4 ms·

When I said correlation isn't causation I meant it. Neither high nor low correlation is evidence for causation. There's a more advanced form of being bad at th

When I said correlation isn't causation I meant it. Neither high nor low correlation is evidence for causation. There's a more advanced form of being bad at this where you think you can show causation by controlling for everything in the environment. This is also wrong; it produces something called collider bias. > Though if you find whatever gene is most correlated with something, what are the options for it not to be causation? 1. It's a coincidence and it's never causal. Imagine an OSS project releases a bugfix and you diff the old and new versions. (This is basically GWAS.) The bugfix part of the diff caused it to be fixed. The updates to the copyright dates or changelogs didn't. 2. It's causal, but the causal chain involves a specific environmental factor, and we should change that instead. For instance, you can say every human has a genetic disease that prevents them from producing their own vitamin C, which most other mammals can do. But instead of calling scurvy a genetic disease we just eat fruits and vegetables.

Dylan16807 3y ago
> 1. It's a coincidence and it's never causal. Half the point of analyzing statistics is to filter out coincidences, and that applies to correlations too. If something is a coincidence, it won't hold up as a proper correlation under reasonable amounts of analysis. So when the premise is we're starting with correlations, I think it's alright to assume they're mostly not coincidences. > The bugfix part of the diff caused it to be fixed. The updates to the copyright dates or changelogs didn't. In that case the bugfix is the "gene most correlated", isn't it? Give it a few generations to randomly spread, and the signal will be far stronger on the bugfix gene than on the copyright gene. (And if it hasn't been spreading for generations then you won't have enough samples to find either gene.) > 2. It's causal, but the causal chain involves a specific environmental factor, and we should change that instead. > For instance, you can say every human has a genetic disease that prevents them from producing their own vitamin C, which most other mammals can do. But instead of calling scurvy a genetic disease we just eat fruits and vegetables. If you're testing just humans, you'll get a 0% correlation because everyone has that gene. If you're testing across mammals, then "WARNING: Prone to scurvy". Which is a completely correct and causal result about a genetic problem, with easily accessible treatments. So, I don't understand your example at all.

Dylan16807 3y ago

> 1. It's a coincidence and it's never causal. Half the point of analyzing statistics is to filter out coincidences, and that applies to correlations too. If something is a coincidence, it won't hold up as a proper correlation under reasonable amounts of analysis. So when the premise is we're starting with correlations, I think it's alright to assume they're mostly not coincidences. > The bugfix part of the diff caused it to be fixed. The updates to the copyright dates or changelogs didn't. In that case the bugfix is the "gene most correlated", isn't it? Give it a few generations to randomly spread, and the signal will be far stronger on the bugfix gene than on the copyright gene. (And if it hasn't been spreading for generations then you won't have enough samples to find either gene.) > 2. It's causal, but the causal chain involves a specific environmental factor, and we should change that instead. > For instance, you can say every human has a genetic disease that prevents them from producing their own vitamin C, which most other mammals can do. But instead of calling scurvy a genetic disease we just eat fruits and vegetables. If you're testing just humans, you'll get a 0% correlation because everyone has that gene. If you're testing across mammals, then "WARNING: Prone to scurvy". Which is a completely correct and causal result about a genetic problem, with easily accessible treatments. So, I don't understand your example at all.