Addressing the 'problem' of temporal correlations in MVPA analysis

摘要

The use of multivariate pattern analysis (MVPA) has grown substantially over the past few years. Many studies using MVPA estimate the response of individual trial activity and perform hypothesis testing using a non-parametric approach. Here we show that the default auto regression model of order 1 used for temporal whitening of BOLD data is problematic in that it leads to biased permutation tests. We show that the correlation of activity estimates across trials can cause extreme bias in non-parametric hypothesis testing so that the proportion of type I or type II errors are inflated. Crucially for MVPA, this inflation increases with sphere size. The error magnitude is such that in our data set, strong univariate effects are completely missed. By whitening the data with a more general auto regression (AR) model, one can correct the bias in permutation testing for better signal detection. Applying higher order AR models is already implemented in many neuroimaging software packages as a non-default option. The use of more aggressive temporal whitening may also prove crucial for valid MVPA inference in fast event related designs.