Purchase this article with an account.
Lúcia Garrido, Maryam Vaziri-Pashkam, Ken Nakayama, Jeremy Wilmer; The case against normalization in fMRI pattern analysis. Journal of Vision 2012;12(9):516. doi: https://doi.org/10.1167/12.9.516.
Download citation file:
© ARVO (1962-2015); The Authors (2016-present)
Multivariate pattern analysis of fMRI responses has become widely used in vision research. Haxby et al. (2001) introduced an influential method to investigate patterns of responses across voxels: we can correctly identify condition X if the pattern of responses to X in even runs correlates higher with the pattern of responses to X in odd runs than with the pattern of responses to another condition Y in odd runs. Before running these correlations, many researchers normalize voxel responses (separately for even and odd runs) by subtracting each voxel’s mean response across conditions from its response to each condition, a step intended to remove the influence of voxel effects (e.g. higher overall responsiveness of some voxels than others).
Mathematically, this step distributes variance across conditions and therefore introduces contaminating dependencies between them (i.e. normalization codes responses to each condition relatively to all other conditions). We illustrate this by running simulations in which one condition (X) is positively correlated between even and odd runs, but all other conditions (Y and Z) have a correlation of approximately zero between even and odd runs. Our simulations demonstrate that we can obtain above chance classification for all conditions after normalization. Above chance classification of both condition Y and condition Z therefore results not from information specific to either Y or Z, but rather from the absence of evidence for condition X.
In many occasions, however, we want to examine whether the pattern of responses in our region of interest carries some information about X relatively to the other conditions in the study. To move beyond classification per se and quantify the relative information contained in our region of interest for one condition versus another, we should feed non-normalized data into our pattern analysis.
Meeting abstract presented at VSS 2012
This PDF is available to Subscribers Only