Is coffee good for you and what research says about its health effects
Coffee is associated with several favourable health outcomes in observational research, but current evidence does not establish that drinking it causes those benefits. Among 402,260 American adults followed for a median of 13.6 years, men drinking six or more cups a day appeared to die at 1.60 times the rate of non-drinkers; in the same men, after smoking and other habits entered the model, the figure was 0.90 (Freedman et al., 2012). Nothing about the coffee changed, only the arithmetic placed around it.
This explainer follows what the studies can and cannot distinguish: the shape of the pooled evidence, what happens when genes stand in for questionnaires, what randomised coffee trials have measured, how brewing method affects cholesterol, and why a cup is not a fixed exposure. Related work is gathered in the research digest.
The result that changed sign when smoking entered the model
The NIH-AARP cohort is the clearest demonstration in the field because it published both versions of its own answer. Among 229,119 men and 173,141 women aged 50 to 71, with 33,731 and 18,784 deaths recorded over 5.1 million person-years, the age-adjusted hazard ratios rose steadily with intake: 1.21 at four or five cups a day and 1.60 at six or more in men, 1.13 and 1.51 in women. After adjustment for smoking and other characteristics, the same categories read 0.88 and 0.90 in men, 0.84 and 0.85 in women (Freedman et al., 2012).
Coffee consumption and mortality before and after adjustment
Group | Coffee intake | Age-adjusted HR | Fully adjusted HR |
Men | 4–5 cups/day | 1.21 | 0.88 |
Men | 6+ cups/day | 1.60 | 0.90 |
Women | 4–5 cups/day | 1.13 | 0.84 |
Women | 6+ cups/day | 1.51 | 0.85 |
Note: these are hazard ratios from Freedman et al. (2012). HR = hazard ratio relative to non-coffee drinkers. Fully adjusted estimates account for smoking and other measured characteristics.
A sign change of that size is not a technicality. The change indicates substantial confounding, particularly by smoking, which was more common among coffee drinkers in the cohort. However, statistical adjustment cannot determine how much of the remaining association reflects coffee itself rather than other measured or unmeasured factors (Freedman et al., 2012). Every cohort in this literature faces the same problem, and the adjusted figure is only as good as the smoking history behind it.
Similar questions arise in the Seven Countries Study, which compared diet, blood cholesterol and coronary deaths across populations. Its observational associations also require care when distinguishing measured relationships from causal claims.
The umbrella review of the field shows what happens when that adjustment is incomplete. Pooled across studies, high coffee intake against low was associated with lung cancer at an odds ratio of 1.59. For any coffee against none, the estimate in people who had never smoked was 0.92 and not statistically significant, and 1.03 across studies that adjusted for smoking (Poole et al., 2017). The apparent harm was a shadow cast by tobacco, and the authors note that harmful associations outside pregnancy were largely nullified once smoking was properly handled.
That cuts both ways, which is the part usually left out. If inadequate adjustment can manufacture a harm, it can also manufacture a benefit when the confounders run the other way, and in later cohorts coffee drinkers tend to be employed, socially active and less likely to be in poor health at baseline.
What the pooled picture actually looks like
The largest synthesis gathered 201 meta-analyses of observational research covering 67 outcomes, alongside 17 meta-analyses of randomised trials covering nine. Against no coffee, three cups a day was associated with all-cause mortality of 0.83, cardiovascular mortality of 0.81 and incident stroke of 0.80. High intake against low was associated with type 2 diabetes at 0.70 and chronic liver disease at 0.35, while any coffee against none gave cirrhosis at 0.61 (Poole et al., 2017).
The breadth is the striking feature, and it is also the warning. A single beverage sits on the favourable side of almost every outcome anyone has examined, which is the signature of a population rather than of a drug. Pharmacological effects are usually narrow and come with matching harms; this pattern is broad, with few harmful associations outside pregnancy, and the same review graded roughly three quarters of its outcomes as very low quality, with a median methodological score of 5 out of 11 (Poole et al., 2017).
A drink that appears to protect against nearly everything is telling you about the people who drink it as much as about the drink.
The harms that did appear cluster in two places. In women, high intake was associated with fracture at 1.14 while in men it ran at 0.76, a split with no obvious mechanism. In pregnancy, high intake was associated with low birth weight at 1.31, pregnancy loss at 1.46 and childhood leukaemia at 1.57 (Poole et al., 2017).
A later cohort of 171,616 adults added a detail that complicates the clean reading. Unsweetened coffee showed the familiar U-shape against mortality, with hazard ratios between 0.71 and 0.84 across intake bands, but sugar-sweetened coffee produced much the same curve, and so did instant, ground and decaffeinated preparations (Liu et al., 2022). An effect that survives the removal of caffeine and the addition of sugar is difficult to pin on either.
What happens when genes stand in for questionnaires
Mendelian randomisation was built for exactly this impasse. Genetic variants are allocated at conception rather than by habit, so if coffee lowers mortality, the variants that raise coffee intake should lower mortality too. In 95,366 Danes the observational U-shape appeared as expected; a two-variant score raised coffee intake by 42%, and in 112,509 Danes hazard ratios per allele came back at 1.02 for ischaemic heart disease, 1.02 for cardiovascular mortality and 1.01 for all-cause mortality (Nordestgaard & Nordestgaard, 2016).
As a comparison, the researchers examined an ApoE variant associated with a 5% lower cholesterol level and an odds ratio of 0.93 for ischaemic heart disease (Nordestgaard & Nordestgaard, 2016). The coffee-related genetic analysis did not identify a comparable protective association. However, a positive result for cholesterol does not establish that genetic instruments for coffee consumption have sufficient power to detect effects of similar magnitude.
The same approach has been pointed at the pregnancy findings. Using coffee-related variants against roughly 50,000 miscarriages, 60,453 stillbirth records and birthweight data on more than 500,000 people, neither the two-sample analysis nor the genetic risk score showed any change in miscarriage, stillbirth, preterm birth or gestational age (Brito Nunes et al., 2023). Observational studies and the genetic analysis reached different conclusions, and neither shows that coffee is harmless in pregnancy. The 2023 findings sit alongside the observational evidence rather than above it (van Dam et al., 2020; Brito Nunes et al., 2023).
There is a catch worth naming, because it is rarely mentioned when these studies are reported as verdicts. The genome-wide analysis behind those instruments found eight loci explaining between 0.03 and 0.14 cups a day each, sitting in genes for caffeine handling and for signalling, and the alleles associated with drinking more coffee were also associated with smoking initiation, higher adiposity and higher fasting insulin, and in earlier work with lower blood pressure and more favourable lipid and liver profiles (Cornelis et al., 2015). An instrument entangled with smoking is an imperfect escape from a confounder that is smoking.
What a randomised coffee trial measured
In 2023, Marcus and colleagues randomised caffeinated coffee consumption during everyday life and measured several short-term physiological outcomes. A hundred adults wore a continuous electrocardiogram, an accelerometer and a glucose monitor for 14 days and were told by daily text message either to drink caffeinated coffee or to avoid caffeine, with adherence verified against dated receipts and geofencing of coffee shop visits (Marcus et al., 2023).
The primary outcome was negative. Premature atrial contractions averaged 58 a day on coffee against 53 on avoidance, a rate ratio of 1.09 with a confidence interval from 0.98 to 1.20. The secondary findings were not negative: premature ventricular contractions ran at 154 against 102, a rate ratio of 1.51; participants walked 1,058 more steps; and they slept 36 fewer minutes a night, 397 against 432. Serum glucose was indistinguishable (Marcus et al., 2023).
This trial illustrates the gap between short-term randomised measurements and the long-term health outcomes investigated in observational studies. Fourteen days of ectopic beats and sleep minutes are not mortality over 13 years. The umbrella review's arithmetic makes the imbalance plain: 201 pooled observational analyses against 17 pooled trial analyses, none of the latter reaching a hard endpoint (Poole et al., 2017).
Why brewing method matters for cholesterol
Evidence on coffee preparation is more direct for serum cholesterol than for many long-term disease outcomes, with experimental findings showing clear differences between filtered and unfiltered preparations. A meta-analysis of 14 randomised trials found a dose-response relation between coffee and both total and LDL cholesterol, with the increases concentrated in trials using boiled or otherwise unfiltered coffee; trials using filtered coffee showed very little change (Jee et al., 2001).
Paper filters retain much of the cholesterol-raising diterpenes, making brewing method an important determinant of exposure. However, the amount consumed also matters, as the trials identified a dose-response relationship between coffee intake and changes in serum cholesterol (Jee et al., 2001). Where the bean comes from, and how it is handled before roasting, is covered in coffee, cocoa and tea cultivation.
A Norwegian cohort of 508,747 adults followed for an average of 20 years, with 46,341 deaths, carried that chemistry through to outcomes. Against no coffee, filtered brew was associated with all-cause mortality of 0.85 in both men and women, while unfiltered brew came in at 0.96 in men, not statistically significant, and 0.91 in women. Among men aged 60 and over, unfiltered brew against no coffee gave cardiovascular mortality of 1.19, with a confidence interval from 1.00 to 1.41 that leaves it at the edge of significance (Tverdal et al., 2020).
The telling analytic detail is what happened to those estimates when total cholesterol was dropped from the model: the hazard ratios for cardiovascular and ischaemic heart disease mortality rose, most sharply among people drinking nine or more cups of unfiltered coffee a day. The pattern fits cholesterol contributing to the association, though dropping a variable cannot establish mediation. Serum cholesterol has its own long measurement history, set out in the 1965 Hegsted study on dietary fat and serum cholesterol.
The evidence supports an effect of brewing method on serum cholesterol, while the Norwegian cohort provides observational evidence that mortality associations differ by preparation. Taken together, these findings make filtration relevant to interpreting coffee and cardiovascular health, but they do not prove that the cholesterol changes caused the mortality differences (Jee et al., 2001; Tverdal et al., 2020).
Why a cup is not a fixed exposure
Every cohort measures intake in cups, and the cup is a poor instrument. When specialty coffees were assayed by gas chromatography, brewed samples averaged 188 milligrams of caffeine for a 16-ounce serving, with a range across caffeinated coffees from 58 to 259 milligrams. The same beverage bought from the same outlet on six consecutive days measured between 259 and 564 milligrams (McCusker et al., 2003). A twofold variation inside one product is larger than the differences between the intake categories that cohorts compare.
The person receiving that caffeine varies as well. In a Costa Rican study of 2,014 heart attack cases matched to 2,014 controls, carriers of the slow-metabolising CYP1A2 variant showed odds ratios of 1.36 at two to three cups a day and 1.64 at four or more, while people with the rapid genotype showed 0.78 and 0.99, with a gene by coffee interaction at P = 0.04. Below the median age of 59 the slow metabolisers reached 2.33 at four or more cups (Cornelis et al., 2006). Averaging those two groups together produces a number that describes neither.
Timescale matters too. Pooled trials in people with raised blood pressure found that 200 to 300 milligrams of caffeine lifted systolic pressure by 8.1 mmHg and diastolic by 5.7 mmHg within the first hour, with the rise lasting at least three hours. Over two weeks, no increase remained against a caffeine-free diet or against decaffeinated coffee, and seven cohorts found no higher cardiovascular risk in habitual drinkers with hypertension (Mesas et al., 2011). An acute effect and a chronic effect answer different questions.
One association has largely escaped the smoking objection by design. Over 30 years of follow-up in 8,004 Japanese-American men, the incidence of Parkinson disease fell from 10.4 per 10,000 person-years in those drinking no coffee to 1.9 in the highest intake group, with the trend holding separately among men who had never smoked and tracking total dietary caffeine rather than other components of coffee (Ross et al., 2000).
What the evidence supports and what it does not
Supported: the brewing method affects serum cholesterol, with trials showing greater increases after unfiltered coffee than after filtered preparations (Jee et al., 2001). Caffeine can raise blood pressure in the short term, whereas the rise is not sustained with habitual intake in the studies reviewed by Mesas et al. (2011). Large observational syntheses also report favourable associations for several health outcomes, although these findings do not establish causation (Poole et al., 2017).
Not supported: that drinking more coffee causes the lower mortality seen in cohorts. The Mendelian-randomisation analysis did not identify a protective association with mortality, but its null findings do not conclusively rule out a causal effect of coffee consumption (Nordestgaard & Nordestgaard, 2016). The review that mapped the literature concluded that current evidence does not justify recommending coffee intake for the prevention of disease, while moderate consumption fits a reasonable diet for most adults who are not pregnant (van Dam et al., 2020).
Unsettled: whether the association describes the drink or the drinker, and why preparations without caffeine produce a similar curve (Liu et al., 2022). The pregnancy question remains genuinely open, with observational harms on one side and a null genetic analysis on the other (Brito Nunes et al., 2023).
The most defensible conclusion is that observational research generally links coffee consumption with favourable long-term outcomes, but causality remains uncertain. Randomised trials have nevertheless demonstrated specific short-term effects on sleep, cardiac ectopy and serum cholesterol, with results depending on caffeine exposure and brewing method (Marcus et al., 2023; Jee et al., 2001).
Frequently asked questions
Why do coffee studies so often contradict each other?
Because the result depends on what else is in the model. One cohort reported hazard ratios rising to 1.60 before adjustment and falling to 0.90 after it, in the same participants. Headlines usually quote one version without the other, so two accurate reports can appear to disagree completely. (Freedman et al., 2012)
Is decaffeinated coffee associated with the same lower risk?
Largely yes, which is awkward for any caffeine-based explanation. In a cohort of 171,616 adults the U-shaped curve against mortality appeared for instant, ground and decaffeinated preparations alike. If removing the active compound leaves the association intact, the association may not be about the compound. (Liu et al., 2022)
Does coffee raise blood pressure?
Acutely, yes. Pooled trials in people with hypertension recorded a systolic rise of 8.1 mmHg within an hour, lasting at least three hours. Over a fortnight of regular intake that difference disappeared against both a caffeine-free diet and decaffeinated coffee, which is consistent with tolerance developing over that period. (Mesas et al., 2011)
Is there a number of cups the evidence points to?
The pooled curve bottoms out near three cups a day, at 0.83 against none, but that is a description of where risk was lowest among people who chose their intake. It is not evidence that moving to three cups changes anything, and the genetic analyses did not detect such an effect (Poole et al., 2017; Nordestgaard and Nordestgaard, 2016).
Does coffee cause extra heartbeats?
The randomised trial found no significant increase in premature atrial contractions, its primary outcome. Premature ventricular contractions did rise, 154 a day against 102. Participants also walked further and slept 36 minutes less, so the trial measured several things at once over 14 days. (Marcus et al., 2023)
Sources
- Freedman ND, Park Y, Abnet CC, Hollenbeck AR, Sinha R. Association of coffee drinking with total and cause-specific mortality. The New England Journal of Medicine. 2012;366(20):1891–1904. DOI 10.1056/NEJMoa1112010.
- Poole R, Kennedy OJ, Roderick P, Fallowfield JA, Hayes PC, Parkes J. Coffee consumption and health: umbrella review of meta-analyses of multiple health outcomes. BMJ. 2017;359:j5024. DOI 10.1136/bmj.j5024.
- van Dam RM, Hu FB, Willett WC. Coffee, caffeine, and health. The New England Journal of Medicine. 2020;383(4):369–378. DOI 10.1056/NEJMra1816604.
- Marcus GM, Rosenthal DG, Nah G, et al. Acute effects of coffee consumption on health among ambulatory adults. The New England Journal of Medicine. 2023;388(12):1092–1100. DOI 10.1056/NEJMoa2204737.
- Cornelis MC, El-Sohemy A, Kabagambe EK, Campos H. Coffee, CYP1A2 genotype, and risk of myocardial infarction. JAMA. 2006;295(10):1135–1141. DOI 10.1001/jama.295.10.1135.
- Nordestgaard AT, Nordestgaard BG. Coffee intake, cardiovascular disease and all-cause mortality: observational and Mendelian randomization analyses in 95,000–223,000 individuals. International Journal of Epidemiology. 2016;45(6):1938–1952. DOI 10.1093/ije/dyw325.
- Tverdal A, Selmer R, Cohen JM, Thelle DS. Coffee consumption and mortality from cardiovascular diseases and total mortality: does the brewing method matter? European Journal of Preventive Cardiology. 2020;27(18):1986–1993. DOI 10.1177/2047487320914443.
- Jee SH, He J, Appel LJ, Whelton PK, Suh I, Klag MJ. Coffee consumption and serum lipids: a meta-analysis of randomized controlled clinical trials. American Journal of Epidemiology. 2001;153(4):353–362. DOI 10.1093/aje/153.4.353.
- McCusker RR, Goldberger BA, Cone EJ. Caffeine content of specialty coffees. Journal of Analytical Toxicology. 2003;27(7):520–522. DOI 10.1093/jat/27.7.520.
- Liu D, Li ZH, Shen D, et al. Association of sugar-sweetened, artificially sweetened, and unsweetened coffee consumption with all-cause and cause-specific mortality: a large prospective cohort study. Annals of Internal Medicine. 2022;175(7):909–917. DOI 10.7326/M21-2977.
- Ross GW, Abbott RD, Petrovitch H, et al. Association of coffee and caffeine intake with the risk of Parkinson disease. JAMA. 2000;283(20):2674–2679. DOI 10.1001/jama.283.20.2674.
- Mesas AE, Leon-Munoz LM, Rodriguez-Artalejo F, Lopez-Garcia E. The effect of coffee on blood pressure and cardiovascular disease in hypertensive individuals: a systematic review and meta-analysis. The American Journal of Clinical Nutrition. 2011;94(4):1113–1126. DOI 10.3945/ajcn.111.016667.
- Coffee and Caffeine Genetics Consortium, Cornelis MC, Byrne EM, Esko T, et al. Genome-wide meta-analysis identifies six novel loci associated with habitual coffee consumption. Molecular Psychiatry. 2015;20(5):647–656. DOI 10.1038/mp.2014.107.
- Brito Nunes C, Huang P, Wang G, et al. Mendelian randomization study of maternal coffee consumption and its influence on birthweight, stillbirth, miscarriage, gestational age and pre-term birth. International Journal of Epidemiology. 2023;52(1):165–177. DOI 10.1093/ije/dyac121.
Related
Coffee, Cocoa and Tea: Cultivation, Management and Production
Dietary Fat and Serum Cholesterol: Hegsted Study Findings
The Seven Countries Study: Diet and Heart Deaths in 1980
How much water you should drink each day and why needs vary
Impostor syndrome and why some people doubt their success
Travel in Later Life: What an Eight-Person Study Found
Adult Literacy Classes in Kenya: What a 1980s Study Found