Comparing performance of different subspace estimation methods, for Weak-Contrast data.

For weak contrast (TaskB vs. TaskA), model performance is shown for different subspace estimation methods, relative to full-dimensionality data (i.e. retaining all PCs). The median, [minimum, maximum] changes are shown for prediction (ΔP), reproducibility (ΔR) and distance ΔD from (P = 1,R = 1), over all single-subject results. Significance is given by Wilcoxon tests, with * indicating significant improvement. We show results for combinations of ICA = MELODIC subspace estimation, PCAsplit = optimized PC subspace on each data split-half, and PCAfull = retaining 35% of PCs from the full data matrix. Note that (PCAfull+PCAsplit) is the subspace selection method used for the rest of the manuscript. Results are shown for optimal fixed preprocessing: motion correction and 4th-order detrending.