Despite our efforts to select variants that are as independent of each other as possible, not all of the pair-wise correlations would be null. This means that signals from the crude analysis could have been confounded by some other SNPs because of LD, which is high in the ADH region. In order to identify independent genetic effects, Bayesian models were fitted to the data for each of the three alcohol phenotypes. These models account for LD patterns in the data and also for multiple testing. SNPs are kept in the model based on their probability of explaining variation in the outcome in relation to the performance of all the other SNPs, which can enter the model individually or in combination (73). The variable selection follows a reversible-jump algorithm based on a binomial prior distribution, with parameter for the number of trials being the total number of predictors included in the model, and the parameter for probability of ‘success’ equal to 0.5 (73). This ensures all possible models are equally likely a priori. Ten thousand iterations were completed, with the first 5000 specified as burn-in.
