Supplements · Metabolic & Cardiometabolic
coffee decreases type 2 diabetes risk
In plain terms: Does coffee lower your risk of type 2 diabetes?
Strong support Supplements 🔬 Includes disconfirming
Part of: • Coffee
RefutedContestedStrong support
consensus score 0.89
Most likely yes, a little. Decades of large studies consistently link regular coffee to lower diabetes risk (roughly 6% less per daily cup), and decaf works too, so it isn't just caffeine. It's observational evidence, so it can't prove cause, but it's one of the most reliable coffee findings we have.
📅 Last reviewed: 2026-07-15 ⓘ
Evidence ladder
How far up the ladder this claim has climbed. A high consensus on a low rung means "consistent so far," not "proven in people."
Top evidence so far: All trials, pooled (Meta-analysis)
MechanismIn-vitroAnimalObservationalRCTMeta-analysis
How the studies fall
12 support 1 contradict 0 tested null 1 mixed · 14 sources, 12 independent groups
What the evidence shows
Habitual coffee drinking is consistently linked to a lower risk of developing type 2 diabetes — one of the most reproducible findings in nutritional epidemiology. Many independent dose-response meta-analyses (pooling millions of people) agree, with roughly 6% lower risk per daily cup, and decaffeinated coffee shows a similar effect, so it is not simply the caffeine.
The evidence (17)
| Source | Grade | Stance | Quality | Finding |
|---|---|---|---|---|
| Huxley R et al. 2009 · Arch Intern Med | meta-analysis | supports | moderate | Meta-analysis of 18 PROSPECTIVE OBSERVATIONAL cohorts, 457,922 participants: 'every additional cup of coffee consumed in a day was associated with a 7% reduction in the excess risk of diabetes relative risk, 0.93 [95% confidence interval, 0.91-0.95]) after adjustment for potential confounders.' Authors caveat: 'Owing to the presence of small-study bias, our results may represent an overestimate of the true magnitude of the association.' No randomised data pooled. |
| Ding Ping et al. 2024 · Journal of Diabetes and Metabolic Disorders | meta-analysis | supports | low | Incidence exposure is POOLED, not coffee: 'The results showed that HR was 0.89 (95%CI,0.81-0.98; P = 0.016; Fig. 3 ) for tea or coffee', I2 = 87.2%, 13 observational studies. The authors separated the two beverages when they could - for MORTALITY they report '0.75 (95% CI, 0.65-0.87; P < 0.001; Fig. 2 ) for coffee, 0.84 (95% Cl, 0.75-0.94; P = 0.002; Fig. 2 ) for tea' - so the combined incidence category reflects the available data, not a coffee estimate. Off-scope for a coffee-specific incidence claim. |
| Su X, Hu D, Tan J 2026 · International Journal of Cardiology. Cardiovascular Risk and Prevention | observational | supports | low | NHANES 2007-2018 CROSS-SECTIONAL, 6311 US women aged 20+ with a pregnancy history (1313 with diabetes mellitus, 574 with GDM). Outcome is PREVALENT diabetes mellitus of ALL types - self-report of 'yes' or 'borderline', or diabetes medication, or HbA1c >=6.5% / FPG >=126 mg/dL / 2hPG >=200 mg/dL - not incident T2D: 'the absence of detailed diagnostic information prevented differentiation between diabetes subtypes (type 1 diabetes, type 2 diabetes, latent autoimmune diabetes in adults, etc.), potentially introducing misclassification bias', though the authors argue T2D dominates. TOTAL coffee, the claim's own subject, did NOT survive adjustment: >3 cups/d OR 0.74 (0.57-0.94, P=0.015) before full adjustment, then 'this association became marginally non-significant after full covariates adjustment (OR = 0.76; 95% CI: 0.57-1.02; P = 0.069)'. The positive result is the CAFFEINATED-coffee stratum: adjusted ORs vs non-consumers of 0.80 (0.66-0.99), 0.73 (0.60-0.88), 0.86 (0.69-1.07, P=0.171, null) and 0.74 (0.57-0.96) across <=1 / 1-2 / 2-3 / >3 cups/d - NON-MONOTONE, with the 2-3 cup band null; the significant dose-response (P<0.001) belongs to the caffeine-QUINTILE analysis (Q5 >=240 mg/d OR 0.72, 0.57-0.90), a different exposure variable, not to the cup categories. 'No significant associations were observed for decaffeinated coffee' (5372 of 6311 reported no decaf at all, so that null is thinly powered). No association with GDM (574 cases, recalled diagnosis from a past pregnancy while exposure was measured today). Quality low on the authors' own account: 'as this is a cross-sectional study, it is inherently limited in its ability to infer causality and remains susceptible to reverse causality', plus self-reported two-day dietary recall and no brewing-method data. |
| Kwapien E et al 2025 · Cureus | observational | supports | low | Narrative review, no pooling and no new computation of its own ('no new statistical computations were performed for this article'). Its entire T2DM statement: 'both coffee and caffeine intake are significantly inversely associated with the incidence of T2DM', with the correlation 'more pronounced in women than in men' - no risk estimate, no cup threshold, no J-shape is given for diabetes. (The 3-4 cups/day figure and the J-shaped curve belong to the paper's mortality/cardiovascular section; Table 1's '>3 cups/day' for the metabolic domain traces to a fatty-liver citation.) Zero weight: its cited primaries already vote in the vault. |
| Di Maso M et al. 2021 · Adv Nutr | meta-analysis | supports | high | Dose-response meta-analysis of prospective cohorts (26 studies / 42 cohorts / 3,713,932 participants pooled across six outcomes; T2D-specific n not stated in the abstract): 'In any coffee consumers, there was a significant inverse association with the risk of ... T2D (RR = 0.90; 95% CI: 0.85, 0.96)'. Risk decreased linearly across the whole caffeinated-coffee range for T2D (nonlinearity was significant only for CVD). |
| Yu Q et al 2026 · J Diabetes Investig | observational | contradicts | moderate | Two-sample Mendelian randomization (IVW primary; exposures from published GWAS, outcomes from FinnGen 2024, European ancestry only). COFFEE and CAFFEINE are two SEPARATE exposures here, with separate instruments, and only the coffee arm measures this claim's subject. Coffee consumption (UK Biobank GWAS, n=375,833, 15 SNPs, unit = 50% change in consumption) vs type 2 diabetes (FinnGen, 71,728 cases / 369,007 controls): 'coffee consumption (IVW OR, 1.01; 95% CI: 1.00-1.02; adjusted P-value <0.001) ... associated with increased odds of type 2 diabetes' - the WRONG direction for this claim, though at a magnitude indistinguishable from null (lower CI bound 1.00 on a per-50%-change unit), and the authors caution MR estimates 'are better suited for testing causality than quantifying precise effect sizes'. Contradicts on both routes: wrong-signed point estimate, and a null on a directional claim is contradicts. The protective OR 0.72 (0.64-0.81) is the CAFFEINE exposure - a different GWAS (n=9,876, only 2 SNPs), not coffee - and so votes on a caffeine claim, not this one; the claim page itself separates the two (decaf shows the same effect). Sensitivity statistics (Egger intercept, Q, MR-PRESSO) are reported in the text only for the smoking/neovascular-glaucoma pair; for coffee-T2D they sit in supplementary tables - cannot-tell from available text. Prior extract borrowed the caffeine estimate to describe 'diabetes-related outcomes' and never gave the coffee number. |
| Grosso G, Godos J, Galvano F, Giovannucci EL 2017 · Annu Rev Nutr | meta-analysis | supports | high | 'we performed an umbrella review of the evidence from meta-analyses of observational studies and randomized controlled trials (RCTs)' — 59 unique outcomes across 112 meta-analyses of observational studies. Coffee showed 'a probable decreased risk of ... type-2 diabetes'. Consistent with the claim, but a credibility re-grading of existing meta-analyses; the underlying pooled estimates belong to those meta-analyses, several of which the vault appraises separately. |
| Ding M et al. 2014 · Diabetes Care | meta-analysis | supports | high | Dose-response meta-analysis of 28 prospective studies (1,109,272 participants, 45,335 cases). Spline vs no coffee: RR 0.92 (0.90-0.94) at 1 cup/day falling to 0.67 (0.61-0.74) at 6 cups/day (27 studies). Per 1 cup/day increase, in the 11-study caffeinated-vs-decaf subset: RR 0.91 (0.89-0.94) caffeinated, 0.94 (0.91-0.98) decaf, P for difference 0.17. Observational: 'a causal relationship cannot be established with these data alone.' |
| Bae JM 2021 · World J Diabetes | meta-analysis | supports | moderate | Asian populations only (7 prospective cohorts, 12 sex strata; 6348 T2DM cases in 141813 participants; Japan x4, Singapore, Taiwan, Korea). Own fixed-effect pooling of primary cohort RRs, not an umbrella review. Highest vs lowest coffee category: 'The risk of T2DM in Asian populations was significantly reduced in the highest coffee dose group compared to the lowest dose group (sRR = 0.73, 95%CI: 0.66-0.82;' I-squared 0.0%. Dose-response: 'drinking a cup of coffee per day reduced the risk of T2DM in Asian populations by approximately 8% (RR = 0.92, 95%CI: 0.90-0.95) with a linear relationship'. Effect held across subgroups by sex and by level of adjustment for smoking, alcohol and family history. Authors detect publication bias (Egger P = 0.01); after trim-and-fill the estimate holds but loses precision - 'Furthermore, coffee consumption still prevented T2DM in Asian populations (sRR = 0.73, 95%CI: 0.54-0.91;'. Author-stated limit: exposure self-reported once at baseline with differing categories and reference groups across cohorts. |
| Lin WY et al. 2011 · Eur J Clin Invest | observational | supports | low | Taichung Community Health Study, Taiwan: cross-sectional analysis of 2,332 adults aged 40+ (T2DM prevalence 14.0% men / 10.4% women). Adjusted ORs for PREVALENT T2DM vs non-drinkers were 0.77(0.52-1.13) at under 1/wk, 0.46(0.28-0.76) at 1-6/wk and 0.37(0.16-0.83) at 7 or more times/wk (Results text: 0.76 / 0.45 / 0.36), p<0.001 for trend; the 0.37 figure is the top-frequency stratum alone, not habitual drinkers as a group, who were OR 0.62(0.41-0.92) in men and 0.54(0.31-0.92) in women. Authors: 'First, the cross-sectional design does not clarify causality. Future prospective cohort studies are necessary to establish causative links.' Exposure was FFQ frequency per week, not cups per day; decaffeinated coffee was not assessed, coffee type 'was not avaliable', and 'Most coffee drinkers added sugar and milk (or cream) into their coffee, so separate analyses were not conducted to assess the influence of coffee additives.' |
| Carlström M, Larsson SC 2018 · Nutr Rev | meta-analysis | supports | high | Meta-analysis of 30 prospective cohort / nested case-control studies, 1,185,210 participants and 53,018 incident T2D cases (PubMed searched to 1 Dec 2017): highest coffee category (median 5 cups/d) vs lowest (median 0 cups/d) RR 0.71 (95%CI 0.67-0.76); risk fell 6% per additional cup/day (RR 0.94, 95%CI 0.93-0.95). Similar for caffeinated (RR 0.93, 95%CI 0.90-0.96) and decaffeinated (RR 0.94, 95%CI 0.90-0.98) per cup/day. All pooled inputs are observational; authors state coffee is 'inversely associated with' T2D risk, not causal. Abstract-grade read. |
| van Dam RM, Willett WC, Manson JE, Hu FB 2006 · Diabetes Care | observational | supports | high | Nurses' Health Study II prospective cohort: 88,259 U.S. women aged 26-46 without diabetes at baseline, coffee assessed 1991/1995/1999, '1,263 incident cases of confirmed type 2 diabetes between 1991 and 2001'. Adjusted RR vs nondrinkers: 0.87 (0.73-1.03) at 1 cup/d, 0.58 (0.49-0.68) at 2-3 cups/d, 0.53 (0.41-0.68) at >=4 cups/d, P for trend <0.0001 - the 0.53 is the >=4 cups/d STRATUM alone, not habitual drinkers as a group. The caffeinated-vs-decaf comparison is a DIFFERENT contrast, a per-one-cup increment rather than the >=4 cups/d category: caffeinated 0.87 (0.83-0.91) and decaffeinated 0.81 (0.73-0.90) per cup/day; filtered 0.86 (0.82-0.90), instant 0.83 (0.74-0.93). Tea was not associated (0.88, 0.64-1.23, for >=4 vs 0 cups/d; P for trend 0.81). Authors conclude coffee 'may lower risk of type 2 diabetes in younger and middle-aged women' - observational, association not causation, single occupational cohort of young women. Abstract-grade. |
| Sevilla-González M et al 2026 · medRxiv | observational | supports | moderate | medRxiv PREPRINT (posted 2026-07-08), NOT peer reviewed; abstract-grade read via the Crossref-deposited abstract (doi.org resolved to medrxiv.org but returned HTTP 403). Two separate analyses, and only ONE of them measured this claim's outcome. The T2D vote is the UK Biobank arm alone: 'Prospective analyses among 333,053 UK Biobank participants without T2D at baseline who had dietary and genetic data and were followed for a median of 13.3 years.' 'In UK Biobank, higher coffee intake was associated with lower T2D incidence (hazard ratio per cup/day, 0.96; 95% CI, 0.95-0.97), lower triglyceride-to-HDL cholesterol ratio (β,−0.01; P = 2.51 × 10^-19), and lower visceral adipose tissue mass (β, −0.01; P = 4.28 × 10^-9).' The number of incident T2D cases is not reported in the abstract (cannot-tell, not asserted absent). The VITAL arm is a DIFFERENT denominator and a DIFFERENT endpoint - 'Cross-sectional analyses among 806 participants without T2D in the VITamin D and OmegA-3 TriaL (VITAL) clinical sub-cohort, who underwent repeated dietary assessment, clinical phenotyping, and dual-energy X-ray absorptiometry imaging at baseline and year-2.' - and it did not measure T2D at all: 'The primary outcomes, in VITAL, were HbA1c, oral glucose tolerance test-derived measures of glucose response and insulin sensitivity, β-cell function, and overall, truncal, and visceral adiposity; in UK Biobank, was incident T2D.' Its findings ('In VITAL, higher coffee intake was associated with higher insulin sensitivity (standardized β per cup/day, 0.046; P = .004) and lower visceral adipose tissue mass (β, −0.047; P = .006), after adjusting for demographic, lifestyle, and clinical factors, including body mass index.') are mechanism-adjacent, not a second vote on T2D risk. Exposure: 'Coffee intake assessed by food frequency questionnaires.' Observational throughout; the authors write 'associated with', not causal. INDEPENDENCE FLAG (not acted on): the author block includes Frank B. Hu, senior author of the cof-ding-2014 meta-analysis, and JoAnn E. Manson, co-author of cof-salazar-2004, so one Boston/Harvard network sits behind at least three separately-grouped rows on this claim; UK Biobank may also overlap the coffee-consumption instrument behind the s42046344 Mendelian-randomization row. |
| Osama H, Abdelrahman MA, Madney YM, Harb HS, Saeed H, Abdelrahim MEA 2021 · Int J Clin Pract | meta-analysis | supports | low | Systematic review + meta-analysis that pools primary studies itself (not an umbrella). The T2D estimate rests on a 31-article subset of a 69-study corpus, not on all 69: 'A total of 69 cross-sectional and cohort studies were included and divided as follows: 31 articles for T2D risk, 15 studies for adiponectin, 6 studies for leptin, 12 studies for CRP and 5 studies for IL-6.' Highest vs lowest coffee: 'Overall, coffee consumption was inversely associated with T2D risk with an estimated pooled RR of 0.73 (95% confidence interval [0.68, 0.80] for the highest vs lowest coffee consumption categories.' Participant and case counts are not reported in the abstract. Quality is low on the paper's own account of its methods: the pool mixes cross-sectional with cohort designs and treats 'odds ratios (OR) or relative risk (RR)' as interchangeable effect sizes, so prevalence ORs open to reverse causation sit beside incidence RRs; and the results sentence writes a non-significant pooled CRP result as a direction of effect - 'lower level of leptin (P = .04) and CRP (P = .2)'. Heterogeneity, publication-bias testing and the design mix of the 31 T2D articles are not reported in the abstract (cannot-tell, not asserted absent). The pooled 0.73 sits on largely the same primary cohorts as Jiang 2014 (0.71), Carlstrom 2018 (0.71) and Bae 2021 (0.73), so it is a re-pooling rather than an independent observation. |
| Salazar-Martinez E, Willett WC, Ascherio A, Manson JE, Leitzmann MF, Stampfer MJ, Hu FB 2004 · Ann Intern Med | observational | supports | high | Nurses' Health Study (84,276 women, 1980-1998) + Health Professionals' Follow-up Study (41,934 men, 1986-1998), all free of diabetes, cancer and CVD at baseline; coffee 'assessed every 2 to 4 years through validated questionnaires'; 1,333 incident T2D cases in men and 4,085 in women. Multivariate RRs across categories (0, <1, 1 to 3, 4 to 5, >=6 cups/d) were 1.00, 0.98, 0.93, 0.71, 0.46 (95% CI 0.26-0.82; P = 0.007 for trend) in men and 1.00, 1.16, 0.99, 0.70, 0.71 (CI 0.56-0.89; P < 0.001 for trend) in women. The 0.46 / 0.71 figures are the >=6 cups/d stratum ALONE, not habitual drinkers as a group: the adjacent 4-5 cups/d band was 0.71 (men) / 0.70 (women), and in women the gradient is not monotone below 4 cups/d (1.16 at <1 cup/d). Decaffeinated coffee, >=4 cups/d vs nondrinkers: 0.74 (CI 0.48-1.12) men and 0.85 (CI 0.61-1.17) women - both non-significant in this cohort. Authors conclude coffee is 'associated with' lower risk; observational, abstract-grade. |
| Greenberg JA, Axen KV, Schnoll R, Boozer CN 2005 · Int J Obes (Lond) | observational | mixed | moderate | NHEFS cohort, 7006 adults, 8.4 y, 301 self-reported diabetes cases + 8 diabetes deaths. In the <=60-y stratum (the only one showing an effect), a 2 cups/d increment gave HR 0.86 (0.75-0.99) for ground-caffeinated, 0.58 (0.34-0.99) for ground-decaffeinated coffee and 0.77 (0.59-1.00) for regular tea; instant coffee (caffeinated or decaf) and herbal tea showed no significant association. Authors' own conclusion: the inverse relationship 'actually only applied to nonelderly adults who had previously lost weight'. |
| Jiang X et al. 2014 · Eur J Nutr | meta-analysis | supports | high | Dose-response meta-analysis of prospective cohorts: highest vs lowest coffee intake RR 0.71 (95% CI 0.67-0.76), from the 26-article coffee subset (50,595 T2DM cases, 1,096,647 participants); T2DM incidence decreased by 12% [0.88 (0.86-0.90)] per 2 cups/day increment. Decaffeinated coffee showed 0.79 (0.69-0.91), so the signal is not caffeine alone. |
Disagree, or know a study we missed?
We grade by evidence, not opinions. The way to weigh in is to point us to a study we haven't cited (check the evidence table above first), or to flag a problem with one we have. Every submission is reviewed; if it holds up, the grade updates and shows in Science Changes Its Mind.
Opens a short form. You'll sign in with Google so submissions are tied to a real account — we don't display your identity, and we only accept a link we can verify (PubMed, DOI, ClinicalTrials.gov).
Educational only, not medical advice. Grades and scores reflect published evidence weighted by study design and quality; see the methodology.