Ross Pomeroy at Real Clear Science discusses five logical fallacies that often get misidentified and abused in arguments. Identified by Steven Novella in his book The Skeptic’s Guide to the Universe, one of these is the old Correlation and Causation fallacy:
2. Correlation and Causation. Correlation does not prove causation. To say that it does is a logical fallacy. However, correlation absolutely can be evidence for causation, the quality of which depends upon, for example, whether the correlation is actually feasible, how strong studies show the link to be (effect size), and whether or not the variables in question demonstrate a dose response (if X fluctuates, does Y also change in a predictable way?)
I often see this one expressed as, “Correlation does not imply causation,” which I think is wrong. As Pomeroy is careful to stipulate, correlation does imply causation. It just doesn’t, by itself, prove it.
But the question is, what does demonstrate causation? David Hume pointed out that we never observe causation, ever. We only ever observe correlation. (Which is related to the infamous problem of induction.) We can only infer causation. But then how do we discriminate between a scenario where two variables are merely correlated from one where one factor causes another?
Pomeroy above talks about whether the variables in question fluctuate together, and I think that’s usually a promising sign, but not always. I know in troubleshooting IT systems, I’ve encountered situations where two variables did fluctuate together but were actually both caused by a third variable that was later uncovered.
So then what tells us that one thing causes another? I think the answer is that we have to isolate the correlation down to one essential item that, if missing, the effect doesn’t occur.
So when establishing a link between, say, an increase in occurrences of a disease and a certain lifestyle or dietary choice, the statistical increase in the occurrence of the disease has to be isolated to one necessary and sufficient factor, with all other factors, such as genetics, eliminated or controlled for. It’s why we can say that smoking causes lung cancer, but have no grounds to say that artificial sweeteners cause cancer.
But is that right? Are there other criteria we should bring to this? How do we know when we’ve found a true cause?