Friday, August 16, 2013

Causal Inference: Coffee Edition

Aigh! Coffee won't kill me, but bad science (and science reporting) might. 
"In a study of more than 40,000 individuals, researchers found that people who drink more than 28 cups per week (that's about four a day) have a 21 percent increased mortality risk and a more than 50 percent increased risk if under 55."
Kudos to the reporter for helpfully dividing 28 by 7 for the reader! I might have severely misinterpreted this research otherwise.

Now, how about we give the magnitudes of our estimates a sniff test before publishing? Four cups of coffee a day increases mortality by 50%? This is enormous. Note that this is an increase in mortality from all causes - not just mortality from conditions that might have a plausible, theoretical link to coffee consumption, such as bladder cancer, uncontrolled tremors, and concussion from bouncing off walls. 

Here is a sniff test of the magnitude of this estimate: a similar, correlational analysis showed that light smoking (less than half a pack of day) is associated with an increase in all-cause mortality of 30%. Heavy smoking (more than half a pack a day), an increase of 80%.  These magnitudes are in the same ballpark as the coffee study, which immediately suggests to me that the coffee estimates are absurd. 

I would hazard that maybe, just maybe, there are confounding factors that the coffee study did not pick up. For example, the authors did not control for physical activity, education or marital status. If inactive, single, high school dropouts drink more coffee, we would get the inflated estimates we see in this paper, since this population dies younger than those who are active, married and better educated.

Now, the smoking estimates have the same weakness as the coffee estimates: they are conditional correlations that do not imply cause and effect. The critical difference is that in the case of cigarettes we have plausible theoretical links between exposure and mortality (lung cancer, emphysema) and decades of clinical and lab science that shows convincingly that cigarettes kill people. Coffee? Nada, though the bluestockings have been trying for a long time to show that it must be harmful. With this lack of theory and evidence linking coffee and mortality, the coffee researchers should be especially cautious in interpreting their correlations as causal.

So, file this one under lousy research as well as lousy reporting. This study gives the cocoa case a run for its money. We are fast building a hall of shame of hot-beverage research and reporting.