, 18 tweets, 7 min read Read on Twitter
1/ Alright, time for round 3 of #EconBookClub! In this iteration I will cover chapters 2.4 to 2.9 of Guido Imbens' new working paper on PO vs. #DAGs (link: arxiv.org/pdf/1907.07271…). #BookofWhy #Causality #AI #MachineLearning
2/ The Twitter thread will probably be much shorter this time compared to the last ones, simply because Guido discusses a lot of terminology and technical details in these section, which I don't have much to comment on.
3/ But I'm extremely happy that Guido provides an introduction into the basics of DAGs and calculus to a broader audience in economics. Finally, I have a paper to cite when I use the terms "confounder" or "mediator", since this type of terminology,
4/ which relates to the relative position of nodes in a graph, and which is hard to capture in pure probability terms, is still not very common in economics. Also interesting to see that Guido includes an, albeit brief, discussion of causal discovery methods in his paper
5/ (see the book by Peters, Janzing and Schölkopf mitpress.mit.edu/books/elements…), as this line of work developed somewhat independently from the "Pearlian" way of doing causal inference.
6/ One paragraph nonetheless caught my eye, because I think it's very telling for the overall idea of the paper. It's when Guido discusses how realistic the frontdoor criterion is. We had a very similar argument here on Twitter before. See e.g. the cited tweet by @Jabaluck.
7/ Frankly, I don't know in how many applied settings the FDC will be useful. And I agree broadly with the paper that the enthusiasm for the FDC portayed in the #BookofWhy might be a bit overblown.
8/ Although I also find the work by Adam Glynn on variants of the FDC to be super interesting. And who knows how this specific new angle might prove to be fruitful for applied work in the future even in econ. scholar.harvard.edu/files/aglynn/f…
9/ But when Guido engages in a long discussion about how useful the FDC really is for econ and whether it's justified to include it in the "canon" of identification strategies, he misses the point, in my view.
10/ It's not about whether DAGs give us one specific new identification strategy, that we have overlooked so far. Instead, DAGs (as way to impose structure on the data generating process) and do-calculus (which provides the algebraic rules for
11/ manipulating do-expression like P(y|do(x))) actually provide us with a myriad of new identification strategies, because they offer a pricipled way of solving *any* identification problem in a given recursive model.
12/ And it doesn't stop with just solving the confounding problem. Do-calculus can also be used to tackle selection bias, incorporate auxilliary experimental knowledge and achieve transportability across heterogeneous populations.
13/ These challenges ubiquitous in applied empirical work – also in econ – and only three simple inference rules bring us a long way of knowing once and for all how to deal with them.
14/ Do-calculus is so powerful, because it has been shown to be complete for many of these tasks. This means that if there is a solution, do-calculus will be able to find it. On the other hand, if do-calculus fails to deliver a solution, we can be sure none exists.
15/ Completeness is such a big step over the piecemeal approaches that we currently use for solving identification problems. Because we don't have to find a fitting nail for our hammer anymore, but can now choose the appropriate hammer for our nail.
16/ On top of that, all this can be automated, and CS folks have developed many powerful algorithms that carry out the identification task for us. So we don't have to bother anymore with the technical details of the econometrics (once we know what's happening "under the hood").
17/ Instead, we can purely focus on the subject matter, collecting high-quality data, and getting our model of the DGP right, based on the prior knowledge accumulated in our field. Isn't that what applied science should ultimately be about?
18/ If you want to learn a bit more about this, you can have a look at my slides from the talk I gave at the AAAI symposium this spring. Link can be found here (3rd to last paragraph): p-hunermund.com/2019/04/04/bey… In the end, the thread turned out to be not much shorter, I realize. :)
Missing some Tweet in this thread?
You can try to force a refresh.

Like this thread? Get email updates or save it to PDF!

Subscribe to Paul Hünermund
Profile picture

Get real-time email alerts when new unrolls are available from this author!

This content may be removed anytime!

Twitter may remove this content at anytime, convert it as a PDF, save and print for later use!

Try unrolling a thread yourself!

how to unroll video

1) Follow Thread Reader App on Twitter so you can easily mention us!

2) Go to a Twitter thread (series of Tweets by the same owner) and mention us with a keyword "unroll" @threadreaderapp unroll

You can practice here first or read more on our help page!

Follow Us on Twitter!

Did Thread Reader help you today?

Support us! We are indie developers!


This site is made by just three indie developers on a laptop doing marketing, support and development! Read more about the story.

Become a Premium Member ($3.00/month or $30.00/year) and get exclusive features!

Become Premium

Too expensive? Make a small donation by buying us coffee ($5) or help with server cost ($10)

Donate via Paypal Become our Patreon

Thank you for your support!