I promised you a post on causation and correlation way back when we reviewed The Art of Thinking Clearly, and as you longtime readers know, I usually eventually get around to keeping those kinds of promises. In light of our upcoming review for Bernoulli’s Fallacy, I was reminded that I had not done it yet, so that caused me to decide to get around to writing the promised post. Or maybe it was just a correlation.
Mistakes of causation and correlation are among the most pernicious, pervasive, and potentially damaging errors in thinking possible, and they are also among the most common. I suspect this derives from our neurophysiological predilections for pattern identification. Neurologically, we want to be able to explain things, we want to be able to understand things, and we want to have reasons for things, because it helps make the world predictable, and a predictable world is a safe world. Plus, it’s usually better to over-identify the pattern of a saber-toothed tiger stalking you than to under-identify it.
Causation, as you have probably gathered by now if you did not already know, is when one event precipitates the occurrence of a subsequent event. Event A does not simply lead to event B, it does not simply happen to occur in a similar time and/or place to event B – it is the source, the initiator, the precipitator. Defined in the negative, causation exists when event B could not have occurred without the prior occurrence of event A.
Correlation, on the other hand, does not require that event B follow event A, or that event B be dependent upon event A, or any other way of phrasing an argument of causation. Instead, correlation simply claims that there is a relationship between event A and event B, however distant or casual. It could be that event A and event B occur in the same county, the same state, the same country, the same planet. It could be that events A and B occur within seconds, minutes, hours, days, years, decades, centuries, millennia of each other. It could simply be that event A and event B are both attended by the same participants, or that they have similar color schemes.
You can see, then, that correlation is a much, much lower bar to clear, and that with a little imagination almost anything can be correlated to almost anything else. The danger is that when most of us see a correlation, we are inclined to assume or infer the existence of causation, instead of mere correlation. Even being aware of this tendency does not entirely prevent it; I am somewhat militant about keeping causation versus correlation in mind, and I still recently made such an error in diagnosing the source of a satellite anomaly (a solar flare happened to occur just a little before the problem was detected, and since solar flares can cause symptoms like I observed, I assumed that the solar flare was responsible – after deeper analysis, that assessment turned out to be flawed). This is one of those concepts for which you will have to be constantly on guard, both in your own thinking and in others’.
There are all kinds of dramatic examples of places where correlation exists without causation, things like the rate of sour cream consumption in Canada tracking with the population growth of koalas in Australia, or the number of onion rings consumed tracking with consumer price index. Entire websites are dedicated to compiling statistics and finding inane correlations like these, ostensibly to help show how correlation does not imply causation. However amusing they might be, I don’t think these examples are very useful for really building an understanding of how to identify cases of causation and distinguish them from cases of mere correlation. Such examples are simply too obvious, too extreme. Most people can recognize that it’s highly unlikely that having Canadians eat more sour cream will lead to more koalas, or that eating more onion rings will lead to higher inflation.
I assert that being able to identify when causation does not exist is rather beside the point, because it implies that you are approaching the problem with the assumption that there is causation, and seeking to disapprove that conclusion. Instead, I urge approaching cases from the opposite direction: assuming that there is no causation, and seeking to prove that there is. This way, we can establish clear criteria for causation with which to analyze cases of correlation, and minimize the chances of inferring causation inappropriately. This is based on a founding assumption that a false positive case of causation is more damaging than a false negative case of causation.
We must, therefore, identify clear criteria by which we can establish causation from correlation. As we’ve already established, mere spatial or temporal proximity, or superficial similarity, are all inadequate to demonstrate causation. Ideally, causation would be identified through controlled experimental setups in which variables are isolated and tested independently and in permutations until the observed effect is achieved, but this is not practical for most circumstances, and this post is not about describing a scientific protocol by which to determine causation. Rather, this post is about exploring a useful mode of immediate analysis that anyone can employ to prevent the thinking-error of mistaken causation.
In my process, the first step is the possible/impossible step. Based on my understanding of physics and other sciences, and how confident I am in that understanding and knowledge, I examine if event A could conceivably have caused event B. If event A could not conceivably have caused event B – perhaps it would violate thermodynamics, or the lightspeed barrier, or something less extreme and fundamental, but very well established – then I dismiss the possibility of causation. If event A could realistically have caused event B, I move onto the next step in my test.
This is the likelihood of the explanation step, and involves a qualitative analysis of prior probabilities (look at me, using my new Bayesian statistics vocabulary). This is where Occam’s Razor comes in (although that tool can sometimes be misleading). We’ve already established in step 1 that event A could have possibly caused event B, but is it likely? This step seeks to answer that question. I’ll look at prior, similar circumstances or alternative explanations. Perhaps event A could have caused event B, but only if ten other things happen in exactly the right way, while there was also an event C that could have caused event B without those conditions being true. Based on this step, I’m inclined to not assign causation between events A and B. Alternatively, maybe there are no other reasonable explanations for how event B could have happened. In that case, I’ll be inclined to accept that there might be a causal relationship between events A and B.
Whether or not I would go further than those two steps depends on how important the analysis is. For most purposes involving cases of everyday correlation and causation, these two steps are sufficient for me to gain a reasonable hypothesis about the causation question with which to proceed and operate. The biggest key, more so than the two-step test, is assuming that there is not a causal relationship, and approaching it as a matter of needing to prove that causation exists, not that I need to disprove it. This is contrary to how our brains are naturally inclined to function, but it is worth the effort.
Of course, even with this post, even with a conscious effort, I know that causation and correlation will continue to be confused, including by me. It’s too easy to let our assumptions get ahead of us, and end up inferring causation where there is only correlation. The point is not so much that we need to expunge this entirely, as that we need to be aware of it, so that when we find ourselves confronted with situations of importance, we know that we need to slow down and consider if what we are seeing can really constitute causation. Whether this post can actually cause any of that, however, remains to be seen.