For the artificial data sets where ground truth is available, we follow [Lange et al., 1999] and report the quality of source identification over a range of possible threshold levels in the form of ROC curves, i.e. as a plot of the false positive rate versus the true positive rate at different threshold levels. We report the temporal accuracy as the (normalised) correlation between the estimated time course and the 'true' timecourse after projection into the same signal space as defined by the PPCA decomposition, i.e. correlation between the estimated time course and , where and index the corresponding columns in the estimated mixing matrix and the true mixing matrix and where is the matrix of the major eigenvectors of the data covariance. By calculating the temporal correlation with respect to projected rather than the 'true' original time courses we ensure that the measure of temporal accuracy is unconfounded by the dimensionality reduction itself.