Hello Freesurfer statistical experts,
I need some assistance in explaining the output generated by mri_glmfit (then followed by correction for multiple comparisions, mri_glmfit-sim). We fit one model and got some statistically significant clusters. Then, we ran another model, the same plus an additional nuisance variable. It had similar significant clusters. How do we compare the results from the 2 different models? Which model is better?
In other statistic programs, we could examine an R^2 value, which can be interpreted as the percent of the response variable variation that is explained by the model. The model with the higher R^2 would be considered better since it fits the data better. But the output from mri_glmfit has rvar.mgh - the sum of the squares of the residual error divided by the DOF. So it would seem that a lower rvar would indicate a better fit. Can rvar.mgh be used to compute an R^2, because the R^2 concept is more widely understood. Or, is there a way to explain rvar in lay terms similar to R^2?
Thanks, Peggy
You can get the SSE from rvar*DOF (eg, fscalc rvar.mgh mul 100 -o sse.mgh). You can get the SST using fscalc y.mgh sqr -o sst.mgh mri_concat sst.mgh --sum --o sst.mgh Then the unadjusted r-squared as fscalc sse.mgh div sst.mgh sub 1 mul -1 -o rsquared.mgh The final calculation is a little obtuse but it realizes r2 = 1 - sse/sst
For judging the model, you can compare the rvars. I think you can also test the significance of the regressor you added.
doug
On 01/06/2016 01:00 PM, Peggy Skelly wrote:
Hello Freesurfer statistical experts,
I need some assistance in explaining the output generated by mri_glmfit (then followed by correction for multiple comparisions, mri_glmfit-sim). We fit one model and got some statistically significant clusters. Then, we ran another model, the same plus an additional nuisance variable. It had similar significant clusters. How do we compare the results from the 2 different models? Which model is better?
In other statistic programs, we could examine an R^2 value, which can be interpreted as the percent of the response variable variation that is explained by the model. The model with the higher R^2 would be considered better since it fits the data better. But the output from mri_glmfit has rvar.mgh - the sum of the squares of the residual error divided by the DOF. So it would seem that a lower rvar would indicate a better fit. Can rvar.mgh be used to compute an R^2, because the R^2 concept is more widely understood. Or, is there a way to explain rvar in lay terms similar to R^2?
Thanks, Peggy
Freesurfer mailing list Freesurfer@nmr.mgh.harvard.edu https://mail.nmr.mgh.harvard.edu/mailman/listinfo/freesurfer
freesurfer@nmr.mgh.harvard.edu
-
Douglas N Greve -
Peggy Skelly