Discover and read the best of Twitter Threads about #CausalInference

Most recents (23)

Finally, @eliasbareinboim and I are able to share a working paper that we've been working on for quite a while now:

"Causal Inference and Data-Fusion in Econometrics"

🚨 #DAG #CausalInference #Econometrics #DataScience #MachineLearning #AI 1/
2/ The paper provides a comprehensive overview of recent developments in the causal AI field and discusses how DAGs can contribute to econometrics. One of our main motivations for starting this project was that the handful of papers that were written on DAGs by economists
3/ seem to be based on a state of the literature in around 2000, when "Causality" by @yudapearl was first published. But DAGs have a lot more to offer than just backdoor and frontdoor adjustment. And the size of the literature has literally exploded in the last 8 years or so.
Read 11 tweets
Award-winning, free material on #causalinference – here’s an overview on the 2013-2018 winners of the „Causality in Statistics Education Award“, awarded by .@AmstatNews, and initiated by .@yudapearl
In 2018, the award went to Jonas Peters, Dominik Janzing, and @bschoelkopf for their open-access textbook “Elements of Causal Inference”. This is the best textbook on the connection between machine learning and causal inference. It’s almost self-contained.
I think it is very useful for students in ML, stats, and physics, and is certainly very interesting for social scientists, because it views familiar topics from a different angle.…
Read 12 tweets
Hello #epitwitter! Time for an @epiellie @AmJEpi tweetorial.

Today’s topic is the Target Trial Framework for #causalinference and how to apply it to improving observational studies.

So, what is the #targettrial framework?

Well it’s not a new method! Instead think of it as pedagogical device that provides a structured way to build your research question and study design for observational studies and minimizes the potential for bias.
What does that mean?

To design an observational study, we first think about what the ideal hypothetical randomized trial (target trial) is that would let us answer our research question.

Then, we try to match our observational study as closely as possible to that trial design.
Read 20 tweets
1/ Alright folks, this will be the last round of #EconBookClub, discussing Guido Imbens' new paper on PO vs. DAGs (link: #EconTwitter #BookofWhy #Econometrics #Causality #CausalInference #MachineLearning #AI
2/ Today I will go through the final chapter 4.6 and the paper's conclusion. Maybe you will notice that I skipped over section 4.5. I had some thoughts on this chapter, but I feel that I didn't understand Guido's point well enough. So I prefer to leave it aside for the moment.
3/ In chapter 4.6, Guido discusses the returns to education example, which is a classic question in economics and which is also closely connected with the development of PO techniques in econometrics.
Read 19 tweets
2/ In this chapter, Guido responds to the main criticism that is brought forward towards the potential outcomes framework by proponents of DAGs – the choice of covariates or how to justify ignorability.
3/ We're all familiar with the ignorability / unconfoundedness assumption that underlies classical matching estimators, for example. The corresponding DAG is depcited in Figure 8a of the paper.
Read 17 tweets
Happy Tuesday, #epitwitter! There’s been a lot of discussion #onhere about #causalinference lately, but that’s not all that epi is about.

For this week’s #epichat with @epiellie, let’s talk about descriptive epi!

To kick things off, an #epiquiz:

How often do you do descriptive epidemiology in your work or research?

(click to see poll options)
Next question — this one’s for discussion so share your thoughts!

What differentiates descriptive epi from causal inference epi?

My take: descriptive epi is about who, what, where, & when; causal inference epi is about why, how, and what would happen if we changed something.
Read 11 tweets
1/3 I am curious, do you know of any paper that states explicitly the asymptotic equivalence of PS and the "adjustment formula"?? I have only seen it in Causality,
ch 11, not elsewhere, Why? Perhaps someone is interested in marketing it as a magic wand? I am also curious to know
2/3 how many readers hear this equivalence for the first time. It says that regardless of what covariates you use, as the number of samples increases, the bias of the PS estimator converges to the bias of the adjustment estimator. If many hear about it for the first time, it will
3/3 serve as an example of information stifling that CI needs to liberate itself from. Anyone knows what the new CI textbooks say about PS? I would use it as a litmus test for authors understanding of modern CI. #causalinference #Bookofwhy @causalinf
Read 3 tweets
At the #SER2019 plenary, @sandrogalea & @_MiguelHernan are *not* debating about #socialepi and #causalinference because, despite what everyone thinks, they don’t actually disagree!

So, instead, they’re here to set the record straight.
If we’re talking about #socialepi, we need a definition. @sandrogalea starts broad: “social epi is the health effects of forces ‘above the skin’” aka everything outside your body
And similarly, #causalinference is about drawing conclusions about causal effects.
Read 20 tweets
Hey everyone! Here’s a #tweetorial on our new paper on why we often can’t make #causalinference using #distancetocare as an exposure or instrument! cc: @epiellie
There are 3 problems with #distancetocare as an exposure for estimating causal effects. Let’s walk through them.
The first problem that probably comes to mind is #confounding 🙀Choices about where to live and where to locate care facilities are complicated and depend on a lot of things we might not be able to measure (e.g. socioeconomic status).
Read 10 tweets
Inspired by a question from @Econdoor , a #tweetorial about instrumental variables and the exclusion restriction.

#econtwitter #epitwitter #FOAMed
First, instrumental variables and causal graphs:

An instrument is a special type of variable that is associated (causally or otherwise) with the exposure you’re interested in.

The directed acyclic graph (DAG) below shows the basic required structure.
There are three key requirements for a variable to be an instrument:
•it is associated with the exposure
•it is not associated with the outcome EXCEPT through exposure
•it does not share any causes with the outcome (ie unconfounded).

We have these 3 properties in our DAG👇🏼
Read 26 tweets
The last days, a fascinating discussion has been happening on #econtwitter & #epitwitter abt #causalinference, pot outcomes, & dir acyclical graphs.

Since #AcademicTwitter is great for open discourse, & bad at keeping all in 1 place, I thought I provide this public good..

1st off, no idea where it all began; so this isn't chronological as much as topical.

From an econ perspective, a good start to get acquainted w/ the idea behind DAGs may be @yudapearl's #BookofWhy or this book

(Disclaimer: Haven't read them yet.)

Also, see this earlier paper for a more technical intro:

As a 1st approximation, I found @PHuenermund's and @juli_schuess's slides helpful; to be found here: & here:

Read 18 tweets
Today’s #tweetorial—positivity for #causalinference: what it is and why it matters.

First, what it’s not: the type of positivity Im going to talk about is not a positive mental attitude!
The technical definition of positivity is that the probability of having a particular level of exposure, conditional on your covariates, is greater than 0 and less than 1, for all strata and exposure levels of interest:

0<P(A=a|L)<1 for all a in A and L
But what does that actually mean?

If you want to compare two types of treatment, then you have to have people in your data who are able to & sometimes will receive all relevant treatment options!
Read 20 tweets
Your composition formula is right ONLY if we assume no M--Y confounding.
But in general, even if M is the only mediator between X and Y, the formula does not hold. This is in fact a beautiful example how modern #causalinference deconstructs strongly held intuitions #Bookofwhy
Intuition says: We have two RCT studies, one gives P(Y|do(M)), the second gives P(M|do(X)), So, if we have a new study with X--->M--->Y the ONLY directed path from X to Y, we should forget confounding and chain the two causal effects together to get P(Y|do(X)). WRONG! We do
need to worry about confounding, but only Y--M, not X--M. It is also beautiful because it is not easy to show that the intuition is wrong, so I bet most traditional researchers would go wrong here. Funny, some researchers still refuse to believe in the revolution #Bookofwhy
Read 3 tweets
Since the word "estimand" seems to be giving readers problems, let me share my definition. First, it is indeed that which needs to be estimated, ..However, in the era of #causalinference, one needs to add something, to avoid confusion. #Bookofwhy
favorite text, or encyclopedia, or handbook. Why? Because those were written by statisticians prior to the causal era. For them, everything was properties of P, including the research question. The were not interested in causal questions which are not defined by P but by M.
One needs to add that an "estimand" is a property of the distribution that governs the observed data,e.g. E[X], var[Y], E[Y|X=x],. We need it because causal quantities e.g., E[Y|do(x)] or E[Y_x|Z=z] may also need to be estimated, but they are not expressed as properties of P.
Read 4 tweets
In the past two days readers had a chance to carefully examine (Appendix) and verify that health scientists communicate in terms of states of variables, as opposed to manipulations of variables. They talk about agents and substances being ``present''
or ``absent'', being at high concentration or low concentration, smaller particles or larger particles; they talk about variables ``enabling,''``disabling,'' ``promoting,''``leading to'' ``contributing to,'' etc. Branding these causal relationships "hopelessly ill-defined",
"extrascientific" or "White Magic" does not do justice to the bulk of scientific discourse. What for? Just because PO started with RCT perceived as the mother of all knowledge? #Bookofwhy #causalinference #EpiTweeter
Read 3 tweets
Your observation is super insightful. One of the major communication obstacles I have encountered with potential outcome (PO) folks is the notion that causal effects and counterfactuals, Y_x, are PROPERTIES OF OUR MODEL. They cannot swallow it because, in the PO framework
of Rubin (1974) there is no such a beast as a MODEL. All he had were conditional probabilities of potential outcomes {Y(0),Y(1)}.
Subsequently, those who entered #causalinference through a PO-
education suffer from the same deficiency -- no model.
#Bookofwhy. (cont....)
Instead, the potential outcome Y(1) is defined in the context of a real-life RCT, not a model of how a population responds to RCT. Hard-core PO folks, including Rubin's disciples in economics continue to operate in this model-blind conception. DAG-using Epidemiologists
Read 6 tweets
Your suggestion to start a conversation was taken seriously in (2018),which provides rigorous analysis of each of your
2016 arguments. In particular, it analyzes the logic of "consistency," the difficulties of defining "well-defined interventions,"
the utility of defining ideal mathematical constructs and ideal manipulations,
the semantics of multi-version interventions, and the practical benefits of attributing causal qualities to non-manipulable variables, from blood-pressure and temperature to gender and obesity.
The Appendix further examines ordinary conversations among health scientists, points out the ubiquity of non-manipulable causes and their communicational benefits. I have received no objection to any of these arguments and assumed their power and transparency convinced you
Read 4 tweets
I love fries; you love fries; but should we only eat 6 per serving? Less is more is good advice but why 6 & not 5, or 7?

Unfortunately, there’s prob no way to know if 5 or 6 is better!

Why? Here’s a #tweetorial on estimating causal effects for nutrition. Grab a 🥗 & get comfy!
Imagine you want to reduce your intake of French fries with the specific goal of reducing your chance of a heart attack.

You need to make 2 decisions: how often should I eat any fries; and how many fries should I eat in a serving?
To help you live your best life (ie eating max safe # of fries), researchers need to ask a pair of causal questions:

•what is the best frequency of French fry consumption to prevent heart attacks?
•what is the best serving size of French fries to prevent heart attacks?
Read 27 tweets
1/3 Readers ask: Why is intervention (Rung-2) different from counterfactual (Rung-3)? Doesn't intervening negate some aspects of the observed world?
Ans. Interventions change but do not contradict the observed world, because the world before and after the intervention entails ...
2/3 ... time-distinct variables. In contrast, "Had I been dead" contradicts known
facts. For a recent discussion, see <>
Remark: Both Harvard's #causalinference group and Rubin's potential outcome framework do not distinguish Rung-2 from Rung-3. (cont.)
3/3 This, I believe, is a culturally rooted resistance that will be rectified in the future. It stems from the origin of both frameworks in the "as if randomized" metaphor, as opposed to the physical "listening" metaphor of #Bookofwhy
Read 3 tweets
I recently finished 'The Book of Why' by @yudapearl
I would highly recommend it to everyone interested in causality and scientific explanations. /THREAD
The book is rather accessible, so you don't need to be a #causalinference nerd to follow the arguments. At the same time, there is a lot to learn even if you have struggled your way through Causality (CUP, 2009) and the technical papers
I have been an early converter and have used the 'structural' approach to causality to organize my own book on research design in political science.
Now I am glad I have an accessible book to point students to for more info and background.
Read 10 tweets
Interested in thinking more deeply about what a “cause” is? Causation: A Very Short Introduction from @ranilillanjum and @SDMumford is a great starting point.

Thread recap starting now👇🏽#twittereview #epibookclub #causation #causalinference #epimethodsclub
Chapter 1 sets up the problem: causation is hard to define, both in general & for specific events. It’s more than just temporal ordering, but is it a separate thing?

An ex., a town gets sick after flood of rats. Did the rats cause sickness? Maybe it was actually a sick visitor.
Chapter 2 introduces us to Hume’s theories of causation.

1st, regularity: sometimes things are *regularly* followed by other things.

Is this causation? If so, how regular is regular enough? How many times do we have to observe steam after heating water to say it’s causal?
Read 23 tweets
Wow, #epibookclub, we’re officially halfway through the #bookofwhy!

Chapter 6 is a doozy... we’re going to learn all about paradoxes!! Get ready to have your mind blown!! 🤯🤯🤯
Given last week’s tragic fail on paradox gifs, I made you our very own #epibookclub gif! It’s paradox time!!

(Please be kind this is my very first homemade gif 😂😂)
I like the chapter set up this week: @yudapearl explains importance of spending time working out the paradoxes...they can give us insight into how people process causal information. He says these are paradoxes b/c they straddle the rungs of the #ladderofcausation

So, up we go!
Read 25 tweets
Judea Pearl has a new book (with Dana Mackenzie).…

To me, Judea is an intellectual hero. My life changed after hearing him at Harvard over 20 years ago. Like many of us in #causalinference, I owe so much to him.

And yet I disagree with him on a key issue.
Pearl believes that any causal effect we can name must also exist.

To him, the meaning of “the causal effect of A on death” is self-evident. He says we can quantify, say, the causal effect of race or the causal effect of obesity.

I don't think we can.
We cannot estimate "the causal effect of obesity" because we don't know what that means.

For the causal effect of A to be well defined, we need a common understanding of the interventions that we would use to change A. Otherwise, the effect is undefined.…
Read 6 tweets

Related hashtags

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!