Dear list,
I did a comparison between linear regression and SPM2 on the same
simulated data, and found that SPM has much better performance than
linear regression in ROC curve level.
For linear regression, the canonical HRF convolved stimulus function
was used as the regressor. The t value for estimated beta at each
voxel was used to build the ROC curve.
In SPM (SPM2), the basis function "hrf" was used. In specifiying
design, no the parametric modulation was conducted, and no additional
regressors were included. In specifiying data, "remove Global
effects", "High-pass filter" and "Correct for serial correlation" were
turned off. After estimation, data in the spmT_0003.{img/hdr} files
were used to build the ROC curve.
My understand is that if the data were subjected to additional
processes (e.g. high-pass filter, correct for seral correlation, and
etc.), linear regression and SPM2 should yield similar results.
Could anyone explain what magic things SPM has done to improve its
performance? Thank you very much.
Regards,
Wuming Gong
|