First, all frequency × channel samples were correlated with age. Samples that survived the initial test (i.e., the uncorrected p-value was less than .01) were clustered based on spatio-spectral proximity. Clustering was performed separately for samples showing positive and negative associations with age. An additional constraint was that within each cluster, each sensor had at least three neighboring sensors. Cluster-level statistics were obtained by summing the frequency × channel sample statistics within each cluster. The maximum of these was used to test the significance of our results against a randomization distribution. This distribution was obtained by randomly permuting the original data, taking the maximum cluster-level statistic and repeating this process for 1000 times. The probability of obtaining a statistic from this distribution larger than our actual cluster statistic was tested.
