For two and three-way interaction models, group-specific odds ratios were reported in tables and text to facilitate interpretation. In logit models, Chow-type tests of the equality of coefficients across groups may be unreliable since they confound the magnitude of the effect for each group with group differences in residual variation (Allison 1999). Predicted probabilities across groups, however, are unaffected by the confounding of the slope coefficients and error variances (Long 2009). The statistical significance of differences in predicted probabilities were examined using Long’s delta method (Long 2009), providing a conservative assessment of the significance of moderating effects in logit models. Finally, interaction models were re-estimated using a subsample that excluded probands with alcohol dependence (n = 1,996) to test whether findings were robust in a non-clinical sample. Because of the reduction in statistical power associated with interaction modeling, adjustments for multiple testing were not made (Brookes et al. 2004). A post hoc estimate of statistical power based on simulation analysis suggests that with a 0.05 error probability, the observed effect sizes, and a high-risk genotype as common as SNP rs279871
