FBIRN:FBIRNCalibFriedmanMon2006
Comment: Noise measures in voxels ranges as 4th power of the signal (??) and so in some cases voxels end up with very low noise relative to the others, which can raise the t- or z-values...
Lee's method: include SFNR as a covariate and use Proportion of Variance Accounted For (r^2 * 100) as the dep. variable. Beta, effect size, number of activated voxels all should be carried forward to the 2nd order analysis.
Question: Does this method throw out signal you want? Using respiratory measures might be cleaner, but we haven't got them and many standard fMRI studies don't. This method has everything you need in the data you have. But it's easy to reduce noise by reducing signal as well. Need to test it on the the patient/normal subject.
But how sensitive is it to the ROIs chosen, as opposed to voxel by voxel applications?
Before this approach, we have focused on trying to change the signal (eg. beta weights) rather than looking at the noise. This goes in from the other direction.
Should we try to make adjustments to make all sites the same, or acknowledge that sites are different and weight the evidence from each site? E.g, differences from Stanford where SNR is very good might be more heavily weighted than differences from UCSD's old 1.5T siemens scanner that was not so good.
We have theories re: BH adjustments; but we don't have a theory to adjust for noise changes. Should we use noise measures as a way to weight the evidence from different sites?
--would that cause a problem with studies of imaging genetics, for example, where the push to actually *merging* the data on a subject by subject basis is needed?
--Lee trying to push interchangeability to its limit; argues this with BH will allow data to be merged/interchangeable.
How does this apply to the Phase II data? In the phase II analyses, we'd get an intercept and a residual, and a scalar as a covariate.
Are we assuming that a subject's signal to noise will scale the same way at every site? That if we had studied them at a different site we could make some predictions?
Sz literature tends to dislike covariate approach since the covariate correlates in many cases with the group difference. This is also a problem in the BH scaling, since Sz smoking affects the BOLD signal and correlates with the group difference.
This approach might decrease the Sz/normal difference at a site, but this removes the physiological noise. If physiological noise differs across groups, then the group difference will decrease--but you don't want to include heartrate or motion differences in your interpretation of the group differences.
Interpretation of PVAF: We have an interpretation of amplitude of response or number of voxels activated. What does PVAF mean? This homogenizes the number of voxels activated...?
