GLM model comparison

Hi everyone,

I was wondering if there is an option in 3dDeconvolve or some other AFNI program that can calculate goodness of fit for GLM. For example, one can use TENT functions instead of a single BLOCK to model the response to a task, and would like to consider whether the TENTs or the BLOCK work better by comparing their goodness of fit. I haven’t seen this is performed anywhere, and there might be an obvious reason why people don’t or rarely do this. The comparison of goodness of fit seems more straightforward for each voxel. What about for the entire brain. I would appreciate if someone can please explain.

Thanks a lot,
Duong

Add option -rout to your 3dDeconvolve script so you will obtain the determination coefficient R^2 for the whole model and each regressor as well as each GLT. You can use R^2 to check the percentage of variance accounted for.

Thanks Gang for your answer.
Is the R^2 generated with -rout already adjusted for the number of regressors in the model, or do I have to calculate the adjusted R^2 manually?
Also, is the R^2 the mean of all voxel-wise R^2?

Duong

Note that -fout is applied by default in afni_proc.py…

  • rick

Is the R^2 generated with -rout already adjusted for the number of regressors in the model,
or do I have to calculate the adjusted R^2 manually?

You may have to adjust the R^2 yourself.

is the R^2 the mean of all voxel-wise R^2?

It’s at the voxel level; that is, one R^2 per voxel.