The domain-opening summary for coffee, built on one gold-tier umbrella review (Poole et al., 2017) that assimilated 201 observational meta-analyses (67 outcomes) + 17 RCT meta-analyses (9 outcomes). The one-line verdict Poole reaches: coffee is «generally safe within usual levels of intake… and more likely to benefit health than harm», with «largest risk reduction for various health outcomes at three to four cups a day». (Poole et al., 2017)
But the whole page rests on a single load-bearing caveat, stated up front so nothing below reads as established causation: almost every estimate here is observational, GRADE-rated low (~25%) or very low (~75%), and the two Mendelian-randomisation studies Poole cites found no genetic evidence for a causal coffee->mortality or coffee->T2D relation — «suggesting residual confounding could result in the observed associations in other studies». (Poole et al., 2017) The coffee->mortality MR is now held primary here (Nordestgaard 2016, ingested as the study Poole cited secondhand) — its instruments, magnitudes, and its two bounding caveats are in The Mendelian-randomization check below. (Nordestgaard & Nordestgaard, 2016) So read every RR below as an association net of whatever smoking/SES confounding survived adjustment, not as an effect. The friction on how much of the protective arm is real is on The U-Shaped Association Artifact. Grosso 2016 has now performed the smoker referent-correction (the smoking-stratified dose-response detail below) (Grosso et al., 2016).
The dose-response shape — a plateau, not a harmful upper arm type-C
Where Poole found non-linearity (all-cause mortality, CV mortality, CVD, heart failure), the shape is a reverse-J that flattens: risk falls to a nadir near 3-4 cups/day, and «increase in consumption beyond this intake does not seem to be associated with increased risk of harm, rather the magnitude of the benefit is reduced». (Poole et al., 2017) Two decision consequences:
- The nadir is a region, not a target — the curve is flat around it, so 2 vs 4 cups barely differs, and the burden is on anyone claiming a sharp optimum. This is the corpus’s standing dose-response default (every reduction/increment near a flat region costs little either way).
- No harmful upper arm for mortality within studied intakes — unlike the alcohol J-curve, the high end is attenuated benefit, not risk. So the U/J-artifact question here is primarily about the lower (protective) arm: is the benefit-vs-none real, or confounded? -> The U-Shaped Association Artifact. (Refinement below: removing smokers cleanly exposes a smoking artifact for the cancer arm; for all-cause/CVD the stratified spline does not support a smoking-driven upper arm — the never-smoker CVD curve itself plateaus at ~0.75 from ~3-4 cups. Corrected 2026-10-06: in never-smokers the curve is linear-monotone with no plateau -> “linear” is Grosso’s per-cup linear model, not the spline; self-critique.) For T2D the dose-response is monotone («risk was still lower for each dose… between one and six cups»), not even a plateau.
Smoking confounds the mortality curve — Grosso’s referent correction type-F
Grosso 2016 (a dose-response MA, 31 cohorts, 1,610,543 individuals) is the smoker/non-smoker referent correction Poole’s page left pending. Two facts frame it: the overall (smoker-mixed) all-cause curve is a J that plateaus then rises — nadir RR 0.83 at 3 cups, back up to 0.90 (0.85-0.96) at 7 cups (Grosso et al., 2016) — while in never-smokers Grosso’s per-cup linear model gives a steady decrement: «a linear dose-response analysis showed a significant decreased risk by 6 % for each additional cup of coffee per day consumed for all- cause and CVD mortality (RR = 0.94, 95 % CI = 0.93, 0.96 and RR = 0.94, 95 % CI = 0.91, 0.97, respectively) and significant decreased risk of 2 % for cancer mortality (RR = 0.98, 95 % CI = 0.96, 1.00).» (Grosso et al., 2016) The fitted spline (Table 3) is not linear throughout: never-smoker all-cause keeps falling slowly after 3 cups (to 0.79, 0.71-0.87, at 7), never-smoker CVD plateaus (0.76, 0.69-0.83, at 4 cups; 0.75 at 5; 0.74 at 7), and the smoker curves do not rise either (all-cause 0.87, 0.77-0.98; CVD 0.71, 0.61-0.83, at 7) — while each stratified model pools only ~5 studies («9 (5)», «10 (5)») against 24 for the pooled curve (Grosso et al., 2016). Grosso’s prose (a J in smokers, slope «linear for all outcomes») and its table disagree here. Heterogeneity fell in every smoking-stratified model — consistent with smoking status as a between-study variance source, though pooling ~5 studies instead of 24 would also lower it.
- The CANCER arm is the clean, Grosso-attributed confounding demonstration. It flips sign: «cancer mortality was significantly decreased only when considering non-smokers, while increased in smokers» (Grosso et al., 2016). Grosso’s discussion attributes it to residual confounding by smoking: «it is hardly plausible that any biological effect of coffee causally differs by smoking status… residual confounding by smoking is the most likely the explanation». (Grosso et al., 2016) — although its conclusion uses effect-modification language for the same contrast: «The contrasting results for higher intake of coffee among smokers depend most likely on the effect modification of smoking habit and more realistic association with mortality risk is provided in non- smokers.» (Grosso et al., 2016). The confounding reading is the better-argued of the two (corrected 2026-10-06: Grosso pins it to confounding, not effect-modification -> Grosso uses both framings; self-critique). The mechanism — heavy coffee drinkers are enriched for smokers, and smoking is the dominant cancer/mortality risk factor — is (inferred from Grosso et al., 2016).
- For all-cause/CVD the artifact reading is NOT supported by the stratified table, and Grosso himself does not make it. Neither stratum’s spline shows the pooled curve’s rise (0.83 -> 0.90): the smoker curves plateau or keep falling, and the never-smoker CVD curve plateaus — so the pooled J’s upper arm cannot come from the smokers’ curve, and more plausibly reflects the different study set (24 vs ~5 studies). Grosso reports «No differences were found between smokers and non-smokers for all-cause and CVD mortality risk… both significantly reduced» (Grosso et al., 2016). So he attributes the smoking artifact explicitly only to cancer (corrected 2026-10-06: removing smokers does remove the upper-arm attenuation … suggestive -> not supported by Grosso Table 3; self-critique).
- What it does NOT fix — the lower arm stays only partly adjudicated. Grosso is observational: it removes the dominant confounder (smoking) but not SES / reverse causation / other residuals — and Poole’s Mendelian-randomisation citations found no genetic causal signal for coffee->mortality. So the two are consistent, not in tension: the per-cup benefit survives the smoking referent-correction (Grosso) yet finds no support from the genetic instrument (Poole’s MR = the held Nordestgaard, which cannot exclude an effect of the observed size) — leaving residual non-smoking confounding (SES, reverse causation) and a modest causal effect that the MR is too small to exclude as the two live explanations (corrected 2026-10-06: the remaining live explanation -> two live explanations; self-critique). Net: smoking is not the whole story of the association, but neither is the surviving linear benefit established as causal (~6% relative per cup is small, and the source gives no baseline risk — relative-only). -> The U-Shaped Association Artifact.
What coffee moves, by evidence state
| Outcome | Direction + magnitude (vs none unless noted) | State | Note |
|---|---|---|---|
| Liver disease (cirrhosis, fibrosis, NAFLD, chronic liver disease, HCC) | benefit, LARGEST + most consistent: chronic liver disease high-vs-low RR 0.35; liver cancer high-vs-low 0.50 (0.43-0.58); cirrhosis (any vs none) 0.61; NAFLD 0.71 | benefit | liver cancer + chronic liver disease are the only outcomes reaching GRADE-upgradeable magnitude (<0.5) |
| All-cause mortality | 0.83 (0.79-0.88) at 3 cups | benefit (assoc.) | MR-null; confounding caveat. CI per Grosso Table 2 (Grosso et al., 2016); Poole prints «0.83 to 0.88», a transcription slip (corrected 2026-10-06: 0.83-0.88 -> 0.79-0.88) |
| CV mortality / CVD | CV mort 0.81 (0.72-0.90); incident CVD 0.85 (0.80-0.90) at 3-5 cups | benefit (assoc.) | MR-null |
| Type 2 diabetes | high-vs-low 0.70 (0.65-0.75); monotone 1-6 cups (Ding) | benefit (assoc.) | decaf lowers risk too — details below |
| Total cancer incidence | high-vs-low 0.82 (0.74-0.89) | benefit (assoc.) | most single sites null |
| Parkinson’s, depression, Alzheimer’s | lower risk, consistent | benefit (assoc.) | Parkinson’s survives smoking adjustment |
| Gallstones, gout, renal stones, metabolic syndrome | lower risk | benefit (assoc.) | — |
| Blood pressure | RCTs marginal, non-significant; obs null | no meaningful effect | — |
| Lung cancer | apparent harm OR 1.59 high-vs-low | confounded to null | «not seen in never smokers» — residual smoking confounding |
| Most cancer sites (gastric, colorectal, breast, ovarian, pancreatic…) | no significant association | no meaningful effect / insufficient | — |
| Pregnancy (low birth weight, preterm, pregnancy loss) | harm: LBW OR 1.31; loss 1.46; 1st-tri preterm 1.22 | harm signal (assoc.) | survives smoking adjustment; single MA per outcome; the one RCT was null; mechanism below |
| Fracture in women | high-vs-low RR 1.14 (1.05-1.24); men 0.76 | harm signal (women only, assoc.) | sex effect-modifier, P<0.001; many fracture studies unadjusted for BMI/smoking/calcium |
| Sleep, respiratory | — | insufficient (no MA existed) | named gaps, not nulls |
The liver row is the big rock. Poole: «The beneficial associations between consumption and liver conditions stand out as consistently having the highest magnitude compared with other outcomes across exposure categories.» (Poole et al., 2017) Liver cancer and chronic liver disease are the only outcomes whose effect size (<0.5) is large enough to permit a GRADE upgrade of observational evidence -> Upgrading Observational Evidence. Antioxidant/anti-inflammatory mechanism (chlorogenic acids, caffeine) plus a direct antifibrotic action on hepatic stellate cells is the proposed pathway (inferred from Poole et al., 2017).
The two harm signals — stratum-specific, and they gate the recommendation
(corrected 2026-10-06: heading The two real harms -> harm signals; both are observational like the benefits; self-critique.)
- Pregnancy is the one place harm survives smoking adjustment — on thin evidence: «These associations were seen in subgroup analyses from articles investigating total caffeine exposure, which showed similar associations, and from a single meta-analysis for each outcome.» (Poole et al., 2017), and the one RCT on birth outcomes was null («none of the outcomes reached significance»). Mechanism is dose-amplification, not a new pathway: «The half life of caffeine is known to double during pregnancy… Caffeine is also known to easily cross the placenta, and activity of the caffeine metabolising enzyme, CYP1A2, is low in the fetus, resulting in prolonged fetal exposure.» (Poole et al., 2017) So the same per-cup intake delivers a higher effective fetal dose. This is a precautionary contraindication (route c) — a consistent signal plus a mechanism, not an established harm — not a shift in the general estimate.
- Fracture in women — no overall association, but sex is an effect modifier: «high versus low consumption was associated with an increased risk of fracture in women (relative risk 1.14, 95% confidence interval 1.05 to 1.24) and a decreased risk in men (0.76, 0.62 to 0.94)… test of interaction… 1.50, 1.20 to 1.88; P<0.001).» (Poole et al., 2017) Two attenuators Poole flags: a caffeine SR found 400 mg/day (~4 cups) not associated with fracture/BMD harm, and «only a small amount of milk added to coffee would be needed to offset any negative effects on calcium absorption» — so the harm may be confined to women with inadequate calcium intake; the milk remark concerns calcium absorption and has not been tested against fracture. And the signal is weakly adjusted: «Notably, many of the studies included in the meta-analyses of coffee consumption and risk of fracture did not adjust for important confounders such as body mass index (BMI), smoking, or intakes of calcium, vitamin D, and alcohol.» (Poole et al., 2017)
Brewing method IS load-bearing — the diterpene / lipid channel type-C
The one within-”coffee” boundary that carries a real, mechanistic decision: filtered vs unfiltered. Coffee overall raises serum lipids in RCT meta-analysis — total cholesterol +0.19 mmol/L (0.10-0.28), LDL +0.14, TG +0.14 — via the diterpenes cafestol and kahweol, and the effect tracks the unfiltered preparations (boiled, cafetière, espresso contain the diterpenes; instant and filtered are «negligible»). A paper filter removes most of it: «The increases in cholesterol concentration were mitigated with filtered coffee, with a marginal rise in concentration (mean difference 0.09 mmol/L, 0.02 to 0.17) and no significant changes to low density lipoprotein cholesterol or triglycerides compared with unfiltered (boiled) coffee.» (Poole et al., 2017)
But do not over-weight the lipid signal: Poole notes the changes reverse with abstinence, are small, and coffee «does not seem to be associated with adverse cardiovascular outcomes» despite them — i.e. the LDL surrogate moves the wrong way while hard CV outcomes do not, a Surrogate Outcomes disconnect. The actionable residual: someone drinking large volumes of unfiltered coffee with high LDL / established ASCVD risk has a cheap lever — switch to filtered — while for everyone else the diterpene effect is marginal. -> Is the Food Category Doing Any Work (brewing method as the load-bearing sub-boundary).
Additives are a second uncontrolled axis — added sugar most of all [INFERRED, 2026-08-07 maintainer challenge]
The cohort “coffee” exposure pools drinks taken black, with milk, and with sugar — additives are as uncontrolled as brew method is for the T2D estimate (below). Milk’s effect is minor (the calcium-offset note above); added sugar changes the exposure — a sugar-sweetened coffee drink imports the held free-sugars / SSB harm -> Free Sugars Intake, so the mortality/T2D benefit here attaches to coffee the beverage, not to a sugar-loaded coffee drink. Same within-category boundary as filtered-vs-unfiltered -> Is the Food Category Doing Any Work: “3-4 cups” is not one exposure — brew, cup size, caffeine, milk and added sugar all vary it. The wiki holds no cohort separating sweetened from unsweetened coffee on hard outcomes — a named acquire-gap, not a held finding.
Caffeine is (mostly) NOT the active agent type-C
Decaffeinated coffee reproduces the main benefits: it lowered all-cause and CV mortality (similar magnitude, nadir 2-4 cups), and for T2D «Consumption of decaffeinated coffee also seemed to have similar associations of comparable magnitude». (Poole et al., 2017) Poole chose coffee, not caffeine, as the exposure precisely because coffee’s ~1000+ bioactives «could be different to effects of caffeine from other sources». (Poole et al., 2017) So for mortality and T2D (where decaf data exist) the exposure is the coffee matrix, not the caffeine — which decouples those benefits from the one component (caffeine) that drives the pregnancy harm and the sleep/BP/anxiety physiology. Not for the liver: Poole reports no significant decaf liver association — its decaf figure (Fig 5) lists a per-cup liver-cancer estimate of 0.93 (0.86 to 1.00) from 3 studies (ref 10), which falls among «The other outcomes investigated for decaffeinated coffee showed no significant associations, though it should be noted that meta-analyses of consumption would have much lower power to detect an effect.» (Poole et al., 2017) — and Poole’s proposed liver mechanism includes caffeine — «Additionally, caffeine could have direct antifibrotic effects by preventing hepatic stellate cell adhesion and activation.111» (Poole et al., 2017). Decaf comparisons also carry a selection caveat: «People who drink decaffeinated coffee might be different from those who drink caffeinated coffee» (Poole et al., 2017) (corrected 2026-10-06: mortality/metabolic/liver -> mortality and T2D; self-critique). -> Is the Food Category Doing Any Work (the decaf test).
Where caffeine specifically DOES matter: pregnancy (fetal caffeine exposure), and a CYP1A2 gene-dose effect on hypertension — «Those with alleles for slow caffeine metabolism were at increased risk of hypertension compared with those with alleles for fast caffeine metabolism.» (Poole et al., 2017) (a candidate effect-modifier, not yet an actionable stratifier).
The Mendelian-randomization check — primary evidence type-F
The disconfirming genetic signal was previously held only secondhand (Poole citing “two MR studies”). Nordestgaard 2016 is now held primary, so the null carries its own instruments and magnitudes rather than a borrowed sentence. Across 95,000-223,000 individuals (Copenhagen General Population Study + two Copenhagen cohorts + Cardiogram/C4D consortia for IHD): «observation- ally coffee intake was associated with U-shaped low risk of cardiovascular disease and all-cause mortality; however, genetically coffee intake was not associated with risk of cardiovascular disease or all-cause mortality.» — «The latter are novel findings.» (Nordestgaard & Nordestgaard, 2016)
- Instrument. A caffeine-intake allele score (near AHR and CYP1A1/CYP1A2); per-allele ~8% higher intake, 0 vs 4 alleles = 2.2 vs 3.1 cups/day (42% higher). First-stage strength is high — «the statistical F-value for allele score was very high at 827». Per-allele genetic HRs/ORs cluster at ~1.00 with CIs spanning 1.0. (Nordestgaard & Nordestgaard, 2016)
- Positive control (the instrument is not dead). The same design recovers a known causal chain: «low cholesterol level caused by ApoE genotype associated causally with the ex- pected low risk of IHD.» (Nordestgaard & Nordestgaard, 2016)
- Confounding named. «U-shaped associations are less prominent in never smokers compared with former and current smokers» — so the authors read their genetic null as evidence the observational ‘benefit’ is confounded (smoking), not causal. (Nordestgaard & Nordestgaard, 2016)
Two caveats bound the null — it finds no genetic signal, but it is not powered to exclude causation at the OBSERVED size, and it does not prove zero. (corrected 2026-10-06: refutes causation at the OBSERVED size -> not powered to exclude it, Nordestgaard chunk 01)
- Power. The instrument «could exclude odds ratio per allele of 0.97 for 8% higher coffee intake» (80% power) — a per-unit effect far steeper than the observational «odds ratio of 0.86 for approximately 350% higher coffee intake», which the authors say would need «approximately 225 000 cases and 225 000 controls» to exclude; «even more individuals and events are required to thoroughly exclude a causal associ- ation». So: no genetic signal, and a causal effect much larger per cup than the observed one is excluded; a causal effect at the observational magnitude is not excluded — insufficient evidence, not no effect (corrected 2026-10-06: rejects a causal effect as large as the observational 0.86 / no meaningful causal effect at the observational magnitude -> not excluded, Nordestgaard chunk 01). (Nordestgaard & Nordestgaard, 2016)
- Linearity — the MR is structurally blind to a true U. «if U-shaped associations … indeed are true, then a Mendelian randomization approach might not detect an eventual as- sociation, since this approach is based on the assumption of linearity … and thus will not be capturing non-linear differences between very low and very high coffee intakes.» (Nordestgaard & Nordestgaard, 2016)
Parameter table (BLOCKING — matched quantities before any cross-source claim):
| Parameter | Nordestgaard 2016 (primary MR) | Incumbent claim / Poole 2017 (obs + secondary MR) | Same quantity? |
|---|---|---|---|
| Genetic (causal) coffee->all-cause mortality | per-allele HR ~1.00, CIs span 1.0; F=827; «genetically coffee intake was not associated with … all-cause mortality» | Poole secondhand: «no genetic evidence for a causal … relation» (no numbers) | YES — same causal-null quantity; Poole reports THIS study secondhand -> type-F firming, not independent-E |
| Observational coffee->all-cause mortality | HR 0.86 for ~350% higher intake (0 vs 4-5 cups) | Poole/Grosso: nadir RR 0.83 (0.79-0.88; Poole misprints the lower bound as 0.83) at 3 cups vs none | ROUGHLY — same association type, different cohorts + contrast (whole-range HR vs per-category nadir); directional agreement only, not a matched estimate |
| Genetic vs observational (WITHIN Nordestgaard) | causal OR/allele 0.97 (8% intake) excludable vs observed 0.86 (350% intake) | — | NO — different exposure scale AND causal-vs-associational; this divergence IS the finding, not a tension to file |
Independence verdict. Nordestgaard IS the coffee->mortality MR Poole cited: Poole’s mortality MR
citation is its ref 123, Nordestgaard AT, Nordestgaard BG … Int J Epidemiol 2016;45:1938-52 — the
held study (Poole et al., 2017) (corrected 2026-10-06:
very likely -> definite, Poole ref list chunk 02). This is primary-sourcing of an already-held claim (type-F), not independent
corroboration (type-E) — so it firms the caveat’s warrant without lifting confidence: (stays
medium: the null converges with the existing “plausibly confounded” reading rather than adding a
new route, and its own power/linearity caveats keep it short of a confident “no effect”).
How much to believe it — the appraisal floor
- Design: overwhelmingly observational cohort; GRADE ~25% low, ~75% very low; AMSTAR median 5/11. Even the RCT meta-analyses graded low.
- Confounding by smoking is the dominant threat — coffee and smoking co-occur, so residual smoking confounding can manufacture an apparent harm (lung cancer, cancer mortality in smokers) or mask a benefit. The apparent harms «were largely nullified by adequate adjustment for smoking, except in pregnancy». This is exactly the referent/confounding machinery -> The U-Shaped Association Artifact.
- Mendelian randomisation gives no support for a causal effect, but the held mortality MR is underpowered (the T2D MR’s power is not held) — held for mortality, secondhand for T2D. Mortality (held, primary): no genetic causal evidence for coffee->mortality (Nordestgaard 2016 — The Mendelian-randomization check above); the instrument is strong (F=827) but the study is not powered to exclude a causal effect at the observational size (~225,000 cases needed), and by its linearity assumption «will not be capturing non-linear differences» — so it refutes neither a non-linear nor an observational-sized linear effect; it excludes only a much steeper one (corrected 2026-10-06: null holds at the observational effect size -> not powered to exclude it) (Nordestgaard & Nordestgaard, 2016). T2D (unheld, secondhand): Poole reports a separate coffee->T2D MR (its ref 122, an earlier Nordestgaard 2015 paper) as also finding no genetic evidence for a causal relation (Poole et al., 2017); its instrument strength and power are not held, so the F=827 / observational-size reading above does not transfer to it (corrected 2026-10-06: mortality and T2D MRs separated, Poole chunk 01). Net: treat the benefits as plausibly confounded associations pending an RCT, which is Poole’s own conclusion (liver disease the best RCT target).
- Measurement: no standard cup size; bean/roast/grind/brew all vary the dose, so cup-based exposure is coarse. Non-differential misclassification biases toward the null -> the true gradients could be steeper, not shallower -> Measurement Error in Dietary Assessment.
Decision summary (layer-1 framing)
- For most non-pregnant adults, coffee is not a big rock — it is a low-cost, likely-net-favourable or neutral habit, not a lever with a large certain effect. «More likely to benefit than harm» + a flat dose-response means there is no strong reason to start, increase, or quit for health within ~3-4 cups/day. Reporting that the lever is small and uncertain is itself the decision-change.
- The genuinely actionable, stratum-specific calls: (1) pregnancy / trying to conceive — limit (precautionary contraindication); (2) women at high fracture risk with low calcium — a small caution; Poole cites evidence that a little milk offsets caffeine’s effect on calcium absorption, which has not been tested against fracture; (3) high LDL / ASCVD risk drinking large volumes of unfiltered coffee — switch to filtered.
- The one large-magnitude, mechanistically-supported benefit worth an RCT is liver disease — but it is not yet a recommendation, only the best causal candidate.
Gaps (G) + attachment points for the cluster
- No coffee-sleep meta-analysis existed at review time (SR only) — so the caffeine/sleep-timing question the deliverable needs is a named gap here, not a null. AWAITS a coffee/caffeine-and-sleep MA (no specific source named yet — class-form gap, not a cashable hold) -> would connect to Sleep and Metabolic Health / Sleep Duration and Mortality.
- Respiratory outcomes, and the natural history of established disease (only 1 MA, post-MI): insufficient evidence.
- IARC 2016 (coffee removed from Group 2B “possibly carcinogenic”) is not cited by Poole — AWAITS an IARC coffee-monograph source (class-form gap — register a placeholder row if acquisition is ever pointed here) before any claim is written.
- Cluster attachment: Grosso 2016 has landed — it refines the mortality arm via the smoker/non-smoker referent split (the U-shape correction, section above) (Grosso et al., 2016); Ding 2014 has landed — it refines the monotone T2D gradient and the decaf/caffeine split (section below) (Ding et al., 2014). Both attach to this nucleus.
The T2D dose-response, refined — Ding 2014 IS the MA under Poole’s T2D headline type-F
Poole’s T2D row (high-vs-low 0.70 (0.65-0.75)) is a compressed summary of Ding’s gold-tier dose-response MA — 28 prospective cohorts, 1,109,272 participants, 45,335 incident cases, median 11-year follow-up. Poole cites Ding as the source of that row (its ref 21), and Ding prints the figure itself: «Compared with the lowest category (median consumption 0 cups/day), the pooled RR for incident type 2 diabetes for individuals in the highest category of consumption (5 cups/day) was 0.70 (95% CI 0.65–0.75, I2 = 50%, P for heterogeneity = 0.001).» (Ding et al., 2014) Superseded 2026-10-07: a 2026-10-06 note (Gate 52 cold re-audit) said the 0.70 does not print in Ding and might be Poole’s re-pooling; it does print, so Poole’s row is Ding’s own estimate. So this is F-refinement of the borrowed summary, not independent corroboration (Ding is inside Poole’s umbrella evidence base — shared cohorts, not a second route). Per the gold-gate rule, cite Ding, not the umbrella, for the T2D effect.
The granular gradient (cubic-spline, vs no coffee): RR 0.92 / 0.85 / 0.79 / 0.75 / 0.71 / 0.67 for 1-6 cups/day — «6 cups/day of coffee was associated with a 33% lower risk of type 2 diabetes». (Ding et al., 2014) The shape is monotone-decreasing, mildly concave (decreasing per-cup returns: -0.07, -0.06, -0.04, -0.04, -0.04), with no plateau or upper arm within 1-6 cups — unlike the mortality arm, whose pooled curve plateaus/rises. Nonlinearity was detected («A cubic spline model accounted for more variance in the outcome than did a linear model… suggesting that the association was not fully linear») (Ding et al., 2014) — so this shape is a genuine spline fit, not a single-coefficient display artifact -> Energy Adjustment and What a Diet Coefficient Means. Magnitudes are relative-only — Ding gives no absolute risk (a person’s absolute benefit scales with their baseline T2D risk: route (a)).
Caffeine is not the driver — the decaf datum, quantified. Per 1-cup/day increase, RR 0.91 (0.89-0.94) caffeinated vs 0.94 (0.91-0.98) decaffeinated, P for difference = 0.17 (NS) (Ding et al., 2014). Ding: «These results suggest that components of coffee other than caffeine are responsible for this putative beneficial effect» (Ding et al., 2014). Even the caffeine-alone association is confounded — «none of the included studies controlled for coffee intake when modeling caffeine intake», and it is «likely to be confounded by other components of coffee because of the collinearity» (Ding et al., 2014). Caveat kept: categorically the caffeinated arm is slightly stronger («P = 0.03 for the second highest group, P = 0.07 for the highest») (Ding et al., 2014) — decaf clearly works, caffeine may add a marginal increment. -> Is the Food Category Doing Any Work (the decaf test, T2D version).
The T2D benefit sits on firmer OBSERVATIONAL footing than the mortality benefit — but causality is still open. Two Ding features the mortality arm lacks: adjusted ≈ unadjusted spline («adjustment for potential confounders minimally affected effect estimates»), and the confounding direction runs toward the null — «higher coffee consumption was generally associated with a less healthy lifestyle… Thus, the true association between coffee and diabetes risk might be stronger than observed» (Ding et al., 2014). This is the opposite confounding direction from the mortality/cancer arm (where coffee-smoking correlation manufactures apparent effect — section above): for T2D the adverse-lifestyle correlation works against the observed benefit, so residual confounding is a weaker escape. But not a closed one — Ding: «it is difficult to establish the causality… solely based on observational evidence», and Poole’s Mendelian-randomisation citations found no genetic causal signal for coffee->T2D either. So: robust, dose-dependent, not-caffeine-driven association, biased if anything toward the null — but the MR-null keeps it a plausibly-confounded association pending an RCT, consistent with the whole page’s appraisal floor. Mechanism (chlorogenic acid on hepatic glucose output / intestinal glucose absorption; lignans, trigonelline) is in-vitro/animal only (inferred from Ding et al., 2014).
Brew method is uncontrolled here — Ding did not assess filtered vs unfiltered («most coffee is likely to be filtered»), so the T2D estimate pools brew methods, unlike the lipid outcome where the filtered/unfiltered split is load-bearing.