The consequences of subtracting the mean pattern in fMRI multivariate correlation analyses

Multivariate pattern analyses of fMRI responses have become widely used in cognitive neuroscience. A popular method introduced by Haxby et al. (2001) is to correlate the patterns of responses to each condition across separate fMRI runs. These correlation analyses are applied both to (1) determine whether it is possible to discriminate/classify patterns of responses to two or more conditions, and (2) examine the relationships between patterns of responses to two or more conditions. Before computing correlations between conditions, many researchers subtract each voxel’s overall mean response to all conditions from its response to each condition1. This is typically done independently for separate fMRI runs or datasets, for example even and odd runs. We refer to this step as “subtracting the mean pattern,” but it has also been called “normalization,” “subtraction of cocktail mean pattern,” or “cocktail blank normalization” (MacEvoy and Epstein, 2009; Op de Beeck, 2010). Here, we discuss the effects of subtracting the mean pattern separately for