The study is limited by the lack of suitable samples for independent replication to validate these findings. The scoring functions for each of the two subtypes, which use cocaine use and related behaviors to yield a probabilistic membership score for each subject in each cluster, are available from the first author upon request to allow other investigators to use the Supplementary Material to attempt to replicate our findings for Groups 4 and 5. The permutation tests required to estimate empirically the significance threshold for type I error for the tests of epistatic effects were very computationally intensive. Advances in statistical methods are required to address this problem, which is often encountered in the examination of higher-order interactions.
