The impact of study design on pattern estimation for single-trial multivariate pattern analysis

A prerequisite for a pattern analysis using functional magnetic resonance imaging (fMRI) data is estimating the patterns from time series data, which then are input into the pattern analysis. Here we focus on how the combination of study design (order and spacing of trials) with pattern estimator impacts the Type I error rate of the subsequent pattern analysis. When Type I errors are inflated, the results are no longer valid, so this work serves as a guide for designing and analyzing MVPA studies with controlled false positive rates. The MVPA strategies examined are pattern classification and similarity, utilizing single trial activation patterns from the same functional run. Primarily focusing on the Least Squares Single and Least Square All pattern estimators, we show that collinearities in the models, along with temporal autocorrelation, can cause false positive correlations between activation pattern estimates that adversely impact the false positive rates of pattern similarity and classification analyses. It may seem intuitive that increasing the interstimulus interval (ISI) would alleviate this issue, but remaining weak correlations between activation patterns persist and have a strong influence in pattern similarity analyses. Pattern similarity analyses using only activation patterns estimated from the same functional run of data are susceptible to inflated false positives unless trials are randomly ordered, with a different randomization for each subject. In other cases, where there is any structure to trial order, valid pattern similarity analysis results can only be obtained if similarity computations are restricted to pairs of activation patterns from independent runs. Likewise, for pattern classification, false positives are minimized when the testing and training sets in cross validation do not contain patterns estimated from the same run.