Critique 143: A Mendelian randomization assessment of alcohol and cardiovascular disease — 20 July 2014

Holmes MV, Dale CE, Zuccolo L, et al (a total of 155 authors).  Association between alcohol and cardiovascular disease: Mendelian randomisation analysis based on individual participant data.  BMJ 2014;349:g4164 doi: 10.1136/bmj.g4164.

Authors’ Abstract

Objective  To use the rs1229984 variant in the alcohol dehydrogenase 1B gene (ADH1B) as an instrument to investigate the causal role of alcohol in cardiovascular disease.

Design  Mendelian randomisation meta-analysis of 56 epidemiological studies.

Participants  261 991 individuals of European descent, including 20 259 coronary heart disease cases and 10 164 stroke events. Data were available on ADH1B rs1229984 variant, alcohol phenotypes, and cardiovascular biomarkers.

Main outcome measures  Odds ratio for coronary heart disease and stroke associated with the ADH1B variant in all individuals and by categories of alcohol consumption.

Results  Carriers of the A-allele of ADH1B rs1229984 consumed 17.2% fewer units of alcohol per week (95% confidence interval 15.6% to 18.9%), had a lower prevalence of binge drinking (odds ratio 0.78 (95% CI 0.73 to 0.84)), and had higher abstention (odds ratio 1.27 (1.21 to 1.34)) than non-carriers. Rs1229984 A-allele carriers had lower systolic blood pressure (−0.88 (−1.19 to −0.56) mm Hg), interleukin-6 levels (−5.2% (−7.8 to −2.4%)), waist circumference (−0.3 (−0.6 to −0.1) cm), and body mass index (−0.17 (−0.24 to −0.10) kg/m2). Rs1229984 A-allele carriers had lower odds of coronary heart disease (odds ratio 0.90 (0.84 to 0.96)). The protective association of the ADH1B rs1229984 A-allele variant remained the same across all categories of alcohol consumption (P=0.83 for heterogeneity). Although no association of rs1229984 was identified with the combined subtypes of stroke, carriers of the A-allele had lower odds of ischaemic stroke (odds ratio 0.83 (0.72 to 0.95)).

Conclusions  Individuals with a genetic variant associated with non-drinking and lower alcohol consumption had a more favourable cardiovascular profile and a reduced risk of coronary heart disease than those without the genetic variant. This suggests that reduction of alcohol consumption, even for light to moderate drinkers, is beneficial for cardiovascular health.

Forum Comments

The analysis is based on Mendelian randomization, using rs1229984 variant of ADH1B gene as the instrumental variable.  As background, while dozens of genes have been found to relate to alcohol metabolism, the A-allele of rs1229984 SNP is associated with less clearing of alcohol after drinking, and carriers of this allele tend to have flushing and other uncomfortable effects of alcohol.  Hence, carriers of such variant (A allele) tend to less frequently be consumers of any alcohol, and are less likely to be heavy drinkers.  The frequency of A allele of this SNP is very low in Northern Europeans (1% to 4% in this study), but more common in Asians and central Europeans and others.  In the present study, overall 7% of subjects had this allele.

The best recent summary article on alcohol metabolism by Edenberg points out that there are many genes affecting alcohol metabolism, some directly and others (and non-coding areas of genes) that modify the expression of single genes.  Also, “With any gene that affects a person’s risk of developing a complex disease, the protective effect of this genetic variant can be modulated by the environment” (Edenberg, The Genetics of Alcohol Metabolism, Alcohol Res Health 2007;30:5-13).  This raises questions about making generalized statements about the effects of alcohol on disease based on results from the analysis of a single nucleotide polymorphism of a gene. 

Problems in using Mendelian randomization for evaluating the complex association between alcohol and disease:  Forum members have real concern that the allele upon which the present study is based may account for only a small percentage of the effects of genetic and environmental factors related to the consumption of alcohol in the population (this percentage is not mentioned in the present paper).  Another fairly recent study on another important gene related to alcohol metabolism, ALDH2, stated that its alleles “explained only 3% of the variance in reported alcohol intake.”(Au Yeung SL, et al.  Am J Epidemiol 2012; DOI: 10.1093/aje/kwr462), which was reviewed by our Forum in 2012 (available at www.bu.edu/alcohol-forum/critique-076).  Thus, any single allele of a gene modifying alcohol metabolism may be a poor instrumental variable by which to provide general statements on alcohol consumption and disease.

As described by Forum member Djoussé, “Mendelian randomization (i.e., using an instrumental variable, IV, to judge effects) is a great tool for epidemiology only when used appropriately.  Consequently, there is a real danger of misusing a good tool.  There are three major assumptions that an ideal instrumental variable must have:

1. Instrumental variable (here ADH1B SNP) must be related to the exposure (here alcohol intake); the paper provides evidence that carriers of the A allele of that SNP drank fewer alcohol units, so this seems satisfactory from this aspect to use as an IV.
2. The instrumental variable must be unrelated to confounders of the relation of alcohol use with cardiovascular disease (CVD), meaning no relation between ADH1B allele and typical confounders of the alcohol-CVD relation; there are some results in the paper indicating that this gene may have other (non-alcohol-related) effects on CVD and its risk factors. 
3. Conditional on confounders and exposure (alcohol intake), ADH1B must be independent of CVD risk; in other words, ADH1B SNP does not provide additional information once we have measured confounders and we know the drinking status; this is equivalent to ‘No direct causal association between the instrumental variable and CVD (that does not go through alcohol use).’

“Violation of any of those 3 assumptions could lead to biased results, and suggests that the proposed IV would not be appropriate to use in Mendelian randomization.  While assumption #1 above can be easily measured in a study, as the authors did, it is not the case for assumptions 2 and 3.  For #2, one must first measure all relevant confounders, and all epidemiologists know that this is seldom possible; there will always be unknown and unmeasured confounders that could influence the results.  Further, data presented in the text indicate an association between this SNP and CVD risk factors such as blood pressure, body mass index, inflammatory markers, and lipids.

“For item #3, it is possible that another SNP (or SNPs) or genetic loci may be in linkage disequilibrium  (LD) with ADH1B and provide direct causal relation with CVD.  In such a case, ADH1B SNP would not be suitable to serve as an instrumental variable, and use of Mendelian randomization with this ADH1B SNP would be inappropriate.”

Djoussé continues: “My other concern is a lack of relation between ADH1B allele A and HDL-cholesterol  demonstrated in this paper; although possible that CVD benefits provided by alcohol could be independent of HDL (hemostatic factors for example), it may be indicative of a poor choice of instrumental variable. Besides LD and limited knowledge about confounders, other potential limitations of Mendelian randomization not addressed by this paper include canalization: the same phenotype regardless of genotype or environment.  References to this include the following: Nitsch et al. AJE 2006;163:397–403; Lawlor et al.  Stat Med 27: 1133–1328; Smith et al. Int J Epidemiol 2003;32:1–22; and Hernán MA & Robins JM.  Epidemiology 2006;17: 360–372.  At least, the authors should have shown a level of uncertainty by using appropriate tone in their conclusions.”  Added reviewer Finkel: “The last sentence of the conclusion is an unwarranted overreach, an unscientific liberty.”  As stated by Keil, “The major weakness of this paper is that it draws conclusions from a special and presumably small genetic subgroup of the population and applies it to the total population.”

Invited reviewer Kenneth J. Rothman stated: “I don’t place much stock in IV analyses.  The conclusions are heavily dependent on the assumption that you have an instrument, of which you can rarely be confident.  I think of the method as a very blunt instrument and subject to considerable bias for variables that are weakly correlated with exposure and not true instruments.  Those concerns would be applicable to the Holmes paper.”

Forum member McCormick wrote: “This study is OK apart from the fact that the sample in each arm of the study is consistently systematically biased.  In one arm every participant has the rs 1229984 variant of the alcohol dehydrogenase 1B gene (ADH1B).  In the other arm nobody does.  If there is only one effect of the gene, i.e., unpleasant effects after alcohol drinking leading to a reduction in drinking, and it otherwise sits in the body doing nothing and it is not associated with other abnormal genes, then this study would be fine.  However that statement is a rather tall order, as we do not know all of the genes associated with cardiac disease, or with alcohol drinking, or even whether the rs 1229984 variant has some direct non-alcohol effect on cardiac disease or another gene closely associated with it does.  So at first sight this study could be said to be comparing apples with oranges.”

McCormick continued: “Given the impracticality of a very large, long-term blinded randomised controlled trial of alcohol use in humans, the scientific community is forced to use other less valid methods to try to establish what is likely to be correct.  This study is one of those less valid methods.  It can’t be ignored, but needs to be considered in conjunction with other studies, e.g., animal model studies, cohort studies, etc., and the totality of evidence used to judge alcohol’s effects on CVD.  Over the years this process has not rejected the J-shaped curve for the cardio-protective effect of alcohol.”

Forum member Zhang added: “Besides assumptions of IV and weak association of the gene with alcohol, there is also the possibility of pleiotropic effects of the ADH1B variant, a potential violation of one of assumptions of IV.  In other words, ADH1B polymorphisms influence alcohol metabolism, and therefore influence exposure to both alcohol and its metabolites.  If these metabolites influence risk of cardiovascular disease, one of the core assumptions underlying Mendelian randomization is violated.  Further, the authors only examined the direction of association of their IV with the outcome, but did not assess the magnitude of effect of association.”  Zhang concludes: “A true IV is difficult to find, and a weak association between IV and risk factor (that may be the case here) amplifies both the estimate of effect and bias.  This is especially the case if the genetic factor not only has an effect on alcohol consumption but also has an effect on CVD through other mechanism, which violates the assumption of the IV method.” 

Reviewer De Gaetano addressed another weakness of the paper: “This Mendelian paper shows that the protective allele induces persons to drink less, not necessarily to completely avoid drinking.  The protection observed in most studies shows a non-linear gradient; this means it is better manifested at low alcohol consumption, while it is less strong as the consumption levels reach far above optimum, and finally disappears and is transformed into harm at very heavy levels of drinking.  This does not contradict but is in agreement instead with the conclusions of most prospective studies and meta-analyses that drinking alcohol at low doses offers the best protection against CVD (a J-shaped curve).”

De Gaetano continued: “For example, in our meta-analysis (Di Castelnuovo et al. Alcohol dosing and total mortality in men and women: an updated meta-analysis of 34 prospective studies.  Arch Intern Med 2006; 166:2437-2445), we concluded: ‘The association with a lower mortality was apparent up to 42 g/d (about 4 drinks per day), and the lowest mortality was seen at 6 g/d, or about half a drink daily (RR, 0.81 [95% CI, 0.80-0.83]).’  The concluding statement of the authors of the present paper does not reflect such a finding.”

Forum Summary

Using a very large dataset from subjects of European descent, the authors have carried out a Mendelian randomization analysis to estimate the effects of alcohol consumption on cardiovascular disease (CVD), using as the instrumental variable an uncommon allele affecting alcohol metabolism, the ADH1B rs1229984 variant.  People with this variant are unusually sensitive to alcohol (developing flushing and other uncomfortable symptoms from alcohol) and it has been well demonstrated that they are less likely than people without this variant to consume alcohol, and are unlikely to be heavy drinkers.  The authors conclude that since the 7% of their subjects with the uncommon allele had less cardiovascular disease, this indicates that alcohol consumption is unrelated to CVD.

Forum members agree that the analyses were done correctly, but strongly disagree with the premise of the study and the conclusions of the authors, and consider the genetic factor chosen as inappropriate to use as the instrumental variable in Mendelian randomization.  The gene studied explains only a fraction of alcohol consumption in the population (not stated in the paper) and it may have effects on CVD beyond those explained by alcohol consumption.  Thus, this ADH1b allele violates the assumptions required for a variable for Mendelian randomization and would be inappropriate for judging the effects of alcohol on CVD.  Further, in the present study, the authors report little relation of their estimate of alcohol consumption with HDL-cholesterol, while essentially all observational studies, clinical trials, and experimental studies have shown that alcohol is an important determinant of HDL.  This further suggests that the use of the ADH1B allele provides an inadequate estimate of alcohol consumption.

Even without the weakness of the allele chosen to reflect alcohol effects, conclusions derived from Mendelian randomization in general are heavily dependent on the assumption that the instrumental chosen is appropriate, of which you can rarely be confident.  Forum reviewers consider this method to be a very blunt instrument and subject to considerable bias for variables that are weakly correlated with exposure and not true instruments. 

Some reviewers were concerned by the conclusion of the authors that this Mendelian randomization paper shows that all persons should drink less; this does not necessarily mean for people to completely avoid drinking.  The protection of moderate drinking against CVD that is generally observed shows a non-linear gradient: it is better manifested at a low alcohol consumption, while it is less strong as the consumption levels rise above optimum and finally disappears and is transformed into harm with heavy alcohol consumption.  This would be in agreement with the conclusions of most prospective studies and meta-analyses that drinking alcohol at low doses offers the best protection against CVD, as there is a J-shaped curve.

                                                  *                       *                      *

Comments on this critique by the International Scientific Forum on Alcohol Research were provided by the following members:

Luc Djoussé, MD, DSc, Dept. of Medicine, Division of Aging, Brigham & Women’s Hospital and Harvard Medical School, Boston, MA, USA 

Ulrich Keil, MD, PhD, Institute of Epidemiology and Social Medicine,  University of Münster, Münster, Germany

Harvey Finkel, MD, Hematology/Oncology, Boston University Medical Center, Boston, MA, USA

Maritha J. Kotze, PhD, Human Genetics, Dept of Pathology, University of Stellenbosch, Tygerberg, South Africa

Dominique Lanzmann-Petithory,MD, PhD, Nutrition/Cardiology, Praticien Hospitalier Hôpital Emile Roux, Paris, France

Yuqing Zhang, MD, DSc, Epidemiology, Boston University School of Medicine, Boston, MA, USA

Pierre-Louis Teissedre, PhD, Faculty of Oenology – ISVV, University Victor Segalen Bordeaux 2, Bordeaux, France

Dag S. Thelle, MD, PhD, Senior Professor of Cardiovascular Epidemiology and Prevention, University of Gothenburg, Sweden; Senior Professor of Quantitative Medicine at the University of Oslo, Norway

Ross McCormick PhD, MSC, MBChB, Associate Dean, Faculty of Medical and Health Sciences, The University of Auckland, Auckland, New Zealand

Giovanni de Gaetano, MD, PhD, Department of Epidemiology and Prevention, IRCCS Istituto Neurologico Mediterraneo NEUROMED, Pozzilli, Italy

David Van Velden, MD, Dept. of Pathology, Stellenbosch University, Stellenbosch, South Africa

Creina Stockley, PhD, MBA, Clinical Pharmacology, Health and Regulatory Information Manager, Australian Wine Research Institute, Glen Osmond, South Australia, Australia

Andrew L. Waterhouse, PhD, Marvin Sands Professor, Department of Viticulture and Enology, University of California, Davis; Davis, CA, USA

Erik Skovenborg, MD, Scandinavian Medical Alcohol Board, Practitioner, Aarhus, Denmark

Fulvio Ursini, MD, Dept. of Biological Chemistry, University of Padova, Padova, Italy;

In addition, a comment on the paper was requested from Kenneth J. Rothman, DrPH, Boston University School of Public Health, who is a leading methodologist in epidemiology. 

Note:  A member of the Forum, R. Curtis Ellison, was one of the 155 co-authors on this paper and has not provided comments in this critique.  The preparation of the critique was under the direction of Luc Djoussé, MD, DSc, of Harvard University.