[slymon99]

Another nutritional epidemiology study with an absurdly large effect size that is almost certainly due to confounding factors.

Yes, the researchers did an admirable job trying to control for confounders. Doesn't matter. Confounding factors for health are impossible to adjust for, whether it's healthy user bias or nutritional choices or a million other possible factors. There's just no way to take this sort of causality seriously when humans are living infinitely complex lives. Stop wasting money and do an RCT.

Also, traditional media is still embarrassingly bad at communicating correlation vs causation. How does this get past a science editorial board.

[beebmam]

That's why RCTs are the gold standard. You can test nearly exactly the effect of an intervention.

[vintermann]

You can't do RCTs on something as common as coffee consumption.

What you can do is mendelian randomization studies. There are genes which influence bitter taste which seem like obvious candidates: If coffee tastes like pure quinine to you, you're probably drinking less of it than you otherwise would.

A quick search shows that there have been M.R. studies on coffee, and they do suggest the relationships between coffee drinking and various good health outcomes aren't causal, just like the ones on alcohol. At this point, we should be suspicious of any "U-shaped" effect curves of extremely common and popular habits.

[beebmam]

RCTs have been done on ethanol consumption, a beverage nearly just as common. This example contradicts your first statement.

[vintermann]

I'm not aware of any RCT that has been done on alcohol consumption. The ethics board must have been crazy if that happened.

[kelseyfrog]

Meta analyses > RCTs

[raffraffraff]

Yes, but go to the Cochrane library and try to get sensible analysis on anything that doesn't have a few RCTs under its belt, and Cochrane just shrugs.