Isegoria

Most correlations are not causal, Crémieux reminds us:

Public health advice provides us with plenty of examples to understand ‘Why?’

Consider sugar. It’s notably declined as a share of the American diet ever since dietary recommendations went out against it.

The people who adopted the advice to drop sugar were disproportionately those who were well-off. That’s sensible, because people who are well-off have more of the time, resources, and wherewithal to follow new advice. We see this all the time. For example, the advice to have babies sleep on their backs to prevent SIDS reduced rates of SIDS death, but also made them more socioeconomically stratified because the people who took the advice last and less well were less well-off parents.

For sugar, this meant much the same: that, over time, sugar in the diet became less associated with education, income, exercising habits, and more associated with bad things like smoking.

In fact, the sugar share of the diet — despite ample coverage of sugar as a problem — wasn’t even associated with BMI until after the issue got really popular with Gary Taubes’ books going viral.

Simply selectively following advice, going along with fads, believing what’s popular and acting on it, and so on, can lead to correlations that ‘make sense’ from some research reference frame, but are not actually true.

[…]

Healthy Adherers.

In the Coronary Drug Project, people who faithfully took an inert placebo had markedly lower mortality than poor adherers—pure selection on being a conscientious, health-seeking person. The pill cannot explain this, so it did not explain this. And, perhaps, some of the available adherence predictors could explain this.

Drinking J-Curves.

In causal studies, drinking appears to be linearly bad. More drink, less health. And this makes total sense because alcohol is poison. But in observational studies, there’s often elevated risk associated with not drinking at all. The reason for this is selection in a few ways, but one of the most important ways is “sick-quitters”: people who quit drinking because they were unwell. Quitting doesn’t eliminate their issues, so they show up as a lump of high-risk non-drinkers and they distort the truly linear relationship between alcohol consumption and harms to health.

Coffee “Was Bad”.

Nowadays, mainstream news outlets frequently report that coffee is linked to better health. These headlines are everywhere, but you would be surprised if you remembered the headlines from a few decades ago. Those headlines routinely linked coffee to worse health. What changed? Smoking declined!

Coffee and smoking go hand-in-hand.

[…]

In older cohorts: more coffee, higher mortality.

In modern cohorts: more coffee, lower mortality.

In older cohorts where we have detailed smoking histories: post-adjustment, the risk goes away and coffee becomes associated with lower mortality.

[…]

If you want to improve your reasoning about the world, then assume selection explains correlations by default. Selection may not explain everything about a correlation, but it could explain a lot of it.

Posted in Science | 2 Comments »