Objective Smoking has been widely studied as a susceptibility factor for amyotrophic lateral sclerosis (ALS), but results are conflicting and at risk of confounding bias. We used the results of recently published large genome-wide association studies and Mendelian randomisation methods to reduce confounding to assess the relationship between smoking and ALS.
Methods Two genome-wide association studies investigating lifetime smoking (n=463 003) and ever smoking (n=1 232 091) were identified and used to define instrumental variables for smoking. A genome-wide association study of ALS (20 806 cases; 59 804 controls) was used as the outcome for inverse variance weighted Mendelian randomisation, and four other Mendelian randomisation methods, to test whether smoking is causal for ALS. Analyses were bidirectional to assess reverse causality.
Results There was no strong evidence for a causal or reverse causal relationship between smoking and ALS. The results of Mendelian randomisation using the inverse variance weighted method were: lifetime smoking OR 0.94 (95% CI 0.74 to 1.19), p value 0.59; ever smoking OR 1.10 (95% CI 1 to 1.23), p value 0.05.
Conclusions Using multiple methods, large sample sizes and sensitivity analyses, we find no evidence with Mendelian randomisation techniques that smoking causes ALS. Other smoking phenotypes, such as current smoking, may be suitable for future Mendelian randomisation studies
Statistics from Altmetric.com
Amyotrophic lateral sclerosis (ALS) is a neurodegenerative disease of motor neurons, resulting in progressive paralysis of skeletal and bulbar muscles, with death from neuromuscular respiratory failure typically occurring within 2–3 years of symptom onset.1 ALS has an incidence of 1–2 per 100 000 person-years and a lifetime risk of about 1 in 300.2 3 It has a peak age of onset of 58 years and affects men slightly more frequently than women.2 In 5%, there is a family history of ALS in a first-degree relative, but twin and other studies have shown that apparently sporadic ALS has a heritability of 60%, leaving the possibility that up to 40% of the contribution could be environmental.4 There are currently no agreed environmental risk factors for ALS, although smoking has been widely studied with mixed results.5–10
Summary statistics from genome-wide association studies allow us to use genetic predisposition for environmental risk factors to investigate causality, using Mendelian randomisation.11 12 Mendelian randomisation is based on Mendel’s laws of inheritance, allowing genotype to be used as an instrumental variable when studying the effect of an environmental exposure on an outcome.13 Genotype is considerably less likely to be confounded with other exposures that may bias the results of observational studies.14 Mendelian randomisation can also help to reduce bias from reverse causation, because the genetic variants one is born with are unchanged through a lifetime. This means the outcome, for example ALS, cannot change an individual’s genetic predisposition for the exposure, for example, smoking. In population-based genetic association studies, as opposed to parent-offspring or between-sibling studies, the randomisation is only approximate, and horizontal pleiotropy (where the genetic variants being tested increase the risk of both the environmental exposure and of ALS) can in any setting distort findings. A series of sensitivity analyses are available that can uncover such biases.11 15 16
Mendelian randomisation analysis has previously been used to assess the causal relationship between smoking and ALS with conflicting results.17 18 Since then, updated genome-wide association studies for smoking and ALS have been published, with much larger numbers, allowing a new analysis with the advantages of sufficient power and updated methods. We performed two-sample Mendelian randomisation analyses to assess whether there was evidence for smoking being causal for ALS and, as a sensitivity analysis, in the other direction to test if ALS liability might be causal for smoking.
Two-sample Mendelian randomisation analysis enables the summary statistics of genome-wide association studies to be used to estimate the causal effect of an exposure on an outcome based on the effect sizes of genetic variations on the exposure and on the outcome in the separate samples.19 The effect estimate from a Mendelian randomisation study is an estimation of the true causal effect of an exposure on the outcome of interest. In the case of ALS diagnosis, this will be expressed as an OR.
We defined an instrument for lifetime smoking index, a continuous measure of smoking exposure from a genome-wide association study of 463 003 people, with 126 independently associated single nucleotide polymorphisms (SNPs) of genome-wide significance that explained 0.31% of the variance in lifetime smoking.20 We also defined an instrument for ‘Ever smoking’, a binary measure of smoking exposure from a genome-wide association study of 1 232 091 individuals, with 378 genome-wide significant SNPs accounting for 4% of variance.21 The variance explained in both studies is in line with genome-wide association studies that have been used to assess the causality of smoking for other conditions.
More details of how the phenotypes were defined can be found in the online supplementary file 1.
We used summary data from the most recently published genome-wide association study for ALS.22 The study reported 10 SNPs to be independently associated with risk of ALS, in a population of 80 610 (20 806 cases and 59 804 controls).
To perform the Mendelian randomisation analysis, we used the ‘TwoSampleMR’ package, an R package and genome-wide association study summary data library developed as a platform for performing Mendelian randomisation tests and sensitivity analyses.23
We applied five different Mendelian randomisation methods: inverse-variance weighted, MR Egger, weighted median, weighted mode and MR using robust associated profile score. Each of these methods makes different assumptions about pleiotropy and instrument strength so a consistent effect across the multiple methods gives us the strongest evidence for causality.24 For details of these and other sensitivity analyses, please see supplementary materials.
As an additional, supportive analysis, we defined genetic risk scores for smoking and used the UK Biobank data to test whether genetic risk score for smoking predicts ALS case control status. Details of the methods can be found in the supplementary file under the heading ‘Genetic risk score analysis’.
Details of sensitivity analyses performed can be found in supplementary information (online supplementary tables S1–S5). All instruments passed sensitivity analyses except for heterogeneity tested using Cochran’s Q (online supplemenatry table S3) and a minority of SNPs did not pass Steiger filtering analysis (online supplemenatry table S4); however, rerunning the Mendelian randomisation analyses without these SNPs did not change the results (online supplemenatry table S5). After quality control, the number of SNPs making up each instrument were n=119 for lifetime smoking exposure and n=353 for ever smoking. Genome-wide association study summary statistics for each SNP making up the instruments are found in online supplementary file 2. Graphs showing scatter plots of effect sizes of SNP on exposure and outcome variable, leave-one-out analyses and single SNP analyses are shown in online supplemenatry figures S1–S12.
The results of the Mendelian randomisation analyses for each instrumental variable are shown in figure 1. We did not find strong evidence that lifetime smoking or ever smoking were causal for ALS. The result of the inverse variance weighted method for lifetime smoking index was OR 0.94 (95% CIs 0.74 to 1.19), p value=0.59, and for ever smoking, OR 1.10 (95% CI 1.00 to 1.23), p value=0.05. We did not find that ALS liability was causal of smoking status (online supplemenatry table S6). ORs tended to be >1 for the instruments testing if having ever smoked was associated with ALS, and <1 for the lifetime smoking instrument. A forest plot of results is shown in figure 1.
Using genetic risk score analysis we found no association between smoking and ALS case control status (online supplemenatry table S7).
Using instruments defined from recently published, large-scale genome-wide association studies, and numerous Mendelian randomisation methodologies and sensitivity analyses, we found no evidence that smoking causes ALS. Our result is supported by a lack of association found when performing genetic risk score analysis. We also found no relationship between genetic liability to ALS and likelihood of smoking, suggesting that reverse causality is not driving the association between smoking and risk of ALS reported in some epidemiological studies.
Two previous studies have used two-sample Mendelian randomisation analysis to assess the causal relationship between smoking and ALS. One reported a positive association between ever smoking and ALS using inverse variance weighted Mendelian randomisation analysis, but the result was not replicated with other Mendelian randomisation analyses and no other sensitivity analyses were reported. The ever smoking instrument was defined from the Social Science Genetic Association Consortium genome-wide association study, which has a smaller sample size than the GWAS & Sequencing Consortium of Alcohol and Nicotine use (GSCAN) study that we used. The outcome SNPs were from an ALS genome-wide association study published previously to the one used here.25 With a larger genome-wide association study, we replicate the inverse variance weighted result with borderline significance, p value 0.05 (figure 1), but do not consider this evidence of causality in the context of the results of the other Mendelian randomisation analyses presented. The other Mendelian randomisation study found no association using smoking instruments from a smaller smoking genome-wide association study and the same ALS genome-wide association study used here.17 18 Our study is necessary to analyse the larger genome-wide association studies available and to report all sensitivity analyses needed to interpret Mendelian randomisation results.
Large numbers of SNPs and high F statistic values mean the genetic instruments had enough strength to detect associations using the inverse variance weighted method.26 The I2 statistic quantifies regression dilution, which can be caused by measurement error.27 When using linear regression analysis (as is the case with MR Egger), measurement error in the exposure variable will cause the effect size to tend to the null and, in the outcome, will reduce statistical power.28 All smoking instruments had I2 values of <0.9, indicating some regression dilution. In cases where I2 >0.6, SIMEX modelling was undertaken to estimate regression values, and the results support the findings from other Mendelian randomisation models used.
For Mendelian randomisation to be valid, the genetic variants used as instrumental variables must mediate an effect only through the exposure of interest (the risk of ALS is only increased due to smoking, not through another effect of the variants), that is, there should be no horizontal pleiotropy. A limitation of this study is that we are unable to fully discount this pleiotropy using statistical techniques. All genetic instruments tested positive for heterogeneity using Cochran’s Q statistic, which may be caused by pleiotropy.29 MR Egger intercept analysis did not find evidence to support the presence of directional pleiotropy. However, low I2 values may invalidate the intercept estimation from MR Egger regression. A previous study used linkage disequilibrium regression score analysis and found that exposure to tobacco smoke in the home and being a light smoker (<100 cigarettes in a lifetime) are genetically related to ALS, which may support pleiotropy.17 Horizontal pleiotropy can cause false positives and false negatives. We used five Mendelian randomisation models that vary in their assumptions of pleiotropy to try account for these potential errors, although future models of Mendelian randomisation and how they account for pleiotropy may be better suited to identifying association between ALS and smoking. Use of a binary outcome measure (which is the case in this study) or otherwise invalidated modelling assumptions may cause heterogeneity tests to produce positive results.24
Inverse variance weighted methods will estimate the true causal effect of an exposure if all Mendelian randomisation assumptions hold; if not, other methods have been developed.27 30–33 It is suggested best practice to use multiple Mendelian randomisation methods to check for consistency of estimated effect.24 Following this approach, we find a consistent lack of relationship between smoking and ALS. However, UK Biobank data contribute to the study cohorts of the lifetime smoking index and ever smoking instrument so the exposure phenotype populations are not completely independent.
The advantage of the instrumental variables we used is that they can be used in an unstratified population (we do not need to know the smoking status of people in the outcome genome-wide association study).20 However, ever smoking is not consistently associated with ALS in epidemiology studies. A meta-analysis of case-control and cohort studies did not find strong supportive evidence of risk of ALS in people who had ever smoked (OR 1.12, 95% CI 0.98 to 1.27).5 Since then, an association has been reported in some studies but not others.6 9 The lifetime smoking index can be used to assess dose dependency, important contributory evidence to showing causality. Evidence of a dose-dependent effect of smoking on ALS risk is rarely shown, although a dose-dependent effect of reduced risk of ALS with increased time since smoking cessation when comparing former to current smokers has recently been reported.34 The only ALS risk study to use the lifetime smoking index did not find an association (unpublished data).
The most powerful Mendelian randomisation evidence on a potential effect of heaviness of smoking on ALS risk would require individual level data on a large sample, in which the CHRNA5 variant—related to heaviness of smoking among smokers—can be related to ALS risk by strata of smoking behaviour.35 There is currently no adequately powered study allowing such analyses.
Triangulation of multiple strands of epidemiological evidence makes findings more robust.36 Since the 1960s, smoking rates in many countries globally have been falling. It may be possible in the future to detect decreased rates of ALS in the population if smoking is a causal risk factor.
Using robust methods to detect association and estimate causal effects with summary statistics from genome-wide association studies, we do not find strong evidence to support a relationship between smoking and ALS.
Contributors AA-C and SO-M conceived and planned the study. AB-A conducted genetic risk score (GRS) analysis, S-OM conducted all other statistical analysis supported by REW. AA-C, GD-S, AB-A and REW provided intellectual input for data interpretation. AA-C and SO-M wrote the first draft of the manuscript. All authors reviewed and approved the final manuscript.
Funding The project is supported through the following funding organisations under the egis of JPND—www.jpnd.eu (UK, Medical Research Council (MR/L501529/1; MR/R024804/1) and Economic and Social Research Council (ES/ L008238/1)). The work leading up to this publication was funded by the European Community’s Health Seventh Framework Program (FP7/2007–2013; grant agreement number 259 867), Horizon 2020 Framework Programme (H2020-PHC-2014-two-stage; grant agreement number 633 413) and Programme Grants for Applied Research. This project is also supported by the National Institute for Health Research (NIHR) Maudsley Biomedical Research Centre at South London Maudsley Foundation Trust and King's College London. This research has been conducted using data from the UK Biobank Resource (application number 19278). We have also received funding from the Motor Neurone Disease Association, ALS Association, Patients Like Me and the Psychiatry Research Trust. This research was also supported by the NIHR Bristol Biomedical Research Centre at University Hospitals Bristol National Health Service (NHS) Foundation Trust and the University of Bristol.
Disclaimer The views expressed in this publication are those of the authors and not necessarily those of the NHS, the National Institute for Health Research or the Department of Health and Social Care.
Competing interests None declared.
Patient consent for publication Not required.
Ethics approval This research involves analysing publicly available data, with no collection of new data. Ethical approval had been obtained by the original study authors.
Provenance and peer review Not commissioned; externally peer reviewed.
Data availability statement This study used publicly available summary statistics from genome-wide association studies to define instruments for Mendelian Randomisation:
Lifetime smoking index: https://www.cambridge.org/core/journals/psychological-medicine/article/evidence-for-causal-effects-of-lifetime-smoking-on-risk-for-depression-and-schizophrenia-a-mendelian-randomisation-study/AA82945360EC59FEC4331A7A567309C9%23fndtn-supplementary-materials
Ever smoking (GSCAN): https://conservancy.umn.edu/handle/11299/201564
The single nucleotide polymorphisms that passed QC and were used in the MR analysis are in supplementary file 2.
Genetic risk score analysis used data from the UK Biobank (application number 19278)
If you wish to reuse any or all of this article please use the link below which will take you to the Copyright Clearance Center’s RightsLink service. You will be able to get a quick price and instant permission to reuse the content in many different ways.