Sunday, November 11, 2007

Correlation, Not Causation

One of my favorite rants is the one about correlation not being equal to causation. That is, simply because two trends seem to be correlated does not imply that one trend causes the other. The folks over at the Church of the Flying Spaghetti Monster demonstrate this fallacious reasoning excellently with their assertion that a decline in the number of pirates causes global warming (scroll down a bit to see the graph). Unfortunately, this idea seemed to be lost on behavioral psychologists, or at least on the reporters who write the articles on their findings.

Enter this fascinating article from the Washington Post on whether or not teen sex causes delinquency. The two variables are clearly correlated, but the causal link was weak. A PhD student at the University of Virginia put the final nail in the causality coffin by using something called "behavioral genetics." Essentially, she studies pairs of twins - whose DNA and external influences are fairly similar - to investigate the correlation between teen sex and delinquency. This method effectively equalizes the effects of all the other variables present, so if the twin that engages in sex earlier is usually more delinquent, the two can said to be causally linked. However, this was not the case - if anything, the twin that had sex earlier was less likely to be delinquent. So much for early sexual activity causing delinquency.

The proper conclusion here is not, of course, that we have to encourage kids to have sex earlier. It is merely that some other variable is obviously influencing both teen sex and delinquency. The UVA researchers point to genes - that explanation seems like a cop-out to me.

Either way, this new analysis method looks like a powerful tool to address the correlation-causation problem. Let's hope that more and more behavioral psychologists (and health scientists too - their work is notoriously littered with shaky causal links) begin to use this sort of analysis to test their theories. And let's hope that some of the more bombastic conclusions out there about causal links between behaviors can be tested too. There's too many crap conclusions being foisted on the American public, and these conclusions often help direct policy even if they are weakly supported. Attributing causation to correlated data is easy, but it's often not true, and it's important that the public and our policymakers know the truth, not the easy answer.


Barzelay said...

That isn't really a new method, but it is a good one. I remember reading about a 1990 study that compared rates of homosexuality between sets of twins who had been separated at birth. So they separated out the environment factor, and were able to focus only on whether there is a genetic predisposition to homosexuality. They found that if one twin was homosexual, the other was something like five times more likely to be homosexual than a random person. Pretty significant causal link.

Mike said...

To use another example of correlation vs. causation, it's important to point out that my miserable Sunday was merely correlated to the fact that the Redskins suck, but not directly caused by it.

Strike that: it actually was directly caused by the fact that the Redskins suck. I mean, I've seen teams suck before, but they are the suckiest bunch of sucks who ever sucked.

By the way, I think lack of teenage sex is much more likely to lead to delinquency.