The TAG conducted a GWAS of smoking behaviour on 74,053 individuals, and a replication sample of over 140,000 individuals. A binary ever/never measure of smoking initiation was used as our exposure of interest as opposed to smoking heaviness, as smoking heaviness requires knowledge of participants’ smoking status, which was not available in the PGC2 summary data. Eight SNPs met genomewide significance (p < 5 × 10−8) for smoking initiation, all located in the BDNF gene region. A number of these SNPs were in very high linkage disequilibrium with each other. Using r2 values obtained from SNAP we identified all pairs of SNPs correlated at ≥0.9 and randomly selected one from each pair to keep, dropping the others from the score in a stepwise manner. Very highly correlated SNPs do not give extra information, but can make the model unstable. The remaining four SNPs (see Table 1) were also correlated, so a correlation matrix was created (see Supplementary Table 2) which required modification of the likelihood-based method for estimation of a causal effect (see below)18. Given the correlation between these SNPs,
