Abstract
The increased sensitivity afforded by multivariate pattern analysis methods (MVPA) has led to their widespread application in neuroscience. Recently, similarity-based multivariate methods seeking not only to detect information regarding a dimension of interest, say, an object's rotational angle, but to describe the underlying representational structure, have flourished under the name of Representational Similarity Analysis (RSA). However, data pre-processing steps implemented before conducting RSA can significantly change the correlation (and covariance) structure of the data, hence possibly leading to representational confusion—i.e., concluding that brain area X encodes information according to representational scheme A, and not B, when the opposite is true. Here, we demonstrate with computer simulations and empirical fMRI-data that time series demeaning (including z-scoring) can lead to representational confusion. Further, we expose a complex interaction between the effects of data demeaning and how the brain happens to encode information—usually the question under study—hence incurring a form of circularity. These findings should foster reflection on implicit assumptions bearing on the interpretation of MVPA and RSA, and awareness of the possible impact of data demeaning on inferences regarding representational structure and neural coding
Meeting abstract presented at VSS 2017