Seems like an ambiguity issue. The phrase is reasonable when it means "correlation does not prove causation". But yes, I agree it should not be used to mean "correlation is no argument for causation"
Causation is counterfactual correlation, but since you can't directly observe counterfactuals, sometimes you can have causation without observed correlation. For instance, in a controlled system, a factor might be constantly adjusted to maintain stability, hiding its causal effect from view. Think of a very good limo driver constantly making small adjustments on the accelerator, on a hilly road with no traffic.
Who ever said or implied that correlation isn't prima facie evidence of causation? I've heard the statement "correlation is not causation" thousands of times, but never in that sense
Luckily, there have been one or two philosophers that attacked the concepts of inference besides Hume. Some even deny causation entirely.
Correlation is not causation. This is absolutely true, because they are not two terms for the same thing. All causation implies correlation, so it is possible that your argument is just that one shouldn’t go so far as to presume correlation and causation are mutually exclusive, but that needs far less statistical notation.
Very clever argument. I can’t think of a useful real world lesson though. Experimental design always takes this account. LLMs may fall for this trap on occasion.
I hate those websites that show graphs of spurious correlations to demonstrate the point that correlation doesn't imply causation. Often, the correlations they show aren't even statistically significant, and when they are, it's because of p-hacking (after all, all these sites do is go through thousands of random data series to find the pairs that happen to look correlated). This gives the misleading impression that "correlation doesn't imply causation" is just a statement of the fact that the apparent correlation could be a statistical fluke. Aside from completely missing the point ("correlation doesn't imply causation" is really a statement about how multiple different causal structures could all lead to correlation), this makes it seem like you can just ignore correlations as meaningless even when they are significant, which is the exact opposite of what you should be teaching in a statistics class.
Uh this really depends on your priors on the causal structures on your relevant variables. If you have a universal confound (a variable U that perfectly causes all others) then no amount of correlation will prove causation, as all correlation can be explained by the causal path through U. Imagine watching a traffic light with no orange light: the green light being off will correlate perfectly with the red one being off, yet the common cause is the computer controlling the traffic light and there is no direct causal link between green and red. Also Leibniz’s notion of pre-established harmony, where U is God. Clearly the DAG A <-U-> B with unobserved U is less parsimonious than A->B though.
Seems like an ambiguity issue. The phrase is reasonable when it means "correlation does not prove causation". But yes, I agree it should not be used to mean "correlation is no argument for causation"
This is how social scientists already think about correlation re: causation, though.
Most evil clickbait of all time
https://imgs.xkcd.com/comics/correlation_2x.png
Causation is counterfactual correlation, but since you can't directly observe counterfactuals, sometimes you can have causation without observed correlation. For instance, in a controlled system, a factor might be constantly adjusted to maintain stability, hiding its causal effect from view. Think of a very good limo driver constantly making small adjustments on the accelerator, on a hilly road with no traffic.
I like your essay more than the title :) I think your subtitle is more apt.
Yeah, the title perhaps overstates the case slightly:)
Admitting to the click bait in paragraph three is diabolical work
Hit them with the quick 1 2
Who ever said or implied that correlation isn't prima facie evidence of causation? I've heard the statement "correlation is not causation" thousands of times, but never in that sense
Luckily, there have been one or two philosophers that attacked the concepts of inference besides Hume. Some even deny causation entirely.
Correlation is not causation. This is absolutely true, because they are not two terms for the same thing. All causation implies correlation, so it is possible that your argument is just that one shouldn’t go so far as to presume correlation and causation are mutually exclusive, but that needs far less statistical notation.
Very clever argument. I can’t think of a useful real world lesson though. Experimental design always takes this account. LLMs may fall for this trap on occasion.
I hate those websites that show graphs of spurious correlations to demonstrate the point that correlation doesn't imply causation. Often, the correlations they show aren't even statistically significant, and when they are, it's because of p-hacking (after all, all these sites do is go through thousands of random data series to find the pairs that happen to look correlated). This gives the misleading impression that "correlation doesn't imply causation" is just a statement of the fact that the apparent correlation could be a statistical fluke. Aside from completely missing the point ("correlation doesn't imply causation" is really a statement about how multiple different causal structures could all lead to correlation), this makes it seem like you can just ignore correlations as meaningless even when they are significant, which is the exact opposite of what you should be teaching in a statistics class.
Uh this really depends on your priors on the causal structures on your relevant variables. If you have a universal confound (a variable U that perfectly causes all others) then no amount of correlation will prove causation, as all correlation can be explained by the causal path through U. Imagine watching a traffic light with no orange light: the green light being off will correlate perfectly with the red one being off, yet the common cause is the computer controlling the traffic light and there is no direct causal link between green and red. Also Leibniz’s notion of pre-established harmony, where U is God. Clearly the DAG A <-U-> B with unobserved U is less parsimonious than A->B though.