Nucleus of the supplements cluster. For a generally-nourished adult without a known deficiency,
supplementing vitamins, minerals, multivitamins, or marine omega-3 to prevent cardiovascular disease,
cancer, fractures, or death is — on the best evidence — a null lever, and for one named agent (beta
carotene) a harmful one. The decisive caveat runs through everything below: these are trials and reviews of
adding a supplement to people who are already replete, not of correcting a deficiency — the two
are different exposures answering different questions.
The landmark RCT — VITAL (n=25,871): four co-primary endpoints, all null
VITAL randomized 25,871 US adults (men >=50, women >=55, selected by age only, no CVD/cancer at entry) in a 2x2 factorial design to vitamin D3 2000 IU/day and/or marine n-3 1 g/day, median follow-up 5.3 years (Manson et al., 2019). Both agents missed both of their co-primary endpoints:
| Agent | Primary endpoint | HR (95% CI) | State |
|---|---|---|---|
| Vitamin D3 | Invasive cancer | 0.96 (0.88-1.06) | no meaningful effect |
| Vitamin D3 | Major CV events (MACE) | 0.97 (0.85-1.12) | no meaningful effect |
| Marine n-3 | Major CV events (MACE) | 0.92 (0.80-1.06) | no meaningful effect |
| Marine n-3 | Invasive cancer | 1.03 (0.93-1.13) | no meaningful effect |
«Supplementation with vitamin D was not associated with a lower risk of either of the primary end points.» (Manson et al., 2019)
«Supplementation with n-3 fatty acids did not result in a lower incidence of major cardiovascular events or cancer than placebo.» (Manson, Cook, Lee, Christen, Bassuk, Mora, Gibson, Albert, et al., 2019)
Fractures were null too. The VITAL ancillary fracture trial (LeBoff 2022) found vitamin D3 did not reduce total (HR 0.98, 0.89-1.08), nonvertebral (0.97), or hip (1.01, 0.70-1.47) fractures (LeBoff et al., 2022):
«Vitamin D3 supplementation did not result in a significantly lower risk of fractures than placebo among generally healthy midlife and older adults who were not selected for vitamin D deficiency, low bone mass, or osteoporosis.» (LeBoff et al., 2022)
An independent trial set reaches the same fracture-null (Kahwati 2018, USPSTF gold SR) [E-independent].
Kahwati pooled 11 RCTs, N=51,419 of community-dwelling adults (deficient/institutionalized/osteoporotic/
prior-fracture strata excluded) and — using no VITAL trial (its search closed March 2017, before
VITAL published) — found no fracture benefit: vitamin D + calcium total fracture ARD -0.35% (-1.02 to 0.31),
hip -0.14%; vitamin D alone hip pooled ARD -0.01% (RR 1.08); calcium alone insufficient. «Vitamin D
supplementation alone or with calcium was not associated with reduced fracture incidence among
community-dwelling adults without known vitamin D deficiency, osteoporosis, or prior fracture»
(Kahwati et al., 2018). Because Kahwati and VITAL
share no trials, no group, no lineage, this is a genuine independent same-outcome witness of the
fracture enhancement-null — strictly that claim, not the point estimates
(inferred from Kahwati et al., 2018; LeBoff et al., 2022). And
Kahwati names the repletion split itself — «The fracture benefits overall appear to be largely attributable
to benefits among the high-risk populations, with little to no benefit in lower-risk populations (1 fewer
hip fracture per 1000 community-dwelling adults per year [95% CI, 0 to 2])»
(Kahwati et al., 2018) (reporting the 2014
Cochrane review, which included high-risk strata) -> Deficiency Repletion vs Enhancement,
Exercise for Preventing Falls in Older Adults (where exercise, not the pill, is the community-dwelling
fracture lever).
The OTHER fracture arm — Yao 2019 shows D + calcium reduces hip fracture in the high-risk/institutionalized stratum (directly). Where Kahwati (community-dwelling only) finds the enhancement null, Yao’s gold SR+MA (6 RCTs, 49,282) finds combined D+Ca cuts hip fracture — «a 6% reduced risk of any fracture (RR, 0.94; 95% CI, 0.89-0.99) and a 16% reduced risk of hip fracture (RR, 0.84; 95% CI, 0.72-0.97)» (Yao et al., 2019) — the benefit «more extreme … in the RCTs of older participants (ie, aged 80 years) living in an institution» (Yao et al., 2019). Vitamin D alone stays null (hip RR 1.14, 0.98-1.32), so calcium is the active co-ingredient. This is not an independent second witness of Kahwati (shared trials: WHI/Jackson, RECORD) — it is the direct evidence for the high-risk-benefit arm Kahwati reported second-hand, and the full net-effect (benefit vs CV + stone harm, by stratum) has its own home -> Vitamin D and Calcium Supplementation for Fracture Prevention.
D-Health — a second large RCT of vitamin D vs major CV events, null-crossing on a TERTIARY outcome [2026-09-02, Thompson]
The D-Health Trial (Thompson 2023) is the second large independent RCT of vitamin D and hard cardiovascular events after VITAL: 21,315 Australians aged 60-84, randomised to 60,000 IU/month D3 (monthly bolus) vs placebo for up to 5 years, with major CV events (MI, stroke, coronary revascularisation) captured by administrative-data linkage (Thompson et al., 2023). Its headline is essentially the VITAL vit-D result again — no clear benefit — but the discipline the authors themselves apply is load-bearing and must be carried, not laundered into vitamin D reduces CVD.
The effect is not statistically distinguishable from null, on a TERTIARY outcome not corrected for multiplicity. Major CV events HR 0.91 (0.81-1.01) — the CI crosses 1:
«Compared with the placebo group, the rate of major cardiovascular events was lower in the vitamin D group (hazard ratio 0.91, 95% confidence interval 0.81 to 1.01), although the upper bound of the confidence interval is consistent with there being no effect» (Thompson et al., 2023)
CV events were a prespecified tertiary outcome — one of 45 (Thompson et al., 2023), and «Although this is one of several outcomes analysed, we have not adjusted for multiple testing» (Thompson et al., 2023); the trial’s primary outcome (all-cause mortality) and CV mortality were both null in the prior D-Health analysis. So a 0.91 among 45 tertiary outcomes, with the CI touching 1.01, is weak evidence — and the absolute layer says the same: -5.8 events/1000 at 5 years (95% CI -12.2 to 0.5/1000) — the absolute-difference CI also crosses zero (Thompson et al., 2023). The reported NNT 172 is derived from a point estimate that is not significantly different from null, so it is not a reliable number; carry it only with that caveat (inferred from Thompson et al., 2023).
The MI and CV-drug-user signals are hypothesis-generating, not findings — the same discipline the VITAL n-3 total-MI secondary gets below. MI 0.81 (0.67-0.98) and coronary revascularisation 0.89 (0.78-1.01) were the nominally-favourable secondaries; stroke was flat 0.99 (0.80-1.23) (Thompson et al., 2023). Among 45 tertiary plus multiple secondary outcomes with no multiplicity control, the single nominally-significant MI arm is a candidate, not a demonstrated effect. The much-cited subgroup — a larger effect among those taking cardiovascular drugs at baseline, HR 0.84 (0.74-0.97) — carries P-for-interaction 0.12 (NOT significant), and the authors say so: «The interactions were not significant at P<0.05, and it is plausible that these are chance findings» (Thompson et al., 2023). This is an exploratory, underpowered effect-modification signal (route (b) territory without a significant interaction), so it is a HYPOTHESIS for a future trial, not an established stratum modifier — and baseline 25(OH)D was modelled, not measured, a further limit on the predicted-status subgroups (Thompson et al., 2023).
What D-Health adds beyond VITAL — a different regimen reaching the same null (type-F replication, NOT
[E-independent]). VITAL and D-Health are different cohorts, countries, authors, and dosing, so the
DATA are independent — but they are the same design (RCT) answering the same question, which is
replication that raises confidence, not the different-method backing type-E requires (an RCT and a
Mendelian-randomisation study reaching one claim). Two RCTs converging on a null is type-F. The
parameter table shows what converges and what differs:
| Parameter | VITAL vit-D arm (Manson 2019, held above) | D-Health (Thompson 2023) | Same quantity? |
|---|---|---|---|
| Exposure | D3 2000 IU/day (daily) (Manson, Cook, Lee, Christen, Bassuk, Mora, Gibson, Gordon, et al., 2019) | D3 60,000 IU/month (monthly bolus) (Thompson et al., 2023) | NO — daily vs monthly bolus |
| Population | US, age >=50/55, prior CVD excluded, replete (mean 30.8 ng/mL) (Manson, Cook, Lee, Christen, Bassuk, Mora, Gibson, Gordon, et al., 2019) | Australia, 60-84, prior CVD NOT excluded, largely replete (placebo mean 77 nmol/L ~31 ng/mL) (Thompson et al., 2023) | partial — both older & replete; D-Health includes prior CVD |
| Composite | MACE = MI + stroke + CV death (Manson, Cook, Lee, Christen, Bassuk, Mora, Gibson, Gordon, et al., 2019) | MI + stroke + coronary revascularisation (Thompson et al., 2023) | NO — revasc vs CV death |
| Prespecification | co-primary | tertiary (1 of 45), no multiplicity adj (Thompson et al., 2023) | NO |
| Result | HR 0.97 (0.85-1.12), null (Manson, Cook, Lee, Christen, Bassuk, Mora, Gibson, Gordon, et al., 2019) | HR 0.91 (0.81-1.01), null-crossing (Thompson et al., 2023) | both null-crossing |
The fourth column is mostly NO, so this is not a joined tension and not an identical-quantity re-derivation — it is a robustness reading: despite a different dose schedule (monthly bolus vs daily), a different composite, and a stratum that includes the prior-CVD patients VITAL excluded, D-Health lands in the same null-crossing region (Manson, Cook, Lee, Christen, Bassuk, Mora, Gibson, Gordon, et al., 2019; inferred from Thompson et al., 2023). Two consequences for the page’s core:
- It weakens the wrong dose / wrong regimen rescue of the VITAL null. A common defence of the vit-D-CV null is that VITAL’s daily 2000 IU was the wrong regimen. D-Health used monthly 60,000 IU — «unlike D-Health and ViDA, VITAL used a daily dosing regimen of 2000 IU/day» (Thompson et al., 2023) — with reported adherence higher than VITAL’s, and still showed no clear CV benefit. So the CV null holds under both daily (VITAL) and monthly-bolus (D-Health) regimens for the replete older stratum (inferred from Thompson et al., 2023).
- It is still enhancement in the replete. D-Health’s cohort was largely replete — the placebo 25(OH)D «was 77 nmol/L, and the proportion with 25(OH)D <50 nmol/L was 13%», and «the findings cannot be generalised to populations with a greater prevalence of vitamin D deficiency» (Thompson et al., 2023) — 77 nmol/L is ~31 ng/mL, essentially VITAL’s baseline. The same load-bearing caveat holds: this tests adding D to an already-adequate person, not deficiency correction -> Deficiency Repletion vs Enhancement.
The authors resist declaring the null dead — read symmetrically. Thompson concludes «Vitamin D
supplementation might reduce the risk of major cardiovascular events, although the absolute risk
difference was small and the confidence interval was consistent with a null finding»
(Thompson et al., 2023) and, on the strength of
the non-significant CV-drug subgroup, «conclusions that vitamin D supplementation does not alter risk of
cardiovascular disease are premature»
(Thompson et al., 2023). Under symmetric
standards this is a null-crossing tertiary result whose most favourable reading rests on an unadjusted
subgroup with a non-significant interaction — it firms the page’s no clear CV benefit from
supplementing the replete position rather than reopening it, while leaving the CV-drug-user interaction
as an explicit candidate lever for a future adequately-powered trial
(inferred from Thompson et al., 2023). The
source itself triangulates against the Mendelian-randomisation evidence it cites — «one reported an
inverse association between genetically predicted 25(OH)D concentration up to 50 nmol/L and
cardiovascular disease» while the other two found none
(Thompson et al., 2023) — i.e. the genetic
evidence is mixed (a below-50-nmol/L deficiency-range signal, null above), consistent with the
repletion-vs-enhancement split this page holds, not a clean RCT+MR convergence on benefit. Held as
context, not stamped [E-independent].
The non-CV mortality anchor: pooled all-cause null, cancer-DEATH benefit [2026-08-23, Zhang MA]
VITAL is one trial; Zhang 2019 pools 50 RCTs of vitamin D ALONE (no calcium co-intervention), 74,655 participants, 7993 deaths — the fabric’s pooled mortality anchor, GRADE high, every secondary outcome I2=0%. It separates the endpoints the single-trial nulls could not:
| Outcome | Zhang pooled RR (95% CI) | State |
|---|---|---|
| All-cause mortality (50 trials) | 0.98 (0.95-1.02) | no meaningful effect |
| Cancer mortality (5 trials) | 0.85 (0.74-0.97) | benefit — 15% lower |
| Cardiovascular mortality (12 trials) | 0.98 (0.88-1.08) | no meaningful effect |
| Non-cancer, non-CV mortality (4 trials) | 1.05 (0.93-1.18) | no meaningful effect |
(Zhang et al., 2019). «Vitamin D supplementation alone was not associated with all cause mortality in adults compared with placebo or no treatment» and «Vitamin D supplementation reduced the risk of cancer death by 15%» (Zhang et al., 2019). The all-cause null is not underpower — trial-sequential analysis met the information size for the 10% and 7.5% relative-risk- reduction thresholds (a real null, not an empty one) (Zhang et al., 2019).
Cancer DEATH is not cancer INCIDENCE — a distinction, not a contradiction with the VITAL null above. This page already holds VITAL’s cancer incidence null; Zhang reports a cancer mortality benefit. They are different quantities, so no tension is joined:
| Parameter | VITAL (held above) | Zhang MA | Same quantity? |
|---|---|---|---|
| Endpoint | invasive cancer incidence | cancer death (mortality) | NO — diagnosis vs death |
| Vitamin D3 effect | 0.96 (0.88-1.06) (Manson, Cook, Lee, Christen, Bassuk, Mora, Gibson, Gordon, et al., 2019) | 0.85 (0.74-0.97) (Zhang et al., 2019) | — |
| Design | single RCT | pooled 5 RCTs incl. VITAL | not independent |
Because Zhang pools VITAL among its constituent trials, the two are not independent witnesses:
this is type-F refinement (Zhang bounds the incidence null with a downstream mortality signal),
NOT [E-independent]. The signal is latency-gated — the cancer-death benefit concentrated in trials
with >3-year follow-up, and VITAL’s own cancer mortality moved from 0.83 (0.67-1.02) overall to
0.75 (0.59-0.96) after excluding the first 1-2 years
(Zhang et al., 2019). Mechanism-plausible (cancer
latency), but held with the authors’ own guard that the subgroup «requires additional evidence».
D3 and D2 are different exposures. All-cause mortality was «significantly lower among trials with vitamin D3 supplementation than in trials with vitamin D2 supplementation (P for interaction=0.04 … although neither group was associated with all cause mortality)» (Zhang et al., 2019) — D3 40 trials RR 0.95 (0.91-1.00) vs D2 10 trials 1.03 (0.98-1.09). And «benefit was only seen in participants receiving vitamin D3 supplementation, and no participants received vitamin D2 supplementation» (Zhang et al., 2019) — so the cancer-mortality signal is a D3-only claim. D3 raises 25(OH)D more efficiently than D2, so the compounds are not interchangeable. (inferred from Zhang et al., 2019) the subgroups are observational, not randomised comparisons — a specification hint, not a settled effect modifier.
The earlier small all-cause benefit rode partly on calcium. Prior reviews found a marginal all-cause reduction (2014 Cochrane RR 0.97, 0.94-0.99; Bolland 2014 RR 0.96, 0.93-1.00); Zhang got a null by excluding vitamin-D-plus-calcium trials: «we excluded more than 10 trials totalling approximately 50 000 participants of vitamin D administered with calcium» (Zhang et al., 2019). Isolating native vitamin D from its frequent calcium co-agent removes a confound the earlier pooled figure carried.
The load-bearing move: this is ENHANCEMENT in the replete, NOT deficiency correction
VITAL’s cohort was not deficient — mean baseline 25(OH)D was 30.8 ng/mL (above the 20 ng/mL adequacy threshold), with only 12.7% below 20 ng/mL (Manson, Cook, Lee, Christen, Bassuk, Mora, Gibson, Gordon, et al., 2019) and just 2.4% frankly deficient at <12 ng/mL (LeBoff et al., 2022). So every null reads as adding D to an already-adequate person buys nothing — the authors say so directly:
«vitamin D requirements for cardiovascular health may have already been met for most participants.» (Manson, Cook, Lee, Christen, Bassuk, Mora, Gibson, Gordon, et al., 2019)
And the fracture trial states its own scope limit — participants «were not recruited on the basis of vitamin D deficiency, low bone mass, or osteoporosis» (LeBoff et al., 2022). Two things follow, and they must be kept apart:
- Within the range VITAL could study, effects did NOT vary by baseline 25(OH)D — the null is not rescued by it only helps the low-D people for the strata present. This weakens the helps-only-the-deficient hypothesis inside the studied range.
- But the frankly-deficient tail (<12 ng/mL, 2.4%; [@leboff2022]) was underpowered, and the authors note a trial in people «well below the 20 ng per milliliter» might show stronger effects (Manson, Cook, Lee, Christen, Bassuk, Mora, Gibson, Gordon, et al., 2019) — so deficiency correction remains a live, separate question this trial does not answer. Repletion is route (a)/(c) on Baseline Risk and the Relative-Absolute Split; enhancement-in-the-replete is what tested null here. The distinction has its own home: -> Deficiency Repletion vs Enhancement (the two-arm concept this page instantiates).
The guideline map — USPSTF 2022 (general adults, RCT-pooled hard endpoints)
USPSTF graded supplementation for CVD/cancer prevention in community-dwelling adults without known deficiency (deficiency, pregnancy, and chronic illness are explicitly carved out) (US Preventive Services Task Force, 2022). Its efficacy estimates are pooled from the RCT literature (its harms estimates also draw on cohorts), so — unlike the observational supplement literature — the healthy-user confound is largely designed out; that is a strength of this evidence base :
| Supplement | USPSTF grade | Basis |
|---|---|---|
| Beta carotene | D — recommend against | HARM outweighs benefit |
| Vitamin E | D — recommend against | no net benefit |
| Multivitamins | I — insufficient | evidence lacking; balance undeterminable |
| Single/paired nutrients (other) | I — insufficient | evidence lacking |
(US Preventive Services Task Force, 2022)
«there is no net benefit of supplementation with vitamin E … the evidence is insufficient to determine the balance of benefits and harms of supplementation with multivitamins … Evidence is lacking and the balance of benefits and harms cannot be determined.» (US Preventive Services Task Force, 2022)
Keep the two nulls apart: vitamin E is no meaningful effect (a graded null on hard outcomes); multivitamins are insufficient evidence (a cancer-incidence signal, OR 0.93, discordant with cancer mortality, on short follow-up) — not the same state. Neither is benefit.
Not an independent second witness of VITAL. USPSTF’s vitamin-D pooled analysis includes VITAL as
one of its largest trials, so the guideline and the trial are not independent backing (no
[E-independent]) — the guideline is the RCT base synthesized, one level up
.
The newest guideline consolidates the trial base — Endocrine Society (Demay 2024) [2026-09-02, Demay]
The Endocrine Society’s 2024 GRADE guideline is the newest authoritative stance on this page’s exact
question, and it lands where the trial base above does: against empiric vitamin D above the DRI
(600 IU/day, 15 mcg) to prevent disease in healthy adults younger than 75, and against routine
25(OH)D testing in every population it considered — «The panel suggests against empiric vitamin D
supplementation above the current DRI to lower the risk of disease in healthy adults younger than 75
years» (Demay et al., 2024) (with the
panel’s own caveat that trial doses varied and many participants continued personal supplements —
a precision limit, not a reversal). It is not an independent witness: it rests on the same
VITAL / D2d / etc. RCTs USPSTF and the mortality MAs above already pool, so this is shared-evidence
agreement — type-F consolidation, NOT [E-independent] (inferred from Demay et al., 2024). It
defeats the guidance null by agreement-with-better-warrant: a second guidance family reaches the
same null and publishes calibrated GRADE certainty and the strata where it stops.
The 50-74 consolidation is the genuine beyond-VITAL move. Where this page held VITAL’s single-trial nulls, Demay’s commissioned meta-analyses upgrade the 50-74 stratum to high certainty of little-or-no effect across four hard endpoints at once — «high certainty of evidence for fractures, CVD events, cancer and mortality, the panel judged that vitamin D supplementation appears to have little or no beneficial impact on the outcomes analyzed in healthy populations aged 50 to 74 years» (Demay et al., 2024). That is a certainty grade the isolated VITAL nulls could not carry on their own (inferred from Demay et al., 2024).
The age-75 carve-out — a route-(a) mortality signal, held with its own hedges. For adults >=75, the panel suggests empiric vitamin D (weighted-avg ~900 IU/day, 23 mcg) for a potential mortality benefit — conditional (GRADE 2), moderate certainty (⊕⊕⊕◯) — on a pooled all-cause mortality MA (25 trials, 49,879 participants): «Meta-analysis suggested that vitamin D lowers mortality compared to placebo (RR 0.96 [95% CI, 0.93-1.00]), with an estimated absolute effect size of 6 fewer deaths per 1000 people» (Demay et al., 2024). Three bounds keep it honest: the CI upper bound touches 1.00 (a borderline signal); Demay’s pool includes calcium-combined trials that the Zhang MA above excluded to reach its 0.98 all-cause null, so the two figures rest on different trial sets — a distinction, not a re-derivation and not a joined tension (inferred from Demay et al., 2024; Zhang et al., 2019); and the panel flags the age bands as «somewhat arbitrary» (Demay et al., 2024). Decision-change: for the oldest stratum a low daily dose is a defensible small lever; for 50-74 it is not.
Prediabetes — the other positive carve-out (route-(a), now grounded in the primary IPD-MA). For high-risk prediabetes, Demay suggests empiric vitamin D (weighted-avg ~3500 IU/day, 88 mcg) to slow progression to diabetes — conditional, moderate certainty. The effect traces not to the guideline but to the underlying individual-participant-data meta-analysis, now held: Pittas pooled IPD from 3 low-risk-of-bias RCTs specifically designed for diabetes prevention (D2d, Tromsø, DPVD; n=4190 adults with prediabetes, mean age 61, mean BMI 30, mean baseline 25(OH)D 63 nmol/L) and found «Vitamin D reduced risk for diabetes by 15% (hazard ratio, 0.85 [95% CI, 0.75 to 0.96]) in adjusted analyses, with a 3-year absolute risk reduction of 3.3% (CI, 0.6% to 6.0%). The effect of vitamin D did not differ in prespecified subgroups.» (Pittas et al., 2023) (unadjusted ITT HR 0.88, 0.77-0.99; NNT 30).
Why the IPD-MA sees an effect the constituent trials and VITAL could not (type-F over the aggregate MAs). Each individual trial was an underpowered null: «The individual trials included in this meta-analysis, which were powered to detect diabetes risk reductions no lower than 25%, reported nearly identical, non–statistically significant risk reductions of 10% (Tromsø), 12% (D2d), and 13% (DPVD)» — pooling harmonized IPD recovered the power to detect the small true effect (Pittas et al., 2023). This is the baseline-risk reconciliation (route-(a)) of the page’s own general-population null: VITAL selected on age alone (replete, average-risk) and found nothing; Pittas selected prediabetes (higher baseline diabetes risk, lower median 25(OH)D), where a small relative effect on a high baseline risk becomes a non-trivial absolute one — «the results do not apply to the general healthy population and should not be extrapolated to those who are at average risk» (Pittas et al., 2023). So Pittas does not contradict the null-in-the-replete thesis; it is the same thesis read at a higher-risk, less-replete stratum (inferred from Pittas et al., 2023). -> Baseline Risk and the Relative-Absolute Split, Deficiency Repletion vs Enhancement.
The lever is small at the margin (Layer-1 sizing, from Pittas’ own comparison). The relative effect «is small (15%) compared with other diabetes prevention strat- egies (58% with intensive lifestyle modification and 31% with metformin in the Diabetes Prevention Program study)» and the absolute gap is wide: «a number of persons with prediabetes needed to treat of 30 (compared with 7 with intensive lifestyle modification and 14 with metformin in the Diabetes Prevention Program study)» (Pittas et al., 2023). So the prevention rock for this stratum is lifestyle first, then metformin; vitamin D is a small, cheap, low-harm add-on, not a substitute -> Lifestyle vs Metformin for Diabetes Prevention, Layer 1 - Ranking Interventions for a Stratum.
Route-(b) effect modification — cholecalciferol works in the leaner, and a named mechanism plus an internal negative control make it credible. Prespecified subgroups were null except a BMI interaction in the two cholecalciferol trials: it «reduced risk for diabetes in participants with a baseline BMI below the median of 31.3 kg/m2 but not in those with a BMI at or above the median (hazard ratios, 0.74 [CI, 0.60 to 0.90] and 1.01 [CI, 0.84 to 1.22], respectively; P for interaction= 0.023). In contrast, in the trial that used eldecalcitol, an active analogue of vitamin D that does not require hydroxylation by CYP2R1, there was no effect modifica- tion by baseline BMI (P for interaction= 0.82).» (Pittas et al., 2023) The mechanism is pre-stated: «obesity represses vitamin D bioactivation by CYP2R1 … leading to reduced production of 25-hydroxyvitamin D, and that weight loss upregulates CYP2R1 expression» (Pittas et al., 2023) — so the pro-drug cholecalciferol fails in the obese who cannot activate it, while the active analogue that bypasses CYP2R1 shows no such gradient. A positive interaction carrying a named mechanism AND an internal negative control is the rare route-(b) evidence that clears the effect-modification bar -> Baseline Risk and the Relative-Absolute Split.
The 76% figure is NOT the randomized effect — it is a within-arm achieved-level analysis. Among cholecalciferol-assigned participants, those maintaining ≥125 nmol/L (≥50 ng/mL) 25(OH)D had «the greatest risk reduction (76%) occurring at intratrial se- rum 25-hydroxyvitamin D levels of 125 nmol/L or higher» (HR 0.24 [CI, 0.16 to 0.36]) versus those at 50-74 nmol/L (Pittas et al., 2023). But achieved 25(OH)D is post-randomization — it tracks adherence, absorption, and adiposity — so this contrast recreates the observational confounding structure inside the RCT and is hypothesis-generating, not the causal randomized effect (which stays the 15%). The authors offer one partial internal control (no such gradient in the placebo arm) and call explicitly for a treat-to-target RCT to test it. Read the 76% as a reason to run that trial, not as an attainable effect size -> Surrogate Outcomes, The U-Shaped Association Artifact. (inferred from Pittas et al., 2023)
A second benefit direction, and the safety null. Vitamin D also raised regression to normoglycemia — «Vitamin D increased the likelihood of regression to normal glucose regulation by 30% (rate ratio, 1.30 [CI, 1.16 to 1.46])» — a patient-relevant direction beyond delaying incidence (Pittas et al., 2023). Safety was null across the prespecified events: «There was no evidence of difference in the rate ratios for adverse events (kidney stones: 1.17 [CI, 0.69 to 1.99]; hypercalcemia: 2.34 [CI, 0.83 to 6.66]; hypercalciuria: 1.65 [CI, 0.83 to 3.28]; death: 0.85 [CI, 0.31 to 2.36]).» (Pittas et al., 2023) The wide hypercalcemia interval and the trials’ exclusion of people at risk of adverse events make this no signal, not proven safe — contrast the combination-with-calcium stone harm above (inferred from Pittas et al., 2023). Pregnancy and children 1-18 carry weaker suggest-empiric recs (low / very-low certainty).
Every recommendation is conditional — the GRADE-strength finding. All 14 Demay recommendations are graded “2” (conditional/weak); certainty spans very-low (⊕◯◯◯) to moderate (⊕⊕⊕◯), none strong and none high-certainty for a recommendation (Demay et al., 2024). So the newest guideline’s own instrument reports what this page does: the evidence for supplementing the healthy is weak, and where it does act it acts conditionally (inferred from Demay et al., 2024).
Two dosing/target refinements. (1) Where Demay suggests empiric D, it prefers daily, lower-dose over nondaily, higher-dose — «For empiric supplementation, daily, lower-dose vitamin D is preferred over nondaily, higher doses» (Demay et al., 2024) — the same daily-not-bolus direction the Martineau ARI section below holds (bolus was null even in the deficient). (2) The Society no longer endorses the 30 ng/mL 25(OH)D target it set in 2011 — «the Endocrine Society no longer endorses the target 25(OH)D level of 30 ng/mL (75 nmol/L) suggested in the previous guideline» (Demay et al., 2024) — and no longer defines deficiency/insufficiency cutoffs, consistent with this page’s treatment of a 25(OH)D target as a surrogate whose transmission to hard outcomes is unproven -> Surrogate Outcomes (inferred from Demay et al., 2024).
The harm arm — more is better is false (the upper-bound)
Beta carotene is the paradigm case that a supplement can harm: in smokers/asbestos-exposed it raised lung cancer, and across trials it raised CV-disease mortality (US Preventive Services Task Force, 2022):
«Two trials conducted in persons who smoke or were exposed to asbestos in the workplace found a significantly increased risk of lung cancer with beta carotene supplementation (risk ratio [RR], 1.18 [95% CI, 1.03-1.36]) … a statistically significant increased risk for cardiovascular disease mortality associated with beta carotene supplementation … (Peto OR, 1.10 [95% CI, 1.02-1.19]).» (US Preventive Services Task Force, 2022)
A single dose above repletion moved a hard outcome the wrong way — consistent with an upper-bound harm region, the more is better fallacy realized on cancer and CV mortality -> The U-Shaped Association Artifact. USPSTF also catalogues dose-dependent harms for vitamin E (haemorrhagic stroke), vitamin A (reduced BMD, hip fracture), vitamin D (hypercalcaemia / kidney stones at >=1000 IU/day), and vitamin C (kidney stones) (US Preventive Services Task Force, 2022).
The kidney-stone harm is quantified and RCT-pooled (Kahwati 2018). Vitamin D combined with calcium significantly raised incident kidney stones — pooled ARD 0.33% (95% CI 0.06 to 0.60), RR 1.18 (1.04-1.35), 3 RCTs, n=39,213, and this was «a statistically significant association for increase in incidence» (Kahwati et al., 2018) — SoE moderate, the only above-low grade in that review. The harm is specific to the combination: calcium alone showed no association (pooled ARD 0.00%, -0.88 to 0.87) and vitamin D alone was unstudied on stones (Kahwati et al., 2018). So a null-benefit supplement here also carries a small, real, moderate-certainty harm — a route-c contraindication signal for stone-formers, not merely an opportunity cost.
A second harm channel for supplemental calcium — cardiovascular (Bolland 2011, high, CONTESTED).
Beyond kidney stones, calcium supplements with or without vitamin D carry a disputed cardiovascular harm
signal: «Calcium or calcium and vitamin D supplements increased the risk of myocardial infarction (relative
risk 1.24, P=0.004) and the composite of myocardial infarction/stroke (1.15, P=0.009)»
(Bolland et al., 2011). Two guards bind it. (1) It is
contested and subgroup-derived (a WHI reanalysis restricted to women not on personal calcium; WHI
investigators and later reviews dispute it) — held as a counterweight, direction credible, magnitude/sign
disputed. (2) The exposure is the supplement bolus, not food calcium: the effect is «independent of
dietary calcium intake» with the proposed mechanism «the abrupt change in plasma calcium concentration
after supplement ingestion … rather than … the total calcium load ingested»
(Bolland et al., 2011), so it does not transport to
dietary/dairy calcium -> Is the Food Category Doing Any Work. This CV harm plus the kidney-stone harm are
the net-effect counterweights on the D+Ca fracture benefit, worked by stratum on
Vitamin D and Calcium Supplementation for Fracture Prevention.
A second pooled MA firms the CV-harm direction — Myung 2021 (high, type-F). A later meta-analysis of
13 double-blind placebo-controlled RCTs (28,935 participants) took the CV endpoints as pre-specified
primary outcomes and reproduced Bolland’s harm almost exactly: CVD composite RR 1.15 (1.06-1.25), CHD RR
1.16 (1.05-1.28), MI RR 1.25 (1.07-1.45), each I2=0% — calcium «significantly increased the risks of CVD»
(Myung et al., 2021) — with the signal concentrated in the
low-risk-of-bias trials (the opposite of an artifact pattern) and stroke null (1.13, 0.97-1.31). It is
not an independent witness: Myung re-pools Bolland’s trials and used Bolland’s unpublished trial data,
so it firms the direction (type-F) without an [E-independent] lift, and the composite significance still
leans on the WHI reanalysis (leave-one-out) (Bolland et al., 2011; inferred from Myung et al., 2021). The full net-effect
appraisal by stratum lives on Vitamin D and Calcium Supplementation for Fracture Prevention.
The primary trial behind the pooled figure — CARET, stopped early for harm. The pooled RR rests on actual trials, chief among them CARET, which randomized 18,314 smokers and asbestos-exposed workers to beta carotene (30 mg/day) plus retinol vs placebo and was halted 21 months early (Omenn et al., 1996):
«The active-treatment group had a relative risk of lung cancer of 1.28 (95 percent confidence interval, 1.04 to 1.57 … the relative risk of death from any cause was 1.17 (95 percent confidence interval, 1.03 to 1.33); of death from lung cancer, 1.46 (95 percent confidence interval, 1.07 to 2.00) … the randomized trial was stopped 21 months earlier than planned» (Omenn et al., 1996)
CARET is not an independent second witness of the USPSTF figure — it is one of the trials inside
that pool (no [E-independent]); what it adds is the stratum-specific magnitude (a lung-cancer death RR
1.46 in this smoker/asbestos population) and the fact the harm was strong enough to stop a trial, both of which
a pooled point estimate flattens.
The harm is broader than beta carotene and broader than cancer — the antioxidant-CLASS mortality signal. A meta-analysis of 68 antioxidant RCTs (232,606 participants) found that, in the methodologically-sound (low-risk-of-bias) trials, supplemental antioxidants raised all-cause mortality (Bjelakovic et al., 2007):
«In 47 low-bias trials with 180 938 participants, the antioxidant supplements significantly increased mortality (RR, 1.05; 95% CI, 1.02-1.08). In low-bias risk trials, after exclusion of selenium trials, beta carotene (RR, 1.07; 95% CI, 1.02-1.11), vitamin A (RR, 1.16; 95% CI, 1.10-1.24), and vitamin E (RR, 1.04; 95% CI, 1.01-1.07), singly or combined, significantly increased mortality.» (Bjelakovic et al., 2007)
So the harm arm is broader than beta-carotene-in-smokers: in the low-risk-of-bias trials, supplemental beta carotene, vitamin A, and vitamin E each raise all-cause mortality, while «Vitamin C and selenium had no significant effect on mortality» (Bjelakovic et al., 2007) — the four evidence states again, one per compound. Two caveats bind it: the all-trials pool is null (RR 1.02, 95% CI 0.98-1.06 (Bjelakovic et al., 2007)) — the harm surfaces only after restricting to low-bias trials; and vitamin A’s 1.16 (the largest number) rests on this single meta-analysis’s low-bias subset alone — USPSTF graded only beta carotene and vitamin E, so it is the least-corroborated claim here.
The two beta-carotene figures do not independently reinforce each other: USPSTF’s beta-carotene pool and Bjelakovic’s beta-carotene arm share CARET/ATBC as constituent trials, so the genuinely new content is vitamin A, vitamin E, and the all-cause endpoint — not a second, independent beta-carotene signal.
Vitamin E — the SAME endpoint, reconciled by bias restriction (not a tension). USPSTF also examined vitamin E on all-cause mortality and found it null: OR 1.02 (95% CI 0.97-1.07; 9 RCTs; n = 107,772) (US Preventive Services Task Force, 2022) — the same outcome on which Bjelakovic’s low-bias subset shows a small increase (1.04). So this is one endpoint, reconciled the same way as the rest: the broad pool is null, the low-bias restriction surfaces a small signal.
Two method lessons the pooled estimate carries: (1) risk-of-bias restriction surfaces the signal, it does not manufacture it — the all-trials pool is null (1.02), the high-bias trials alone show apparent benefit (RR 0.91, 95% CI 0.83-1.00 (Bjelakovic et al., 2007)), and the low-bias trials show harm (1.05), so pooling everything lets the spuriously-favourable high-bias trials cancel the harm — risk-of-bias appraisal was decision-determining -> Rating Certainty of Evidence; (2) the finding is about synthetic isolates, not food — the authors state it «should not be translated to potential effects of fruits and vegetables» (Bjelakovic et al., 2007) -> Is the Food Category Doing Any Work.
A cancer authority formalises the harm arm — and adds a calcium wrinkle [2026-08-05, WCRF]
WCRF’s Third Expert Report issues a formal recommendation — «Do not use supplements for cancer
prevention» — on exactly the harm evidence above: «Taking high-dose beta-carotene supplements is a
cause of lung cancer in current and former smokers.»
(World Cancer Research Fund & American Institute for Cancer Research, 2018) Not an
independent witness — WCRF’s judgement rests on the same CARET/ATBC trials the USPSTF pool does, so it
is a same-evidence echo, not [E-independent]; what it adds is a cancer-prevention body converting the
RCT harm into an action. (inferred from World Cancer Research Fund & American Institute for Cancer Research, 2018)
One genuinely new item — a supplement that helps at one site and may harm at others. WCRF grades calcium supplements (>200 mg/day) a probable protector against colorectal cancer, but «some trials for other cancer sites have shown potential for unexpected adverse effects». (World Cancer Research Fund & American Institute for Cancer Research, 2018) This is the net-effect-not-the-intended-effect rule in one supplement: a real site-specific benefit does not license the supplement, because the whole-body ledger (other sites, plus the general-population nulls above) is what the recommendation weighs. WCRF’s overall stance matches this page’s core: «for most people … consumption of the right food and drink is more likely to protect against cancer than consumption of dietary supplements.» (World Cancer Research Fund & American Institute for Cancer Research, 2018)
The one notable non-primary signal — n-3 and myocardial infarction
VITAL’s marine n-3 arm carried a secondary signal on total MI even as its MACE composite stayed null (stroke and CV-death were flat) (Manson, Cook, Lee, Christen, Bassuk, Mora, Gibson, Albert, et al., 2019):
«for total myocardial infarction, 0.72 (95% CI, 0.59 to 0.90)» (Manson, Cook, Lee, Christen, Bassuk, Mora, Gibson, Albert, et al., 2019)
Read it as hypothesis-generating, not a finding: it is a secondary endpoint with no multiplicity control, the composite it sits in was null, and a prespecified low-fish-intake subgroup (MACE HR 0.81) suggests any benefit may be repletion of an inadequate baseline intake rather than a pharmacologic effect of the pill. A separate exposure — high-dose (4 g/day) prescription EPA in high-triglyceride statin-treated patients — is a different question with its own trial evidence (REDUCE-IT), resolved in the section below; it does not transport to a general-population 1 g/day capsule. VITAL also notes the n-3 dose ceiling: «The dose-response curve for most effects plateaus at 1 g or less of n-3 fatty acids per day» (Manson, Cook, Lee, Christen, Bassuk, Mora, Gibson, Albert, et al., 2019) — one of the few explicitly located plateaus in the corpus’s dose-response holdings.
Marine n-3 and cognition — also null (MAPT) [2026-08-07]
Adding cognition to the outcome menu does not rescue marine omega-3. MAPT — «the longest and largest randomised controlled trial so far testing the efficacy of omega 3 polyunsaturated fatty acids on cognitive decline in elderly adults» (800 mg DHA + 225 mg EPA/day, 3 yr, n≈760 in the supplement arms) — found the omega-3 arm flat vs placebo on a cognitive-composite (between-group 0·011, 95% CI –0·081 to 0·103; adjusted p=0·812) and additive-free inside a lifestyle bundle. (Andrieu et al., 2017) A marker-vs-intervention gap worth naming: low baseline red-blood-cell DHA+EPA marked more 3-year decline, but supplementing omega-3 did not prevent it (inferred from Andrieu et al., 2017) — low status is a marker of decline, not shown to be a lever, exactly as the low-fish-subgroup reading of VITAL’s MI signal above suggests repletion rather than a pill effect -> Multidomain Lifestyle Intervention and Cognitive Decline.
COSMOS — the cocoa-flavanol supplement, VITAL’s sister trial, primary CVD outcome null [2026-08-31, Sesso]
COSMOS (Sesso 2022) is a cocoa-extract supplement RCT built in the VITAL mold — same PI (Manson), recruited partly from VITAL contacts, 2x2 factorial (cocoa extract x multivitamin), 21,442 older US adults free of major CVD, median 3.6 y. It is the closest thing the corpus holds to a randomized physical isolation of a food-borne flavanol on a hard CV outcome, and on its primary endpoint it is null:
| Endpoint | HR (95% CI) | Prespecified? | State |
|---|---|---|---|
| Total CVD events (PRIMARY) | 0.90 (0.78, 1.02); P=0.11 | primary | no meaningful effect |
| CVD death | 0.73 (0.54, 0.98) | secondary | benefit — but see caveat |
| Total MI | 0.87 (0.66, 1.16) | secondary | no meaningful effect |
| Stroke | 0.91 (0.70, 1.17) | secondary | no meaningful effect |
| All-cause mortality | 0.89 (0.77, 1.03) | secondary | no meaningful effect |
| Total cancer | 1.10 (0.97, 1.24) | secondary | no meaningful effect (point adverse) |
Absolute layer. «The annualized rates of total cardiovascular events were 1.08% and 1.20% in the active and placebo cocoa extract groups, respectively» (Sesso et al., 2022) — a 0.12 percentage-point/yr difference, non-significant, with «410 participants taking cocoa extract and 456 taking placebo» having events (HR 0.90; 95% CI 0.78, 1.02; P=0.11) (Sesso et al., 2022).
The primary was null; do not let the secondary CVD-death signal stand in for it. CVD death 0.73 (0.54, 0.98) was the only secondary reaching nominal significance, but it is hypothesis-generating, not a finding (inferred from Sesso et al., 2022): it carries no multiplicity control, the composite it sits inside was null, and a per-protocol total-CVD figure (HR 0.85, 0.72-0.99) the authors themselves flag «should be interpreted with caution» (Sesso et al., 2022). This is the same discipline applied to VITAL’s total-MI secondary above. Cancer, the other secondary, was null with a slightly adverse point estimate (1.10).
The null is not an adherence artifact. Treatment produced «a more than 3-fold increase compared with placebo in gVLM concentrations, with an overall ratio of geometric means (95% CI) of 3.23 (2.84, 3.67; P < 0.001)» (Sesso et al., 2022) (gVLM = a validated flavanol-intake biomarker) — the flavanol was delivered and absorbed; the composite still did not move.
This is enhancement in the replete, and the exposure is an EXTRACT, not the flavanol and not the food. The 500 mg/d cocoa flavanols dosed «substantially exceeds the mean intake reported in Europe of 105 mg/d» (Sesso et al., 2022) — ~5x the dietary mean, so this joins the page’s core framing (adding a supplement to an already-adequate population). And even this RCT does not isolate the flavanol: «we tested a cocoa extract supplement … containing all naturally occurring bioactive components of the cocoa bean, including cocoa flavanols, (-)-epicatechin, and theobromine … thus, we cannot disentangle the effects of its individual components» (Sesso et al., 2022). So the primary null is consistent with any of three readings held together (inferred from Sesso et al., 2022): the component is not causal; the observational flavonoid signal is confounding; or the extract/dose/matrix differs from food-borne flavanols. The food-vs-component payload lives on Is the Food Category Doing Any Work and Flavonoid Intake and Mortality; here it lands as another supplement with a surrogate/mechanistic rationale that fails to move hard CV events in a low-risk, replete population.
COI (symmetric standards). Cocoa extract supplied by Mars Edge; NIH + Mars co-funded, with blinded assays «completed independently» (Sesso et al., 2022). Note it — but the primary endpoint is null (against the sponsor’s interest), so the COI does not obviously inflate the headline; the caution attaches to the favourably-read secondaries.
Layer-1 placement — a low/absent lever for the well-nourished
For a generally-nourished adult, supplementation to prevent CVD/cancer/fractures/death ranks at or near the bottom of the intervention hierarchy: the proven big rocks (not smoking, activity, weight, sleep, a decent dietary pattern) are elsewhere, and this is a worked instance of the ceiling is itself a finding -> Layer 1 - Ranking Interventions for a Stratum. It also extends the low-risk cardiometabolic ceiling — a supplement is one more intervention with a surrogate/mechanistic rationale that fails to reduce hard events in a low-risk population -> Cardiometabolic Interventions and Hard CV Outcomes in Low-Risk People. The nutrients are better obtained from food (the isolate-vs-food discipline -> Is the Food Category Doing Any Work); USPSTF does not price cost or opportunity cost (US Preventive Services Task Force, 2022), so the expense of a null stack is a real layer-3 constraint this evidence leaves to the person. Scope: this page is general-population disease prevention; it is not deficiency correction, not ergogenic/performance supplements (creatine -> Creatine Supplementation, protein -> Protein and Resistance Training for Muscle and Strength), and not pregnancy folic acid (which USPSTF separately recommends) — those are different decisions, answered on surrogate performance outcomes, not the hard disease-prevention endpoints this page holds.
The counter-example boundary — periconceptional folic acid DOES work, now held as PRIMARY evidence [2026-08-04, De-Regil]
This page’s null is general-population disease prevention — enhancement in the replete on chronic endpoints. The standing counter-example is periconceptional folic acid for neural-tube-defect (NTD) prevention, which this page previously acknowledged only second-hand (via USPSTF’s separate recommendation). De-Regil 2015 (a gold Cochrane-review SR/MA) upgrades that mention to held PRIMARY evidence with a magnitude: periconceptional folate cut NTDs to RR 0.31 (95% CI 0.17-0.58); 5 RCTs; 6708 births; GRADE high, recurrence RR 0.34, absolute 35 -> 11 per 1000, no harm (miscarriage RR 1.10) (De-Regil et al., 2015). Authors: «Folic acid … prevents NTDs» [EXTRACTED chunk 01].
So supplements are not uniformly a distraction: for the pregnancy-planning stratum, on a HARD outcome (a birth defect, not a surrogate), a supplement clearly works — the clean opposite pole to this page’s replete-adult null (inferred from De-Regil et al., 2015). It is a different decision (a life-stage-defined stratum, a critical periconceptional window, a supra-adequate outcome-specific target ~9x above the deficiency threshold), placed precisely as a distinct window-gated flavor of repletion — not the enhancement-in-the-replete this page holds -> Deficiency Repletion vs Enhancement (operative detail
- the distinct-flavor appraisal). This refines the scope line below: pregnancy folic acid is answered on a hard outcome, unlike the performance-surrogate carve-outs (creatine/protein).
Two boundaries the general-population null does not settle (deliverable-critique, 2026-08-01)
A reasonable surprise (deliverable-critique) - “no multivitamin indication? omega-3 no benefit?” Two scope limits keep the null honest without overturning it:
- Multivitamins: USPSTF’s verdict is “I - insufficient”, not “no benefit”. Insufficient evidence is the fourth evidence-state, not a demonstrated null -> Deficiency Repletion vs Enhancement. Contested positive signals (a cognition trial; a modest total-cancer signal in male physicians) are a separate open question the fabric does not yet hold, flagged for acquisition. “Half-dosing a multivitamin” has no evidence base either way.
- Omega-3: the null is standard-dose (1 g/day) enhancement in a general, replete population. It does NOT transport to high-dose prescription EPA (4 g/day) in a high-triglyceride / high-CV-risk stratum - a different exposure, dose and stratum (and whose landmark trial carries a contested comparator). That stratum question is now held (REDUCE-IT) and resolved as a distinction below; the general-population finding stands.
REDUCE-IT cashes the high-dose stratum — and it is a DIFFERENT exposure, not a contradiction of VITAL [2026-08-04, Bhatt]
The reflex on seeing REDUCE-IT (benefit) beside VITAL (null) is to file a tension: does omega-3 prevent CV events or not? It is not a tension — it is a distinction. The two trials give the same word (“omega-3”) to different compounds, at a 4-fold dose gap, in different strata, on different composites. Run the parameter table before any comparison:
| Parameter | REDUCE-IT (Bhatt 2019) | VITAL n-3 arm (Manson 2019) | Same quantity? |
|---|---|---|---|
| Analyte / form | purified EPA ethyl ester, no DHA (Bhatt et al., 2019) | EPA+DHA mix (460 EPA + 380 DHA) (Manson, Cook, Lee, Christen, Bassuk, Mora, Gibson, Albert, et al., 2019) | NO |
| Dose | 4 g/day (Bhatt et al., 2019) | 1 g/day (840 mg) (Manson, Cook, Lee, Christen, Bassuk, Mora, Gibson, Albert, et al., 2019) | NO (4x) |
| Population | statin-treated, established CVD (70.7%) or DM+risk; TG 135-499 (median 216); high CV risk (Bhatt et al., 2019) | general primary prevention (men ≥50 / women ≥55), not TG-selected, replete (Manson, Cook, Lee, Christen, Bassuk, Mora, Gibson, Albert, et al., 2019) | NO |
| Primary composite | 5-point (CV death, MI, stroke, coronary revasc, unstable angina) (Bhatt et al., 2019) | 3-point MACE (CV death, MI, stroke) (Manson, Cook, Lee, Christen, Bassuk, Mora, Gibson, Albert, et al., 2019) | NO |
| Comparator | mineral oil (contested; LDL rose ~10% in placebo) (Bhatt et al., 2019) | placebo (not characterized in held extract) | unknown |
| Primary result | HR 0.75 (0.68-0.83), ARR 4.8 pp, NNT 21 (Bhatt et al., 2019) | HR 0.92 (0.80-1.06), null (Manson, Cook, Lee, Christen, Bassuk, Mora, Gibson, Albert, et al., 2019) | — |
Every input row is NO. The not-joined check (ii) fires — different scope/exposure/population, consistent once matched: a different compound at 4x the dose in a high-risk stratum producing a different result is exactly what transportability predicts, not a clash about one exposure. No tension is filed. Bhatt says the same in his own voice — the prior n-3 nulls may reflect «the low dose or… the low ratio of EPA to docosahexaenoic acid (DHA)», and the results «should not be generalized to other n−3 fatty acid preparations — in particular, dietary-supplement preparations of n−3 fatty acid mixtures» (Bhatt et al., 2019). -> Is the Food Category Doing Any Work (the label-is-not-the-exposure diagnostic). (inferred from Bhatt et al., 2019; Manson, Cook, Lee, Christen, Bassuk, Mora, Gibson, Albert, et al., 2019)
What REDUCE-IT establishes, decision-form (route (a)/(b), high baseline risk). For a statin-treated patient with established CVD or diabetes and elevated triglycerides (135-499 mg/dL), 4 g/day icosapent ethyl is a real lever: over 4.9 y, NNT 21 for the primary composite and a significant 20% lower CV death (HR 0.80, 0.66-0.98) (Bhatt et al., 2019). Three bounds keep it honest, held to the same bar as any favourable finding:
- It is a PRESCRIPTION drug, not the supplement aisle. Icosapent ethyl is a regulated, dosed, purified drug — the realistic alternative here is pharmacotherapy, not an OTC fish-oil capsule (which is the null VITAL exposure). This is a pharmacotherapy finding admitted because a drug is often the realistic comparator to a lifestyle change.
- The comparator is contested. The mineral-oil placebo raised LDL ~10%, which may have inflated the effect; the authors argue the small LDL gap cannot explain a 25% RRR and that a post-hoc analysis was robust to it, and cite open-label JELIS (no mineral oil, 19% RRR) (Bhatt et al., 2019). Unresolved by this trial; a clean-comparator trial (STRENGTH, corn-oil placebo) was null — not held here, flagged for acquisition. So the magnitude carries a discount even though the stratum-benefit direction is credible.
- Harms travel with it: AFib/flutter hospitalization 3.1% vs 2.1% (P=0.004) and a non-significant bleeding signal (2.7% vs 2.1%) (Bhatt et al., 2019) — real layer-3 considerations, not dealbreakers, for the person weighing it.
All-cause mortality was NOT significantly reduced (HR 0.87, 0.74-1.02) (Bhatt et al., 2019) — so this is a hard-CV-event lever with a favourable but unproven survival effect, not a demonstrated longevity intervention. Scope firewall: this section is the high-risk-stratum prescription boundary; it does NOT loosen this page’s general-population supplement null, which is 1 g/day OTC enhancement in the replete and stands unchanged.
Lombardi 2020 generalizes the split across 14 trials — the dose axis, not just VITAL-vs-REDUCE-IT. A pairwise + network meta-analysis (14 RCTs, 125,763 patients, 78.6% statin-treated) splitting trials at a 1 g/day cutoff found high-dose (> 1 g/d) O3FA vs control lowered ischemic endpoints — major vascular events IRR 0.78 (0.71-0.85), MI 0.71 (0.62-0.82), cardiac death 0.79 (0.65-0.96) — while «no benefit of LD treatment compared to HD in any of the efficacy endpoints» and total death, sudden cardiac death, and stroke were null even at high dose. (Lombardi et al., 2020) This firms VITAL beyond a single trial: the low-dose stratum is near-null across the pooled RCT evidence, and pooling every trial regardless of dose washes the ischemic signal toward null (the low-dose trials dilute the high-dose signal). But it does not resolve the distinction into a clean dose law — the high-dose arm is dominated by EPA-only trials in higher-risk populations, so «it remains unknown whether it was the dose or formulation that drove the observed CV benefit», on «an arbitrary cut-off of 1 g/day» with «no direct comparison be-tween HD and LD O3FA». (Lombardi et al., 2020) So Lombardi is a type-F refinement of the parameter table above (adds that the dose axis tracks benefit across the whole trial set), not a filed tension and not independent type-E (it re-pools VITAL, REDUCE-IT, JELIS, ASCEND et al. already held). High-dose benefit is bought with excess bleeding (1.49) and AFib (1.35); and the n-3-vs-control gap was smaller under olive-oil placebos, which the authors read as evidence that «olive oil is not an inert control» — a comparator threat across the literature mirroring the REDUCE-IT mineral-oil concern. (Lombardi et al., 2020) The REDUCE-IT AFib-hospitalization signal (3.1% vs 2.1%) above is now the high-dose edge of a pooled dose-dependent AF harm: Gencer 2021 (7 trials, 81,210 pts) puts overall incident-AF HR at 1.25 (1.07-1.46), rising to 1.49 (1.04-2.15) in the >1 g/d trials vs 1.12 (1.03-1.22) at <=1 g/d (P-interaction <0.001), and the signal survives excluding REDUCE-IT (1.23) — so it is not a mineral-oil artifact -> Omega-3 Supplementation and Atrial Fibrillation. (inferred from Gencer et al., 2021) (inferred from Lombardi et al., 2020)
The sharpest boundary — the deficient stratum on a DIFFERENT outcome (Martineau, vitamin D x ARI)
The general-population null above is enhancement in the replete, on fractures/cancer/CVD. It does not extend to repletion in the deficient, and Martineau 2017 (gold IPD meta-analysis, 25 RCTs, 10,933 participants, GRADE high) is the held demonstration of the other arm — on a different outcome, acute respiratory tract infection (ARI). Vitamin D cut ARI overall («adjusted odds ratio 0.88, 95% confidence interval 0.81 to 0.96» (Martineau et al., 2017), NNT=33), and — the load-bearing part — the benefit was modified by baseline status: in the deficient (<25 nmol/L) on daily/weekly dosing, OR 0.30 (0.17-0.53), NNT=4, versus OR 0.75 in the replete-ish (P for interaction 0.006) (Martineau et al., 2017). Two constraints keep this from loosening the general-population verdict:
- It is a different outcome and a different stratum. ARI is not fracture/cancer/CVD, and the benefit concentrates in the deficient — so this sharpens the null (it is specifically enhancement-in-the- replete on hard chronic endpoints) rather than overturning it. See Deficiency Repletion vs Enhancement for the two-arm structure this instantiates.
- The exposure that works is daily/weekly, not bolus — bolus dosing was null even in the deficient (OR 0.82, 0.51-1.33) (Martineau et al., 2017). And this is NOT independent corroboration of VITAL: it is a different outcome composing the fuller arc (deficient benefits on ARI / replete does not on chronic disease), not a second witness of the same claim (Manson, Cook, Lee, Christen, Bassuk, Mora, Gibson, Gordon, et al., 2019; inferred from Martineau et al., 2017).
Decision-change: for a confirmed vitamin-D-deficient person, daily/weekly D is a real lever on ARI — route (a)/(b) on Baseline Risk and the Relative-Absolute Split, with absolute benefit scaling to deficiency (NNT 4 vs 33). This does not license broad supplementation of the replete, whose ARI benefit was much smaller and whose chronic-disease benefit was null.
The same enhancement-in-the-replete / effect-modification structure recurs for vitamin C on the common cold — regular high-dose C is a precise incidence null in the general community (RR 0.97, 0.94-1.00) but halves incidence in a heavy-acute-physical-stress stratum (RR 0.48, 0.35-0.64), a route-(b) exception -> Vitamin C and the Common Cold.
A second deficient-stratum instance — iron -> cognition (Falkingham 2010, gold MA). The same
enhancement-vs-repletion split holds for a different nutrient: iron supplementation raised IQ in the
anaemic by +2.5 points (95% CI 1.24 to 3.76) but had «no effect on non-anaemic participants»,
and the replete null was well-powered (>1500) hence reliable
(Falkingham et al., 2010),
(Falkingham et al., 2010). So iron, like
vitamin D, is a repletion lever for a deficient stratum and an enhancement null for the
replete — a second nutrient x outcome confirming the deficient-benefits / replete-does-not structure
this page’s general null instantiates -> Deficiency Repletion vs Enhancement (where the two
instances lift the principle to [E-independent]). Caveats: the iron RCTs were «generally small,
short and methodologically weak» with modest publication bias — hold it below the GRADE-high
vitamin-D evidence (Falkingham et al., 2010).
A third deficient-stratum instance — iodine -> child mental development (Bougma 2013, gold MA).
Repleting iodine-deficient mothers/children raised child IQ by +7.4 points (range 6.9-10.2,
d 0.49) (Bougma et al., 2013) — the
largest repletion effect of the three, tracking the deepest deficiency (severe iodine deficiency ->
cretinism). This is the iodine analogue of the same enhancement-vs-repletion split: a repletion lever
for a deficient stratum, and — for a well-nourished adult buying an iodine-containing multivitamin —
enhancement with no such demonstrated gain. Unlike the vitamin-D and iron instances, Bougma directly
tests only the deficient-benefit leg (no supplement-the-replete arm), and its benefit is gated by
developmental timing (early pregnancy d 0.51 vs late 0.17) — so the decision-change is
narrow: repletion of a confirmed iodine-deficient pregnant/young stratum, in the fetal-brain window,
not general-population supplementation -> Deficiency Repletion vs Enhancement (three independent
nutrients now lift the principle to [E-independent]). Hold below the GRADE-high vitamin-D evidence:
only 2 RCTs, small samples, uncontrolled confounding incl. a maternal-behaviour channel
(Bougma et al., 2013).
A diet-DEFINED deficient stratum — B12 in vegetarians/vegans (Pawlak 2013), NOT a fourth quantified instance. Where the three instances above modify a supplement’s benefit by measured baseline status, a vegetarian/vegan diet creates a genuine B12 shortfall by construction (B12 has essentially no plant source) — so B12 supplementation is close to non-optional for this stratum rather than the bottom-of-hierarchy enhancement the general null describes -> Vitamin B12 Status in Vegetarian and Vegan Diets. It is held separately and more weakly, because the source is a descriptive prevalence review (11-90% biochemically deficient, cutoff-driven) with no patient-important-outcome effect and a two-thirds-deficient / zero-clinical-symptoms finding — so the repletion benefit here is a surrogate-flagged gap, not a demonstrated effect like vit-D/iron/iodine (inferred from Pawlak et al., 2013).
Self-critique [run 2026-09-01, before commit — Lombardi 2020 NMA added to the REDUCE-IT/VITAL section]
- Not overclaimed. The Lombardi bullet states the dose finding as an association (HD > 1 g/d associated with lower ischemic events) and twice withholds the causal dose reading — dose is confounded with EPA-only form and high-risk stratum, on an arbitrary cutoff with no closed loop. The decision-relevant claim is the negative one (OTC 1 g/d near-null across pooled trials), which firms the page’s existing general-population supplement null rather than loosening it.
- Not a fake tension. Lombardi is filed as a type-F refinement of the existing VITAL-vs-REDUCE-IT
parameter table (which already resolved the pair as a distinction, same-quantity NO on every row), not
as a new
[[tension]]. - Not laundered-E. Explicitly marked not independent — it re-pools VITAL, REDUCE-IT, JELIS,
ASCEND et al. already held here, so it raises certainty/detail on the supplement leg, not
independence; the page’s
confidence: highrests on the VITAL/USPSTF/Bjelakovic evidence, unchanged. - Coherence, not validity (R1): the loop is open; the olive-oil/mineral-oil comparator threat is named as an unresolved bias in the underlying trials.
Self-critique [run 2026-09-02, before commit — Demay 2024 Endocrine Society guideline woven]
- Not laundered-E. The Demay section is explicitly marked not independent — same VITAL/D2d
trial base as the held USPSTF/Zhang evidence, so it is type-F consolidation, and the page’s
confidence: highis unchanged (a guideline resting on already-held trials adds authority and GRADE calibration, not a second independent backing). - Guideline-tier discipline held. The one effect claim that traces to an underlying SR — the prediabetes HR 0.85 — was held pending its IPD-MA and is now grounded in Pittas directly (the Pittas ingest of 2026-09-04 cashed that hold; see the prediabetes carve-out and the run below), not asserted as Demay’s own; the mortality RR 0.96 is presented as Demay’s own commissioned pooled MA (a guideline-produced estimate), flagged with its CI touching 1.00.
- Not a fake tension with Zhang. Demay’s RR 0.96 vs Zhang’s 0.98 is filed as a distinction (different trial sets — Demay includes calcium-combined trials Zhang excluded), not a joined tension; both CIs border 1.
- Thresholds carry their load. The age-75 cut is stated with the panel’s own «somewhat arbitrary» hedge; doses are given IU/day AND mcg; each stance carries GRADE strength (all conditional “2”) and certainty. No bare cutpoint launders in.
- No [PRIOR] verdict scored here (ingest builds fabric, does not adjudicate): the EtD/weighting observation is appraised on Which Objective Moved This Recommendation, not scored against the weighting [PRIOR] in this ingest.
Self-critique [run 2026-09-04, before commit — Pittas 2023 vitamin-D IPD-MA woven; AWAITS cashed]
- AWAITS cashed correctly. The prediabetes HR 0.85, previously borrowed from Demay’s guideline and
held, now traces to the primary IPD-MA; the token is struck in the same
commit and Pittas is added to
sources:/authors:. No orphan AWAITS remains for this slug. - Not a fake tension with the page thesis. Pittas’ positive prediabetes effect does not contradict the page’s null-in-the-replete headline: it is a different stratum (high diabetes risk, lower median 25(OH)D), read as the baseline-risk route-(a) reconciliation, not a joined tension. Filed as consistency, not friction.
- The 76% is quarantined. The within-arm achieved-25(OH)D HR 0.24 is explicitly marked non-randomized (post-randomization exposure recreates observational confounding), with the causal claim held to the 15% ITT effect and the placebo-arm-null named as the authors’ partial control. No laundering of the observational gradient into a randomized effect.
- Route-(b) meets the bar, stated as such. The BMI×cholecalciferol interaction is admitted as a genuine effect-modification instance because it carries a pre-stated mechanism (CYP2R1 bioactivation) AND an internal negative control (eldecalcitol, P-interaction 0.82) — not on mechanistic plausibility alone.
- Not [E-independent]. Pittas shares the D2d and Tromsø trials with the held aggregate MAs
(Zhang/Barbarawi), so it is type-F consolidation/upgrade over aggregate data, not an independent
second backing;
confidence: highunchanged. - Coherence, not validity (R1): the loop is open; the treat-to-target question and the general-population non-transportability are named as gaps, not resolved.