Commentary: Tobacco consumption and body weight: Mendelian randomization across a range of exposure

Tobacco consumption is consistently associated with reduced body weight, creating an incentive to initiate smoking and a disincentive to cease, although the health risks associated with the habit outweigh the benefits of reduced weight. Among smokers however, increasing consumption has been associated with increased body weight. To determine whether this contradiction reflects causal processes, Winsløw et al. have applied Mendelian randomization (MR) in testing the association of a genetic variant, rs1051730 in CHRNA3, with measures of body weight among 80 342 members of the Copenhagen General Population Study. Among smokers, each minor (T) allele carried was associated with an increase of about one cigarette per day, but with a decrease in several measures of body weight, in contrast to the observational results. These results, in line with other recent studies, suggest that increased tobacco consumption causes reduced body weight, as does smoking itself. Here I remark on two aspects of this study that may recur in other MR studies of this type: the restriction of genetic effects to current smokers, and the change, in the observational data, from decreasing to increasing body weight as cigarette consumption increases. The associations of rs1051730 with decreased body weight are present in current smokers, but not in former or never smokers, suggesting that the gene acts on body weight only through its effect on smoking. However, by stratifying on smoking status, the results are potentially prone to collider bias, whereby an association is induced between the gene and confounders of the exposureoutcome association, creating a non-causal association between the gene and the outcome. The question is whether smoking status should be considered as derived from the quantity consumed (one who smokes zero cigarettes per day is a non-smoker, others are smokers, Figure 1 arrow a), or as an exogenous variable whose value constrains the possible consumption (non-smokers must smoke no cigarettes per day, smokers must smoke a positive number, Figure 1 arrow b). Under the former, stratifying on smoking status would entail a collider bias, and an association would be seen between genotype and smoking status; but such an association was not observed by Winsløw et al., nor was any association between rs1051730 and known confounders when stratifying on smoking status. Thus the data support the view that smoking status is distinct (i.e. has distinct determinants) from tobacco consumption (Figure 1 arrow b). Indeed it seems reasonable that individuals would generally decide whether or not to smoke as a precursor to developing their usual consumption. Nevertheless, collider bias can only be completely ruled out if the whole sample is analysed, and here the genetic associations remained significant owing to the sufficiently high proportion of smokers among them. The lack of genetic association among never smokers could be interpreted as evidence against alternative pathways to that through smoking, in line with MR assumptions. However, Taylor et al. reported positive association between rs1051730 and body mass index (BMI) (P1⁄4 6.4 10 ) in a sample of 66 809 never smokers. This trend is apparent in the 32 937 never smokers in Winsløw et al. (P1⁄4 0.07, their Figure 4) and would have reached significance (P1⁄4 0.01) if the same trend were observed in 66 809 subjects. Thus the exclusion restriction seems to be violated, but as the effect is in the opposite direction to