Have Another Drink! - Nassim Nicholas Taleb
Nassim Nicholas Taleb
SubscribeSign in
Have Another Drink!<br>Government sponsored study on alcohol doesn't stand up to scrutiny.
Nassim Nicholas Taleb<br>Aug 15, 2026
148
13
Share
I have been finding flaws in studies critical of alcohol consumption in low to moderate doses. The one I will examine is the following one used as “definitive” argument in the medical discussions.<br>George S, Naimi TS, Keyes K, Martinez-Matyszczyk P, et al. (including Shield KD). “Alcohol Intake and Health Study: No Protective Effect at Low Levels, With Mortality Increasing to 1 in 25 at 14 Drinks Per Week.” Journal of Studies on Alcohol and Drugs, 2026 (Vol. 87, Issue 4; DOI: 10.15288/jsad.25-00435). This analysis was commissioned by the US goverment. It has a deceiving title and a deceiving summary conclusion: “At low levels of consumption, no protective net effect of alcohol consumption on health was observed.”. Their claim is that lifetime alcohol-attributable mortality risk were >1 in 1,000 at around 7 drinks per week and reached ~1 in 25 (4%) at 14 drinks/week (the established upper guideline limit). It offers a guidance of no more than 1 drink per day (or less).<br>Table 1 in the paper shows data that doesn’t fit the narrative (there are obvious gains from alcohol at low levels, and high standard errors) and, worse, some methodology that is obscure enough to be suspicious.
The central result.<br>Primo, they use mysterious asymmetric CIs (confidence intervals). The risk for 7 drinks for >70, here −5.47, is not halfway in [−28.59 14.57] (should be −7.01). Nothing appears to fit a standard probability distribution. The lognormal must be ruled out since there are negative risk numbers. Monte Carlo (MC)? needs an explanation. The supplemental material is highly amateurish and silent on calculation.<br>Secundo, the total risk is an exact unweighted sum of all age segments. However the CI are not added. The total interval is tighter than the sums of subparts. Naive summation gives 35.81, quadrature (independence) gives 22.42, and TABLE 1 reports 31.86, much nearer dependence. This is a MC propagating the uncertainty using orrelated draws for every age bracket.<br>This is sloppy. We have no idea about the correlation structure. Their baby equations in the text and their supplementary bloviation do not explain it.<br>Terso, one age bracket dominates the study —the one where standard error is maximal, where cause of death is almost never isolated. For males consuming 21 drinks, the >70 contributes 27.30 of 68.92 , [5.58, 48.33]. So the entire claim rest on the most stochastic segment: mortality is highest, exposure is recalled and subjected to survivor-filtering, the cause-of-death attribution is almost never unique. The underlying subjects are frail. There is no sensitivity analysis excluding it/bootstrapping, etc.<br>For a paper claiming to overturn decades of alcohol drinking recommendations, the statistical foundation is underdocumented and, to be mild, statistically strange. There is nothing definitive there.
148
13
Share
TopLatest
No posts
Ready for more?
Subscribe
© 2026 Nassim Nicholas Taleb · Privacy ∙ Terms ∙ Collection notice<br>Start your SubstackGet the app<br>Substack is the home for great culture
This site requires JavaScript to run correctly. Please turn on JavaScript or unblock scripts