Nonlinear fitting using 3dNLfim?

AFNI Message Board
1

Dear AFNI Experts,

I am trying to do nonlinear fitting to βs I got from 3dDeconvolve using 3dNLfim. Could anybody give me an example code to do that?
For better understanding my question:
I did linear regression using AM2 option in 3dDeconvolve, and fitted the βs(BOLD response to stimuli) with behavior data. Now I want to see whether nonlinear fitting works better or worse than linear regression. But I do not know where to input the behavior data when I tried to do this using 3dNLfim.
Any suggestion is welcome!

Thanks a lot!

Sally.

2

Hi Sally,

What is the exact non-linear model that you want to fit to the betas? One option is to use -stim_times_IM to generate a set of betas, one per event, and then fit a model to that. That is easier than setting up a new model for 3dNLfim. If there are too many events, particularly with other conditions, then IM is not a great option.

But in any case, can you describe the model you would like to fit to?

  • rick
3

Hi Rick,

Sorry for the late response, it seems that I could not get email alerts.
The model I need to use is quadratic model and I do have a lot of events(300 totally). I am trying to fit fMRI response towards parametrized faces (300) with the behavior data.
Do you have any suggestions about doing this?

Many thanks!

Sally

4

Hi Rick,

Just want to make sure that this topic has not been buried!

Thanks,
Sally

5

Hi Sally,

So you expect a quadratic fluctuation of the betas corresponding to some external behavioral parameter? If that is true, this might be more easily solved using amplitude modulation in 3dDeconvolve/3dREMLfit. You could add linear and quadratic modulators (using Legendre polynomials) to the stimulus times, and the result would be 3 betas, for the mean, linear and quadratic magnitudes. That would be more straightforward than trying to implement something in 3dNLfim.

Does that seem like a reasonable option?

  • rick
6

Hi Rick,

Yes, you are right. Your suggestion sounds a great option! But I am still not sure how to do this. It is said in the 3dDeconvolve documentation that"3dDeconvolve does LINEAR regression’', and " If you want NONLINEAR regression, see program 3dNLfim ". Would you please provide a sample about how to achieve this? Where should add the quadratic modulators in the script please? I did see an option "-legendre " but it is for baseline model.

Many thanks!

Sally