributed to the modelling but that the contribution of the noise
> inherent in such data would be minimised. Does anyone have any suggestions
> about the best way to do this? At present I suppose that I will fit the RT
> data to an exponential function and hope for the best. Of course, as well
> as getting the best of both worlds I may be getting the worst of both. If
> anyone has a better idea that they are willing to share I would be ever
> grateful.
There is one other perspective to add to this issue. You are assuming
that you will get a 'better' statistical model if it is constrained by
some behavioural correlate of learning and, therefore, you want to
enter the RT data as independent variables. You could however adopt
another approach and use the RT as dependent variables in a post hoc
analysis having determined the neurphysiological dynamics of learning.
This may be more compelling: For example use a temporal basis set to
model the time x condition interactions (e.g. bi-exponetial decays).
Then demonstrate the time-dependent changes in activation predict RTs.
The advantage of this is that you can establish a constuct validity for
the imaging results in bahavioural terms. You cannot do this if you
put the RT data into the design matrix (directly or through some
parameterised model of learning).
With very best wishes - karl
%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
|