Randomised trials and cohort studies bear on one decision question: does about one drink a day, against none, raise blood pressure? They measure different quantities. Roerecke’s trial meta-analysis finds no significant BP change in its lowest band (two drinks a day or fewer; weeks to months, by weight mostly one 4-week crossover in women) and reads it as an apparent threshold. Di Federico’s cohort meta-analysis finds BP change over 4-12 years higher at 12 g/day than with none (SBP +1.25, 0.49 to 2.01), a difference its authors call «rather small» at one serving a day yet read as making «any alcohol» a risk factor for high SBP, with no threshold shown over 0 to ~55 g/day. Matched on quantity they do not clash. One is a weeks-scale level response, the other a years-scale divergence in BP trajectory by baseline intake, and both can hold at once (not-joined check ii). A numeric clash appears only if the cohort difference is read as a level effect that stopping would reverse within weeks, which neither source shows. Even then it appears only on overall DBP; in women only (the trials have no men-only <=2 stratum), the intervals touch. Filed as a distinction with a conditional clash and a type-G gap: held evidence cannot say whether the cohort gradient is causal, or whether a real low-dose effect needs years to show. Synthesis mode: configurative (the two designs share no common metric to pool). (Di Federico et al., 2023; inferred from Roerecke et al., 2017)
View A: the randomised low-dose band shows no BP effect (Roerecke 2017)
Roerecke pooled 36 alcohol-reduction trials and stratified them by baseline intake. In the band of two drinks a day or fewer, 5 trials gave SBP -0.18 (-1.02 to 0.66) and DBP +0.61 (-0.04 to 1.26) mmHg; the alcohol arm averaged 15 g/day, and at this level alcohol «was mostly compared with abstinence or near abstinence». (Roerecke et al., 2017)
«A reduction in alcohol consumption reduced blood pressure in a dose-dependent manner with an apparent threshold eff ect at two drinks per day. People drinking two drinks or fewer per day did not have a signifi cant reduction in blood pressure when they reduced their alcohol consumption to near abstinence, suggesting that this amount of alcohol intake does not increase blood pressure.»
The next sentence keeps the dose-response above the band: «However, the more people drink beyond this level, the higher the subsequent reduction in blood pressure.» The paper’s conclusion softens the threshold to «a possible threshold eff ect». (Roerecke et al., 2017)
What the five trials were. Three gave abstainers or near-abstainers wine against (near) abstinence, water or non-alcoholic beer: Cordain (10 weeks, overweight women), Gepner 2015 (104 weeks, n=224, type 2 diabetes) and Shai (12 weeks, type 2 diabetes). Droste gave carotid-plaque outpatients red wine against abstinence (20 weeks). Mori 2015 gave regular drinkers red wine against dealcoholised wine (4 weeks). Three are parallel-group, two crossover. By weight the band is mostly Mori 2015 (about 8 g/day), which carries 52% of the SBP and 69% of the DBP pooled weight; Cordain adds 39% of the SBP weight but only 10% of DBP, level with Droste (9%) and Gepner (11%). The two type 2 diabetes trials carry 3% (SBP) and 12% (DBP). (Roerecke et al., 2017)
View B: the cohort SBP curve shows no threshold over 0 to ~55 g/day (Di Federico 2023)
Di Federico pooled 7 cohorts (19,548 adults; 3 North American, 4 Japanese or Korean; median follow-up
5.3 y, range 4-12), relating baseline intake to BP change over follow-up in a 1-stage spline model.
SBP was 1.25 (0.49 to 2.01) and DBP 1.14 (0.60 to 1.68) mmHg higher at 12 g/day than with «no alcohol
consumption», rising to SBP 4.90 (3.71 to 6.08) at 48 g/day. The plotted curves run 0 to ~55 g/day
(figure read); no formal non-linearity test or certainty rating is reported
[searched: linearity / P for / GRADE / certainty across chunks 01-02 (main text; supplement read at prep but not chunked), srcgrep 0 / 0 / ref-list only / «uncertainty» only].
(Di Federico et al., 2023)
The authors first grant that the change at one serving a day was «rather small», then draw the no-threshold conclusion:
«The main finding of our meta-analysis was that usual alcohol consumption at baseline was positively associated with change in SBP over time in both men and women, with a substantially linear pattern, although the amount of change associated with a limited amount of alcohol intake (such as one serving/day) was rather small. Our findings suggest that in longitudinal analysis there is no threshold below which no association exists between alcohol consumption and a higher SBP. Consequently, consumption of any alcohol should be considered a risk factor for high SBP, with important preventive and therapeutic implications.»
In the abstract they limit the linear, no-threshold reading to SBP: «Our results suggest the association between alcohol consumption and SBP is direct and linear with no evidence of a threshold for the association, while for DBP the association is modified by sex and geographic location.» The Discussion’s closing paragraph states it for BP without the split: «Overall, our review indicates that the relationship between alcohol intake and BP is almost linear with no evidence of a threshold for the association.» (Di Federico et al., 2023)
The paper sets itself against the trial band. Its introduction, citing Roerecke, says «there is uncertainty regarding the relationship between alcohol consumption and BP at or below an intake of 2 standard alcoholic drinks per day. Whether this results from a lack of statistical power in studies of low levels of alcohol intake or reflects biological differences at lower compared with higher intakes of alcohol is unclear.» (Di Federico et al., 2023) In women (2 cohorts) and North America (3), the 12 g/day SBP intervals include 0. (Di Federico et al., 2023)
The parameter table
| Parameter | View A: Roerecke, <=2-drink band | View B: Di Federico, 12 g/day | Same quantity? |
|---|---|---|---|
| Answer | «this amount of alcohol intake does not increase blood pressure» | «any alcohol should be considered a risk factor for high SBP» | no: verbal conclusions only. Roerecke’s rests on a non-significant short-run band; Di Federico’s on a years-scale association it calls «rather small» at one serving/day. They oppose in wording, not in a shared measured quantity |
| Contrast | randomised, ~15 g/day vs abstinence or near-abstinence | observed, baseline 12 g/day vs none | partly (similar dose contrast; randomised vs observed) |
| Who changes intake | 3 of 5 trials start abstainers drinking; only Mori (n=24, 4-week crossover) has regular drinkers drop to dealcoholised wine | nobody; baseline intake only | no; habitual light drinkers cutting down appear in one 4-week trial (Mori 2015, about 8 g/day), but it carries 52% of the SBP and 69% of the DBP weight |
| Horizon | BP at the time point Roerecke used (it «preferred shorter time periods over longer time periods»); trials ran 4 to 104 weeks, but Table 2 puts only 3 of 5 at >=10 weeks | BP change over 4-12 years | no; even the longest trial (2 y) is short of the shortest cohort follow-up (4 y) |
| Population | 2 trials in type 2 diabetes, 1 in carotid plaque, 2 in women without disease | healthy adults (diagnosed CVD/diabetes excluded per stated criteria; ARIC and CABL nonetheless adjust for diabetes), US and Japan/Korea, 65% men | no |
| SBP | -0.18 (-1.02 to 0.66) | +1.25 (0.49 to 2.01) | no (level difference at the trial time point Roerecke used vs difference in change over 4-12 y); read as a reversible effect, compatible only at ~0.5-1.0 mmHg |
| DBP | +0.61 (-0.04 to 1.26) | +1.14 (0.60 to 1.68) | no (same split); read as a reversible effect, the overall intervals do not overlap (the trial excludes a DBP fall beyond ~0.04 mmHg). Women only, they touch at ~0.7 mmHg (trial 0.87, -0.70 to 2.44; cohort 1.45, 0.70 to 2.20) |
(Roerecke et al., 2017) (Di Federico et al., 2023)
Not-joined checks: a distinction, with a conditional clash
- (i) Same observable, differing only in words? No. The trial band predicts no BP change from ~15 g/day over weeks; the cohort curve predicts about 1 mmHg more BP change over years at 12 g/day.
- (ii) Different scope, consistent once matched? Yes. The two differ in horizon (4-104 weeks vs 4-12 years), estimand (BP level after a change in intake vs difference in BP change by baseline intake), population (mostly women and people with diabetes vs mostly men) and design. Both findings can hold at once: no detectable short-run level response at ~1 drink/day, and a slow divergence in BP over years. A numeric clash appears only if the cohort’s difference in change is read as a level effect that stopping reverses within weeks. Neither source shows that (Di Federico infers only «the potential for bigger reductions in BP with prolonged reduction or abstinence from alcohol», a reversal over a prolonged horizon, not within weeks), and the change-score design subtracts a steady level effect out. Women only (Roerecke has no men-only <=2 trials), even that reading leaves the DBP intervals touching. The decision does not split either: under both readings the harm-side default holds and the BP payment runs from zero to small (Decision relevance below).
- (iii) Part of a multi-position author? No, but both positions are hedged. Roerecke’s whole Discussion and conclusion were read: an «apparent», then «possible», threshold, and «Because only three trials reported results for women,28,50,53 we have less confi dence in the pooled eff ect estimates». Di Federico’s whole Introduction and Discussion were read: one serving/day «rather small», and it names horizon as its advantage over the trial meta-analyses: «The trials considered were however generally of short duration (4—8 weeks). In contrast, follow-up in the cohort studies included in our meta-analysis ranged from 4 to 12 years, a major strength compared the meta-analyses based on experimental findings.» (The 4-8 weeks follows its account of Xin 2001; Roerecke’s <=2 band ran 4-104 weeks.)
(Roerecke et al., 2017) (Di Federico et al., 2023) (Di Federico et al., 2023)
Distinction (not-joined check ii), not a joined type-D. The conditional clash and the designs that would settle it are filed as type-G below. (Di Federico et al., 2023; inferred from Roerecke et al., 2017)
Synthesis: neither the threshold nor the no-threshold is a shape the data can fix
The reflex reading is that the randomised evidence wins. It does not settle this question, because
neither side measures the shape directly. Roerecke’s threshold is a comparison between trials: bands
are built from different trials (the only trial straddling the <=2 boundary is Mori 2015, whose low- and
high-volume arms sit in the <=2 and 3-drink bands; its within-trial contrast, SBP -0.40 at ~8 g/day vs
-2.00 at ~26 g/day and DBP +0.30 (-0.48 to 1.08) vs -1.20 (-1.98 to -0.42), leans toward a threshold
on DBP point estimates, but both intervals admit a straight line), and baseline intake explains most of
the between-study variance («Alcohol intake at baseline explained 75·4% of the between-study variance
in systolic blood pressure»), while the lowest band also cut the least alcohol (-15 g/day vs -30, -40
and -32 in the higher bands). Di Federico’s no-threshold is a spline pinned at a referent whose make-up
(never vs former drinkers) is not reported in the main text, with no formal non-linearity test
[searched: former / never / linearity across Di Federico chunks 01-02, srcgrep 0 / 0 / 0].
(Roerecke et al., 2017)
(Di Federico et al., 2023; inferred from Roerecke et al., 2017)
Four readings reconcile the cohort gradient with the trial null, one with no real low-dose effect (a) and three with one (b-d), and the held evidence does not pick among them.
(a) The cohort gradient is confounded or referent-biased. Every point on the curve, 48 g/day included, is a contrast against «no alcohol consumption»; a 3-knot spline pinned at zero cannot isolate a step there, so a contaminated referent could tilt the whole curve. The authors write that «the nonexperimental design of the studies did not allow us to exclude the possibility of residual confounding, although a relevant bias due to key unmeasured confounders did not appear to be likely», and credit healthy-adult longitudinal data with «reducing the risk of reverse causation»; nothing in the paper tests unmeasured confounding (its sensitivity analyses only drop studies not adjusted for smoking or BMI).
(b) The effect needs years. Cohort follow-up was 4-12 years, trials weeks to months. The support is thin on both sides: Di Federico finds only a «slightly positive association of duration of follow-up», steeper for SBP than DBP, and Roerecke reports trial length «associated with the eff ect size» while its dose-response held «independent of the length of the trial». Within the trial band the >=10-week stratum does not trend toward a fall: DBP +1.38 (0.00 to 2.77), SBP -0.36 (-3.06 to 2.35); only the one 2-year trial’s SBP point estimate leans that way (Gepner 2015, -4.67, -10.89 to 1.55), and whether that is a 2-year reading is not shown, since Roerecke preferred shorter time periods. In the three trials with repeated measurements, one of them (Droste) in this band, Roerecke found «almost no diff erence from the fi rst measurement to the last, which suggests that the eff ect of a reduction in alcohol consumption on blood pressure is sustained when lower alcohol consumption is sustained».
(c) The trial band does not transport. By trial count it mostly starts abstainers drinking, two trials in type 2 diabetes. By weight it is mostly a 4-week crossover in 24 healthy premenopausal women who already drank, given about 8 g/day against dealcoholised wine. So the pooled number is short-horizon and mostly female (its weight-averaged contrast, about 11-13 g/day, matches the cohorts’ 12 g/day, so dose is not the gap), while the cohorts are mostly men followed 4-12 years. A light drinker cutting down for longer than 4 weeks is in neither.
(d) Power. On SBP the trial interval admits a fall of ~1 mmHg, so a small real effect could hide in 5 trials, a possibility Di Federico’s introduction raises alongside biological differences at low intakes. On the overall pooled DBP the trials exclude a fall of that size; women only, the intervals touch at ~0.7 mmHg, so power is not ruled out there.
(Di Federico et al., 2023) (Roerecke et al., 2017) (Di Federico et al., 2023; inferred from Roerecke et al., 2017)
The no-threshold trial result Di Federico calls «generally consistent» with its own, Xin 2001, adds little: by Di Federico’s account it «relied on a meta-regression linear model», which cannot show a threshold, and it shares an author (Whelton). Xin is not held. (Di Federico et al., 2023)
What would join the issue (type-G)
The two bodies would meet on one quantity only in a design that measures BP as habitual light drinkers change intake, over the cohorts’ horizon. Either of two would do:
- a randomised trial in habitual drinkers of two drinks a day or fewer who cut down, powered for ~1 mmHg SBP and run for months to years;
- a cohort analysis of within-person changes in intake against changes in BP.
Di Federico cites two trajectory cohorts of the second kind: Qiu 2022 (China Health and Nutrition Survey) — would show whether BP change tracks intake change at low doses; and Jung 2021 (Korean MRCohort, one of the 7 pooled cohorts) — the same, in non-hypertensive adults. Neither is held. If such data showed BP falling when light drinkers cut down, the cohort reading would gain a causal leg; if not, the trial null would extend past 4 weeks. (Di Federico et al., 2023)
Decision relevance
For a light drinker (about one drink a day) the expected BP gain from stopping is between zero and about 1.25 mmHg SBP (cohort interval 0.5 to 2): zero on the trial reading, ~1.25 mmHg on the cohort reading if it is causal and reverses on cutting. On this wiki’s arithmetic, carrying BPLTTC’s drug-trial hazard ratio (~0.91 per 5 mmHg of pharmacological lowering, Blood Pressure Lowering and Cardiovascular Events) over to an alcohol-associated BP difference, which neither source models, that is roughly a 2% relative reduction in major CV events (about 1-4% across the cohort interval), modelled, not measured. The harm-side default holds either way, every reduction pays, but through BP the payment is small at best. Which reading holds changes the reason to cut light drinking more than the decision: the cancer curve rises from zero with no protective arm (The U-Shaped Association Artifact), and the mortality J-curve’s protective arm is largely artifact (Alcohol and Mortality and Vascular Disease). Above two drinks a day there is no disagreement: both bodies show BP rising with intake. (Di Federico et al., 2023; inferred from Roerecke et al., 2017)