Causality

What Hacker News says

1. kqrAug 2021
   I can almost guarantee you that if you're on Hacker News, you have the prerequisites needed for Judea Pearl.

   I read Causality which I understand is the more technical of their books, and everything was presented surprisingly intuitively. Sure, I had to go over some things twice, but that's to be expected when you learn something new.

   If you're worried, start with one of the more pop-aimed books? You'll be fine.

   (Pearl did change the way I look at causality and correlation, fundamentally for the better, so I do strongly recommend getting familiar with it. I also liked Willful Ignorance which is sort of one the same theme but also not and takes a wider approach.

   Show morethread
2. wavegeekMay 2021
   I don't put any credence on this. As the authors say this study merely suggests a possible relationship that needs to be investigated in a proper RCT.

   What I don't like

   1. Passive observational studies only. Therefore a huge risk of confounders (e.g. people who generally live a healthy lifestyle).

   2. The effect size is not very large, making it much harder to winnow out confounders.

   3. Invariably attempts to 'control' for confounders use simple linear models that that come nowhere close to actually controlling for confounders.

   4. They generally only control for known confounders. If there is an unknown confounder it will get through and mess up the results.

   5. It is really hard to tell what is a confounder and what is a legitimate node in the causal chain. As a simplistic example consider 'controlling' for lung cancer in a study of whether smoking is bad for you. Of course this would take away much of the reason why smoking is bad for you. You need a proper causal model informed by data - something that is hard to do. See Judea Pearl "Causality".

   6. Note the heterogeneity of results. This suggests that things are going on that are not captured by the studies. Note also that the types of mushrooms don't seem to matter - a strange thing if there were some magic factor in mushrooms, given how mushrooms vary greatly in nutrient content.

   7. Misuse of the term "significant" usually a shorthand for statistically significant. A very different thing from "practically significant, large enough to matter".

   Show morethread
3. andyxorMar 2021
   Causality, i.e. Causal inference & Graphical Models, see the work by Judea Pearl, he pretty much singlehandedly pioneered the field.

   https://www.amazon.com/Causality-Reasoning-Inference-Judea-P...

   https://en.wikipedia.org/wiki/Causality#Theories

   https://en.wikipedia.org/wiki/Graphical_model

   thread
4. epistasisJul 2017
   Edit: sorry, instrumental variables do not describe what the original poster was talking about, but could be useful for understanding the connections between correlation and causation in other ways.

   The concept of an instrumental variable

   https://en.m.wikipedia.org/wiki/Instrumental_variable

   covers some of it. Judea Pearl's book Causality summarizes lots of recent (past few decades) work in the area, with a focus on his theoretical framework. There are lots of algorithms for discovering the structural of causality for sets of variables larger than three. Having interventional data always makes these things easier, but it's not the only way to get at causality.

   Show morethread
5. data_hopeJun 2017
   The effect in the data (unless it isn't random fluctuation. At the moment I'd go with the assumtion that it is a genuine correlation, as the effect is present over all subgroups), might (1) not be monocausal, i.e. a combination of contributing factors like development experience, age of IDEs/tooling, etc. might play a role as well as other aspects. (2) a cause might not be "explained away" because the control variable was considered. Let me elaborate

   The years of experience variable might be taken to explain away the effect of experience, and then conclude that tabs vs. spaces must be due to another effect than years of experience. But chances are, that "years of experience" and "tabs vs. spaces" are just correlated to a common, causal property (like "programming proficiency" or however you want to call it). Both "years of experience" and "tabs vs. spaces" are then just incomplete reflections of the underlying cause, both rendering the effect of the underlying cause incompletely.

   What I am trying to say is: Its complicated, probably you won't be able to find the one true cause for the effect in the data. If this were physics, one could come up with a predictive theory to put this to the test. In social studies, we just cannot control the parameters well enough.

   If you are interested in reading more on this, "Causality" by Judea Pearl is a good (but exhausting read).

   Show morethread
6. axplusbFeb 2017
   There are actually two sides of what is referred to as causal inference. Either (a) inferring a causal graph from the data, or (b) given a graph and data, measuring the causal effect of variables among each other.

   The broad idea in (a) is to start with a fully connected graph, and eliminate edges between nodes that can be tested as independent, or independent conditionally on other nodes. This gives you a non-directed graph which can be oriented by several methods (identifying V-structures, looking at residuals of regressions of X on Y vs Y on X).

   The theory in (b) actually generalizes instrumental variables and lays out graphical configurations where you can measure the causal effect of a variable onto another variable, and how to compute that effect.

   A great reference: https://www.amazon.com/Causality-Reasoning-Inference-Judea-P...

   A nice introduction: https://www.youtube.com/watch?v=RPgvfSeQB8A

   Show morethread