The cluster analysis approach employed in the current study was refined from the one used in Kranzler et al. [2008], from which it differs in three major respects. First, in contrast to the approach used in Kranzler et al., the clustering process in the current analysis consisted of two layers of a cascaded process, as depicted in the supplemental material. This allowed us to identify subtypes that are both homogeneous and highly heritable. Second, in Kranzler et al., k-means cluster analysis was used to generate intermediate clusters. This method is an iterative procedure that is initialized with randomly chosen cluster centers. It is sensitive to outliers, with different initialization known to yield different clusters. In the present study, we used smart k-medoids cluster analysis, which is more robust to variation in the sample distribution. The supplementary material contains additional information about the cluster approach that we used in the present study. Third, the sample used here is much larger than the 1,393 subjects in Kranzler et al. and it yielded larger and more clearly differentiated clusters. Kranzler et al. found
