those from simulated null distributions from varying number of clusters, and the largest difference indicates the optimal number of clusters. In our analysis, ABC peaked at 2 and 3 clusters from full and pruned (r2 = 0.1) sets of SNPs, respectively. Profiles of clusters from 3-means model were similar to 3-class LPA (Table 5). Clusters correspond to the Null, MD, and AD+ASP classes from 3-class LPA were identified, although the Null cluster was much smaller making other classes larger in 3-means result, compared to 3-class LPA result. Importantly, cross-tabulating cluster memberships across 2- to 4-means models showed consistent results as in LPA. In Table 6, the majority of SNPs in the null class of 2-means model stays in the null cluster of 3-means model, while SNPs in the Signal cluster were mostly split into either AD+ASP or MD cluster resulted from 3-means model. AD+ASP class was further split into ASP or AD clusters of 4-means model, and SNPs in MD cluster mostly remained in the MD cluster of 4-means model. Results of LPA from split-half samples were consistent with each other and with the result from overall sample. In both halves, classes with profiles and proportions similar to those of
