Commentary on Carrasquilla et al.: Smoking and obesity; uncovering causal mechanisms through triangulation of different methods

While observational studies suggest smokers have less abdominal fat than non-smokers, Carrasquilla et al. find opposite effects using Mendelian Randomization. We discuss how different methods leading to different conclusions highlights that we need rigorously conducted Mendelian Randomization in combination with complimentary methods to obtain reliable conclusions about causal effects. Carrasquilla et al. [1] conducted a Mendelian randomization (MR) study to test for causal associations between different smoking measures and abdominal obesity. MR is a powerful method for causal inference where genetic variants are used as instrumental variables, but it relies on several assumptions that require careful testing and interpretation. Interestingly, the authors find that smoking initiation and lifetime smoking may causally increase abdominal adiposity. This is in contrast with different lines of evidence suggesting that nicotine suppresses appetite and that smokers have lower body weight than non-smokers [2] and with a previous MR study [3] suggesting that cigarette smoking decreases adiposity measures. The study by Carrasquilla and colleagues differs from prior MR research in two important ways. First, although three prior studies used a single genetic variant in the CHRNA3/5 locus to instrument tobacco consumption [3-5], Carrasquilla and colleagues use genome-wide data from large genome-wide association studies on smoking initiation and lifetime smoking (unavailable at the time of the other studies). Second, Carrasquilla and colleagues use a combination of different, relatively novel, two-sample MR methods that better control for pleiotropy and latent heritable confounders, which assess bidirectional causal effects, in addition to more traditional MR sensitivity tests. This strength is particularly important in light of the increased popularity of MR across different disciplines, which is accompanied by concerns around the quality and, therefore validity of these studies. For example, a recent systematic review on MR studies on risk factors for cancer found that the vast majority of studies failed to include proper sensitivity analyses, with only a limited number of associations supported by robust evidence for causality [6]. The findings presented here by Carrasquilla and colleagues further underline the importance of reducing our reliance on any one set of assumptions underlying specific forms of MR analysis by performing multiple, complimentary sensitivity analyses. Another method used in Carrasquilla and colleagues is multivariable MR, which allows one to examine the effects of multiple exposures on a single outcome simultaneously (to assess potential mediation/confounding). Based on these analyses, the association between smoking and abdominal obesity seemed to be independent of different smoking-related traits: socioeconomic status, alcohol consumption, attention-deficit/hyperactivity disorder and risk-taking behaviour. A factor that was not explored, however, is the role that physical activity/exercise might play in the tentative causal effect of smoking on adiposity. Unsurprisingly, observational studies reveal negative associations between smoking and exercise/physical activity [7], and between abdominal fat and exercise/physical activity [8]. However, there is a surprising lack of causally informative studies examining the interrelations between these three factors. Physical activity possibly serves as a confounder in the smoking-adiposity association because increased levels of physical activity may reduce both smoking and adiposity. Moreover, physical activity may also act as a moderator, where higher levels of smoking may lead to lower levels of physical activity, which in turn may lead to increased adiposity. It is worth exploring if exercise/physical activity might play a moderating or confounding role in this association, which could be assessed using multivariable MR or other (non-genetic) types of mediation analyses. In addition to including multiple types of MR methods, evidence on causality could be strengthened further by 'triangulating' the results from an MR analysis with findings from other causally informative approaches. Triangulation refers to the process of integrating different research approaches with different, unrelated sources of bias, which is expected to lead to more reliable evidence for whether or not there is a causal effect [9]. For example, for the relationship of interest here, one could triangulate with analyses from prospective cohorts, which allow investigation of longitudinal effects across varying timeframes. This comparison is particularly interesting as a complement to MR, which estimates the lifetime effect of an exposure through one's genetic liability for the exposure. The effect of smoking on adiposity might vary over time, something that is not captured in an MR study. Another interesting addition in the context of triangulation is within-family (MR) designs [10], which allow one to control for shared familial influences, dynastic effects, assortative mating and population stratification, which cannot be accounted for with conventional MR. If results from these different methods point to the same direction, it would provide more robust evidence for a causal effect. Margot P. van de Weijer: Conceptualization (supporting); writing—original draft (lead). Karin J. H. Verweij: Writing—review and editing (supporting). Jorien L. Treur: Conceptualization (lead); supervision (lead); writing—original draft (supporting). M.P.W. and J.L.T. are funded by the European Union (ERC, UNRAVEL-CAUSALITY), project 101 076 686. The views and opinions expressed are, however, those of the author(s) only and do not necessarily reflect those of the European Union or the European Research Council. Neither the European Union nor the granting authority can be held responsible for them. K.J.H.V. is supported by the Foundation Volksbond Rotterdam and by a grant (406.22.GO.043) from The Dutch Research Council. No financial or other relevant links to companies with an interest in the topic of this article.