attainment, and the negative genetic correlation between AUDIT and both BMI/obesity and ADHD, may not generalize to cohorts with higher levels of alcohol use (Goldman et al. 2005) or AUD populations. Indeed, the selection of the 23andMe cohort (e.g. highly educated, with high socioeconomic background), or that of similar cohorts recruited from the general population such as UK Biobank, could have induced collider bias (Munafo, Tilling, Taylor, Evans & Smith 2017), underlying some of the paradoxical associations between AUDIT scores in our study, and high alcohol consumption in Clarke et al. (2017), and education/obesity. Another limitation of this study is the reliance on self-reported alcohol consumption, which may have induced biases and result in a more heterogeneous sample (Agrawal et al. 2012). Furthermore, AUDIT explicitly asks about alcohol use in the past year (i.e. state rather than trait), and this temporal specificity is not optimal for a genetic study. Nevertheless, AUDIT scores have a high correlation (r2=0.88) with scores on the Michigan Alcohol Screening Test (MAST), which assesses lifetime risk for AUDs, indicating that the temporal specificity of AUDIT may have a limited impact (Bohn, Babor & Kranzler 1995).
