Unbiased Analysis of Item-Specific Multi-Voxel Activation Patterns Across Learning

Recent work has highlighted that multi-voxel pattern analysis (MVPA) can be severely biased when BOLD response estimation involves systematic imbalance in model regressor correlations. This problem occurs in situations where trial types of interest are temporally dependent and the associated BOLD activity overlaps. For example, in learning paradigms early and late learning stage trials are inherently ordered. It has been shown empirically that MVPAs assessing consecutive learning stages can be substantially biased especially when stages are closely spaced. Here, we propose a simple technique that ensures zero bias in item-specific multi-voxel activation patterns for consecutive learning stages with stage being defined by the incremental number of individual item occurrences. For the simpler problem, when MVPA is computed irrespective of learning stage over all item occurrences within a trial sequence, our results confirm that a sufficiently large, randomly selected subset of all possible trial sequence permutations ensures convergence to zero bias - but only when different trial sequences are generated for different subjects. However, this does not help to solve the harder problem to obtain bias-free results for learning-related activation patterns regarding consecutive learning stages. Randomization over all item occurrences fails to ensure zero bias when the full trial sequence is retrospectively divided into item occurrences confined to early and late learning stages. To ensure bias-free MVPA of consecutive learning stages, trial-sequence randomization needs to be done separately for each consecutive learning stage.