As someone who also heavily uses UK Biobank in my research, I would just encourage people to think about this effect size and ask whether it's plausible. If true, coffee has a more powerful effect on your risk of all-cause death than any drug for most diseases. It's wildly implausible. The results are confounded. There is no causal anchor.
Agreed, this doesn't pass the sniff test at all. Makes me think of this article[1]:
> Your estimate will be wrong for a silly, almost tautological reason: if you can only detect large effects, then any effect you detect will be large. If you keep looking for an effect, over and over again, until finally one study gets lucky and sees it, that study will almost necessarily give a wild overestimate of the effect size.
My point is that most drugs don't work this well for your disease, if you are diseased. If this result were mechanistically true, then we could instantly prolong human lifespan dramatically.
But my point is that when you're healthy, the mortality risk is so low, that it doesn't take much to make it even lower.
Even more impressive result, exercise reduces all cause mortality by 40%:
> Using a large nationally representative sample of US adults, we found that those who engaged in both aerobic and muscle strengthening activities consistent with the recommended 2018 physical activity guidelines for Americans showed a reduced risk of all cause mortality (40% reduction).
But that doesn't follow. It's conditional. More accurate follow would be "instantly prolong human lifespan dramatically AS LONG AS YOU DON'T GET DISEASED"
Example: having a fire extinguisher in your house decreases risk of losing your house in a fire by 30% before a fire. But if your house is on fire I bet it's much much lower. There are two different populations, two different distributions.
Luckily, there is no signal for harm from coffee, so we can randomize a bunch of healthy people and see if there is an effect. Personally, I'd want to see a higher quality of evidence before spending my time on it, but if people believe it, this is one of the most important studies in the history of human health waiting to be written.
While it’s closer to speculation than science, past discussion of health benefits from coffee focused on antioxidants. Coffee, green and black tea, and cocoa have relatively high antioxidant content[1], particularly after adjusting for quantity consumed.
I haven’t seen any breakdowns of how the average person ingests antioxidants, but it’s not like everyone eats 3 oz of blueberries a day :-) Beverages might be a major source.
Again, this is pretty close to speculation, but it’s not nothing.
Note that there is a suspected link between anti-oxidants and cancer. The mechanism is pretty simple: oxidative stress is a warning system to the cell that initiates various repair mechanisms, up to apoptosis. Too much anti-oxidants suppress this warning system, thus allowing a damaged cell to go on.
Yours is perfectly fine speculation for a message board. I just think that the result is likely confounded, and there is no reason to reach for a biological mechanism when the study wasn't causal in the first place.
I assume this would fall into the confounding part: culturally we see coffee as a drug that can have adverse effects, and will make efforts to limit its consumption when health deteriorates.
For instance, I would expect people to not drink coffee if they have chronic stomach illness or are under a pretty strict diet, or heavily medicated. This could heavily skew results of how long heavy coffee drinkers tend to live.
As Bayesians, shouldn't this effect size cause us to update in the direction that coffee is good for health, even if we think confounding contributed to the large effect size?
Seems to me that a likely reason for a large effect size is both a causal effect of drinking coffee and confounding, added together.
As an analogy, suppose you and I are talking about a Hollywood star who has made a lot of money. I say: "The star is probably a good actor." You say: "I would just encourage people to think about this wealth level and ask whether it's plausible. The star is probably just physically attractive." Of course, the wealthiest Hollywood stars tend to be both good actors and physically attractive.
Why is a large effect size more plausible for some unspecified confounder than it is for coffee? I think some research suggests coffee induces benefits akin to caloric restriction (autophagy) among other good things: https://www.ncbi.nlm.nih.gov/pmc/articles/PMC4111762/ Should we really have a prior that an unspecified confounder can have such a large effect size? How common is that?
>Compared with nonconsumers, consumers of various amounts of unsweetened coffee (>0 to 1.5, >1.5 to 2.5, >2.5 to 3.5, >3.5 to 4.5, and >4.5 drinks/d) had lower risks for all-cause mortality after adjustment for lifestyle, sociodemographic, and clinical factors, with respective hazard ratios of 0.79 (95% CI, 0.70 to 0.90), 0.84 (CI, 0.74 to 0.95), 0.71 (CI, 0.62 to 0.82), 0.71 (CI, 0.60 to 0.84), and 0.77 (CI, 0.65 to 0.91); the respective estimates for consumption of sugar-sweetened coffee were 0.91 (CI, 0.78 to 1.07), 0.69 (CI, 0.57 to 0.84), 0.72 (CI, 0.57 to 0.91), 0.79 (CI, 0.60 to 1.06), and 1.05 (CI, 0.82 to 1.36). The association between artificially sweetened coffee and mortality was less consistent. The association of coffee drinking with mortality from cancer and CVD was largely consistent with that with all-cause mortality. U-shaped associations were also observed for instant, ground, and decaffeinated coffee.
Positive health effects from drinking sugar-sweetened coffee make me think it's not just confounding, since I wouldn't expect health-conscious people to drink sugar-sweetened coffee. But I'm suspicious regarding the "less consistent" association for artificially sweetened coffee. That makes me think that there is just too much noise in the data to know for sure, or perhaps artificial sweeteners are actually bad for you?
> As Bayesians, shouldn't this effect size cause us to update in the direction that coffee is good for health, even if we think confounding contributed to the large effect size?
If you apply this thought process to alcohol (given what we know now), what would you conclude about this approach to updating your priors based on implausible observational data?
Having said that, coffee is great.