Fact Checking Economists

Causal Inference.
Scott, stop posting fake news. You've been fact checked. This is the original pic.twitter.com/f61zxTHoaz

— http://www.ar-tiste.xyz (@artistexyz) August 10, 2022

DAG extraction from text (DEFT)

This  is  an app that I would love to write.

Consider the narrative “The rooster crowed at dawn”. I would break this into nodes

A. rooster crowed

B. sun starts to rise (dawn)

Note that each node is a noun-verb event.

Then I would write the following DAG: B->A

Could we take a whole novel like To Kill a Mocking Bird, and extract some DAGs from it this way, automatically? Sounds difficult! Could we not collect a large library of these DAGs by processing thousands of books this way? I already have a library of DAGs in my program JudeasRx. I call it a dot_atlas, because the DAGs are written in the dot language of graphviz. However, the DAGs in that library are all generated by hand.)

I got this idea from the following series of Tweets by Bob Thibadeau

/2 My Ph.D. was on this. Memory for such causal structure ('good' narratives) is phenominally good. This is largely unrecognized in the literature. Too obvious to see. All animals recognize causality…all can fall prey to such deceptions.https://t.co/jsyOXCl5Qt

— Bob Thibadeau (@BobThibadeau) August 7, 2022

When is it valid to use “immutable traits” like gender/race in causal inference?

Check this out:

https://stats.stackexchange.com/questions/366301/when-is-it-valid-to-use-race-ethnicity-in-causal-inference

Some people question whether one can use “immutable traits” like gender/race as nodes in a causal DAG.

I think it is fine to use gender or race as a node in a DAG, and it should be used as a root node because it is determined at birth, which usually precedes all other nodes. Remember, the nodes of a causal DAG are ordered by time. I discuss this further here.

For any DAG node X, you can find an individual that will refuse vehemently to change X, which is equivalent to an individual for which X cannot be changed. So if you say gender/race can’t be used as nodes in a DAG, then no X can be used. It’s incorrect to say that a mathematical do(X) intervention is only justified if there is a corresponding physical mechanism whereby each individual of the population of patients being considered, can be made to assume all values X . That is almost never the case for most X.

For example, consider X=Trump follower(Yes, No) or X=smoker(Yes, No). I hate Trump and smoking. The only do(X) intervention that will change my mind about this is if you do a brain transplant on me. In fact, most people are similarly unwilling to change their Trump and smoking stance. Do we filter out individuals like me? No. We don’t care. What is the difference between such an X and gender/race?

Reply to Yann LeCun essay entitled “What can AI tell us about Intelligence”

Check out the following essay by Yann LeCun and Jacob Browning, recently published by Noema magazine.

Noema highlights the essay on AI that Jake Browning and I recently published. https://t.co/swQGCb8Jaf

— Yann LeCun (@ylecun) August 3, 2022

Excerpts:

(Referring to Gary Marcus) More broadly, he holds this extends into our more basic abilities, where there is an underlying symbolic logic behind causal reasoning and reidentifying the same object over time.

…

Is this a call to stop investigating hybrid models (i.e., models with a non-differentiable symbolic manipulator)? Of course not. People should go with what works. But researchers have worked on hybrid models since the 1980s, and they have not proven to be a silver bullet — or, in many cases, even remotely as good as neural networks.

The essay elicited this respose from me:

Nice essay. I am a proponent of causal inference (CI), not of symbolic logic. Symbolic logic reminds me of the Stone Age of AI, i.e., Macsyma and rule based expert systems. Bayesian networks, whereby we express CI, were invented to overcome the limitations of rule based expert systems, so I wish Marcus would stop lumping CI and symbolic logic together, as if they were cojoined twins. They are not. I also wish that LeCun, instead of addressing the issue of hybrid systems of NN and symbolic engines, would address the far more important issue, at least to me,  of hybrid systems of NN and CI engines.

The human brain is a hybrid system of left and right brain, each performing different functions. AI might find it advantageous to imitate Nature in this respect.

By the way, LeCun is patently wrong when he implies that AI has only 2 options: 1. continuous, differentiable systems amenable to back propagation, 2. discrete systems for symbolic logic. He is ignoring tens of thousands of papers on probabilistic networks (such as Judea Pearl’s causal Bayesian networks and Andrew Gelman’s hierarchical models) and MCMC. These systems are neither differentiable nor discrete. Just like NNs come equipped with the powerful tool of backpropagation, probabilistic networks come equipped with the powerful tool of MCMC.

The following posts, critiquing the current state of AI, might also be of interest to the reader.

• Andrew Ng Blames the data
• Stanford championing shaky Foundation Models, a dead end street for AI

What is Causality? How would you define it?

People often debate how to  define causality. To define the concept of causality, the Wikipedia article on Causality quotes philosophers like Aristotle (384BC-322BC), Kant (1724-1804), Hume (1711-1776), and physicists like Alfred Robb (1873-1936) and Max Jammer (1915-2010).  The philosophers trot out mysterious sounding words like ontology and  epistemology. The physicists of course trot out the link of causality to space time geometry and relativity. All these smart men (no women!, WTF) agree that causality is a real thing worthy of our attention, and that it clearly entails much more than mere correlation. Since my PhD is in physics, I prefer to define causality as a purely physical phenomenon. This is how I define it in my free book Bayesuvius(676 pgs) :

Causality is a time-induced ordering between two events, the transmission of information (and its accompanying energy) from the earlier of the two events to the later one, and the physical response of the later event to the reception of that information.

I expand on this definition in my essay entitled; “Bayesian Networks, Causality and the Passage of Time”.https://qbnets.wordpress.com/2022/07/27/bayesian-networks-causality-and-the-passage-of-time/

Artist Harry Clarke’s 1919 illustration for “A Descent into the Maelström”

Bullet ripping through Jack of Hearts. High speed photo by MIT’s Harold (Doc) Edgerton.

From Deep Learning to Deep Understanding

Causal Inference Elevates Data Science from Deep Learning to Deep Understanding, according to Pearl.

Slide from Pearl talk at IJCAI-2022

Full set of slides and video available here

Overwhelmed by the friendly reception of my keynote talk @IJCAIconf today, in Vienna, I am happy to share my slides for our readers:https://t.co/UAP0TbdoMBhttps://t.co/R6fvXnxc3t
A video of the 2nd part of my talk, separately recorded, will be available on @IJCAIconf website. https://t.co/KYkakg7BHw

— Judea Pearl (@yudapearl) July 28, 2022

Bayesian Networks, Causality and the passage of time

Time flies

For those who would prefer not to download my whole free book Bayesuvius (650 pages), I’ve excepted the chapter (5 pages, no equations or code) entitled “Bayesian Networks, Causality and the passage of time”. I’ve also translated it to Spanish.

chapter in Spanish

chapter in English

One line graphical proofs of backdoor, frontdoor and napkin adjustment formulae without using Do-Calculus rules

About eight months ago, I wrote a blog post where I set out on the quixotic quest of simplifying the proofs of Pearl’s backdoor and frontdoor adjustment formulae and of any adjustment formulae associated with Pearl’s identifiability.  The result of that quest was a chapter entiltled “Do Calculus Proofs” in my book Bayesuvius. Today, I am happy to announce that I just overhauled that chapter with some new ideas that I’ve had since the first version of that chapter. Now I can prove the backdoor, frontdoor and napkin adjustment formulae graphically, in just  a few lines (depending of how you count). The new chapter is still entitled “Do Calculus Proofs”. It is currently chapter 18 of Bayesuvius. For your convenience, in case you don’t want to download all of Bayesuvius (650 pgs), I have excerpted that chapter only and placed a pdf of it here. https://github.com/rrtucci/do-calculus-proofs-project/raw/master/main.pdf

Causal Inference Unification

In writing my free open source book Bayesuvius (650 pages) about Bayesian Networks and Causal Inference (CI), it has become painfully obvious to me that CI is currently practiced by several enclave communities that don’t communicate much with each other. I made jest of this in a previous blog post, but I am also trying to do something concrete about it, more than just joking about it, via my book Bayesuvius.

First and foremost, Bayesuvius attempts to unify seamlessly Rubin’s Potential Outcomes Theory and Pearl’s CI. Both theories strongly disagree on the need for DAGs and SCM, yet both are describing the SAME physical phenomenon. So it’s inconceivable to me that these two approaches are irreconcilable.

One Ring to rule them all. One theory for the 3 rungs, not two (Pearl/PO)

One of the original goals of Bayesuvius was to explain all of AI/ML through the lens of Bayesian Networks, so there are dozens of chapters in Bayesuvius that will be of interest to the AI/ML community (e.g., Bayesian networks, decision trees, NNs, message passing, back propagation, MCMC, reinforcement learning, GANs, Diffusion Models, etc.)

Bayesuvius also tries to explain all the pet CI methods currently circulating in the economics community. For that, I read 2 excellent CI books by economists (namely, the mixtape book by Scott Cunningham, and the brave & true book by Matheus Facure). Then I wrote a chapter in Bayesuvius for each of those pet CI methods (e.g., difference-in-differences, synthetic controls, instrumental variables, LATE, regression discontinuity design, etc.)

Bayesuvius also tries to bring in the neuroscience community. For that, I wrote chapters in Bayesuvius on time series analysis and on Granger Causality.

More recently,  Bayesuvius has been catering to the epidemiology/biostatistics community.  About a month ago, I set myself the goal of learning about the following CI related subjects that are trending in that community

1. Survival Analysis
2. Generalized Linear Models
3. Targeted Estimators
4. g-formula
5. Modified Treatment Policy

I am happy to report that I have now written chapters in Bayesuvius on each of those 5 subjects.

Issac and Nora, Latin American Music, amazing two children and father band

https://m.youtube.com/c/IsaacetNora/videos

RCT to determine the effectiveness of parachutes

The more extreme advocates of what is called EBM (Evidence Based Medicine) claim that the only way to prove causal relationships is by doing a RCT (Randomized Controlled Trial). Causal Inference (CI)  advocates like me disagree. We claim that CI provides ways to find causal relationships besides RCTs, and sometimes an RCT is impossible, in which case non-RCT methods given by CI are the only choice possible. We claim that you should first do a well designed survey (a.k.a. observational study), and then, if possible, use the knowledge gained to design a better RCT, because RCTs are more expensive, time consuming and laborious than surveys. Here is a funny example of a case where an RCT is not a good idea.

Source:

https://twitter.com/eliowa/status/1539658116559941632

The Causal Inference Roundup

Bob lifts motorcycle (targeted estimators) with his hard head

I finally finished the first version of a chapter on targeted estimators (TMLE estimators) for my book Bayesuvius. It is currently chapter 74. Check it out!

TMLE estimators is a topic that is  in vogue right now within the causal inference community, especially among biostatisticians and economists. It improves on what is called doubly robust estimators.

This chapter was very difficult to write, because the mathematical calculations involved can be at times very advanced, confusing and long. I tried to make the chapter as mathematically clear and explicit as possible. I used a very intuitive yet precise notation of functional derivatives that I learned from Physics. Contrary to most TMLE references, I do not offer code. But I hope, to compensate, that you find my math clear, easy and fun.

I feel after this exertion like the guy in the video below must feel after lifting a motorcycle with his head.

Calling R from Python

Calling R from Python | Magic of rpy2 by Nevin Baiju https://t.co/PUUlkvMbYp

— http://www.ar-tiste.xyz (@artistexyz) May 30, 2022

Biostatistics OR BUST

Biostatisticians (Bstats) and epidemiologists (Epis) are synonymous in the same way that machine learning (ML) and artificial intelligence (AI) are synonymous. Some allege subtle differences between these two tribes, but I find them indistinguishable, except by title. Some say that Epis do more community outreach and public health, whereas Bstats do more statistical analysis, but there are numerous exceptions to this rule.

Recently, I’ve been reading Twitter messages by Epis/Bstats that use the terms

1. Cox proportional hazards model (used in Survival Analysis),
2. Targeted Maximum Likelihood Estimate (TMLE, used in Targeted Learning). 
3. Modified Treatment Policies

These 3 topics are all closely related to Causal Inference so I’ve decided to write 3 chapters on them for my free book Bayesuvius. Actually, I’ve already written the chapter on topic 1, so I  have 2 chapters to go.