4 ms·

Any further reading recommendations for concept of "correlation does not imply causation"? I always look for a concise way to describe this (there are just too

by LiweiZ 12y ago

Any further reading recommendations for concept of "correlation does not imply causation"? I always look for a concise way to describe this (there are just too many reports concluded in this way, I hope I could let others understand this point sooner to avoid being trapped in the details for too long if not forever) but have not been successful. Your words seems the best I can find.

skylan_q 12y ago
I really like this video. He goes over the difference between clinical and observational studies. https://www.youtube.com/watch?v=y1RXvBveht0 https://www.youtube.com/watch?v=y1RXvBveht0
stolio 12y ago
For a humorous look at what happens if we assume correlation causes causation there's Spurious Correlations: http://www.tylervigen.com/ http://www.tylervigen.com/ My favorite: "Number people who drowned by falling into a swimming-pool correlates with Number of films Nicolas Cage appeared in" With enough variables and noise you'll find things that correlate if you look hard enough. Without experiments where you control one thing and observe the other it's hard to demonstrate causation.
downandout 12y ago
Every time I bring this concept up on HN, I get massive numbers of downvotes....I'm shocked that the parent comment is currently the top comment. Maybe it's me. Anyway, here's one of my favorite applications of the concept: http://www.investopedia.com/terms/s/superbowlindicator.asp http://www.investopedia.com/terms/s/superbowlindicator.asp
- wutbrodo 12y ago
  Hm, this and some of the other comments being mentioned are the much less interesting instances of "correlation != causation". Specifically, these examples (including the Superbowl indicator) are assumed to be cases of coincidence (with enough data, certain slices of noise will look like signal). By contrast, the much more interesting and relevant examples are things like hidden variables: a third factor that has causal relationships (in whatever direction) with both of the correlated variables being described, even though the two correlated variables have no such causal relationship. For an extremely trivial example, the data likely supports the conclusion that you're more likely to get into car accidents when you have an umbrella with you. This obviously doesn't imply that having an umbrella in your car causes more car accidents: but rather that the hidden variable of rainy weather causes both "people have umbrellas with them" and "people get into more car accidents". Controlling for rainy weather (i.e. examining the correlation during sunny days only or rainy days only), you'd likely find that the correlation vanishes.
- LiweiZ 12y ago
  Thank you all for sharing your recommendations. I guess the main reason you got downvotes with this concept is that, in most cases, you have different conclusion/insight from majority's. For this report, it's easier for us to come to similar ones. As an explanation for the possible issue for this report, users tend to upvote the parent comment as most of us accept it.

by LiweiZ 12y ago

skylan_q 12y ago

I really like this video. He goes over the difference between clinical and observational studies. https://www.youtube.com/watch?v=y1RXvBveht0 https://www.youtube.com/watch?v=y1RXvBveht0

stolio 12y ago

For a humorous look at what happens if we assume correlation causes causation there's Spurious Correlations: http://www.tylervigen.com/ http://www.tylervigen.com/ My favorite: "Number people who drowned by falling into a swimming-pool correlates with Number of films Nicolas Cage appeared in" With enough variables and noise you'll find things that correlate if you look hard enough. Without experiments where you control one thing and observe the other it's hard to demonstrate causation.

downandout 12y ago

Every time I bring this concept up on HN, I get massive numbers of downvotes....I'm shocked that the parent comment is currently the top comment. Maybe it's me. Anyway, here's one of my favorite applications of the concept: http://www.investopedia.com/terms/s/superbowlindicator.asp http://www.investopedia.com/terms/s/superbowlindicator.asp

wutbrodo 12y ago
Hm, this and some of the other comments being mentioned are the much less interesting instances of "correlation != causation". Specifically, these examples (including the Superbowl indicator) are assumed to be cases of coincidence (with enough data, certain slices of noise will look like signal). By contrast, the much more interesting and relevant examples are things like hidden variables: a third factor that has causal relationships (in whatever direction) with both of the correlated variables being described, even though the two correlated variables have no such causal relationship. For an extremely trivial example, the data likely supports the conclusion that you're more likely to get into car accidents when you have an umbrella with you. This obviously doesn't imply that having an umbrella in your car causes more car accidents: but rather that the hidden variable of rainy weather causes both "people have umbrellas with them" and "people get into more car accidents". Controlling for rainy weather (i.e. examining the correlation during sunny days only or rainy days only), you'd likely find that the correlation vanishes.
LiweiZ 12y ago
Thank you all for sharing your recommendations. I guess the main reason you got downvotes with this concept is that, in most cases, you have different conclusion/insight from majority's. For this report, it's easier for us to come to similar ones. As an explanation for the possible issue for this report, users tend to upvote the parent comment as most of us accept it.