Hello-
I had a generic question about fMRI data analysis. After applying the standard pre-processing fMRI steps to resting state fMRI (quality assurance, distortion corrections, slice time corrections, motion corrections, temporal filtering, spatial smoothing), it can still be the case the the fMRI data is unit-less. In other words, there is no fixed range where the fMRI intensity values can lie - they can be arbitrarily large or small. How to handle this aspect of fMRi data in the statistical analysis?
Many statistical analyses packages will standardize / z-score the BOLD timeseries to center it and rescale it to unit variance. This is largely for computational convenience, as some modeling packages just tend to work better if all the model terms have similar orders of magnitude. Even if you do not do that, it theoretically shouldn’t matter. Rescaling data is a linear operation. The beta coefficients resultant from the general linear model will be larger for signal with larger values, but the variance will also be larger. When that’s taken into account, the statistics that tend to matter the most (i.e., the t-stat or z-stat) will be scaled accordingly.
This is an interesting topic that has been explored in the past (e.g., this commentary). Is the magnitude of the BOLD response measurable? If so, how important is it to measure and report it to improve transparency and reproducibility? Would relying solely on statistical values, such as t- or Z-scores, without providing real estimates, be comparable to physicists reporting the speed of light without actual measurements?