Skip to content
 

What to think about this new study which says that you should limit your alcohol to 5 drinks a week?

Someone who wishes to remain anonymous points us to a recent article in the Lancet, “Risk thresholds for alcohol consumption: combined analysis of individual-participant data for 599 912 current drinkers in 83 prospective studies,” by Angela Wood et al., that’s received a lot of press coverage; for example:

Terrifying New Study Breaks Down Exactly How Drinking Will Shorten Your Life

Extra glass of wine a day ‘will shorten your life by 30 minutes’

U.S. should lower alcohol recommendation because booze shortens our lives, study says

Here’s the key graph from the research paper:

According to one of the news articles, 100 grams of alcohol per week, which according to the graph above is approximately the maximum safe dose, is equivalent to “five standard 175ml glasses of wine or five pints [of beer] a week.”

The press coverage of this study was uncritical and included this summary from our friend David Spiegelhalter who described it as “massive and very impressive”:

The paper estimates a 40-year-old drinking four units a day above the guidelines [the equivalent of drinking three glasses of wine in a night] has roughly two years’ lower life expectancy, which is around a 20th of their remaining life. This works out at about an hour per day. So it’s as if each unit above guidelines is taking, on average, about 15 minutes of life, about the same as a cigarette.

And the statistics

On one hand, I’m always suspicious of headline-grabbing studies. On the other, I respect Spiegelhalter.

I took a look at the research article, and . . . it’s complicated. They need to do a lot with their data. Most obviously, they need to adjust for pre-treatment differences in the groups, differences of age, sex, smoking status, and other variables. They adjust for whether people are current smokers or not, but they don’t seem to adjust for the level of smoking or past smoking status; maybe these data are not available? The researchers also have a complicated series of decisions to make regarding missing data, inclusion of nonlinear terms and interactions in their models, and other statistical details.

It was not clear to me exactly how they got the above graph, and how much the pattern could be distorted by systematic differences between the groups, not caused by alcohol consumption.

That said, you have to do something. And, from a casual look at the paper, the analyses seem serious and the results seem plausible.

The logical next step is for the data and analysis to be shared. Immediately. Put all the data on a spreadsheet on Github so that anyone can do their own analyses. Maybe the data are already publicly available and easily accessible? I don’t know. There are various questionable steps in the published analysis—that’s fine, no analysis is perfect!—and the topic is important enough that it’s time to let a thousand reanalyses bloom.

P.S. More here in this excellent post by Spiegelhalter.

39 Comments

  1. One of the biggest issues in alcohol research is that people’s responses to surveys have complicated survey error. The “professional” drinker who every night opens a 12 pack of beers and nurses it along from 5pm to 5am constitutes on the order of 1 percent of the population. They have a very clear low noise idea of their drinking, because they count them carefully to keep from killing themselves. It’s sad, but true.

    On the other hand, the college party kids drink anywhere from 0 to 8 units in a day, usually not in clear-cut servings, and widely varying from one day to the next… They have really NO idea what their total consumption for 30 days is for example. Typical 40 year old parent of two kids… they drink maybe zero, maybe one, maybe two, maybe three units on a given night, and it varies a bunch depending on whether they’ve gotten their busy lives together enough to go to the store, or they’re making up for kids soccer going late by ordering take-out… they rarely have 8 drinks like the college party animals, but they also have no idea what their consumption over 30 days has been.

    Comparing total sales in the US to extrapolations from phone surveys shows that in fact there is a serious bias downward in survey estimates.

    So, not having looked at this study *at all yet* my first inclination is to rewrite the x axis label to “reported usual alcohol consumption” rather than actual. But because of the low bias, if you report say 300mg you’re probably drinking 400-600 and so the ground-truth will be very hard to discover unless you run a decent randomized *controlled* trial where people who already drink are given an allotment of their favorite beverage each week, and at the end of the week they report what they actually consumed, or something like that.

  2. Garnett says:

    A doctor friend once told me that he was trained to double the amount of alcohol consumption reported by patients.

    Anyway, in my younger days as a behavioral health research statistician we used a recall calendar to document alcohol use on each day in the past week, month, year, and, believe it or not, _ever_. This measurement technique was the primary endpoint for drug trials of anti-craving medications. I always found that method problematic, but the patients in these studies were very heavy drinkers, and what you say in your first paragraph is consistent with what these patients often reported.

  3. Jonathan says:

    And the mechanism by which this happens is? When you compare anything to cigarettes, you are dealing with a mechanism by which lung and other cancers develop, as well heart disease. Many of these pathways are known down to the cellular change levels. What is the mechanism by which all this mortality occurs? I don’t see one. In fact, what I see is a question: what if many of these people are helped by alcohol? What if the stresses of their lives would have killed them earlier if not for alcohol to relieve their internal pain? We have some knowledge of how stress causes actual disease and people use alcohol because it pushes away their stress, so to me it looks like a graph of human need only partly being quenched by the relatively ineffective use of alcohol. (My hope is that cannaboids in some form work much better because alcohol has so many negative behavioral effects.)

    You used the phrase in the 3rd paragraph from the bottom: ‘not caused by alcohol consumption’. But what is the model by which alcohol consumption suddenly transforms so you can use the word ‘caused’ in that way? You note that there are many questions about how the data was fit, but all that fitting was to examine the variable alcohol in relation to mortality. That isn’t causal. That isn’t a model in which cause exists.

    I don’t know your background. I know you teach at Columbia. Ever been around working people? Working people drink because they need to drink because life is bleeping hard. The Boston Globe did a report about the status of African-Americans here in the most progressive state in the union: they have net worth of $6 per family. As I assume you know, that’s true of over half of all Americans: they have nothing. That’s stress. That’s a really good reason to drink: you worry every week, every month about paying your bills, hoping you don’t have to put money into your car, into new shoes for your kid and a few drinks relaxes you enough to keep going and you’re telling them they should just suck it up and live without relief because statistically drinking more in a hard world kills you sooner? I think the rational response should be: make the world bleeping easier and we won’t need drink.

    • Phil says:

      Jonathan,
      Jonathan,

      I think Andrew works really hard! He certainly works harder than I do, by maybe a factor of 2 or 3. I only work part-time so I wouldn’t necessarily call myself a ‘working person’ but I’d certainly call Andrew a working person. There are lots of kinds of work, of varying degrees of stress and varying degrees of pay. Being an Air Traffic Controller is supposedly pretty stressful, as is being an emergency room doctor. If you ask an ER doctor coming off a shift if she has every known any working people, you deserve what you get.

      A bigger problem is that you seem to ascribe to Andrew a point he never made. Nowhere does he say “you’re telling them they should just suck it up and live without relief”, I don’t see that he offers any advice or judgment whatsoever about how much or whether people should drink alcohol. He sees results of a study and wants the researchers to make the data available so other people can do their own analyses of it, and somehow you take offense. WTF, dude.

      Your point about causality is a good one, though. Some people drink because of the conditions that shorten their lives, but their lives aren’t necessarily shortened by drinking. If suicides are higher among drinkers, is it because drinking increases depression or because more depressed people drink more?

      It would be interesting to see the relationship between drinking and age-adjusted mortality for different causes of death, if those data are available. Also, I wonder about binge vs regular drinking. About twice a week I have two beers or two glasses of wine, but I rarely drink three and nearly never drink four or more. How do people like me compare to people who get plastered every two weeks or so, thereby consuming the same amount of alcohol?

      So many interesting questions. I do hope the data are made available.

    • Dzhaughn says:

      “Accidental injuries” is one simple and clear mechanism.

  4. This paper was published back in April, and was the subject of quite a lot of discussion in the UK. Many headlines said things like “No Safe Level Of Drinking” and in fact David Spiegelhalter (whom I also respect enormously) gave a broader commentary which included something like “There’s no safe level of driving either, but nobody suggests we should give up driving”.
    One thing to note is that they chose to exclude non-drinkers, and you have to look at the “supplementary appendix” (page 31) to find a plot which includes ex-drinkers and never-drinkers as well as drinkers at different levels. Never-drinkers have an ACM hazard ratio of 1.2 compared to baseline (consistent with previous findings of a J-shaped curve, which is accounted for by the protective effects of very small amounts of alcohol against CVD).
    If you’re interested, it’s definitely worth your unpacking the Appendix, where they include more data.
    Thanks for running an interesting blog
    Timandra

  5. Dale Lehman says:

    +100 to your conclusion about the next step being to release the data (regardless of whether or not it comes from public sources). This example is a perfect example of how things should be changed. Assembling and documenting the data should provide rewards to the authors of this paper. I have no problem with them also doing an analysis – even the first analysis. But I think the data itself should receive more credit than the analysis they conducted. The many forking paths can provide ample grounds for many studies, and hopefully, an advancement in our understanding of the relationships between alcohol consumption and health. This one particular study has little value to me – it is one exploration among n (large n) possible studies of this data. So, let the authors receive the credit for obtaining, documenting, and making the data available.

  6. So ‘maximum safe dose’ := ‘maximum dose for which hazard ratio = 1’? Seems like a conservative way to define safety. Not to be trite about it, but: alcohol is enjoyable. People might be willing to trade off a little safety for a little enjoyment.

  7. Keith O’Rourke says:

    I think there is another alchemy here – converting systematic error into random error.

    The major uncertainty in epidemiological studies is by far systematic error but for some reason most meta-analyses of epidemiological studies try to use random effects models in an attempt to convert systematic into random. For a discussion of why this is hopeless see Greenland S, O’Rourke K. Meta-analysis. 2008. or likely better Good practices for quantitative bias analysis. Lash et al https://academic.oup.com/ije/article/43/6/1969/705764

    • Sorry to digress. It’s a curious thing, for which I have no answer. I lose my taste for alcohol [wine] by fall. Drink 3 or 4 glasses of wine per week for July and August when I’m with my friends out on the patio.

      I have wondered whether the sulfite additives & pesticides are causes for concern. I was passing by a neighborhood liquor story where there was a wine & cheese tasting event. So tried to get some information from the wine rep hosting the winetasting. I gatherthat some people are allergic to sulfites.

  8. Kyle C says:

    They adjusted for “alcohol consumption amount and status (ie, non-drinker vs current drinker), … age, sex, history of diabetes and smoking … and … history of cardiovascular disease.” That’s not enough, IMHO. Where’s income/wealth? That’s my problem with every one of these alchohol studies — basically they find that the healthiest drinking pattern is the one that richer people follow. Who drinks about one drink a day, no more, no less, mostly wine? People of the upper middle class and above, who we know are healthier for all kinds of other reasons.

  9. Marcus Kubsch says:

    I wonder how many of the people in the sample are actually in the groups that consume more than 100g alcohol per week. Possibly not that many? In addition, there are just three level of weekly consumption after 100g. I wonder whether the trend implied by these three (!) data points is really there. There could be a lot of variation in between.

  10. Max Griswold says:

    As an epidemiologist, I found this study most problematic due to the use of all-cause mortality as the outcome. However, we found similar results in our recent article, using a very different method to aggregate to “all-cause”: http://dx.doi.org/10.1016/s0140-6736(18)31310-2

    And endeavored to make the data available: https://data.mendeley.com/datasets/5thy2mcwn7/5

    Not sure about this group of authors but we found it hard to get the Lancet to post the data initially (let alone the code).

    • Max, thanks for posting again, we had a brief discussion in comments here on yet another alcohol paper, a while ago. Can you point me at what you consider the “raw” data in your data.mendeley.com archive? I see you put up stuff like the data for your plots that’s outputs of your model.

      The thing that seems mostly likely to be your “input” data is the “attributable_burden_results” but it seems like these too are estimates… outputs of your model. I didn’t download it as it’s 600+MB but if that’s where I should look I can grab that.

      What model raw input data did you actually use? I’m assuming things like maybe survey results or the like, or was this entirely a meta-analysis of other people’s estimates?

      • Max Griswold says:

        No, the file you’re referring to is the results from the analysis.

        Regarding raw input data, given the sheer size of the exposure data (and DUAs for certain datasets), I cannot upload to Mendeley. I instead uploaded the sources used. I could tabulate the microdata and upload that, though I’m not sure if that adds much value above and beyond the source list. One would have to extract those sources individually, which I realize is a tall burden. Still trying to think of a good solution for how to host that dataset (given it’s collectively about 4tb).

        For the meta-analysis, since those numbers were just extracted from published studies, I’m in the process of cleaning the input sheets for Mendeley and will have those up soon. I think this raw data is of more interest than the exposure data, and more open to interpretation, based on modeling approaches, so hoping to have that up soon for other researchers.

        • 4TB hunh? That is an issue. A suggestion, can you load it into a SQLite database and wind up with something more efficient, like maybe a 100GB file or something?

          • Max Griswold says:

            I haven’t tried that; I’ll look into that option. I mostly use csv, Rdata, and HDF5 in my work, depending on the size of data. My thinking at this point is let people investigate specific sources of interest (since most people are interested in country-specific results, rather than global), and if there’s a desire for full raw data/results, I’ll cross that bridge at that point. No inquiries yet, only inquiries about the relative risks and modeling.

            • Honestly I wish more people would use SQLite as a data-transfer method. It’s a full SQL based database in a single file. It’s moderately efficient, and it’s capable of self-documenting in the sense that you can put tables in their with metadata about the other tables in text fields and things like that. An all-in-one data dump.

              • Jens Åström says:

                Just beware of the known compatibility issue with putting SQLite DBs on CIFS shares, which is common in networked Windows environments. Could trip you up.

              • I think these incompatibilities are with locking and multi use but yes I strongly recommend building any SQLite database on a local drive, in this case probably a spinning disk given the size. If you are using a btrfs file system on Linux you also need to disable copy on write for the file, for efficiency reasons.

        • Max if you buy a Western digital red 6tb drive and format as ext4 Linux fs and copy the files and mail it to me together with a second blank drive, I will investigate consolidating the data and hosting it, and make it available back to you as well. Cost to you in hardware is about $350. If interested contact me at contact info on my site
          http://www.lakelandappliedsciences.com

  11. Z says:

    “they need to adjust for pre-treatment differences in the groups”

    Well this is a time varying treatment. Any impact on the outcome would be from sustained drinking over an extended period. And drinking patterns can change over time. Drinking patterns from year 2-3 after baseline might be confounded by variables observed during year 1-2 after baseline. These confounding variables in year 1-2 are post treatment in year 0-1 but pre-treatment in year 2-3. Further, they might be influenced by treatment in year 0-1. Methods for estimating effects of time-varying treatments with treatment confounder feedback are required.

    • Andrew says:

      Z:

      Good point. “Adjust for pre-treatment differences in the groups” is a form of non-model-based advice. As problems get more complicated, such simple rules aren’t enough, and you have to move toward more of an actual model of how the variables relate to each other.

    • This is my biggest problem with alcohol research in general. Alcohol consumption and its good and bad effects are very likely cumulative over decades. Any model for the effects which regresses ultimate endpoint like mortality against a noisy estimate of current consumption is hopelessly misguided right from the start. What’s needed is a timeseries model. It’s entirely possible that the kind of data we have collected is hopelessly useless for the task. This means we need to collect different data. Until we acknowledge the requirements of timeseries modeling, the field will go nowhere.

  12. According to Wikipedia’s page on alcohol consumption per capita, the French consume 12.2 liters of pure alchohol on average per year. That works out to:

    (12.2 liter / year)
      * (33.8 ounce / liter)
      / (12 ounce / beer)
      * (100% pure alcohol / 5% beer)
      / (52 weeks / year)
    = 13 beers / week
    

    Calculating in units of grams (785 grams/liter, the internet tells me for ethyl alcohol) works out to 184 grams/week, which is consistent with what 100 grams being roughly 5 drinks given my rough approximation by American beer. It’s roughly 2 glasses/wine per person per day. I guess they make up for it with the Mediterranean diet.

    Russia is 15.1 liter/year, the UK 11.6, the US 9.2, and Japan 7.2 for comparison.

  13. zbicyclist says:

    It’s either a very good sign, or a very bad sign, that it’s the Lancet press office that asked for the absolute levels which were ignored by the Lancet editorial staff. This is from the Spiegelhalter article linked above:

    “But in spite of the Lancet’s own guidelines for meta-analyses saying
    For risk changes or effect sizes, give absolute values rather than relative changes
    the paper did not report any absolute risks, meaning that readers couldn’t tell how dangerous drinking alcohol really was for them. Fortunately this extraordinarily lax review process was countered by the Lancet press office who asked for absolute risk estimates from the authors. This is truly excellent practice for which the press office deserve sincere congratulations.”

    Spiegelhalter appears frequently as an expert on the BBC stats podcast “More or Less”, a program I recommend.

  14. Max Griswold says:

    I didn’t find this to be a fair criticism from David and a misinterpretation of what happened behind the scenes.

    We didn’t initial produce absolute risks (nor were required by the editors), given our study wasn’t about relative risks or the meta-analysis, ultimately. The study used relative risks, specifically ,for producing the primary variables of interest (population attributable fractions, deaths, and DALYs). Absolute risks would not have made sense within the context of our study. Once David inquired about absolute risk to the press office, we provided the values. There wasn’t any breaking of policy happening at the Lancet.

    Further, his assessment of the absolute risk took the results out of context. At the level he was advocating in his blog post, there would still be 400k deaths. Absolute risk for the specific range he picked might look small but leading to 400k deaths isn’t insignificant, in my subjective assessment.

  15. Stevec says:

    A while back I looked up some papers on drinking red wine, knowing as all good people do, that drinking alcohol is bad for you and I thought this would help me cut down. Hmm.

    (Apologies I don’t have references because I can’t find these papers in my files. But if I see replies asking, I will be sure to redo my research and highlight them).

    Across maybe 5 or 6 papers I found that drinking zero alcohol was associated with the same health problems as drinking 5 units a day of red wine, with 3 creating the optimum. That is, taking this research on board, all doctors should be recommending drinking almost half a bottle of red wine a day because “it’s better than not drinking anything”. Red wine appeared to have different properties than even white wine.

    Later I read a worthy Australian government report on why we should all cut down massively on alcohol, and, being the kind of person who appreciates this blog, I looked up lots of the underlying papers. As far as I could tell, a decent part of the problem is people who drink and drive, drink and use power tools like chainsaws, drink and engage in risky behaviour like casual unprotected sex.. and so on. Boring people like me miss out on these mortality causes..

    It also didn’t appear to break out people who binge drink vs people like me (alcohol dependents) who drink a quality amount (a flagon) of red wine at a slow enough rate that we could probably drive and be under the limit (not in Scotland) – not that I would, by the way. If you drink 50 units of alcohol a week over 7 days x 8 hours, v 50 units a week over 2 nights in the weekend.. I can see a clear medical reason why these would be different.

    So you start to wonder.

    Living during WW2 was bad for your mortality. But only if you aggregate the world population and assume it affects everyone equally. Chileans were probably fine for example.

    Then you ask if drinking alcohol is correlated with something else. What randomised controlled studies exist?

    Finally, being an alcohol dependent, you say, oh, apparently I might live to 90 if I didn’t drink, but only 87 if I carry on with my crazy lifestyle. Giddy up!

  16. Joshua Pritikin says:

    “For cardiovascular disease subtypes other than myocardial infarction,” hm, the linked Lancet article should have cited Stockwell et al (2016): Do “Moderate” Drinkers Have Reduced Mortality Risk? A Systematic Review and Meta-Analysis of Alcohol Consumption and All-Cause Mortality.

    OBJECTIVE:

    Previous meta-analyses of cohort studies indicate a J-shaped relationship between alcohol consumption and allcause mortality, with reduced risk for low-volume drinkers. However, low-volume drinkers may appear healthy only because the “abstainers” with whom they are compared are biased toward ill health. The purpose of this study was to determine whether misclassifying former and occasional drinkers as abstainers and other potentially confounding study characteristics underlie observed positive health outcomes for low volume drinkers in prospective studies of all-cause mortality.

    METHOD:

    A systematic review and meta-regression analysis of studies investigating alcohol use and mortality risk after controlling for quality-related study characteristics was conducted in a population of 3,998,626 individuals, among whom 367,103 deaths were recorded.

    RESULTS:

    Without adjustment, meta-analysis of all 87 included studies replicated the classic J-shaped curve, with low-volume drinkers (1.3-24.9 g ethanol per day) having reduced mortality risk (RR = 0.86, 95% CI [0.83, 0.90]). Occasional drinkers (<1.3 g per day) had similar mortality risk (RR = 0.84, 95% CI [0.79, 0.89]), and former drinkers had elevated risk (RR = 1.22, 95% CI [1.14, 1.31]). After adjustment for abstainer biases and quality-related study characteristics, no significant reduction in mortality risk was observed for low-volume drinkers (RR = 0.97, 95% CI [0.88, 1.07]). Analyses of higher-quality bias-free studies also failed to find reduced mortality risk for low-volume alcohol drinkers. Risk estimates for occasional drinkers were similar to those for low- and medium-volume drinkers.

    CONCLUSIONS:

    Estimates of mortality risk from alcohol are significantly altered by study design and characteristics. Meta-analyses adjusting for these factors find that low-volume alcohol consumption has no net mortality benefit compared with lifetime abstention or occasional drinking. These findings have implications for public policy, the formulation of low-risk drinking guidelines, and future research on alcohol and health.

Leave a Reply