Coffee & caffeine

Coffee is the most-tested food in this section by one measure — four Mendelian randomization analyses, agreeing on a direction — and one of the least convincing. It is worth an entry mostly as a worked example of how to read a genetic study that says what you did not expect.

What the genetic analyses found

Mendelian randomization uses genetic variants associated with a behaviour as proxies for lifetime exposure, which removes reverse causation and much confounding. Applied to coffee and osteoarthritis, it keeps returning the same answer.

The largest analysis found genetically predicted coffee intake raised osteoarthritis risk across joint sites, most strongly at the knee — odds ratio 2.03 (95% CI 1.57–2.61, p < 0.001) — and weakly for self-reported disease (1.03, 1.01–1.05). An earlier, independent analysis found the same direction across total, knee, hip and self-reported osteoarthritis. A meta-analysis pooling eleven Mendelian randomization studies put coffee and arthritis at 1.02 (1.01–1.03). And in a Mendelian randomization of eighteen dietary factors, coffee was among the exposures raising knee osteoarthritis risk.

Four analyses, one direction. Taken at face value that is the strongest causal signal for any food in this section.

The decaffeinated result

The same paper that reported the knee odds ratio of 2.03 also broke coffee down by type — decaffeinated, instant, ground, other. Only one type was causally associated with osteoarthritis: decaffeinated coffee, at an odds ratio of 4.40 (95% CI 1.71–11.33) for knee osteoarthritis and 1.13 (1.02–1.26) for self-reported disease. Ground and instant coffee showed nothing.

That is not a caffeine effect, because the caffeinated types show nothing. It is not a coffee effect, for the same reason. What it looks like is an instrument that has picked up who chooses decaf — a group that skews older, more health-conscious, more likely to have been told by somebody to cut down, and more likely to already have symptoms that prompted the switch. Mendelian randomization is supposed to be immune to that kind of confounding, and this is what it looks like when it is not.

The same paper adds a second reason to discount its own headline: body mass index mediated more than half of the total effect. Bone mineral density mediated under 5%. The separate eighteen-factor analysis reached the same conclusion from another angle, routing coffee's effect through obesity, body mass index, diabetes, hypertension and prolonged standing.

And the pooled figure is small. Across eleven studies, coffee and osteoarthritis came to 1.02, with the lower bound of the confidence interval touching 1.00 — in a paper whose smoking estimate was 1.44 and whose alcohol estimate was 1.02 with an interval crossing 1 in both directions.

What the studies with radiographs found

Five thousand five hundred people in the Korean national survey had knee radiographs and beverage-frequency data. Coffee was associated with less radiographic knee osteoarthritis — in women only, with a significant trend across intake levels — and it was the only beverage in the survey to show anything. In 655 people interviewed in Turkey, coffee showed no association either way.

Neither of those is strong. Both point away from the genetic result.

What is missing underneath

A mechanism, and there isn't one. Nothing in this batch of searches tested a coffee constituent against chondrocytes or cartilage; searching caffeine against cartilage returns growth-plate and bone work rather than articular tissue. No trial of coffee or caffeine with a joint endpoint exists. No cohort has put coffee intake against cartilage on imaging.

So the position is: a consistent genetic signal, of negligible pooled size, mostly mediated by body weight, with its largest component attached to the one coffee type that contains no caffeine, contradicted by the only radiographic survey, and unsupported by any biology.

What would change this entry

A single analysis of coffee against cartilage measured on MRI, in a cohort that already has both. If the association is real it should be visible in tissue; if it is instrument behaviour, it will not be — and either answer is more useful than a fifth Mendelian randomization.