Skip to main content

Thank you for visiting nature.com. You are using a browser version with limited support for CSS. To obtain the best experience, we recommend you use a more up to date browser (or turn off compatibility mode in Internet Explorer). In the meantime, to ensure continued support, we are displaying the site without styles and JavaScript.

A multivariable Mendelian randomization analysis investigating smoking and alcohol consumption in oral and oropharyngeal cancer

Abstract

The independent effects of smoking and alcohol in head and neck cancer are not clear, given the strong association between these risk factors. Their apparent synergistic effect reported in previous observational studies may also underestimate independent effects. Here we report multivariable Mendelian randomization performed in a two-sample approach using summary data on 6,034 oral/oropharyngeal cases and 6,585 controls from a recent genome-wide association study. Our results demonstrate strong evidence for an independent causal effect of smoking on oral/oropharyngeal cancer (IVW OR 2.6, 95% CI = 1.7, 3.9 per standard deviation increase in lifetime smoking behaviour) and an independent causal effect of alcohol consumption when controlling for smoking (IVW OR 2.1, 95% CI = 1.1, 3.8 per standard deviation increase in drinks consumed per week). This suggests the possibility that the causal effect of alcohol may have been underestimated. However, the extent to which alcohol is modified by smoking requires further investigation.

Introduction

Head and neck squamous cell carcinoma (HNSCC), which includes cancers of the oral cavity and oropharynx, is the world’s 6th most common cancer1. Prognosis remains poor, with survival ranging between 19 and 59% at 10 years2. Established risk factors include cigarette smoking and alcohol intake, as well as the human papilloma virus (HPV), which is mainly linked with oropharyngeal cases and thought to be sexually transmitted3,4,5. While a large proportion of cases of head and neck cancer are attributable to the combination of smoking and alcohol, the respective contribution of these risk factors is not clear given the strong association between these behaviours. Better estimation of the known risk effects may help identify or clarify the importance of other potential risk factors. One large pooled analysis6 from 17 European and American case-control studies (11,221 cases and 16,168 controls) found that, of the population attributable risk of smoking and alcohol, 4% could be attributed to alcohol alone and 33% to tobacco alone, with 35% explained by a greater than multiplicative joint effect of alcohol and tobacco combined for oral and oropharyngeal cancer. However, the apparent synergistic effect seen in observational studies may underestimate the independent effects of smoking and alcohol.

Mendelian randomisation (MR) is an approach which attempts to minimise issues of measurement error, reverse causation and confounding by using genetic variants which are randomly distributed at birth and are known to be reliably associated with modifiable risk factors of interest, to obtain causal effect estimates for these risk factors on disease outcomes7,8. To thoroughly evaluate the causal effects of both alcohol consumption and smoking on the risk of oral and oropharyngeal cancer, we first conducted univariable MR analysis. A recent genome-wide association study (GWAS) reported a genetic correlation (rg ~0.34, 95% CI = 0.3, 0.4, p = 6.7 × 10−63) between alcohol use and smoking initiation, suggesting that sequence variations overlap substantially and that there is a causal pathway operating between them9. To account for this correlation between smoking and alcohol, and to simultaneously investigate the independent effects of each, we conducted multivariable Mendelian randomisation10,11. As well as investigating the effects of smoking initiation, we aimed to capture a quantitative lifetime measure of smoking behaviour using the comprehensive smoking index (CSI). This was used to compare the independent causal effect of smoking with the continuous measure of alcoholic drinks per week12. Finally, given evidence of strong genetic correlation for both smoking and alcohol intake with other risky behaviours13, and the established link between sexual behaviour and HPV-driven head and neck cancer14, we conducted additional MR analysis of both risk tolerance and lifetime number of sexual partners on oral cavity (OC) and oropharyngeal cancer (OPC).

In this work we use multivariable MR to demonstrate strong evidence for an independent causal effect of both smoking and alcohol consumption on oral and oropharyngeal cancer. This suggests the possibility that the effect of alcohol may have been previously underestimated. However, the extent to which alcohol is modified by smoking requires further investigation.

Results

Univariable Mendelian randomisation

Using 176 single nucleotide polymorphisms (SNPs) robustly and independently associated with smoking initiation (Supplementary Data 1), univariable MR provided strong evidence that smoking increases risk of oral/ oropharyngeal cancer (IVW Odds Ratio (OR) 2.5, 95% CI = 1.6, 3.9, p = 6.94 × 10−5 per log odds of smoking) (Table 1, Supplementary Figs. 1, 2). The direction of effect was consistent across the four MR methods tested (IVW, weighted median, weighted mode and MR-Egger), although MR-Egger results had wider confidence intervals (CI) and were less reliable due to low I2 values of the SNP-exposure associations, indicating possible violation of the no measurement error (NOME) assumption (Supplementary Table 1). There was some evidence of heterogeneity in the SNP effects (Supplementary Table 2) although no SNP effect outliers were detected from MR-PRESSO and MR-Egger intercepts indicated limited evidence of directional pleiotropy (Supplementary Table 3). Effects were consistent, and ORs slightly larger, when the summary statistics for smoking initiation were obtained from a GWAS meta-analysis in GSCAN which excluded 23andMe and UK Biobank (IVW OR 3.3, 95% CI = 1.8, 6.1) and a GWAS conducted in the UK Biobank only (IVW OR 5.2, 95% CI = 1.8, 14.9) (Table 1).

Table 1 Univariable Mendelian randomisation of smoking and risk of oral and oropharyngeal cancer.

Findings of a causal effect of smoking initiation on oral and oropharyngeal cancer risk were further supported by MR analysis using 108 SNPs associated with a lifetime measure of smoking behaviour, the comprehensive smoking index (CSI) (Supplementary Data 1). Here, a 1 standard deviation (SD) increase in the CSI (equivalent to an individual smoking 20 cigarettes a day for 15 years and stopping 17 years ago, or an individual smoking 60 cigarettes a day for 13 years and stopping 22 years ago) was found to increase oral and oropharyngeal cancer risk combined, with an IVW OR of 3.5 (95% CI = 2.4, 5.0, p = 5.97 × 10−11) (Table 1, Supplementary Figs. 3, 4, Supplementary Tables 13).

Using 60 SNPs robustly and independently associated with number of alcoholic drinks per week, univariable MR provided strong evidence that increased alcohol consumption increases risk of oral/oropharyngeal cancer. Here a 1 SD increase in drinks per week (equivalent to 9 additional drinks per week) was found to increase oral and oropharyngeal cancer risk combined, with an IVW OR of 10.0 (95% CI = 5.3, 18.6, p = 5.64 × 10−13) (Table 2). The direction of effect was consistent across the four MR methods tested and high I2 values of the SNP-exposure associations indicated low bias from regression dilution (i.e., attenuation towards the null of an association as a result of random measurement error) (Supplementary Table 1). However, there was some heterogeneity in the SNP effect estimates (Supplementary Table 2) and MR-Egger intercepts indicated some evidence for directional pleiotropy (Supplementary Table 3). Effects were consistent, although ORs slightly smaller, when the summary statistics for drinks per week were obtained from a GWAS meta-analysis in GSCAN which excluded 23andMe and UK Biobank (IVW OR 8.3, 95% CI = 4.7, 14.4) and a GWAS conducted in the UK Biobank only (IVW OR 5.8, 95% CI = 3.7, 9.0) (Table 1). In the main analysis, MR-PRESSO identified an influential outlier which was the ADH1B variant, rs1229984 (outlier test p-value = 0.024) which was also shown to be a clear outlier in both scatter and leave-one-out plots (Fig. 1). With removal of this outlier, the causal effect estimate was halved but still large (OR 4.5, 95% CI = 2.3, 9.0). The MR-Egger intercept indicated no further directional pleiotropy once this variant was removed (Supplementary Table 3).

Table 2 Univariable Mendelian randomisation of alcohol consumption and risk of oral and oropharyngeal cancer.
Fig. 1: Scatter and leave one out plots demonstrating influential outliers in univariable MR of alcohol consumption and oral and oropharyngeal cancer risk.
figure1

a Scatter plot showing ADH1B (rs1229984), an outlier single nucleotide polymorphism in the analysis of alcohol consumption (drinks per week, n = 941,280) and oral and oropharyngeal cancer risk (n = 6034 cases and 6585 controls). b Leave one out plot again demonstrating the ADH1B (rs1229984) influential outlier. Consistent evidence for a causal effect of alcohol on oral and oropharyngeal cancer risk was found even when this variant was excluded from the analysis. Effect estimates are reported per SD increase in the exposure and error bars represent 95% confidence intervals.

Stratification by cancer subsite

In MR analysis stratified by cancer subsite, we found evidence for a causal effect of smoking initiation in both oral cavity (IVW OR 2.0, 95% CI = 1.2, 3.4) and oropharyngeal cancer (IVW OR 2.8, 95% CI = 1.6, 4.9); comprehensive smoking index in OC (IVW 2.6, 95% CI = 1.6, 4.3) and OPC (IVW 4.0, 95% CI = 2.5, 6.5); and alcoholic drinks per week in OC (IVW OR 5.9, 95% CI 2.4, 14.2) and OPC (IVW OR 3.3, 95% CI = 1.3, 8.1), which were consistent across the MR methods (Fig. 2).

Fig. 2: Univariable Mendelian randomisation of smoking initiation, lifetime smoke exposure and drinks per week on oral and oropharyngeal cancer subsites.
figure2

Univariable effects displayed were obtained using summary-level data from the GWAS of (a) smoking initiation (n = 1,232,091), (b) comprehensive smoking index (n = 462,690) and (c) drinks per week (n = 941,280) on oral and oropharyngeal cancer risk (n = 6034 cases and 6585 controls). Smoking initiation estimates are reported per log odds increase, while comprehensive smoking index and drinks per week are reported per SD increase in drinks per week. Error bars represent 95% confidence intervals. All statistical tests were two-sided.

Multivariable Mendelian randomisation

In the multivariable MR analysis controlling for alcohol consumption, there was strong evidence for a direct causal effect of lifetime smoking behaviour on risk of oral/oropharyngeal cancer (IVW OR 2.6, 95% CI = 1.7, 3.9 per SD increase in the CSI) (Table 3). In multivariable MR analysis controlling for lifetime smoking, there was also strong evidence for a direct causal effect of alcohol consumption on risk of oral/oropharyngeal cancer (IVW OR 5.2, 95% CI = 3.2, 8.6 per SD increase in drinks per week) (Table 3). The independent causal effects estimated from multivariable MR-Egger were consistent with the IVW analysis for both alcohol consumption and lifetime smoking (Table 3). While there was limited evidence for heterogeneity in the SNP effect estimates indicating instrument validity, the MR-Egger intercept deviated from the null, which is suggestive of directional pleiotropy. When we removed the ADH1B variant found to be an outlier in the univariable MR analysis, the independent effects of both smoking and alcohol on risk of oral/oropharyngeal cancer persisted remained, although the magnitude of the effect was reduced for alcohol (IVW OR 2.1, 95% CI = 1.1, 3.8) (Table 3, Fig. 3). The MR-Egger intercept indicated no further directional pleiotropy once this variant was removed (Supplementary Table 3).

Table 3 Multivariable Mendelian randomisation for smoking initiation and alcohol consumption with risk of oral and oropharyngeal cancer.
Fig. 3: Multivariable Mendelian randomisation of lifetime smoke exposure and drinks per week on oral and oropharyngeal subsites.
figure3

Effect estimates (ORs) are reported per SD increase in the exposure of drinks per week (n = 226,223) and the comprehensive smoking index (n = 226,223) on oral and oropharyngeal cancer risk (n = 6034 cases and 6585 controls). Error bars represent 95% confidence intervals. Genetic instrument for drinks per week excludes outlying variant, ADH1B (rs1229984). Univariable and multivariable effects displayed were obtained using summary-level data from the GWAS of comprehensive smoking index in UK Biobank and summary-level data from a GWAS of drinks per week in GSCAN, excluding UK Biobank and 23andMe. All statistical tests were two-sided.

In multivariable MR analysis stratified by cancer subsite, we found evidence for a causal independent effect of lifetime smoking in OC (IVW OR 2.5, 95% CI = 1.5, 4.1) and OPC (IVW 3.7, 95% CI = 2.3, 6.0); and for alcoholic drinks per week in OC (IVW OR 2.2, 95% CI = 1.0, 5.0) and OPC (IVW OR 1.9, 95% CI = 0.9, 4.0) (Fig. 3).

Investigating other risky behaviours

We found limited evidence to suggest there is a causal effect of risk tolerance more generally on risk of oral and oropharyngeal cancer (IVW OR 1.0, 95% CI = 0.6, 1.7 per SD) which was consistent across the MR methods applied (Fig. 4). There was some evidence for a causal effect of lifetime number of sexual partners (IVW OR 1.5, 95% CI = 1.0, 2.3 per SD), which was specific to oropharyngeal (IVW OR 2.2, 95% CI = 1.3, 3.8 per SD) rather than oral cavity cancer risk (IVW OR 1.2, 95% CI = 0.7, 2.0), and reasonably consistent across the MR methods applied (Fig. 4).

Fig. 4: Univariable Mendelian randomisation of risk tolerance and number of sexual partners on oral and oropharyngeal cancer subsites.
figure4

Effect estimates (log odds) are reported per SD increase in (a) risk tolerance (n = 508,782) and (b) number of sexual partners (n = 370,711). There was some evidence for a causal effect of number of sexual partners specific to the oropharyngeal subsite. Error bars represent 95% confidence intervals. All statistical tests were two-sided.

Discussion

In this study, we applied both univariable and multivariable Mendelian Randomisation to determine a causal, independent effect of both smoking and alcohol consumption on the risk of oral and oropharyngeal cancer. The results were largely robust to sensitivity analyses accounting for horizontal pleiotropy. Effects were consistent between both oral cavity and oropharyngeal subsites and unlikely to be strongly influenced by risk tolerance and lifetime number of sexual partners.

Supporting of our findings regarding an independent effect of smoking, average relative risks (RR) of reported tobacco smoking have been found to range from 4.0–5.0 for both the oral cavity and oropharynx across 16 case-control and 3 cohort studies15. Alcohol use has been associated with an increased risk of oral and oropharyngeal cancer in a dose-dependent manner (RR 1.1 (95% CI = 1.0, 1.3) for light drinking to RR 5.1 (95% CI = 4.3, 6.1) for heavy drinking (>50 g alcohol per day)16. Hashibe et al6. investigated the independent effects of smoking and alcohol and found an OR of 2.4 (95% CI = 1.7, 3.4) for ever tobacco use among never alcohol drinkers and a more than multiplicative joint effect of tobacco and alcohol, but no clear effect of ever alcohol use among never tobacco users (OR 1.1, 95% CI = 0.9, 1.3), suggesting that alcohol use on its own may not play an important role6. Limitations of observational studies include recall bias, differential measurement error, as well as potential residual confounding, for example by socio-economic position, HPV status17 and other lifestyle factors18,19,20.

The independent effect for alcohol in our study suggests the possibility that previous observational estimates for alcohol may have been underestimated, but these cannot be directly compared given the differences in the methodological approaches and interpretation of estimates. MR estimates may reflect the effects of lifelong alcohol exposure, in comparison to the short-term effects captured in observational studies. Furthermore, while multivariable MR can estimate the independent (and unconfounded) effect of alcohol on oral/oropharyngeal cancer risk, it cannot determine the extent to which this effect is modified by smoking status, i.e., is it observed among smokers only or the whole population. Further individual level data analysis is required to determine this.

Several mechanisms have been suggested to explain the observed associations between alcohol and risk of head and neck cancer, including contaminants present in alcoholic drinks (e.g., N-nitrosodiethylamine in beer)21, free radical damage and impairment of DNA repair capacity22,23. Ethanol is oxidised to acetaldehyde, which has a direct carcinogenic effect and moreover alcohol may act as a ‘solvent’ for tobacco carcinogens and the induction of carcinogen-metabolising enzymes24,25. Given their strong co-existence, the effects of smoking and alcohol are often difficult to separate out. However, in addition to the reported synergistic effect, previous observational data has highlighted an independent effect of both agents26, with our results supporting this.

Studies which have used genetic variation in alcohol-metabolising genes do give some support to our findings. For example, East Asians who are homozygous for the (*2*2) variant allele of ALDH2 (aldehyde dehydrogenase 2) are unable to metabolise acetaldehyde, which deters them from drinking alcohol, whereas those who are heterozygous (*1*2) have a 6-fold higher blood acetaldehyde concentration post-alcohol consumption with respect to the wild type *1*1. The OR of HNSCC among individuals with *2*2 has been found to be 0.5 (95% CI = 0.3, 1.0) and 1.8 (95% CI = 1.2, 2.8) among those with *1*2, relative to *1*1 individuals. These findings support the theory that alcohol increases HNSCC risk given the protective effect of non-drinking among homozygous individuals, while also indicating the carcinogenic action of acetaldehyde27. While the frequency of the *2 variant of ALDH2 in non-Eastern populations is low, a recent GWAS revealed variation at other sites in the genome which are linked with alcohol consumption9. An MR analysis using these multiple variants robustly associated with alcohol from this recent GWAS9, demonstrated a causal effect of alcohol intake on HNSCC (OR 3.9, 95% CI = 1.3, 11.2) in the UK Biobank, although there were only 856 HNSCC cases included in this cohort28. A more recent multivariable MR study in UK Biobank29 investigated smoking initiation and alcohol consumption with multiple cancers. This study demonstrated that smoking initiation increased odds of head and neck cancer when adjusted for alcohol (IVW OR 1.4, 95% CI = 1.1, 1.8, p = 0.002). There was also a strong positive effect of alcohol consumption on head and neck cancer risk when adjusted for smoking (IVW OR 1.9, 95% CI = 0.8, 4.6, p = 0.142), although with wider confidence intervals29.

The current study applied both univariable and multivariable MR methods, using the largest number of SNPs identified from the latest GWAS for both smoking9,12, alcohol consumption9,13 and head and neck cancer30, that could be identified in the literature. A series of pleiotropy-robust MR methods and outlier detection were applied to rigorously explore the possibility that findings were not biased as a result of pleiotropy. Comparability of smoking initiation and alcohol consumption is an issue given that alcohol consumption is closer to a measure of lifetime use, whereas smoking initiation captures both light and short-term smokers and heavy, long-term smokers. To account for this and to evaluate the causal effects of a quantitative measure of smoke exposure, we performed univariable and multivariable MR using genetic variants identified in relation to a CSI12.

The effect of smoking and alcohol on HNSCC risk is thought to be synergistic, with their combined use interacting in a multiplicative manner31. We were not able to investigate this here as MR interaction analysis relies on individual-level data which we did not have access to32. We were also unable to investigate the potential carcinogenic role of acetaldehyde, distinct from alcohol intake, as has been done previously using alcohol dehydrogenase genotypes27,33, as this again typically relies on individual-level data to assess gene-by-environment interactions. Our finding that the ADH1B variant, rs1229984 was a clear outlier which inflated causal estimates, indicates pleiotropy of this variant via its role in alcohol metabolism. However, we found consistent evidence for a causal effect of alcohol on oral and oropharyngeal cancer risk when this variant was excluded from the analysis.

While we used a large number of genetic variants to serve as stronger instruments for both smoking and alcohol, these variants have not all been well characterised and indeed some of the top loci have been previously associated with other risky behaviours13. Therefore, additional analyses to evaluate the causal effects of risk tolerance and lifetime number of sexual partners on oral/oropharyngeal cancer were conducted. While there was no strong evidence for a causal effect of risk tolerance more generally, an effect of lifetime number of sexual partners was observed, specific only to oropharyngeal cancer risk. This effect is likely mediated through HPV and thought to be sexually transmitted4,5. While further work is required to investigate this effect of HPV status among oropharyngeal sites, it is unlikely to have substantially biased the MR estimates for smoking and alcohol, since no effect of lifetime number of sexual partners was observed in relation to oral cavity cancer risk, whereas similar effect estimates were observed in relation to alcohol and smoking. Finally, as most participants in the GAME-ON network were of European or North American decent, with only 10.8% from South America, more work is required to determine if our results translate to other ancestry groups.

In conclusion, this study used both univariable and multivariable MR analyses to demonstrate an independent causal effect for both smoking and alcohol on oral and oropharyngeal cancer risk. In particular, we observed large effects for alcohol in our multivariable MR analysis, which suggests the possibility that previous observational estimates for alcohol may have been underestimated. However, these cannot be directly compared given the differences in the methodological approaches and interpretation of estimates. Further work using individual-level data could provide more robust evidence for a synergistic effect and the carcinogenic role of acetaldehyde. Findings from this study add to a growing body of evidence from MR studies surrounding the harmful effects of alcohol consumption34 and should be used to guide public health messages regarding the harms of even moderate drinking.

Methods

Univariable and multivariable Mendelian randomisation were applied using summary-level genetic data from the GWAS and Sequencing Consortium of Alcohol and Nicotine use (GSCAN)9, the UK Biobank study12,13, and a GWAS of oral and oropharyngeal cancer conducted by the Genetic Associations and Mechanisms in Oncology (GAME-ON) Network30, in a two-sample MR framework35. Mendelian randomisation is an approach which uses genetic variants as instruments to obtain estimates for the causal effect of these risk factors on disease outcomes. Three assumptions must be satisfied to ensure an MR study is valid which include: (1) genetic variants should be robustly associated with the risk factor of interest (i.e., the relevance assumption), (2) there are no confounders of the genetic variants-outcome association (the independence assumption) and (3) the exclusion restriction assumption. For further detail on the terms used in this study please read the Mendelian randomisation dictionary36.

Summary-level data from GSCAN and UK Biobank

Summary-level genome-wide association studies were obtained for alcohol consumption (drinks per week, n = 941,280) and smoking initiation (a binary phenotype indicating whether an individual had ever smoked regularly) (n = 1,232,091) from the GSCAN study9. The comprehensive smoking index (CSI) was derived by Wootton et al in the UK Biobank (n = 462,690)12. This included information on smoking duration, heaviness and cessation, which were combined into a lifetime smoking index along with a simulated half-life (τ) constant. Summary-level GWAS data were obtained for this phenotype in the UK Biobank12.

Summary-level data from GAME-ON

GWAS was performed on 6,034 HNSCC cases and 6,585 controls from 12 studies which were part of the GAME-ON network30. Cancer cases comprised the following ICD-10 codes: oral cavity (C02.0–C02.9, C03.0–C03.9, C04.0–C04.9 and C05.0–C06.9) and oropharynx (C01.9, C02.4 and C09.0–C10.9). The study population included participants from Europe (45.3%), North America (43.9%) and South America (10.8%). Details of the studies included, as well as the genotyping and imputation performed are published30,37.

Univariable Mendelian randomisation

To assess causal effects of smoking and alcohol consumption, we identified 60 single nucleotide polymorphisms (SNPs) for alcohol consumption and 176 SNPs for smoking initiation reaching genome-wide significance (p < 5 × 10−8) in the GSCAN GWAS which were present in the GAME-ON GWAS (Supplementary Data 1, Supplementary Data 2). We also identified 108 independent SNPs associated with the CSI at p < 5 × 10−8 in the UK Biobank which were present in the GAME-ON GWAS (Supplementary Data 1, Supplementary Data 2).

Instrument strength was determined by the magnitude and precision of association of the genetic instruments with the risk factor, which was considered to be sufficient if the corresponding F-statistic is >10. Linkage disequilibrium statistics (r2) were evaluated to check the overlap between smoking initiation, comprehensive smoking index and drinks per week loci using LDmatrix. All pairwise correlations were low for the SNPs used to instrument smoking initiation and drinks per week (r2 < 0.06) and those used to instrument comprehensive smoking index and drinks per week (r2 < 0.07), with the exception of rs10236149 and rs6962772 which had an r2 of 0.743. For smoking initiation and comprehensive smoking index, there were 6 duplicate SNPs and 11 SNP pairs which were in strong LD (r2 > 0.8).

Two-sample MR analyses were conducted using the TwoSampleMR package (version 0.5.5) in R (version 3.5.3), to extract the SNPs instrumenting the risk factor from the oral/oropharyngeal cancer GWAS38. We next performed harmonisation of the direction of effects between exposure and outcome associations, where palindromic SNPs were aligned when minor allele frequencies (MAFs) were less than 0.3 or were otherwise excluded. SNP-specific Wald estimates were calculated (SNP-outcome estimate divided by SNP-exposure estimate) and an inverse variance weighted (IVW) random effects method applied to meta-analyse these to obtain an estimate for the causal effect of the risk factor on oral/oropharyngeal cancer risk.

To further assess the robustness of findings, we inspected the Cochran’s Q statistic which assesses heterogeneity between individual genetic variants, and which may indicate the presence of invalid instruments (e.g., due to horizontal pleiotropy). Horizontal pleiotropy occurs when the genetic instruments are associated with more than one independent biological pathway, which can result in violation of the MR exclusion restriction assumption (i.e., the variable is not related to the outcome other than via the risk factor of interest). First-order inverse-variance weights were used to calculate both the IVW estimate and Cochran’s Q. A random effects model was used in the presence of heterogeneity39. Scatter and leave-one-out plots were produced to evaluate influential outliers and MR-PRESSO (Mendelian Randomisation Pleiotropy RESidual Sum and Outlier) was used to detect and correct for potential outliers (p < 0.05)40. MR-PRESSO is a unified framework that evaluates genetic pleiotropy in a standard MR model. It attempts to perform outlier removal in order to re-estimate the original exposure-outcome relationship to reduce bias in MR estimates. The IVW method will provide an unbiased estimate in the absence of horizontal pleiotropy or when horizontal pleiotropy is balanced41. To account for pleiotropy, we compared results with three other MR methods, which each make different assumptions about this: MR-Egger42, weighted median43 and weighted mode44. MR-Egger can provide unbiased estimates even when all SNPs in an instrument violate the exclusion restriction assumption. Where there was evidence of violation of negligible measurement error (NOME), assessed based on the I2 statistic, MR-Egger was performed with simulation extrapolation (SIMEX) correction45. SIMEX relies on simulation to estimate or reduce bias due to measurement error, considering additional data sets with increasing measurement error variance. The weighted median stipulates that at least 50% of the weight in the analysis stems from variants that are valid instruments43, while the weighted mode requires that the largest subset of instruments which identify the same causal effect to be valid instruments44.

Stratification by cancer subsite

We further performed MR analysis for alcohol and smoking with stratification by cancer subsite to evaluate potential heterogeneity in the causal effects. For this we used GWAS summary data on a subset of 2641 oropharyngeal cases and 2990 oral cavity cases from the 6034 HNSCC cases in the GAME-ON Network GWAS used in the main analysis and the 6585 common controls30. A total of 954 individuals with HNSCC were excluded from this analysis since these were cases of hypopharynx and overlapping cancers.

We also conducted further sensitivity analyses to estimate causal effects using summary-level data obtained from the GWAS of smoking initiation and drinks per week in GSCAN excluding UK Biobank and 23andMe, and summary-level data obtained from a GWAS of smoking initiation38 and drinks per week13 in the UK Biobank only.

Multivariable Mendelian randomisation

We next conducted two-sample multivariable MR analysis which included SNPs which were genome-wide significant in either the GSCAN GWAS of alcohol consumption or the UK Biobank GWAS of comprehensive smoking index. After excluding SNPs with a pairwise r2 greater than 0.001, 168 independent SNPs were used in the analysis. To remove sample overlap between the GWAS for alcohol consumption and CSI in the analysis, we used summary-level data obtained from the GWAS of drinks per week in GSCAN, excluding UK Biobank and 23andMe (n = 226,223) and summary-level data obtained from the GWAS of CSI in the UK Biobank (n = 462,690). Both the IVW and MR-Egger framework have been extended to estimate causal effects in multivariable MR analysis46,47, which was conducted using both the MVMR (version 0.2.0) and Mendelian Randomization48 (version 0.5.0) packages in R. We used generalised versions of Cochran’s Q statistical tests to assess both instrument strength and validity in the two-sample summary data setting, where the covariance between the effects of the genetic variants on each exposure was fixed at zero by using non-overlapping samples for each exposure. This was performed using the MVMR package (version 0.2.0) in R.

Investigating other risky behaviours

We used 124 SNPs associated with risk tolerance and 118 SNPs associated with lifetime number of sexual partners from a large GWAS meta-analysis13. We performed MR using the same univariable methods described in the main analysis and also evaluated the effects of these exposures stratified by cancer subsite.

Statistics and reproducibility

All genome-wide association study (GWAS) data used in this study had been previously replicated in independent datasets (see respective GWAS studies)9,12,30. In our study, MG and RR independently repeated all univariable and multivariable Mendelian randomisation analyses once, both successfully obtaining the same conclusions.

Reporting summary

Further information on research design is available in the Nature Research Reporting Summary linked to this article.

Data availability

GWAS summary statistics for the SNPs used to estimate the causal effects of smoking initiation, lifetime smoking and alcoholic drinks per week in this study are presented in Supplementary Data 1 and 2. Full summary statistics for the GAME-ON GWAS have been deposited in dbGAP (OncoArray: Oral and Pharynx Cancer; study accession number: phs001202.v1.p1)30. Published data from this study can also be found in: Lesseur, C. et al. Genome-wide association analyses identify new susceptibility loci for oral cavity and pharyngeal cancer. Nat Genet. 48, 1544-1550 (2016)30. Smoking initiation and alcohol consumption data (GSCAN study) are published in: Liu, M. Z. et al. Association studies of up to 1.2 million individuals yield new insights into the genetic aetiology of tobacco and alcohol use. Nat Genet. 51, 237 (2019)9. Comprehensive smoking index (CSI) data is published in: Wootton, R. E. et al. Evidence for causal effects of lifetime smoking on risk for depression and schizophrenia: a Mendelian randomisation study. Psychol Med, 1–9 (2019)12. UK Biobank approval was given for this project (ID 40644-Investigating aetiology, associations and causality in diseases of the head and neck). A copy of the data generated in this this analysis is available at: https://github.com/rcrichmond/smoking_alcohol_headandneckcancer49

Code availability

Two-sample MR analyses were conducted using the TwoSampleMR package (version 0.5.5) in R (version 3.5.3). Multivariable MR analysis was conducted using both the MVMR (version 0.2.0) and MendelianRandomization (version 0.5.0)48 packages in R. A copy of the code used in this this analysis is available at: https://github.com/rcrichmond/smoking_alcohol_headandneckcancer49

References

  1. 1.

    Warnakulasuriya, S. Global epidemiology of oral and oropharyngeal cancer. Oral. Oncol. 45, 309–316 (2009).

    PubMed  Article  Google Scholar 

  2. 2.

    Cancer Research UK. Head and Neck Cancer Statistics. https://www.cancerresearchuk.org/health-professional/cancer-statistics/statistics-by-cancer-type/head-and-neck-cancers#heading-Two (2019).

  3. 3.

    Ang, K. K. et al. Human papillomavirus and survival of patients with oropharyngeal cancer. N. Engl. J. Med. 363, 24–35 (2010).

    CAS  PubMed  PubMed Central  Article  Google Scholar 

  4. 4.

    Heck, J. E. et al. Sexual behaviours and the risk of head and neck cancers: a pooled analysis in the International Head and Neck Cancer Epidemiology (INHANCE) consortium. Int. J. Epidemiol. 39, 166–181 (2010).

    PubMed  Article  Google Scholar 

  5. 5.

    Thomas, S. J., Penfold, C. M., Waylen, A. & Ness, A. R. The changing aetiology of head and neck squamous cell cancer: A tale of three cancers? Clin. Otolaryngol. 43, 999–1003 (2018).

    Article  Google Scholar 

  6. 6.

    Hashibe, M. et al. Interaction between tobacco and alcohol use and the risk of head and neck cancer: pooled analysis in the International Head and Neck Cancer Epidemiology Consortium. Cancer Epidem. Biomar. 18, 541–550 (2009).

    CAS  Article  Google Scholar 

  7. 7.

    Smith, G. D. & Ebrahim, S. ‘Mendelian randomization’: can genetic epidemiology contribute to understanding environmental determinants of disease? Int. J. Epidemiol. 32, 1–22 (2003).

    PubMed  Article  PubMed Central  Google Scholar 

  8. 8.

    Davey Smith, G. & Hemani, G. Mendelian randomization: genetic anchors for causal inference in epidemiological studies. Hum. Mol. Genet. 23, R89–R98 (2014).

    CAS  PubMed  PubMed Central  Article  Google Scholar 

  9. 9.

    Liu, M. Z. et al. Association studies of up to 1.2 million individuals yield new insights into the genetic etiology of tobacco and alcohol use. Nat. Genet. 51, 237 (2019).

    CAS  PubMed  PubMed Central  Article  Google Scholar 

  10. 10.

    Burgess, S. & Thompson, S. G. Multivariable Mendelian randomization: the use of pleiotropic genetic variants to estimate causal effects. Am. J. Epidemiol. 181, 251–260 (2015).

    PubMed  PubMed Central  Article  Google Scholar 

  11. 11.

    Sanderson, E., Davey Smith, G., Windmeijer, F. & Bowden, J. An examination of multivariable Mendelian randomization in the single-sample and two-sample summary data settings. Int. J. Epidemiol. 48, 713–727 (2018).

  12. 12.

    Wootton, R. E. et al. Evidence for causal effects of lifetime smoking on risk for depression and schizophrenia: a Mendelian randomisation study. Psychol. Med. 50 2435–2443 (2020).

    PubMed  Article  Google Scholar 

  13. 13.

    Karlsson Linner, R. et al. Genome-wide association analyses of risk tolerance and risky behaviors in over 1 million individuals identify hundreds of loci and shared genetic influences. Nat. Genet. 51, 245–257 (2019).

    CAS  PubMed  Article  Google Scholar 

  14. 14.

    Gillison, M. L., Chaturvedi, A. K., Anderson, W. F. & Fakhry, C. Epidemiology of human papillomavirus-positive head and neck squamous cell carcinoma. J. Clin. Oncol. 33, 3235–3242 (2015).

    CAS  PubMed  PubMed Central  Article  Google Scholar 

  15. 15.

    Vineis, P. et al. Tobacco and cancer: recent epidemiological evidence. J. Natl Cancer I 96, 99–106 (2004).

    CAS  Article  Google Scholar 

  16. 16.

    Bagnardi, V. et al. Alcohol consumption and site-specific cancer risk: a comprehensive dose-response meta-analysis. Br. J. Cancer 112, 580–593 (2015).

    CAS  PubMed  Article  PubMed Central  Google Scholar 

  17. 17.

    D’Souza, G. et al. Case-control study of human papillomavirus and oropharyngeal cancer. N. Engl. J. Med. 356, 1944–1956 (2007).

    PubMed  Article  PubMed Central  Google Scholar 

  18. 18.

    Gaudet, M. M. et al. Anthropometry and head and neck cancer: a pooled analysis of cohort data. Int. J. Epidemiol. 44, 673–681 (2015).

    PubMed  PubMed Central  Article  Google Scholar 

  19. 19.

    Chuang, S. C. et al. Diet and the risk of head and neck cancer: a pooled analysis in the INHANCE consortium. Cancer Res. 23, 69–88 (2009).

  20. 20.

    Khetan, P. et al. Occupations and the risk of head and neck cancer a pooled analysis of the International Head and Neck Cancer Epidemiology (INHANCE) Consortium. J. Occup. Environ. Med. 61, 397–404 (2019).

    PubMed  PubMed Central  Article  Google Scholar 

  21. 21.

    Rogers, M. A., Vaughan, T. L., Davis, S. & Thomas, D. B. Consumption of nitrate, nitrite, and nitrosodimethylamine and the risk of upper aerodigestive tract cancer. Cancer Epidemiol. Biomark. Prev. 4, 29–36 (1995).

    CAS  Google Scholar 

  22. 22.

    Wight, A. J. & Ogden, G. R. Possible mechanisms by which alcohol may influence the development of oral cancer–a review. Oral. Oncol. 34, 441–447 (1998).

    CAS  PubMed  Article  Google Scholar 

  23. 23.

    Olshan, A. F., Weissler, M. C., Watson, M. A. & Bell, D. A. Risk of head and neck cancer and the alcohol dehydrogenase 3 genotype. Carcinogenesis 22, 57–61 (2001).

    CAS  PubMed  Article  PubMed Central  Google Scholar 

  24. 24.

    Dellarco, V. L. A mutagenicity assessment of acetaldehyde. Mutat. Res. 195, 1–20 (1988).

    CAS  PubMed  Article  Google Scholar 

  25. 25.

    Homann, N. et al. Effects of acetaldehyde on cell regeneration and differentiation of the upper gastrointestinal tract mucosa. J. Natl Cancer Inst. 89, 1692–1697 (1997).

    CAS  PubMed  Article  Google Scholar 

  26. 26.

    Rothman, K. & Keller, A. The effect of joint exposure to alcohol and tobacco on risk of cancer of the mouth and pharynx. J. Chronic Dis. 25, 711–716 (1972).

    CAS  PubMed  Article  PubMed Central  Google Scholar 

  27. 27.

    Boccia, S. et al. Aldehyde dehydrogenase 2 and head and neck cancer: a meta-analysis implementing a Mendelian randomization approach. Cancer Epidem. Biomar. 18, 248–254 (2009).

    CAS  Article  Google Scholar 

  28. 28.

    Ingold, N., Amin, H. A. & Drenos, F. Alcohol causes an increased risk of head and neck but not breast cancer in individuals from the UK Biobank study: a Mendelian randomisation analysis. medRxiv 19002832 (2019).

  29. 29.

    Larsson, S. C. et al. Smoking, alcohol consumption, and cancer: a mendelian randomisation study in UK Biobank and international genetic consortia participants. PLoS Med. 17, e1003178 (2020).

    PubMed  PubMed Central  Article  Google Scholar 

  30. 30.

    Lesseur, C. et al. Genome-wide association analyses identify new susceptibility loci for oral cavity and pharyngeal cancer. Nat. Genet. 48, 1544–1550 (2016).

    CAS  PubMed  PubMed Central  Article  Google Scholar 

  31. 31.

    Salaspuro, V. & Salaspuro, M. Synergistic effect of alcohol drinking and smoking on in vivo acetaldehyde concentration in saliva. Int. J. Cancer 111, 480–483 (2004).

    CAS  PubMed  Article  Google Scholar 

  32. 32.

    Rees, J. M. B., Foley, C. N. & Burgess, S. Factorial Mendelian randomization: using genetic variants to assess interactions. Int. J. Epidemiol. (2019).

  33. 33.

    Brennan, P. Commentary: Mendelian randomization and gene-environment interaction. Int. J. Epidemiol. 33, 17–21 (2004).

    PubMed  Article  Google Scholar 

  34. 34.

    Millwood, I. Y. et al. Conventional and genetic evidence on alcohol and vascular disease aetiology: a prospective study of 500 000 men and women in China. Lancet 393, 1831–1842 (2019).

    PubMed  PubMed Central  Article  Google Scholar 

  35. 35.

    Pierce, B. L. & Burgess, S. Efficient design for Mendelian randomization studies: subsample and 2-sample instrumental variable estimators. Am. J. Epidemiol. 178, 1177–1184 (2013).

    PubMed  PubMed Central  Article  Google Scholar 

  36. 36.

    Lawlor, D. A. et al. A Mendelian Randomization dictionary: useful definitions and descriptions for undertaking, understanding and interpreting Mendelian randomization studies. OSF Preprints. https://doi.org/10.31219/osf.io/6yzs7 (2019).

  37. 37.

    Dudding, T. et al. Assessing the causal association between 25-hydroxyvitamin D and the risk of oral and oropharyngeal cancer using Mendelian randomization. Int. J. Cancer 143, 1029–1036 (2018).

    CAS  PubMed  PubMed Central  Article  Google Scholar 

  38. 38.

    Hemani, G. et al. The MR-Base platform supports systematic causal inference across the human phenome. Elife 7, e34408 (2018).

  39. 39.

    Bowden, J. et al. Improving the visualization, interpretation and analysis of two-sample summary data Mendelian randomization via the Radial plot and Radial regression. Int. J. Epidemiol. 47, 1264–1278 (2018).

    PubMed  PubMed Central  Article  Google Scholar 

  40. 40.

    Verbanck, M., Chen, C. Y., Neale, B. & Do, R. Publisher Correction: Detection of widespread horizontal pleiotropy in causal relationships inferred from Mendelian randomization between complex traits and diseases. Nat. Genet. 50, 1196 (2018).

    CAS  PubMed  Article  Google Scholar 

  41. 41.

    Hemani, G., Bowden, J. & Smith, G. D. Evaluating the potential role of pleiotropy in Mendelian randomization studies. Hum. Mol. Genet. 27, R195–R208 (2018).

    CAS  PubMed  PubMed Central  Article  Google Scholar 

  42. 42.

    Bowden, J., Davey Smith, G. & Burgess, S. Mendelian randomization with invalid instruments: effect estimation and bias detection through Egger regression. Int. J. Epidemiol. 44, 512–525 (2015).

    PubMed  PubMed Central  Article  Google Scholar 

  43. 43.

    Bowden, J., Smith, G. D., Haycock, P. C. & Burgess, S. Consistent estimation in mendelian randomization with some invalid instruments using a weighted median estimator. Genet. Epidemiol. 40, 304–314 (2016).

    PubMed  PubMed Central  Article  Google Scholar 

  44. 44.

    Hartwig, F. P., Smith, G. D. & Bowden, J. Robust inference in summary data Mendelian randomization via the zero modal pleiotropy assumption. Int. J. Epidemiol. 46, 1985–1998 (2017).

    PubMed  PubMed Central  Article  Google Scholar 

  45. 45.

    Bowden, J. et al. Assessing the suitability of summary data for two-sample Mendelian randomization analyses using MR-Egger regression: the role of the I-2 statistic. Int. J. Epidemiol. 45, 1961–1974 (2016).

    PubMed  PubMed Central  Article  Google Scholar 

  46. 46.

    Rees, J. M. B., Wood, A. M. & Burgess, S. Extending the MR-Egger method for multivariable Mendelian randomization to correct for both measured and unmeasured pleiotropy. Stat. Med. 36, 4705–4718 (2017).

    MathSciNet  PubMed  PubMed Central  Article  Google Scholar 

  47. 47.

    Burgess, S., Dudbridge, F. & Thompson, S. G. Re: “Multivariable Mendelian randomization: the use of pleiotropic genetic variants to estimate causal effects”. Am. J. Epidemiol. 181, 290–291 (2015).

    PubMed  Article  Google Scholar 

  48. 48.

    Yavorska, O. O. & Burgess, S. MendelianRandomization: an R package for performing Mendelian randomization analyses using summarized data. Int. J. Epidemiol. 46, 1734–1739 (2017).

    PubMed  PubMed Central  Article  Google Scholar 

  49. 49.

    Gormley, M. et al. Multivariable Mendelian randomization to investigate the independent effects of smoking and alcohol on risk of head and neck cancer. GitHub. https://doi.org/10.5281/zenodo.4085102 (2020).

Download references

Acknowledgements

M.G. was a National Institute for Health Research (NIHR) academic clinical fellow and is currently supported by a Wellcome Trust GW4-Clinical Academic Training PhD Fellowship. This research was funded in part, by the Wellcome Trust [Grant number 220530/Z/20/Z]. For the purpose of Open Access, the author has applied a CC BY public copyright licence to any Author Accepted Manuscript version arising from this submission. M.G. also received a bursary from the Association of British Academic Oral and Maxillofacial Surgeons (ABAOMS) to help support this work. R.C.R. is a de Pass VC research fellow at the University of Bristol. J.T. is supported by an Academy of Medical Sciences (AMS) Springboard award, which is supported by the AMS, the Wellcome Trust, Global Challenges Research Fund (GCRF), the Government Department of Business, Energy and Industrial strategy, the British Heart Foundation and Diabetes UK (SBF004\1079). R.M.M. was supported by a Cancer Research UK (C18281/A19169) programme grant (the Integrative Cancer Epidemiology Programme) and is part of the Medical Research Council Integrative Epidemiology Unit at the University of Bristol supported by the Medical Research Council (MC_UU_12013/1, MC_UU_12013/2, and MC_UU_12013/3) and the University of Bristol. R.M.M. and A.R.N. are supported by the National Institute for Health Research (NIHR) Bristol Biomedical Research Centre which is funded by the National Institute for Health Research (NIHR) and is a partnership between University Hospitals Bristol and Weston NHS Foundation Trust and the University of Bristol. This publication presents data from the Head and Neck 5000 study. The study was a component of independent research funded by the NIHR under its Programme Grants for Applied Research scheme (RP-PG-0707-10034). The views expressed in this publication are those of the author(s) and not necessarily those of the NHS, the NIHR or the Department of Health. Core funding was also provided through awards from Above and Beyond, University Hospitals Bristol Research Capability Funding and the NIHR Senior Investigator award to A.R.N. B.D. and the University of Pittsburgh head and neck cancer case-control study are supported by US National Institutes of Health (NIH) grants: P50 CA097190, P30 CA047904 and R01 DE025712. The genotyping of the HNSCC cases and controls was performed at the Center for Inherited Disease Research (CIDR) and funded by the US National Institute of Dental and Craniofacial Research (NIDCR; 1X01HG007780-0). The University of North Carolina (UNC) CHANCE study was supported in part by the National Cancer Institute (R01-CA90731).

Author information

Affiliations

Authors

Contributions

M.G., R.C.R. and T.D. conceived the study and all three carried out data curation and analysis, validating the results separately. Head and neck cancer summary genetic data was obtained through multiple collaborations from studies lead by A.R.N., S.T., A.F.O., R.J.H., G.L., B.D., S.B. and P.B. The initial manuscript was drafted by M.G., R.C.R. and T.D. Expert guidance on multivariable MR methodology was provided by E.S. and M.M. provided specific advice on smoking and alcohol instruments such as the community smoking index analysis. G.D.S., J.T., S.T., M.M., P.B., R.M.M. and A.R.N. provided expertise on both observational and genetic epidemiological methodology and analysis. All authors M.G., T.D., E.S., R.M.M., S.T., J.T., A.R.N., P.B., M.M., M.P., S.B., A.F.O., R.J.H., G.L., B.D., G.D.S. and R.C.R. contributed to the interpretation of the results and critical revision of the manuscript. M.G. and this work was jointly supervised by R.C.R., J.T., A.R.N. and G.D.S.

Corresponding author

Correspondence to Mark Gormley.

Ethics declarations

Competing interests

The authors declare no conflicts of interest. The funders had no role in the design of the study; the collection, analysis, and interpretation of the data; the writing of the manuscript; and the decision to submit the manuscript for publication. The authors alone are responsible for the views expressed in this article.

Additional information

Peer review information Nature Communications thanks Verena Zuber, and the other, anonymous, reviewer for their contribution to the peer review of this work. Peer reviewer reports are available.

Publisher’s note Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations.

Supplementary information

Rights and permissions

Open Access This article is licensed under a Creative Commons Attribution 4.0 International License, which permits use, sharing, adaptation, distribution and reproduction in any medium or format, as long as you give appropriate credit to the original author(s) and the source, provide a link to the Creative Commons license, and indicate if changes were made. The images or other third party material in this article are included in the article’s Creative Commons license, unless indicated otherwise in a credit line to the material. If material is not included in the article’s Creative Commons license and your intended use is not permitted by statutory regulation or exceeds the permitted use, you will need to obtain permission directly from the copyright holder. To view a copy of this license, visit http://creativecommons.org/licenses/by/4.0/.

Reprints and Permissions

About this article

Verify currency and authenticity via CrossMark

Cite this article

Gormley, M., Dudding, T., Sanderson, E. et al. A multivariable Mendelian randomization analysis investigating smoking and alcohol consumption in oral and oropharyngeal cancer. Nat Commun 11, 6071 (2020). https://doi.org/10.1038/s41467-020-19822-6

Download citation

Comments

By submitting a comment you agree to abide by our Terms and Community Guidelines. If you find something abusive or that does not comply with our terms or guidelines please flag it as inappropriate.

Search

Quick links

Nature Briefing

Sign up for the Nature Briefing newsletter — what matters in science, free to your inbox daily.

Get the most important science stories of the day, free in your inbox. Sign up for Nature Briefing