Most coffee research treats coffee in isolation, which is a slightly artificial way to study a habit. Real people drink coffee and tea, often on the same day, and the compounds in each overlap only partly. A UK Biobank analysis published in 2021 is one of the few large studies to look at both together, and it produced the most specific dose answer available in this whole area, along with a shape of relationship that is easy to misreport.
The study
Zhang and colleagues, publishing in PLOS Medicine[1], followed 365,682 UK Biobank participants aged 50 to 74 who joined between 2006 and 2010, tracking them until 2020. Over a median follow-up of 11.4 years, 5,079 participants developed dementia and 10,053 had a stroke.
Their statistical models adjusted for an unusually long list of potential confounders: sex, age, ethnicity, qualifications, income, body mass index, physical activity, alcohol status, smoking status, overall diet pattern, consumption of sugar-sweetened beverages, HDL and LDL cholesterol, and histories of cancer, diabetes, coronary artery disease and hypertension. That is a thorough adjustment set, and it matters, because tea and coffee preferences in the UK correlate with class, income and lifestyle in ways that would otherwise contaminate the result badly.
The numbers
Compared with people who drank neither, those drinking 2 to 3 cups of coffee and 2 to 3 cups of tea per day had a 32% lower risk of stroke, a hazard ratio of 0.68, and a 28% lower risk of dementia, a hazard ratio of 0.72. Both were statistically significant.
The combination also outperformed either drink on its own, and was associated with lower risk of ischaemic stroke and of vascular dementia specifically. A separate finding concerned dementia following a stroke, where the lowest risk appeared at a daily consumption of 3 to 6 cups of coffee and tea combined, a hazard ratio of 0.52, though this was based on a much smaller number of cases and the confidence interval was correspondingly wide.
The shape of the curve is the real finding
The most important result in this paper is not any single hazard ratio but the shape of the relationship. The associations of coffee and tea with stroke and dementia were explicitly non-linear, with statistical tests for non-linearity coming out significant.
In practice that means the benefit does not accumulate with each additional cup. Risk fell as intake rose from zero, reached a minimum, and then rose again at higher intakes. The lowest-risk points were coffee at 2 to 3 cups a day, tea at 3 to 5 cups a day, or a combination of 4 to 6 cups a day. Above that range the curve turned back upward.
This is a U-shaped or J-shaped relationship, and it is the single most commonly mangled feature of nutrition research in popular coverage. A headline saying that coffee reduces dementia risk implies that more is better. This study found the opposite above a threshold. If you take one thing from this article, it should be that the research supports a moderate, specific intake rather than a maximal one.
Note also that this shape does not appear in every study of coffee and brain outcomes. The Parkinson's research covered in our article on coffee and Parkinson's found a monotonic gradient with no upturn at high intakes, and the mortality analysis by Loftfield and colleagues[2] found the association persisting even at 8 or more cups a day. Different outcomes, different curves.
What might explain it
Both coffee and tea contain caffeine, and both are rich in polyphenols, though of different families: chlorogenic acids dominate in coffee, catechins and theaflavins in tea. Because the combination outperformed either drink alone, and because the two share caffeine but differ in polyphenols, the pattern is at least consistent with the polyphenols contributing something additive.
It is also worth noting that this study asked a question most coffee research does not. Nearly all of the literature treats coffee as if it were the only thing a person drinks, which is obviously false, and tea in particular overlaps with coffee both in caffeine and in the polyphenol families of interest. Studying them separately and in combination is methodologically better, and it is why the combination result carries some weight.
The mechanistic candidates split accordingly. Caffeine, shared by both, acts on adenosine receptors and on cerebral blood flow. The polyphenols differ: coffee's chlorogenic acids and tea's catechins and theaflavins are chemically distinct, though both have been studied for effects on vascular function and oxidative stress. If the combination genuinely outperforms either drink alone, the polyphenol families each adding something the other lacks is the most economical explanation, and it is an untested one.
The stroke result is the sturdier half
Stroke and dementia are bundled together in this paper's headline, but the two findings do not deserve equal confidence. Stroke is an acute, well-defined, well-recorded event with a clear date and a clear diagnosis. Dementia is a slow, difficult-to-date syndrome that is under-diagnosed in registry data and frequently recorded years after onset.
That difference matters for reverse causation. A stroke in year eight of follow-up is unlikely to have altered someone's coffee habits in year one. A dementia diagnosis in year eight may well reflect a process already underway at enrolment, and early cognitive change can alter routines, appetite and shopping habits long before anyone notices. The 10,053 strokes are also roughly double the 5,079 dementia cases, giving that half of the analysis more statistical power.
None of this makes the dementia finding wrong. It does mean that if you were going to trust one of the two numbers more, it should be the stroke one.
The finding that the association was specifically strong for ischaemic stroke and vascular dementia is mechanistically suggestive too. Both of those are vascular conditions, driven by blood vessel health rather than by the amyloid and tau pathology of Alzheimer's disease. That points toward a cardiovascular pathway, which would connect this result to the broader cardiovascular associations described in our article on coffee and the heart. This is a coherent story, and it remains a story rather than a demonstrated mechanism.
The limitations, including the ones the authors named
The researchers listed three limitations themselves, and they are substantial. Coffee and tea intake was self-reported at baseline only, so it may not reflect what people drank across the following decade. Unmeasured confounding can still bias effect estimates in any observational study, no matter how long the adjustment list. And UK Biobank participants are not representative of the wider UK population: they are healthier, wealthier and more educated than average, a well-documented issue known as healthy volunteer bias.
Healthy volunteer bias deserves a sentence of its own, because it is specific and well quantified in this cohort. UK Biobank participants were more likely to be older, female, living in less deprived areas and reporting fewer health conditions than the general UK population, and overall mortality in the cohort is substantially lower than the national rate. That does not invalidate internal comparisons between coffee drinkers and non-drinkers within the cohort, which is what this study makes. It does mean the absolute risks are not the general population's, and it can distort associations when the exposure correlates with the same factors that drove participation.
There is a fourth worth adding. Dementia has a very long preclinical phase, and changes in habits and preferences can begin years before diagnosis. As with the Parkinson's literature, reverse causation cannot be excluded, and an 11-year median follow-up is not long enough to fully rule it out for a disease that develops over decades.
It is also worth remembering that these are relative risks applied to conditions with a specific absolute frequency. Across 365,682 people over 11.4 years, 5,079 dementia cases arose. A 28% relative reduction on a baseline risk that is already low for most middle-aged people is a much smaller absolute change than the percentage suggests.
Does it fit the rest of the evidence?
Broadly, yes. The umbrella review by Poole and colleagues[3] found coffee associated with lower risk across a range of neurological outcomes, and cardiovascular disease at a relative risk of 0.85 at three to four cups a day, which is in the same intake region this study identified. It is not universal agreement, though. Other UK Biobank analyses using different endpoints and different modelling have found tea associated with lower dementia risk while coffee was not, which is a reminder that a single well-conducted study in a large cohort is still a single study.
Where analyses of the same cohort disagree, the responsible reading is that the effect is small enough to sit near the resolution limit of the data, not that one team got it right and the other got it wrong. Small effects are exactly where analytical choices show up most.
It is also why single studies, however large, are worth less than the pattern across many of them.
On that basis the coffee and stroke association is currently the better supported of the two findings in this paper.
Both remain observational, and neither establishes cause.
That caveat is not decoration; it is the single most important sentence in the article.
The disagreements are instructive rather than disqualifying. Different analyses of the same cohort can reach different conclusions because they define the outcome differently, adjust for different variables, or model the dose differently. That is not a scandal, it is what happens when a genuine effect is small enough to sit near the edge of what the data can resolve. A finding that survives every analytical choice is rare in nutrition research, and its absence here is a reason for calibration rather than dismissal.
The practical reading
The defensible summary is that in a large UK cohort, moderate consumption of coffee and tea together, roughly 2 to 3 cups of each per day, was associated with meaningfully lower risk of stroke and dementia over the following decade, with the benefit disappearing at higher intakes. It is observational, subject to reverse causation and healthy-volunteer bias, and does not establish that either drink prevents anything.
Putting the numbers in absolute terms
Relative risks obscure how much any of this would change an individual's life, so it is worth doing the arithmetic. Across 365,682 people followed for a median of 11.4 years, 5,079 developed dementia. That is roughly 1.4% of the cohort over more than a decade, in a group aged 50 to 74 at enrolment.
A 28% relative reduction applied to a baseline of that size is a change of well under one percentage point in absolute terms. That is a genuinely meaningful public health difference when applied across a population of millions, and a very small consideration for any one person deciding what to drink. Both of those statements are true simultaneously, and the difference between them is the single most common source of confusion in health reporting.
The same arithmetic applies to the stroke result, where 10,053 events across 365,682 people works out to under 3% over the follow-up period.
If you drink a few cups of coffee and a few of tea a day, this is reassuring and requires no change. If you drink eight cups of coffee a day for brain-health reasons, this specific study does not support that, and neither does any other. And nothing here is a substitute for the things with real evidence behind them for brain health, which remain blood pressure control, physical activity, sleep, hearing, and not smoking. Our coffee and health guide covers the everyday version of all of this.
Those interventions have trial evidence behind them that coffee does not, and the ordering matters more than any individual item on the list.
The framing that survives all the caveats is a modest one: coffee and tea in ordinary quantities do not appear to be doing your brain any harm over a decade, and may be doing a little good, with the benefit concentrated at a moderate intake and disappearing above it. That is not a headline, and it is what the data support.