Anyone read Judea Pearls, Causality book? I’ve started, and Im hoping it’s going to be my next of those rare “oh f??k ... wow ...” moments. It’s a short-ish textbook that needs to be slowly read.
I read a SEM book , with a chapter on casual models ... which made me wonder why the hell I bothered with SEM. Any practitioners care to comment?
FYI there's (at least?) two Judea Pearl books on causality. "Causality" was written in early 2000s, and is filled with theorems + proof + prose detailing all the major results from the theory. It's not a particularly hard book if you studied Mathematics, but it is Mathematics.
"Causal inference in statistics" is a 2017 book that covers the actually important results from the previous book, in an approachable manner, and with a focus on applying the results to actual problems. Some theorems are stated formally, but usually without proof. It's 40$ on Amazon, strongly recommended. https://www.amazon.com/Causal-Inference-Statistics-Judea-Pea...
There's also "The Book of Why", which is a pop-sci-ish book by Pearl and Dana Mackenzie. It contained just enough math and examples to get me really excited for causal inference, so I just bought "Causal inference in statistics" to see the theory in detail. If you want to learn what causal inference is about, but don't necessarily want to wade through a textbook immediately, I highly recommend "The Book of Why".
If anyone's reading the first of those two, make sure to go straight to the last chapter or epilogue or whatever it was. IIRC, it's basically a talk turned into a chapter for the layman. This might only be the in the more recent edition(s).
I have, and I really like it. If you read some of the historical debates between Judea Pearl and Don Rubin (of the Rubin Causal Model fame), and also Andrew Gelman (on his blog: http://andrewgelman.com/category/causal-inference/) you will have a much better appreciation of the nuances that Pearl was trying to push forth.
What I like about SEM and graphical model formulation is that it makes the model explicit and easy to communicate. It compactly encodes many hypotheses that you can test on your data.
The problem with Pearl is that, while his tools seem pristine, the way he tries to convince applied statisticians (econometricians foremost, but others too) to use them is very childishly argumentative.
The best way to have those tools gain popularity with that crow is IMO to find a very cool application and publicize the hell out of it (like alexnet propelled deepnets from obscurity)
His push into social sciences did not seem very childish to me; it was more like a plea to have more explicit models. I haven't come across his editorials for econometricians. Could you post some references that I can read?
Unlike his previous books this is intended for general audiences rather than practitioners. It offers a good overview of Causal Inference, as well as a personal take on why there is such a split between his graphical approach and others such as SEM.
Overall, Pearl is unabashedly optimistic that statistics is finally on the verge of a "causal revolution", and this book tries to describe what that means. I'd recommend it highly, either as standalone or as background to accompany his more technical works.
I've not spent enough time on Pearl, but Migual Hernan and Jamie Robins have a free textbook on causal inference at https://www.hsph.harvard.edu/miguel-hernan/causal-inference-... ; parts of it were pretty useful for a causal inference assignment I did earlier this year.
I have been reading through a lot of his papers on causality. The next evolution of AI will be driven by a better understanding and implementation of causal models. AI/ML today is seriously lacking in this area. For this reason, spending time navigating this relatively unchartered territory will be worth the effort.
Was it Bollen's book? In general, I think the value of SEM is, given some assumptions, in identifying which out of a set of causal theories are the best explanations of some data. In practice this usually turns out to be a set of causal theories that are indistinguishable, given the data, and some that are clearly worse.
With latent variables, they're used all over in psychology to represent constructs (e.g. happiness, extroversion, IQ).
Econometric models typically do not say anything about causality. For example, the coefficients in OLS are just correlation, i.e. one unit increase in X is associated with $\beta$ increase in Y. Let's say that we regress "happiness" on "being married." It's impossible to tell which way does the causal relation flow, or if "happiness" and "being married" are both caused by a third factor altogether.
In contrast, causal models explicitly think about when we can claim that an effect is causal.
From the econometrics I have seen there is a heavy focus on finding correlations and at best argumentations whether such correlations are plausible causal relationships.
Good scientific method suggests working in the opposite direction: first make a hypothesis, then test it. Working backwards from the statistical test to the hypothesis is... troublesome.
Many academic fields are overlapping, yet have a surprising ignorance of each other. Tenure is largely independent of one's knowledge outside one's own field.
I read a SEM book , with a chapter on casual models ... which made me wonder why the hell I bothered with SEM. Any practitioners care to comment?