Representational confusion: the plausible consequence of demeaning your data

Abstract

The use of multivariate pattern analysis (MVPA) methods has enjoyed this past decade a rapid increase in popularity among neuroscientists. More recently, similarity-based multivariate methods aiming not only to extract information regarding the class membership of stimuli from their associated brain patterns, say, decode a face from a potato, but to understand the form of the underlying representational structure associated with stimulus dimensions of interest, say, 2D grating or 3D face orientation, have flourished under the name of Representational Similarity Analysis (RSA). However, data-preprocessing steps implemented prior to RSA can significantly change the covariance (and correlation) structure of the data, hence possibly leading to representational confusion--i.e., a researcher inferring that brain area A encodes information according to representational scheme X, and not Y, when the opposite is true. Here, I demonstrate with simulations that time-series demeaning (including z-scoring) can plausibly lead to representational confusion. Further, I expose potential interactions between the effects of data demeaning and how the brain happens to encode information. Finally, I emphasize the importance in the context of similarity analyses of at least occasionally explicitly considering the direction of pattern vectors in multivariate space, rather than focusing exclusively on the relative location of their endpoints. Overall, I expect this article will promote awareness of the impact of data demeaning on inferences regarding representational structure and neural coding.