A single gold dose-response SR+MA of prospective cohorts (Fang 2016, BMC Medicine): 40 publications / 70 reports, >1 million participants, 67,261 cases, 4-30 year follow-up, FFQ-assessed dietary magnesium (NOT supplements). It quantifies what a 100 mg/day higher dietary magnesium intake is associated with across six outcomes. (Fang et al., 2016)
The effect estimates — outcome-specific, not uniform
Per 100 mg/day higher dietary magnesium (relative risks; the analysis reports no absolute risks — absolute benefit scales with each stratum’s baseline risk, route (a), see Baseline Risk and the Relative-Absolute Split):
| Outcome | RR per +100 mg/day (95% CI) | Verdict |
|---|---|---|
| Type 2 diabetes | 0.81 (0.77-0.86) | benefit — largest and most robust |
| Heart failure | 0.78 (0.69-0.89) | benefit, but thin base (3 datasets / 2 cohorts) |
| Stroke | 0.93 (0.89-0.97) | benefit — ischemic-driven |
| All-cause mortality | 0.90 (0.81-0.99) | benefit — borderline, least robust |
| CHD | 0.92 (0.85-1.01) | no significant association per-increment |
| Total CVD | 0.99 (0.88-1.10) | no association |
The author summary: a 100 mg/day increase is associated with a 7% / 22% / 19% / 10% lower risk of stroke / heart failure / T2D / mortality, and no clear association with CHD or total CVD. (Fang et al., 2016)
Why the split matters: the outcome menu is not moved uniformly (as it never is — Surrogate Outcomes). T2D carries the strongest, largest-magnitude, tightest signal; total CVD is a clean null. A single magnesium is cardioprotective headline would erase that structure.
Curve shape — non-linearity is flagged but NOT located
Restricted-cubic-spline tests reject a purely linear model for CHD (P<0.05), stroke (P<0.001), T2D (P<0.001), and mortality (P<0.01); CVD shows no non-linearity (P=0.097). But the paper gives only a spline P-value and a figure — it does not locate a knee, threshold, or plateau in the text. Per the dose-response discipline, that is a non-linear shape is present over the studied range, not a minimum effective dose sits at X. No optimum is derivable here. (Fang et al., 2016)
- Studied ranges (the extrapolation boundary): roughly 100-550 mg/day across outcomes (CVD/T2D/mortality ~100-500; CHD ~150-450; stroke ~150-550). Any apparent flattening at the top edge is as likely the sampling boundary as a true plateau.
- The RDA (350 mg/day male, 300 mg/day female) is a deficiency-coverage construct — read it as a requirement floor, never as the optimum this curve implies; population intakes in EU/US surveys run below it. (Fang et al., 2016)
The binding uncertainty — is magnesium the lever, or a marker of the diet carrying it?
The magnesium-rich foods are whole grains, green leafy vegetables, nuts, beans, and cocoa — i.e. a whole-food, plant-rich dietary pattern (the same foods behind the DASH mineral story and potassium). The authors themselves cannot rule out that magnesium is a proxy: «we cannot exclude the possibility that other nutrients and/or dietary components correlated with dietary magnesium may have been responsible, either partially or entirely, for the observed associations». (Fang et al., 2016)
This is the component-vs-pattern problem in full: an observational dietary-magnesium gradient is not evidence that the magnesium atom is the active ingredient rather than the food matrix or the overall diet quality it indexes (). Consequence for the recommendation: eat more magnesium-rich whole foods is well-supported directionally; supplement magnesium to cut CVD/mortality is NOT what this evidence shows.
The RCT leg partly resolves this confound — for one endpoint. Dibaba’s supplement RCTs isolate the magnesium atom (an isolated pill/solution, placebo-controlled), so their BP reduction cannot be the food matrix or overall diet quality — for blood pressure in the impaired stratum, the atom itself carries a causal effect. That is a genuine narrowing of the lever-vs-marker uncertainty Fang could not touch. Two limits keep it narrow: it holds only for the BP surrogate (not the hard outcomes Fang measured), and only in the metabolically-impaired / low-Mg stratum (plausibly deficiency-correction, above). The pattern-vs-component confound therefore stays open for magnesium’s hard-outcome signal; it is closed only for supplement -> BP -> impaired stratum.
Dietary is not supplemental — a different exposure
Fang is dietary intake only. The paper cites separate trial evidence that oral magnesium supplements (>=4 months) improve insulin sensitivity and glucose control (Simental-Mendia, via Fang) — but that is a different exposure (isolate vs food matrix), not pooled here, and it reports a surrogate (glycaemia), not hard outcomes. No large RCT has raised magnesium intake to prevent CVD/T2D events; for hard outcomes the base is entirely observational. (Fang et al., 2016)
The supplement leg Fang only gestured at is now held for one surrogate — blood pressure — via two RCT-MAs: Dibaba 2017 (impaired stratum, below) and Zhang 2016 (general population, added 2026-09-25). Both are a different exposure (isolated supplement) on a different endpoint (a BP surrogate, not events) than Fang’s dietary-cohort hard outcomes, so they do not merge with Fang’s numbers — see the three-leg parameter table under Synthesis. Zhang is the parent general-population pool that Dibaba re-slices to the impaired stratum (8/11 shared trials) — so Zhang and Dibaba are the same question at two strata, not independent replications.
The supplement-RCT causal leg — magnesium supplementation lowers BP in the impaired stratum (Dibaba 2017)
A gold MA of RCTs (Dibaba 2017, AJCN): 11 parallel-group trials, 543 participants (278 supplemented), follow-up 1-6 months (mean 3.6). This is the causal leg — randomization isolates the magnesium and breaks the pattern-vs-component confound that limits Fang’s observational dietary gradient (for this one endpoint). (Dibaba et al., 2017)
Stratum — deliberately restricted, and that is the point. Trials were included only in participants with insulin resistance, prediabetes, type 2 diabetes, or cardiovascular disease (no renal/cancer trials were found). Prior MAs mixed apparently-healthy and primary-hypertensive participants; Dibaba argues the impaired stratum responds differently — «Note that individuals with or without preclinical or chronic disease may respond to magnesium supplementation differently» — and pooled only the impaired stratum. (Dibaba et al., 2017) This is a restriction, not a within-study interaction test — it is route-(b)-flavoured (stratum selected on metabolic condition) but does not formally test effect-modification; the larger effect than mixed-population MAs claim is a cross-MA comparison, not an interaction coefficient.
The effect — small, and read the right number. The placebo-controlled (between-group) estimate is the standardized mean difference: SBP SMD -0.20 (95% CI -0.37, -0.03); DBP SMD -0.27 (95% CI -0.52, -0.03). In mm Hg the paper reports two different framings that must not be conflated:
| mm Hg figure | What it is | Value |
|---|---|---|
| Between-group (placebo-controlled, the causal estimate) | supplementation vs control | SBP -2.22 · DBP -2.54 |
| Within-arm (baseline -> end, supplementation group only) | not placebo-subtracted | SBP -4.18 · DBP -2.27 |
The headline «Magnesium supplementation resulted in a mean reduction of 4.18 mm Hg in SBP and 2.27 mm Hg in DBP» is the within-arm before-after change in the treated arm, not the placebo-controlled effect; the controlled between-group decrement is «a weighted mean decrement in SBP by 2.22 mm Hg and in DBP by 2.54 mm Hg when comparing the supplementation group with the control group». Quote the between-group / SMD as the causal effect. (Dibaba et al., 2017)
Robustness — SBP solid, DBP fragile.
- SBP: no significant heterogeneity (chi2(10)=10.21, P=0.42); significant in every leave-one-out except dropping Guerrero-Romero & Rodriguez-Moran (ref 13). Egger P=0.07 (no strong publication-bias signal).
- DBP: significant heterogeneity (chi2(10)=20.03, P=0.03); lost significance on several leave-one-out cuts and when the assumed baseline-end correlation was set to r=0.3 (held at r=0.5/0.7). Egger P=0.49.
- Dose 365-450 mg/d elemental Mg (supra-RDA); trials short (mean 3.6 mo). (Dibaba et al., 2017)
Deficiency-correction, not a general dose-response — the transportability caveat. Several included trials selected hypomagnesemic / low-serum-Mg diabetics, so the effect may be repletion of a deficit rather than a dose-response that transports to magnesium-replete healthy people — the deficiency-vs- enhancement distinction (Vitamin and Mineral Supplements for Disease Prevention). Dibaba themselves flag residual confounding by background dietary magnesium and unadjusted co-medications «thereby possibly confounding the effect of dietary magnesium intake and the heterogeneity between some of the trials, particularly for DBP». (Dibaba et al., 2017) So the transportability target is the metabolically-impaired, plausibly-hypomagnesemic stratum, not the reasonably-healthy default reader.
BP is a surrogate — the events transmission is borrowed, not measured. No Dibaba trial measured CV events; the leap from BP to outcomes rests on antihypertensive-drug trials — «The trial suggested that a reduction of BP by 2-3 mm Hg might account for a difference of stroke rate by 6-12% between antihypertensive medications». That transmission is evidenced for drug-induced BP lowering, not for magnesium, and a surrogate that moves is not an outcome that moves (Surrogate Outcomes, Blood Pressure Lowering and Cardiovascular Events). (Dibaba et al., 2017)
The general-population supplement-RCT arm — Zhang 2016 (the parent pool Dibaba re-sliced)
A gold MA of 34 double-blind placebo-controlled RCTs, 2028 participants (1010 supplemented, 1018 placebo), median 368 mg/d elemental Mg for 3 months, in general-population normotensive OR hypertensive adults — the arm Dibaba deliberately excluded. Placebo-controlled (between-group) pooled effect: SBP -2.00 mm Hg (95% CI -3.58 to -0.43); DBP -1.78 mm Hg (95% CI -2.82 to -0.73), with serum Mg +0.05 mmol/L. «Mg supplementation at a median dose of 368 mg/d for a median duration of 3 months significantly reduced systolic BP by 2.00 mm Hg (95% confidence interval, 0.43–3.58) and diastolic BP by 1.78 mm Hg (95% confidence interval, 0.73–2.82).» So a small general-population BP effect is real — «provision of Mg may slightly lower BP and might be effective in preventing hypertension in the general population». (Zhang et al., 2016)
These figures are the causal (placebo-controlled) estimate — and that resets the stratum contrast. The naive reading is that the impaired stratum responds more (Dibaba’s headline 4.18 mm Hg vs Zhang’s 2.00). But 4.18 is Dibaba’s within-arm change (baseline->end, treated group only, not placebo-subtracted); the like-for-like placebo-controlled Dibaba figure is SBP -2.22 mm Hg, essentially identical to Zhang’s general-population -2.00. Compared as the same quantity, the general-population and impaired-stratum causal effects are the same size — so the impaired responds more story is NOT supported at the level of the placebo-controlled estimate; it rested on a within-vs-between-group quantity mismatch (the parameter-table discipline catching exactly the error it exists to catch). (Dibaba et al., 2017; inferred from Zhang et al., 2016)
The route-(b) effect-modification signal is suggestive but NOT clean — read Zhang’s own interaction tests. Within Zhang (all placebo-controlled, same design, same quantity — the cleanest available test), the point effect IS larger in the low-Mg / treated strata: baseline serum Mg Q1 (<0.71 mmol/L) SBP -4.50 (-7.24 to -1.76), DBP -5.05 (-9.12 to -0.97); treated (on antihypertensive/antidiabetic drugs) SBP -5.69 (-10.4 to -1.00). Zhang’s discussion concludes «the antihypertensive effect of Mg was significant only among the subgroup with Mg deficiency». But the between-subgroup interaction tests were non-significant — baseline-Mg P-interaction 0.21 (SBP) / 0.90 (DBP); medication-history 0.11 / 0.72 — so this is a larger-point-estimate + within-subgroup-significance pattern, not an established effect-modifier. The only significant BP interactions Zhang found were methodological (trial quality P=0.02; dropout rate P=0.002), and prior BP status showed no modification (normotensive vs hypertensive P-interaction 0.79 / 0.59). (Zhang et al., 2016)
Net: deficiency-correction remains the most plausible mechanism for any stratum difference (route-(b), consistent with Dibaba’s design logic and the Vitamin and Mineral Supplements for Disease Prevention deficiency-vs-enhancement rule), but the interaction is unproven on either MA’s clean tests — a route-(b) candidate, held under the U/J-artifact-style caution that a subgroup point estimate must survive an interaction test before it drives a recommendation.
Measurement error and the null arms
Dietary magnesium is FFQ-self-reported, and the authors note «measurement error might occur in dietary assessment, which would likely bias true associations towards a null association» — so the CVD/CHD nulls and the mortality upper bound touching 1.0 are weak evidence of no gradient, not proof of none (Measurement Error in Dietary Assessment; attenuation-toward-null). (Fang et al., 2016)
Robustness notes
- Mortality is the weakest of the four positive findings: per-100mg RR 0.90 (0.81-0.99) with I2=62%, upper CI at 0.99, and it did not survive a subgroup/meta-regression cut where stroke incidence stayed inverse (0.92; 0.89-0.95) but mortality did not (RR 1.07; 0.90-1.28).
- Stroke is ischemic-driven (ischemic 0.93, 0.88-0.98; hemorrhagic null 0.97, 0.88-1.07).
- Heart failure is the largest per-increment effect (22%) but rests on 3 datasets from 2 cohorts — precise (I2=0) but narrow; non-linearity untestable.
- Publication bias: no significant evidence (funnel/Egger/Begg) across outcomes. NOS mean quality 8.2.
All figures in this section: (Fang et al., 2016)
Synthesis — three legs, and which pairs are the same quantity
The page now holds three gold sources. The same-quantity check governs which may be compared and which may only be configured:
| Parameter | Fang 2016 | Dibaba 2017 | Zhang 2016 | Same quantity? |
|---|---|---|---|---|
| Exposure | dietary Mg (FFQ), +100 mg/d | Mg supplement, 365-450 mg/d | Mg supplement, median 368 mg/d | Fang No (food vs isolate); Dibaba ~ Zhang: yes (both isolate supplement) |
| Design | prospective cohort (observational) | MA of RCTs | MA of RCTs | Fang No; Dibaba = Zhang both RCT-MA — but 8/11 trials SHARED (not independent) |
| Outcome | hard endpoints (incidence) | BP surrogate | BP surrogate | Fang No; Dibaba = Zhang both BP |
| Population | general adult cohorts | impaired (IR/prediabetes/T2D/CVD) | general (normotensive + hypertensive) | Dibaba vs Zhang: different STRATA — the contrast |
| Effect metric | RR per +100 mg/d | SMD; between-group -2.22 / within-arm -4.18 mm Hg | between-group WMD -2.00 mm Hg | Compare between-group to between-group only: -2.22 ~ -2.00; NOT -4.18 |
Fang is a type-F composite with the two supplement MAs (different exposure/design/outcome, authorship splits it from the supplement leg): Fang shows a dietary-intake association with hard outcomes but cannot say magnesium is the lever; the supplement RCTs show an isolate causally moves a BP surrogate but cannot reach events. Their union raises confidence that a real magnesium signal exists and that one link (Mg -> BP) is causal — while the decision-critical link (Mg -> hard events, causally) stays unproven. (Dibaba et al., 2017; inferred from Fang et al., 2016; Zhang et al., 2016)
Zhang and Dibaba are the SAME question at two strata — a type-F cross-stratum arm, NOT type-E, and NOT a
clean interaction. Same exposure (isolate supplement), same design (RCT-MA), same outcome (BP), differing
only in stratum (general vs impaired). They are not independent: Dibaba shares 8 of its 11 trials with
Zhang — «Eight of the studies overlapped with the studies included in the present study» — and Song Yiqing
co-authored both, so Zhang is the PARENT general-population pool and Dibaba an impaired-stratum
re-slice of largely the same trials (laundered-E if marked as independent replication — so NOT
[E-independent]). What the pair buys is type-F: Zhang supplies the general-population arm Dibaba excluded,
and the pair bounds the stratum question rather than confirming it. (Dibaba et al., 2017)
And bounding it deflates the stratum story. Comparing the like quantities (both placebo-controlled between-group): Dibaba impaired -2.22 mm Hg ~ Zhang general-population -2.00 mm Hg — the causal effects are the same size across strata. The apparent impaired responds more rested on Dibaba’s within-arm -4.18 (not placebo-subtracted) vs Zhang’s placebo-controlled -2.00 — a not-same-quantity comparison. So the route-(b) effect-modification signal is a candidate, not a finding: the larger effect shows only in Zhang’s own low-Mg/treated subgroups (non-significant interaction) and in a within-vs-between-group artifact, never in a clean interaction test on either MA. Deficiency-correction remains the most plausible mechanism if any modification is real, but neither MA establishes it. (Dibaba et al., 2017; inferred from Zhang et al., 2016)
Layer-1 placement (where this ranks)
A modifiable exposure with a plausible relative signal on T2D and stroke (dietary, observational) and a small causal effect on a BP surrogate (supplement RCT, impaired stratum) — but (i) no hard-outcome RCT, (ii) the dietary hard-outcome signal still confounded by pattern-vs-component, (iii) the supplement BP effect small, short-term, DBP-fragile, and plausibly deficiency-correction, and (iv) absolute benefit unquantified. It is a refinement lever, not a big rock: for a reasonably healthy (magnesium-replete) person, prefer magnesium-rich whole foods is subsumed by the broader whole-food / DASH-pattern recommendation that already carries potassium, fibre, and low-SFA benefits, and supplementing to lower BP has no demonstrated events benefit and is sized against mature, low-harm BP drugs and the bigger sodium/potassium/weight levers. A general-population placebo-controlled BP effect is now held (Zhang, SBP -2.00 / DBP -1.78) — but it is small and roughly the same size across strata on the like-for-like between-group comparison (Zhang -2.00 ~ Dibaba impaired -2.22), so the impaired stratum is where the rock is larger claim is not clean — the deficient/low-Mg subgroup shows a larger point effect but no significant interaction. Net: a small surrogate-only lever whose marginal rank does not clearly rise even in the impaired stratum.
Gaps (G)
- Does repleting magnesium (diet OR supplement) causally reduce hard cardiometabolic events? The RCT base now reaches a BP surrogate (Dibaba) but no trial measured events — the surrogate-to-outcome causal gap is open (insufficient evidence, not a null).
- Does the supplement BP effect hold in magnesium-replete, non-impaired people, or is it deficiency- correction confined to the impaired/low-Mg stratum? PARTLY ANSWERED (Zhang 2016, held): a general-population placebo-controlled effect IS present (SBP -2.00, DBP -1.78, incl. normotensives), and it is ~the same size as Dibaba’s impaired-stratum between-group effect (-2.22), so the effect is not confined to the impaired stratum. Still open: whether it is larger in the deficient stratum — the larger point estimates in Zhang’s low-Mg/treated subgroups did not reach a significant interaction, so effect-modification stays an unproven route-(b) candidate. Needs an RCT that pre-stratifies on baseline serum Mg with an interaction test, or an IPD MA.
- What is the shape of the T2D / stroke curve (knee location, minimum effective dose)? Non-linearity is flagged but unquantified in Fang — needs the spline coordinates or a curve-shape SR.
- Absolute risk reductions by baseline-risk stratum — not derivable from these relative/surrogate-only sources.
Magnesium’s anxiety/stress outcome is a different cell, held separately at Magnesium Supplementation and Subjective Anxiety (low warrant) — do not read a cardiometabolic magnesium benefit as an anxiety benefit, and vice versa.