Wednesday 31 March 2010

Moderate drinking and health

My best read of the evidence is that moderate drinking reduces mortality risk. Most work finds a J-curve relationship: the relative mortality risk of drinking decreases for light to moderate drinking, becomes comparable to teetotalling somewhere around 30-40 grams of alcohol per day, then continues increasing. The curve looks a bit like the letter J, though it never gets quite as vertical as the letter.

There are, of course, lots of studies out there, of varying quality, and they don't always agree.

And, worse, there are some nasty potential confounds out there. One obvious one, which folks dismissive of the J-curve relationship tend to promote, is that many studies confound never-drinkers with ex-drinkers. If someone quits drinking because he's destroyed his liver, it would be a mistake to class him as a non-drinker to evaluate the health effects of moderate drinking. Other potential confounds include omitted variable bias: if moderate drinkers also tend to have better health behaviours than heavy drinkers or non-drinkers, then the relationship could be due to the other health behaviours rather than due to the alcohol.

The best way forward is a decent meta-study that pools results from various studies and assigns them weight according to how well the study was conducted, or takes averages across different sets of studies that control differently for various effects.

The best one of these I've seen is Castelnuovo and Donati's 2006 paper from Archives of Internal Medicine. They strongly endorse the J-curve relationship. They pool 34 studies of a total of more than a million individuals. Some of these studies include former drinkers with abstainers; others carefully separate them out. But both types of studies find a J-curve. Including former drinkers as never drinkers biases things, but not enough to overturn the result: the studies that are more careful still show a strong J-curve relationship, just with the curve cutting the x-axis around 30 grams per day rather than around 60 grams per day.

Castelnuovo and Donati note the problem of omitted variable bias, but suggest it's likely to be pretty limited. They note that the health behaviour covariates included in those studies adjusting for those covariates only have small effect on the overall curve: the unadjusted figures show a maximum risk reduction of 19% while the adjusted figures knock that back to 16%. There will of course be other health behaviours correlated both with drinking and with mortality outcomes, but they would have to have more than five times the effect of the observable health covariates to flatten the J-curve to a hockey stick. That's rather implausible. Note of course I mean a proper hockey stick held with the shaft horizontal - not one of those silly Kiwi lawn game sticks.

In the economics of discrimination literature, we can pretty readily believe that the 5% or so remaining difference between blacks and whites in wage regressions is largely due to unobserved variables that correlate both with wages and with race. If observable characteristics correlated with race and wages knock back the wage gap from 20% to 5%, then it's pretty plausible that unobservables might have a quarter the effect of the observables. But having five times the effect of the observables? Nah.

And so I was pretty surprised to find that the New Zealand Drug Foundation considers the J-curve a "myth". Why do they think it's a myth? Non-randomised trials suffer from omitted variable bias and often conflate never-drinkers with former drinkers (egads, asked and answered!). What do they cite?
  • Chikritzhs et al 2009, the upshot of which seems to be "oh, empirical work based on surveys is hard, so we really can't say anything that might encourage people to drink". For what it's worth, Chikritzhs also spends a fair bit of time investigating the evils of the alcohol industry and pushing for new and higher alcohol taxes;
  • Connor et al in the NZMJ. This is a fun one. They say alcohol's responsible for 1037 deaths but helps prevent 981 deaths, for a net loss of 56 lives in 2000. They also specifically tally more than 4000 life years gained due to reductions in ischaemic heart disease associated with alcohol consumption. Hardly a great source if you want to claim that there's no benefits of moderate alcohol consumption! I wish they'd listed which claims they wished to back up with Connor et al's piece. Connor specifically says that there are health benefits from regular moderate drinking for those middle-aged and older;
  • Mukami et al on beliefs about moderate drinking
  • Rimm and Mukami 2008, the upshot of which is that since moderate alcohol consumption increases some disease risks while decreasing others, we might want to be careful in recommending moderate consumption to folks with a very strong susceptibility to the disorders that alcohol exacerbates (we'd presumably then also want to more strongly recommend it to folks with susceptibility to disorders attenuated by moderate alcohol use);
  • Doug Sellman (et al)'s 2009 "Viewpoint" piece in the NZMJ that alcohol cardioprotection has been talked up (never mind that Corrao 2000 definitely finds a J-curve in cardioprotection for both high and low quality studies, and that Rimm and Moats 2007 are trenchant about recent efforts to downplay the existence of the J-curve)
So they're relying on some op-eds and nebulous worries about the quality of existing studies rather than the best consensus estimate of the literature.

Of course, they're not the only ones. The WHO also spends a bit of time trying to talk down the cardioprotective effects of alcohol. I've noted before the WHO's ongoing war against alcohol, so it's little surprise they play a bit loose with the literature.

WHO briefly notes Corrao et al, blustering on about confounding where non-drinkers are lumped in with former drinkers, but Corrao specifically checks for that and finds that the J-curve on coronary heart disease doesn’t go away when you have studies that split out former and never drinkers. And, they get Corrao simply wrong on another dimension: they cite it as a meta-study of 28 cohort studies: Corrao based its main findings on 28 high quality studies but also presented findings from the 51 overall studies they’d selected: the selected studies showed a smaller J-curve effect than the 51, so they weren’t picking the 28 to get a larger J-curve.

It instead looks like WHO was trying to downplay Corrao relative to Fillmore, making it seem as though the gap in number of studies covered was much larger than in actuality. Corrao is listed as a meta-analysis of 28 while Fillmore is "a recent meta-analysis of 54". Fillmore disagrees with Corrao but Fillmore’s results seem to hinge on the two studies they view as being error-free; Corrao’s results just seem more robust. And, Rimm and Moats (2007, linked above) nicely show Fillmore’s results to be outside of the norm.

While WHO doesn’t bother noting Castelnuovo’s extensive meta-study on overall mortality it cites Jackson’s op-ed in the Lancet, almost every objection in which is answered in Castelnuovo. And, the WHO doesn't even note that their footnote here points to a two-page Lancet op-ed (it's in the "Comment" section) rather than to an empirical study. The WHO paper just isn’t good science. It's motivated reasoning.

The Drug Foundation thus far stands by its Mythbusting article. That's disappointing. If the goal is healthism, the best evidence suggests that a bit less a drink a day has the greatest mortality risk reduction, that there are still health benefits (relative to teetotalling) up to about three or four drinks per day, and that folks ought to be cautious about adverse health effects beyond that. They might also note that folks with family history of cancer drink a bit less while those with family history of heart disease drink a bit more. The economist would then say to weigh the health costs of drinking more than the health-maximizing amount against the consumption benefits.

I'm a bit puzzled why the anti-alcohol folks would want to bury the evidence of a J-curve - I have a very hard time seeing how a truth-seeker could find other than that there's a J-curve on the balance of the evidence. Is it the noble lie: that folks who've heard of the J-curve would use it to rationalize far greater drinking, so it's best to pretend it doesn't exist?


  1. I received a call from Southern Cross today, having got health insurance with my new job. One of the (many) questions asked was: do I drink 3 or less glasses of alcohol per day?

    30mg of alcohol or less, and I qualify for a 10% discount on premiums.

    Fancy that, a private health insurance provider has come to the same conclusion as you ;)

  2. @Herr: Would have been better if it was no discount for no glasses, discount for 1-3, then premium for 5+ :>

  3. I wonder if there's concern that drinking one to three per day could lead (later in life) to more drinking (more than is healthy)?

    The literature might address this (e.g., by using sufficiently long time horizons?), but I have no idea.

  4. Quite possibly. But the "Noble Lie" is a terrible basis for policy or science.

  5. Yeah, I agree. If it is what I'm saying, I'd rather they say so explicitly! (As my spouse does, in the reason for preferring I not have a drink every day.)