Dear Colline,
> Dear Mr Buchel,
> I have read your paper about nonlinear regressors in parametric fMRI
> experiment (Neuroimage, 1998) and I have some questions about
> it. I would like to know if it is possible to implement this
> approach to
> investigate modifications over several sessions of a
> difference of conditions.
>
> In my study I have 2 conditions A and B plus a rest (which is
> not declared
> as a condition in my model). As I was interested to learning
> effects, the
> experiment was reproduced 4 times. Now I am looking for the
> presence or
> absence of modifications in the contrast A minus B over the
> four sessions
> and to characterise these modifications if they exist (no a priori
> hypothesis about the function which could fit the
> modifications). Is it
> possible to perform a parametric then quadratic analysis on a
> conditions
> difference ? Is it possible when there is only 4 points (I am not
> interested by the modifications which could occur within each
> session but
> just between sessions) ? If so, how to implement them ? If
> no, how can I
> process my data to answer my questions ?
Given that you only have 4 time points a 2nd order fit might overfit your
data. I would rather go with a linear time x condition interaction or
possibly with an exponential, as this reflects learning better than a linear
increase.
Nevertheless, should you decide to go for a second order polynomial
expansion, you want to use the F-statistic to infer the effect, or if
desired use a t-test to ask the question of whether the 2nd order regressor
(non-linear) can explain additional variance not accounted for by the linear
term.
In terms of implementation, simply say yes when it comes to "Parametric
modulation" say "time". The next question order allows you to specfy a
simple linear effect (order = 1) or a quadratic ie nonlinear effect (order =
2), and obviously higher order terms, that only make sense if you have many
replications of your condition. If you want to model a biologically more
plausible function that has a ceiling effect you might say "other" and then
enter something like
-exp(-0.9*[1:4])
Do not worry about mean centering, spm will do that for you.
In this example 0.9 is the time-constant i.e. determines how fast or slow
this curve rises to the ceiling. However, keep in mind that having only 4
timepoints there will be no much difference between these models.
-Christian
Dr. Christian Büchel
Neurologische Universitätsklinik, Haus S10, Universitäts-Krankenhaus
Eppendorf
Martinistr. 52, D-20246 Hamburg, Germany
Tel.: +49-40-42803-4726 Fax.: +49-40-42803-9955
[log in to unmask]
http://www.uke.uni-hamburg.de/zentren/neuro/neurologie/mitarbeiter/buechel_c
hristian.html
|