Friday, July 27, 2007

Obesity: correlation, not causation

Wailing. Gnashing of teeth. Rending of garments.

That's my reaction to the continuing inability of even semi-informed humans to remember the distinction between correlation and causation, of fully informed specialists to "forget" the distinction ...
Find Yourself Packing It On? Blame Friends - New York Times:

.... The answer, the researchers report, was that people were most likely to become obese when a friend became obese. That increased a person’s chances of becoming obese by 57 percent. There was no effect when a neighbor gained or lost weight, however, and family members had less influence than friends.

It did not even matter if the friend was hundreds of miles away, the influence remained. And the greatest influence of all was between close mutual friends. There, if one became obese, the other had a 171 percent increased chance of becoming obese, too...

...Dr. Nicholas A. Christakis, a physician and professor of medical sociology at Harvard Medical School and a principal investigator in the new study, said one explanation was that friends affected each others’ perception of fatness. When a close friend becomes obese, obesity may not look so bad.

“You change your idea of what is an acceptable body type by looking at the people around you,” Dr. Christakis said....

Christakis knows better of course, but he also knows how to play the game. He's discovered an interesting correlation in the Framingham data set, and no doubt he thinks he's "controlled" for confounders like education, smoking, socio-economic status, exercise, hobbies, and attitudes towards food. All factors that may play a role in both friendship and obesity (and all of which are affected by genes, which is another topic).

Grr. If the NEJM had better editors they'd publicly slap the hand of researchers who pretend to be unable to tell the difference between causation and correlation. No wonder the media gets confused.

We have a long history of studies like this that demonstrate our ability to control for confounders is weak -- no matter what statisticians say. (Why it's weak is an interesting question.) I am very much doubt that a true study (impossible to do), one that randomized people to be "friends" would find any correlation.

Update 7/27/07: Other variables plausibly correlated with both obesity and friendship include term pregnancy, child rearing, and marital-equivalents. There are likely several others ...

2 comments:

Jacob said...

Well put, John. I often find myself explaining such things to patients in terms the "get" -- such as .. "TICK" doesn't cause "TOCK" does it? Of course not. The lightbulb goes on over the head - and I know we can (at that point) begin to discuss whether MMR will cause their kids to turn into Martians.

John Gordon said...

I love the "tick-tock" analogy. Of course if the patient is a theoretical physicist they may contest that in some complex quantum mechanical entropic arrow-of-time way tick does caust tock ...

Mostly though, it ought to work!