Additional File 1

Additional File 2

Additional File 3

Author response:
Our response aims to address the following:
<blockquote>
The lack of pleiotropy is an unconfirmable assumption of MR, and the addition of those models is therefore quite important, as this is a primary weakness of the MR approach. Given that concern, I read the sensitivity analyses using pleiotropy-robust models as the main result, and in that case, they can't test their hypotheses as these models do not show a BMI instrumental variable association. The other weakness, which might be remedied, is that the power of the tests here is not described. When a hypothesis is tested with an under-powered model, the apparent lack of association could be due to inadequate sample size rather than a true null. Typically, when a statistically significant association is reported, power concerns are discounted as long as the study is not so small as to create spurious findings. That is the case with their primary BMI instrumental variable model - they find an association so we can presume it was adequately powered. But the primary models they share are not the pleiotropy-robust methods MR-Egger, weighted median, and weighted mode. The tests for these models are null, and that could mean a couple of things: (1) the original primary significant association between the BMI genetic instrument was due to pleiotropy, and they therefore don't have a robust model to explore the effects of the tobacco genetic instrument. (2) The power for the sensitivity analysis models (the pleiotropy-robust methods) is inadequate, and the authors share no discussion about the relative power of the different MR approaches. If they do have adequate power, then again, there is no need to explore the tobacco instrument.
</blockquote>
We would like to highlight that post-hoc power calculations are often considered redundant since the statistical power estimated for an observed association is directly related to its p-value[1]. In other words, the uncertainty of the association is already reflected in its 95% confidence interval. However, we understand power calculations may still be of interest to the reader, so we will incorporate them in the revised manuscript.
The reason we use inverse variance weighted (IVW) Mendelian randomization (MR) to obtain our main results rather than the pleiotropy-robust methods mentioned by the reviewer/editors (i.e., MR-Egger, weighted median and weighted mode) is that the former has greater statistical power than the latter[2]. Hence, instead of focussing on the statistical significance of the pleiotropy-robust analyses, we consider it is of more value to compare the consistency of the effect sizes and direction of the effect estimates across methods. Any evidence of such consistency increases our confidence in our main findings, since each method relies on different assumptions. As we cannot be sure about the presence and nature of horizontal pleiotropy, it is useful to compare results across methods even though they are not equally powered. It is true that our results for the genetically predicted effects of body mass index (BMI) on the risk of head and neck cancer (HNC) differ across methods. This is precisely what led us to question the validity of our main finding (suggesting a positive effect of BMI on HNC risk). We will clarify this in the discussion section of the revised manuscript as advised.
We understand that the reviewer/editors are concerned that we do not have a robust model to explore the role of tobacco consumption in the link between BMI and HNC. However, we have a different perspective on the matter. If indeed, the main IVW finding for BMI and HNC is due to pleiotropy (since some of the pleiotropy-robust methods suggest conflicting results), then the IVW multivariable MR method is a way to explore the potential source of this bias[3]. We were particularly interested in exploring the role of smoking in the observed association because smoking and adiposity are known to influence each other [4-9] and share a genetic basis[10, 11].
References:
(1) Heinsberg LW, Weeks DE: Post hoc power is not informative. Genet Epidemiol 2022, 46(7):390-394.
(2) Burgess S, Butterworth A, Thompson SG: Mendelian randomization analysis with multiple genetic variants using summarized data. Genet Epidemiol 2013, 37(7):658-665.
(3) Burgess S, Davey Smith G, Davies NM, Dudbridge F, Gill D, Glymour MM, Hartwig FP, Kutalik Z, Holmes MV, Minelli C et al: Guidelines for performing Mendelian randomization investigations: update for summer 2023. Wellcome Open Res 2019, 4:186.
(4) Morris RW, Taylor AE, Fluharty ME, Bjorngaard JH, Asvold BO, Elvestad Gabrielsen M, Campbell A, Marioni R, Kumari M, Korhonen T et al: Heavier smoking may lead to a relative increase in waist circumference: evidence for a causal relationship from a Mendelian randomisation meta-analysis. The CARTA consortium. BMJ Open 2015, 5(8):e008808.
(5) Taylor AE, Morris RW, Fluharty ME, Bjorngaard JH, Asvold BO, Gabrielsen ME, Campbell A, Marioni R, Kumari M, Hallfors J et al: Stratification by smoking status reveals an association of CHRNA5-A3-B4 genotype with body mass index in never smokers. PLoS Genet 2014, 10(12):e1004799.
(6) Taylor AE, Richmond RC, Palviainen T, Loukola A, Wootton RE, Kaprio J, Relton CL, Davey Smith G, Munafo MR: The effect of body mass index on smoking behaviour and nicotine metabolism: a Mendelian randomization study. Hum Mol Genet 2019, 28(8):1322-1330.
(7) Asvold BO, Bjorngaard JH, Carslake D, Gabrielsen ME, Skorpen F, Smith GD, Romundstad PR: Causal associations of tobacco smoking with cardiovascular risk factors: a Mendelian randomization analysis of the HUNT Study in Norway. Int J Epidemiol 2014, 43(5):1458-1470.
(8) Carreras-Torres R, Johansson M, Haycock PC, Relton CL, Davey Smith G, Brennan P, Martin RM: Role of obesity in smoking behaviour: Mendelian randomisation study in UK Biobank. BMJ 2018, 361:k1767.
(9) Freathy RM, Kazeem GR, Morris RW, Johnson PC, Paternoster L, Ebrahim S, Hattersley AT, Hill A, Hingorani AD, Holst C et al: Genetic variation at CHRNA5-CHRNA3-CHRNB4 interacts with smoking status to influence body mass index. Int J Epidemiol 2011, 40(6):1617-1628.
(10) Thorgeirsson TE, Gudbjartsson DF, Sulem P, Besenbacher S, Styrkarsdottir U, Thorleifsson G, Walters GB, Consortium TAG, Oxford GSKC, consortium E et al: A common biological basis of obesity and nicotine addiction. Transl Psychiatry 2013, 3(10):e308.
(11) Wills AG, Hopfer C: Phenotypic and genetic relationship between BMI and cigarette smoking in a sample of UK adults. Addict Behav 2019, 89:98-103.

Belinda Nicolau

Eduardo Franco

eLife Assessment
The findings represent an important contribution to understanding whether BMI influences head and neck cancer (HNC) risk after accounting for tobacco use. Within the context of the Mendelian Randomization (MR) field, the strength of evidence appears convincing, supported by rigorous methods and a thorough exploration of multiple genetic models of adiposity using diverse MR approaches. Limitations include the absence of associations in sensitivity models designed to better account for pleiotropy, which prevents evaluation of whether incorporating an instrumental variable for tobacco use would alter the findings. Additionally, the lack of a formal power assessment for detecting associations with the instrumental variables employed limits the interpretability and reach of the results.

anonymous

Joint Public Review:
Summary:
The authors have conducted the largest to date Mendelian Randomization (MR) analysis of the association between genetically predicted measures of adiposity and risk of head and neck cancer (HNC) overall and by subsites within HNC. MR uses genetic predictors of an exposure, such as gene variants associated with high BMI or tobacco use, rather than data from individual physical exams or questionnaires, and if it can be done in its idealized state, there should be no problems with confounding. Traditional epidemiologic studies have reported a variety of associations between BMI (and a few other measures of adiposity) and risk of HNC that typically differ by the smoking status of the subjects. Those findings are controversial given the complex relationship between tobacco and both BMI and HNC risk. Tobacco smokers are often thinner than non-smokers, so this could create an artificial ('confounded') association that may not be fully adjusted away in risk models. The findings of a BMI-HNC association are often attributed to residual confounding, and this seems ripe for an MR approach if suitable genetic instrumental variables can be created. Here, the authors built a variety of genetic instrumental variables for BMI and other measures of adiposity, as well as two instrumental variables for smoking habits, and then tested their hypotheses in a large case-control set of HNC and controls with genetic data.
The authors found that the genetic model for BMI was associated with HNC risk in simple models, but this association disappeared when using models that better accounted for pleiotropy, the condition when genetic variants are associated with more than one trait, such as both BMI and tobacco use. When they used both adiposity and tobacco use genetic instruments in a single model, there was a strong association with genetically predicted tobacco use (as is expected), but there was no remaining association with genetic predictors of adiposity. They conclude that high BMI/adiposity is not a risk factor for HNC.
Strengths:
The primary strength was the expansive use of a variety of different genetic instruments for BMI/adiposity/body size, along with employing a variety of MR model types, several of which are known to be less sensitive to pleiotropy. They also used the largest case-control sample size to date.
Weaknesses:
The lack of pleiotropy is an unconfirmable assumption of MR, and the addition of those models is therefore quite important, as this is a primary weakness of the MR approach. Given that concern, I read the sensitivity analyses using pleiotropy-robust models as the main result, and in that case, they can't test their hypotheses as these models do not show a BMI instrumental variable association. The other weakness, which might be remedied, is that the power of the tests here is not described. When a hypothesis is tested with an under-powered model, the apparent lack of association could be due to inadequate sample size rather than a true null. Typically, when a statistically significant association is reported, power concerns are discounted as long as the study is not so small as to create spurious findings. That is the case with their primary BMI instrumental variable model - they find an association so we can presume it was adequately powered. But the primary models they share are not the pleiotropy-robust methods MR-Egger, weighted median, and weighted mode. The tests for these models are null, and that could mean a couple of things: (1) the original primary significant association between the BMI genetic instrument was due to pleiotropy, and they therefore don't have a robust model to explore the effects of the tobacco genetic instrument. (2) The power for the sensitivity analysis models (the pleiotropy-robust methods) is inadequate, and the authors share no discussion about the relative power of the different MR approaches. If they do have adequate power, then again, there is no need to explore the tobacco instrument.
Reviewing Editor Comments:
We suggest that the authors add power estimates to assess whether the sample size is sufficient, given the strength and variability of the genetic instruments. It would also be helpful to present effect estimates for the tobacco instruments alone, to clarify their independent contribution and improve the interpretation of the joint models. In addition, the role of pleiotropy should be addressed more clearly, including which model is considered primary. Stratified analyses by smoking status are encouraged, as prior studies indicate that BMI-HNC associations may differ between smokers and non-smokers. Finally, the comparison with previous studies should be revised, as most reported null findings without accounting for tobacco instruments. If this study finds an association, it should not be framed as a replication.

Reassessing the link between adiposity and head and neck cancer: a Mendelian randomization study

Figures and data

Data sources and instruments for other adiposity-related anthropometric measures

F-statistics and variance explained for other adiposity-related anthropometric measures

Forest plot for the genetically predicted effects of body mass index on the risk of head and neck cancer and its subsites

Forest plot for the genetically predicted effects of waist-to-hip ratio on the risk of head and neck cancer and its subsites

Forest plot for the genetically predicted effects of waist circumference on the risk of head and neck cancer and its subsites

Forest plot for the genetically predicted effects of BMI on the risk of HNC and its subsites, before (univariable-black) and after (multivariable-blue) accounting for comprehensive smoking index (CSI)

Forest plot for the genetically predicted effects of BMI on the risk of HNC and its subsites, before (univariable-black) and after (multivariable-blue) accounting for smoking initiation (SI)

Be the first to read new articles from eLife