While drinking alcohol provides pleasure in several ways, no-one disputes that it is a drug, potentially both addictive and dangerous, so there can be too much of a good thing. But how much is too much? This is a question on which the jury is still out.
I have written about it twice before, in July 2016, following the introduction of revised, lower, drinking guidelines from the UK’s chief medical officer. But new research and new advice keeps coming along. This post is prompted by a recent paper in The Lancet, and the considerable discussion that resulted.
It would be nice to think that both the new research and the new implied advice meet the highest scientific standards. I would certainly heed them if they did, yet there is a research protocol for testing drugs – and for reporting these tests – from which alcohol always seems to be excused.
The paper in question was published by The Lancet in April 2018. It was snappily titled Risk thresholds for alcohol consumption: combined analysis of individual‑participant data for 599 912 current drinkers in 83 prospective studies and was the work of no fewer than 120 authors. (I cannot guarantee that this link to it will grant you access.) The analysis reinforced previously published findings that the u-shaped relationship between intake of red wine and incidence of heart disease does not apply to other alcohol-related conditions such as liver cancer and throat cancer; for these, the relationship is monotonic, so the safest consumption level is zero.
The popular media naturally made up their own headlines when reporting it. The Daily Mail Online had “There is NO safe limit for drinking: Going teetotal is only way to avoid risking health, major study says”. The online Express had “Just one drink a day is too much”.
The Sun disagreed: ”Strict booze rules should be lowered AGAIN — to just five pints a week”, while charmingly translating the 599,912 current drinkers as “600,000 boozers”. It also published an endorsement by Prof. Edoardo Casigilia, conveniently failing to mention that he was one of the 120 authors. The paper in The Lancet did not offer advice: it reported research. This did not stop The Sun’s headline offering its own.
My preferred online newspaper is The Times and The Sunday Times. The Times had the ungrammatical “Wine warning: couple of glasses a night shortens life by two years”. (Cue for the old saw: forgoing your couple of glasses of wine a night will not extend your life by two years, it will only seem like that.) Hundreds of readers posted online comments, generating 367 comments – one of them mine (more on that later). They included: “Looks like being involved in an alcohol study gives you a one in 15 chance of dying!” and ‘I gave up wine, fast women and fatty foods. It was the longest half hour of my life.’ However, my favourite is:
‘Those familiar with the works of the famous Oenologists, Monty Pultchano, Mal Beck and Reece Ling, will be well aware of the enjoyable and life extending qualities of their fine produce and viticulture. Wine rejoices the heart of man, and joy is the mother of all virtues.
My own comment was a bit late going in because I needed time to access and read the Lancet paper (it is quite long). I doubt that many, if any, of the other commentators did so and I doubt whether the reporters at the newspapers had either. Probably all they had to go on was a press release from The Lancet.
Accessing the paper was not helped by the journalists repeating mistakes that they almost always make when reporting new research. They are very bad at identifying their sources. The full name of the article is virtually never given; sometimes not even the journal is named; and quite often the issue date and/or the names of any authors are missing. It can then be quite challenging to track down the article, and the search often ends in two or more possible contenders.
So what does this paper actually say? First it states its data sources. It was a meta-analysis. A meta-analysis is an re-analysis of several existing studies, pooled. So the sample size of 599,912 was the sum of the samples of the 83 existing prospective studies. (A prospective study is one that tracks a fixed cohort of respondents over time.) None of the media reports that I read noted that there was no new sampling or measurement involved, so the usual description as a “study” is somewhat misleading.
Next it states the research methodology. Part of this is given in a supplementary appendix. The printed paper is less than nine pages long, excluding declarations of interest, acknowledgements and references, but the appendix, only available online to save paper and space in the journal, is 49 pages long. You can’t fully understand the methodology unless you are a statistician (like me), and you won’t fully understand it even if you are because it isn’t made clear. A fair summary is that a statistical model was fitted to the data, except there isn’t a good enough description of the model for anyone to be sure what it consists of, to reproduce it, to play with it themselves or to confirm the results.
Trials of potential new drugs are not conducted using meta-analyses or models. The protocols are complex and detailed, but essential elements are (a) use of a control sample receiving either an existing treatment or a placebo (b) allocation of subjects to either the test arm or the control arm at random, and (c) the administration of the drugs and placebo is ‘double-blind’. This means that not only does the patient or subject not know which they are getting (i.e. which arm they are in), but the professional who administers or supplies the drug doesn’t know either, so he or she cannot influence the subject, even unintentionally by length of consultation, body language, tone of voice etc. Naturally, everything is labelled, numbered and documented so the researcher behind the scenes does know who is receiving what.
Alcohol has not been tested this way because it cannot be, for reasons of both practicality and ethics. You know if you are consuming it rather than a placebo, and no-one is going to control your intake but yourself. Unfortunately, this does not make alternative research protocols any more valid. It is a clear case of the best being the enemy of the good.
What then is the value of studies such as this? They are still useful. They cannot in themselves prove or disprove causality, but they can provide valuable correlations. If those who drink a lot get ill more or die younger, and if medical research shows that alcohol can cause these illnesses, then we are more than halfway there. This is how smoking was found to be unhealthy, and exposure to various industrial or environmental substances such as asbestos.
But several problems of interpretation remain. One is that a host of cofactors might be involved in the chain of causality. For example, do those who lead stressful lives (perhaps through poverty, being victims of crime, or poor health), and are going to die young anyway, take to drink for comfort, but are not affected much by it? Conversely, other research has concluded that exercise counters the negative effect of alcohol on health. Cofactors are present regardless of the design of the research, but can be ignored in analysis of drug trials as they affect both arms equally. For observational studies such as those analysed in the Lancet paper, this is not the case. Cofactors can be incorporated into the statistical models, but only if the data were collected in the first place, and only if the researchers decide to do so. Neither stress nor exercise was included in the Lancet paper.
A further problem in measuring the relative risk of different levels of alcohol intake is: relative to what? In drug trials the placebo arm provides the benchmark. In observational studies of alcohol, the usual recourse is to use teetotalers. Unfortunately, these differ from drinkers in many ways other than alcohol consumption. Some don’t drink because they lack the enzymes that metabolise alcohol, and get a flushing reaction. Some abstain on religious grounds, so differ in ethnic profile and other ways, such as choice of diet. Some are reformed alcoholics and other ex-drinkers, with a medical history all of their own, the so-called sick quitters. Others are drinkers in denial. So non-drinkers are a heterogeneous group, no part of which properly matches drinkers. As is apparent from its title, the Lancet paper avoided this problem by analysing only drinkers. So relative risk could only be assessed at different levels of consumption, not for consumption vs. abstinence. It is follows that none of the newspaper headlines can be justified by the analysis.
There is a further problem specific to alcohol studies. In clinical trials, dosage is closely monitored. Indeed, drug trials require three separate stages with separate samples of increasing size, so that toxicity can be detected early, and optimal dosage established later. But observational studies, such as the 83 used for the Lancet paper, rely on self-reporting of consumption, which is both unreliable (inaccurate at random) and biased (under-reported on average, by about 33%-50%). There is a large research literature on this, which papers such as this one in The Lancet consistently ignore. So as to be sure, I wrote to the lead author, Dr. Angela Wood of the Department of Public Health and Primary Care, University of Cambridge, with the following question:
“May I ask whether consideration was given to treatment in analysis for under-reporting of alcohol consumption?”
Her reply was:
“…we agree it is a limitation. … we accept that the self-reported consumption levels may still be prone to under-reporting bias.”
No mention of this is made in the paper. It is so important. If we assume just 33% under-reporting, then those survey respondents reporting consumption of 14 units/week were actually consuming 21 units/week. So the level of risk of contracting, say, liver cancer, that the paper attaches to 14 units/week does not actually occur unless you drink 21 units/week. In other words, risks are substantially overstated.
Papers in The Lancet are peer-reviewed, and this shortcoming should have been picked up and corrected. But the peer-review process, like democracy, is imperfect, merely better than the alternatives. I speak as a reviewer myself, for the International Journal of Market Research, and as an author of peer-reviewed papers, one published with a (non-fatal) error that the reviewers should have spotted.
Finally, the Lancet paper makes another serious howler of interpretation. I reacted to this, and found myself in very good company in doing so. I sent the following comment to The Times:
A global study published in The Lancet this week cautioned that there was no safe level of alcohol consumption, even for those that drink in moderation. That is a statement about absolute risk, and it doesn’t say that. … The only occurrence of the word ‘safe’ is in ‘Our results show that the safest level of drinking is none.’ which is a statement about relative risk.
The Lancet paper shows that the relative risk (including alcohol-related morbidity as well as premature mortality) at 14 units/week is about +10% compared to abstinence and at 21 units/week about +20%. However, just 1.2% of deaths in the UK in 2016 were due to alcohol (7,327 out of 597,206) and it is unlikely that more than a tiny proportion of these were moderate drinkers.
Figures on morbidity are less clear because they are collected by health authorities each using different criteria; and because alcohol is sometimes just one of several factors. But roughly, 1% of the UK population is admitted to hospital each year with alcohol as a factor, and it is the only factor for just a quarter of these. Again, these are mostly heavy drinkers. Many believe that the pleasure from moderate drinking more than counterbalances the small (absolute) risks and will treat the media’s sloppy reporting and scaremongering accordingly.
The confusion of absolute risk with relative risk is widespread. It leads to the well-known saying that the most dangerous part of flying is the journey to the airport. As a statement of relative risk, this is true. As a statement of absolute risk, it is alarmist nonsense. In the case of alcohol studies, the confusion means that the public are unnecessarily alarmed, constantly reminded that moderate drinking raises relative risk even though the absolute risk is still negligible. Thankfully, many moderate drinkers see past this, even if for the wrong motives.
The company in which I found myself was Prof. David Spiegelhalter OBE FRS, Winton Professor of the Public Understanding of Risk in the Statistical Laboratory at the University of Cambridge, Chair of the Winton Centre for Risk and Evidence Communication, current president of the Royal Statistical Society (I am a mere ex-vice president, and a Chartered Statistician) and world champion at Loop. He wrote a long piece on the Lancet paper making mostly the same points as me, though better. (Again I cannot guarantee access.) It starts by admonishing The Lancet for breaking its own guidelines on publishing statements of risk. He describes the statistical model as “large and complex”. Then he re-interprets the analysis in a less alarmist manner. For example, he illustrates the low absolute risk of modest drinking by pointing out that for every 25,000 people getting through 400,000 bottles of gin a year, just one will have a serious alcohol-related condition. He concludes “But claiming there is no ‘safe’ level does not seem an argument for abstention. There is no safe level of driving, but government do not recommend that people avoid driving. Come to think of it, there is no safe level of living, but nobody would recommend abstention.”
His piece was duly picked up by the Twitterati and went viral, but you can read all that yourself. I’m stopping here – I need a drink.