Wonder if I could ask your advice about scaling post fMRIprep analysis before bringing into AFNI for first level analysis.
Should I be removing noise before running the scaling? I don’t see this as a step in this webpage post fMRIprep analysis fMRIPrep Tutorial #4: Additional Preprocessing Steps — Andy's Brain Book 1.0 documentation, but thinking about it, I am wondering if using images without the nuisance regressors excluded might amplify noise in the scaled image (e.g., large motion spikes) that would normally be regressed out during the first-level analysis?
Should I remove the confounds BEFORE scaling with something like this or does this not make any difference?
The question of whether, how, and when to scale fMRI signals is not merely a technical detail; it carries important conceptual implications. The longstanding norm in neuroimaging has been to overlook effect quantification, largely due to a field-wide emphasis on the traditional notion of statistical significance as the dominant framework for inference. In an environment where only statistical evidence takes center stage, the specifics of scaling appears irrelevant.
However, local scaling, as opposed to global or grand scaling, offers several practical and conceptual advantages. AFNI’s recommendation of voxel-wise (local) scaling facilitates interpretation in terms of percent signal change, enables valid group-level analyses by making individual effects more comparable, and supports robust reproducibility efforts such as cross-study meta-analysis.
From a practical standpoint, local scaling is typically performed by dividing the signal at each voxel by its mean prior to individual-level modeling. Ideally, this would be done using a “true” baseline signal at each voxel. But such a baseline is often ill-defined or unknown. One might consider using the intercept from the individual-level regression model as a proxy baseline while accounting for both covariates and task regressors, but even this approach is conceptually ambiguous. In the presence of low-frequency drift, the very definition of a baseline becomes fluid.
This raises the natural question: How much accuracy is lost in percent signal change when local scaling is performed prior to modeling? Fortunately, as demonstrated in Chen et al. (2017), the loss is practically negligible. Back-of-the-envelope calculations and empirical checks show minimal impact on final estimates. For further reassurance, one can directly compare percent signal change outcomes under different scaling schemes to assess their influence.