Hello Alex:
Perhaps I should say a few words on when it is appropriate to use program
3dNLfim vs.program 3dDeconvolve.
Program 3dNLfim is widely used for modeling fMRI drug response data. Program
3dDeconvolve is usually not appropriate for analyzing this type of experiment.
Since 3dDeconvolve estimates an arbitrary IRF (e.g., hemodynamic response
function), the number of parameters to be estimated is a function of the length
of the IRF. Depending on the details of the experiment, this could be anywhere
from 0 to (say) 15 TR, which means estimating 1 to (say) 16 parameters per
input stimulus function.
However, for drug response experiments, it is not the hemodynamic response that
is being estimated; rather, it is the pharmacokinetic response. This could last
for (say) 100 TR. To apply 3dDeconvolve to this type of experiment would
require estimating (say) 100 parameters,
A further distinction is that for 3dDeconvolve-type experiments, there are
typically many trials. However, for drug response experiments, there is
typically only one drug injection. To apply 3dDeconvolve to this type of
experiment would mean that for each parameter there is only 1 observation.
Obviously, this is not good or particularly meaningful. Fortunately, for drug
response experiments, we have a priori models, such as the differential
exponential model, and the gamma variate model. These models have just a few
(say, 4 or 6) parameters that must be estimated. However, since these
parameters enter into the model in a nonlinear fashion, this requires a
different methodology than the multiple linear regression used in 3dDeconvolve.
This nonlinear estimation problem is inherently more difficult; hence, program
3dNLfim takes longer to find the solution.
To summarize:
Program 3dDeconvolve is appropriate if:
(1) The experiment contains multiple trials.
(2) The IRF has short duration (e.g., hemodynamic response)
(3) There is no restriction on the shape of the IRF.
Program 3dNLfim is appropriate if:
(1) The experiment contains 1 (or a few) trials.
(2) The IRF has long duration (e.g., pharmacokinetic response).
(3) There is a priori functional form for the IRF.
One thing that 3dDeconvolve and 3dNLfim have in common is that both have
a corresponding plugin. I always suggest that people use the Deconvolution
plugin to plot the fit on top of the fMRI data, to visually verify that the
model has been set up correctly. Due to the greater opportunity for making
mistakes in defining the 3dNLfim model, it is even more important that
people use the NLfit plugin to check the model and the parameter constraints.
Doug Ward
