Our study has several limitations. We used patient billing data in creating our phenotypes and did not use natural language processing of clinical notes to identify our case cohort. Such data are known to be imperfect, although we were able to validate at least one SNP previously identified using more traditional methods(Song et al., 2018; Wright et al., 2011). For illicit substances such as opioids, it is also possible that cases were missed due to deliberate under-reporting. Because diagnostic codes are more likely to be observed among frequent users of healthcare services, the repeated appearance of codes belonging to various substance abuse disorders in patients’ records could cause the inflation of the apparent genetic correlation among them. In addition, our study cohort was not population-based, although it derived from a large, integrated health care system in a large metropolitan region. The relatively old age our study cohort could suggest a population with some degree of left-censoring, particularly for patients with the greatest severity of OUD. Also, we could not study non-European-Americans, whose genetic diversity tends to be highest, and the
