I know this is a tough issue to answer in the abstract, but would appreciate any comments. I ran an analysis of visual stimulation with n=9 using 3dDeconvolve (TENT) to get IRF functions for n conditions per subject. Then, I ran an ANOVA on one of the IRF timepoints. I compared the output of the ANOVA (and the waveforms) for these IRFs to those generated by simply deviating the BOLD response in each voxel from the 1 image preceding stimulus onset for each post-stimulus time period (i.e., raw change scores), averaging across condition and doing an ANOVA on the same timepoint.

In general, and as expected, the averaged waveforms for the IRFs and the change scores for each condition were similar in shape, although often offset.

But, again in general, statistical tests of visual stimulation vs off showed less reliability with the IRFs (i.e., fewer voxels significantly activated) in expected regions of visual cortex (based on previous studies with these materials), compared to change scores. On the other hand, for differences between conditions, expected differences (e.g. based on previous studies with these materials) were statistically reliable using IRFs but not for change scores.

So -- tests of absolute LEVEL of BOLD change were more reliable when using raw change scores whereas tests of relative DIFFERENCEs between conditions were more reliable with IRFs.

Again, I realize that the IRFs are produced through a more complex procedure than the change scores, making it difficult to directly compare them.

Thanks,
MB