Second, we used cluster analysis, which groups similar subjects together based on their clinical features, to create clusters of subjects. In the present study, we combined the k-medoids clustering method (Kaufman and Rousseeuw, 1990; Theodoridis and Koutroumbas, 2003; van de Laan et al., 2003) consecutively with agglomerative hierarchical clustering (Calinski and Harabasz, 1974; Day and Edelsbrunner, 1984; Milligan, 1979; Tan et al., 2009). The k-medoids method first partitioned the subjects into 100 intermediate clusters. Then hierarchical clustering was used to merge the intermediate clusters to form a hierarchy of clusters based on Ward’s aggregation criterion, yielding a dendrogram and statistics such as cubic clustering criterion (CCC), R2, pseudo F and pseudo t2, which guided the determination of the final number of clusters. To produce more reliable clusters, the clustering approach used here differs in a number of ways from the k-means approach of Chan et al. (2011). Specifically, rather than using the average of subjects in a cluster as the cluster centroid, the k-medoids method groups data by finding the most representative subjects to serve as cluster centroids. Thus, the
