Mass univariate testing biases the detection of interaction effects in whole-brain analysis of variance.

Abstract

Whole-brain analysis of variance (ANOVA) is a common analytic approach in cognitive neuroscience. Researchers are often interested in exploring whether brain activity reflects to the interaction of two factors. Disordinal interactions -- where there is a reversal of the effect of one independent variable at a level of a second independent variable -- are common in the literature. It is well established in power-analyses of factorial ANOVAs that certain patterns of interactions, such as disordinal (e.g., cross-over interactions) require less power than others to detect. This fact, combined with the perils of mass univariate testing suggests that testing for interactions in whole-brain ANOVAs, may be biased towards the detection of disordinal interactions. Here, we report on a series of simulated analysis --including whole-brain fMRI data using realistic multi-source noise parameters-- that demonstrate a bias towards the detection of disordinal interactions in mass-univariate contexts. Moreover, results of these simulations indicated that spurious disordinal interactions are found at common thresholds and cluster sizes at the group level. Moreover, simulations based on implanting true ordinal interaction effects can nevertheless appear like crossover effects at realistic levels of signal-to-noise ratio (SNR) when performing mass univariate testing at the whole-brain level, potentially leading to erroneous conclusions when interpreted as is. Simulations of varying sample sizes and SNR levels show that this bias is driven primarily by SNR and larger sample sizes do little to ameliorate this issue. Together, the results of these simulations argue for caution when searching for ordinal interactions in whole-brain ANOVA.