Optimising analysis choices for multivariate decoding: creating pseudotrials using trial averaging and resampling

Abstract

Multivariate pattern analysis (MVPA) is a popular technique that can distinguish between condition-specific patterns of activation. Applied to neuroimaging data, MVPA decoding for inference uses above chance decoding to identify statistically reliable condition-specific information in neuroimaging data which may be missed by univariate methods. However, several analysis choices influence decoding results, and the combined effects of these choices have not been fully evaluated. In particular, an increasingly popular approach is to average data from several trials together before training an MVPA classifier, but the decision about how much averaging to do is arbitrary and the effect of varying this parameter has not been documented. Here we systematically assessed the influence of trial averaging and resampling on decoding accuracy and subsequent statistical outcome on simulated data. Although the optimal parameters varied with the classifier and cross-validation approach used, we found that modest trial averaging using up to 5-10% of the total number of trials per condition improved decoding accuracy and associated t-statistics. In addition, a small amount of resampling could improve t-statistics and classification performance, but was not always necessary. We provide code to allow researchers to optimise these analysis choices for the parameters of their data.