83
contralateral and ipsilateral response means, weighted by their variance. Because of the
large variance in the individual trial responses, many trials were wrongly classified by
the boundary. The univariate classification performance was calculated by creating a
boundary from training runs, and then measuring accuracy on testing runs, to ensure an
unbiased comparison with the equivalent leave-one-out analysis used for MVPA. The
univariate S2 classification performance was 66%, much less than the 95% accuracy
achieved with multivariate analysis for left S2 in this subject. A similar analysis for left Sl
showed univariate accuracy of 66%, less than the 92% accuracy for MVPA in the same
ROI. Next, we examined the more difficult classification task of experiment 2. In left S2
of a single subject the average response to Dl, D3 and D5 touches was similar in
amplitude (Fig. 6C) and the distributions of the individual trial responses Werealmost
completely overlapping (Fig. 6D); a similar pattern was seen in Sl- Classification
accuracy for the univariate analysis was 43% for left S2 in this subject and 47% for SI,
much less than the MVPA accuracy of 55% for S2 and 69% for SI.