5 ms·

Thinking Clearly About Correlations and Causation (2018) [pdf]

64 points by Anon84 6y ago

RobertoG 6y ago
Please, add [pdf] to the title.
- wombatmobile 6y ago
  Sounded interesting. I found it so challenging to read on my phone I didn’t read it.
  - segfaultbuserr 6y ago
    Is there any mobile PDF reader that can do "smart" reformatting?
    - HuShifang 6y ago
      Xodo [0] has a "Reading" view (at least in its Android app) that seems to do just that. A good selection of other features too. [0]: https://www.xodo.com/ https://www.xodo.com/
vajrabum 6y ago
Those methods were invented by Judea Pearl, a computer science professor at UCLA who won the Turing Award in 2011 partly for that work. He's written several books on the topic. The most recent is The Book of Why. See here for previous discussions of the book including one just two days ago https://news.ycombinator.com/item?id=24487135 https://news.ycombinator.com/item?id=24487135 https://news.ycombinator.com/item?id=20889143 https://news.ycombinator.com/item?id=20889143 ago https://news.ycombinator.com/item?id=18871450 https://news.ycombinator.com/item?id=18871450
pacbard 6y ago
Unfortunately for psychologists, establishing causality within a structural equations framework is really hard (regardless of the common claims that SEM is causal in nature). The approach outlined in the paper is close to [Directed Acyclic Graphs](https://en.wikipedia.org/wiki/Directed_acyclic_graph https://en.wikipedia.org/wiki/Directed_acyclic_graph) and Pearl's approach to causality (see [this comment](https://news.ycombinator.com/item?id=24506593 https://news.ycombinator.com/item?id=24506593) for more info). The basic issue with these approaches is that the structural model needs to be well specified in order to get to an unbiased estimate of the causal effect that you want to estimate. By well-specified, I mean that you have measured all possible variables impacting the relationship of interest and that you have specified the correct relationships among the variables (no controls for mediators or no colliders, as the paper outlines). If you believe that, SEM can give you unbiased estimates of the coefficient of interest. On the other hand, if the model is misspecified, coefficients will be biased and it is difficult to figure out where the bias is. Bollen (1989) pretty much says that any coefficient could be biased in a misspecified model. (Micro) Economists have addressed this issues by moving away from a regression framework and towards experimental and quasi-experimental methods. The same methods are not really used outside of microeconomics, as far as I can tell. SEM is a little bit more common (at least with developmental psychologists that I know of) and that framework doesn't really work well with even basic quasi-experimental methods (e.g., diff-in-diff).
- currymj 6y ago
  well-specified also requires that the functional relationships are actually linear, if you're using linear models, right? this is a common critique of linear SEMs -- some ML researchers refer to a broader class of "structural causal models (SCMs)", which are basically SEMs where the relationships may be drawn from broader function classes.
  - pacbard 6y ago
    Yes. SEM assumes that all relationships are linear and that all variables are normally distributed (for standard errors and fit indices calculations). Those assumptions are baked into the model and are usually never discussed. Non-linear models are not common in social sciences so that's probably why I never used them. There are a few variables that are usually quadratic (like the effect of age on wages). Other than that, linear relationships are good enough most of the time. The only real application of non-linear model that I have seen are generalized structural equation models (GSEM). These allow for the use of link functions in SEM (logits, logs, poisson, exponentials, etc.) and are the multiple equations analogue of generalized linear models. I am not familiar with structural causal models (SCMs), but a quick google search shows that these are a non-parametric version of SEM based on Baysian estimation and are a generalization of Baysian network. They sound cool but I don't think that they will become mainstream in psychology research anytime soon.
tj-teej 6y ago
Reminds me of the XKCD joke: Person A: I used to think correlation implied causation, then I took a statistics class. Now I don't. Person B: Sounds like that class helped. Person A: Well, maybe...

64 points by Anon84 6y ago

RobertoG 6y ago

Please, add [pdf] to the title.

wombatmobile 6y ago
Sounded interesting. I found it so challenging to read on my phone I didn’t read it.
- segfaultbuserr 6y ago
  Is there any mobile PDF reader that can do "smart" reformatting?
  - HuShifang 6y ago
    Xodo [0] has a "Reading" view (at least in its Android app) that seems to do just that. A good selection of other features too. [0]: https://www.xodo.com/ https://www.xodo.com/

vajrabum 6y ago

Those methods were invented by Judea Pearl, a computer science professor at UCLA who won the Turing Award in 2011 partly for that work. He's written several books on the topic. The most recent is The Book of Why. See here for previous discussions of the book including one just two days ago https://news.ycombinator.com/item?id=24487135 https://news.ycombinator.com/item?id=24487135 https://news.ycombinator.com/item?id=20889143 https://news.ycombinator.com/item?id=20889143 ago https://news.ycombinator.com/item?id=18871450 https://news.ycombinator.com/item?id=18871450

pacbard 6y ago

Unfortunately for psychologists, establishing causality within a structural equations framework is really hard (regardless of the common claims that SEM is causal in nature). The approach outlined in the paper is close to [Directed Acyclic Graphs](https://en.wikipedia.org/wiki/Directed_acyclic_graph https://en.wikipedia.org/wiki/Directed_acyclic_graph) and Pearl's approach to causality (see [this comment](https://news.ycombinator.com/item?id=24506593 https://news.ycombinator.com/item?id=24506593) for more info). The basic issue with these approaches is that the structural model needs to be well specified in order to get to an unbiased estimate of the causal effect that you want to estimate. By well-specified, I mean that you have measured all possible variables impacting the relationship of interest and that you have specified the correct relationships among the variables (no controls for mediators or no colliders, as the paper outlines). If you believe that, SEM can give you unbiased estimates of the coefficient of interest. On the other hand, if the model is misspecified, coefficients will be biased and it is difficult to figure out where the bias is. Bollen (1989) pretty much says that any coefficient could be biased in a misspecified model. (Micro) Economists have addressed this issues by moving away from a regression framework and towards experimental and quasi-experimental methods. The same methods are not really used outside of microeconomics, as far as I can tell. SEM is a little bit more common (at least with developmental psychologists that I know of) and that framework doesn't really work well with even basic quasi-experimental methods (e.g., diff-in-diff).

currymj 6y ago
well-specified also requires that the functional relationships are actually linear, if you're using linear models, right? this is a common critique of linear SEMs -- some ML researchers refer to a broader class of "structural causal models (SCMs)", which are basically SEMs where the relationships may be drawn from broader function classes.
- pacbard 6y ago
  Yes. SEM assumes that all relationships are linear and that all variables are normally distributed (for standard errors and fit indices calculations). Those assumptions are baked into the model and are usually never discussed. Non-linear models are not common in social sciences so that's probably why I never used them. There are a few variables that are usually quadratic (like the effect of age on wages). Other than that, linear relationships are good enough most of the time. The only real application of non-linear model that I have seen are generalized structural equation models (GSEM). These allow for the use of link functions in SEM (logits, logs, poisson, exponentials, etc.) and are the multiple equations analogue of generalized linear models. I am not familiar with structural causal models (SCMs), but a quick google search shows that these are a non-parametric version of SEM based on Baysian estimation and are a generalization of Baysian network. They sound cool but I don't think that they will become mainstream in psychology research anytime soon.

tj-teej 6y ago

Reminds me of the XKCD joke: Person A: I used to think correlation implied causation, then I took a statistics class. Now I don't. Person B: Sounds like that class helped. Person A: Well, maybe...