,
18 tweets,
12 min read
Read on Twitter
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..
1/19
Since #AcademicTwitter is great for open discourse, & bad at keeping all in 1 place, I thought I provide this public good..
1/19
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 amzn.to/2WuWf7d.
(Disclaimer: Haven't read them yet.)
2/19
From an econ perspective, a good start to get acquainted w/ the idea behind DAGs may be @yudapearl's #BookofWhy or this book amzn.to/2WuWf7d.
(Disclaimer: Haven't read them yet.)
2/19
Also, see this earlier paper for a more technical intro: bit.ly/2uwoRBa
As a 1st approximation, I found @PHuenermund's and @juli_schuess's slides helpful; to be found here: bit.ly/2uxW3Ze & here: bit.ly/2uwnGBK.
3/19
As a 1st approximation, I found @PHuenermund's and @juli_schuess's slides helpful; to be found here: bit.ly/2uxW3Ze & here: bit.ly/2uwnGBK.
3/19
In my mind, what ppl like @yudapearl, @eliasbareinboim, & @PHuenermund argue is w/ DAGs you:
1st, use formal language of graphs to articulate knowledge abt causal relationships.
2nd, you evaluate the identifiability of the outcome w/ all hands on deck re: assumptions.
4/19
1st, use formal language of graphs to articulate knowledge abt causal relationships.
2nd, you evaluate the identifiability of the outcome w/ all hands on deck re: assumptions.
4/19
There are 2 confounding/selection biases that come up time & again as examples for how DAGs can be helpful:
Collider bias & M-bias.
5/19
Collider bias & M-bias.
5/19
1) Collider bias
A → B ← C
Let A be sb's looks & C be their talent. If either can bring you to Hollywood (B = job), among the stars, they may be neg correlated but could be independent in the pop!
If B is a collider, A & C are uncond indep, but become dep cond on B.
6/19
A → B ← C
Let A be sb's looks & C be their talent. If either can bring you to Hollywood (B = job), among the stars, they may be neg correlated but could be independent in the pop!
If B is a collider, A & C are uncond indep, but become dep cond on B.
6/19
What @causalinf noted: is core of @conjugateprior @jonmummolo, et al's critique of Fryer (2018) on racial bias in police arrests.
@cdsamii notes, Heckman's selection problem is just cond on a collider's descendant: bit.ly/2Ue7tzB (p.19).
7/19
@cdsamii notes, Heckman's selection problem is just cond on a collider's descendant: bit.ly/2Ue7tzB (p.19).
7/19
2) M-bias
As best as I understand, bias introduced by conditioning on a pre-treatment covariate in the presence of a particular M-structure: between a treatment, an outcome, two latent factors, & the covariate as "collider".
(Pic, HT @EpiEllie)
8/19
As best as I understand, bias introduced by conditioning on a pre-treatment covariate in the presence of a particular M-structure: between a treatment, an outcome, two latent factors, & the covariate as "collider".
(Pic, HT @EpiEllie)
8/19
This is sth @EpiEllie pointed out as being one of the main selling points for DAGs to econs; see here: (also, great examples from epidemiology and sociology!)
(Helpful explainer of R package ggdag to play around w/ DAGs: cran.r-project.org/web/packages/g….)
9/19
(Helpful explainer of R package ggdag to play around w/ DAGs: cran.r-project.org/web/packages/g….)
9/19
Both types of biases are not entirely new to econs but have arguably been neglected (e.g., treated as "bad controls" in Mostly Harmless Metrics, mostlyharmlesseconometrics.com/2009/05/commen…).
This allusion by @autoregress to Frisch-Waugh is a helpful econ example: .
10/19
This allusion by @autoregress to Frisch-Waugh is a helpful econ example: .
10/19
What everyone agrees on:
i) DAGs help clarify biases in an intuitive way!
ii) DAGs are helpful in teaching causality w/out the need of abstract math.
ii) DAGs can be helpful in checking the assumption of a model that has been built prior by use of contextual knowledge.
11/19
i) DAGs help clarify biases in an intuitive way!
ii) DAGs are helpful in teaching causality w/out the need of abstract math.
ii) DAGs can be helpful in checking the assumption of a model that has been built prior by use of contextual knowledge.
11/19
What's contested:
i) DAGs are useful, & that would be enough for paper to be seen as novel. But to make it part & parcel of syllabus is quite another thing. So, what are ways in which great econ papers are flawed that DAGs help resolve? (Along the lines of @Jabaluck)
12/19
i) DAGs are useful, & that would be enough for paper to be seen as novel. But to make it part & parcel of syllabus is quite another thing. So, what are ways in which great econ papers are flawed that DAGs help resolve? (Along the lines of @Jabaluck)
12/19
ii) Relatedly, econs seem to use @yudapearl's backdoor criterion extensively. What is a concrete example of the frontdoor criterion put to good use in econ? (Point made by @MartinRavallion , here: )
13/19
13/19
iii) I think @Jabaluck's main critique is one of sequence: Fair if, 1st, we come up w/ "ready-made" DAGs (IV, RDD, Diff-in-diff, etc.) based on contextual knowledge & then use DAG to check assumptions. This adds to what micro econs do today. But don't turn it around!
14/19
14/19
iv) Complexity is an issue. When @eliasbareinboim contends that DAGs are helpful w/ "toy examples" of 4 vars but indispensable w/ a "100-var instance", @Jabaluck retorts that this is not a level of complexity that is useful for causal research ().
16/19
16/19
@causalinf had some helpful summaries, here: & here: .
@cdsamii had a fascinating application of the collider problem to FEs in panel data, here:
17/19
@cdsamii had a fascinating application of the collider problem to FEs in panel data, here:
17/19
V much hope this is a fair representation of everyone's views.
Happy to stand corrected if not.
18/19
Happy to stand corrected if not.
18/19
Bottomline:
1. I'll read much more on this
2. #AcademicTwitter is the best Twitter
3. I'm massively indebted to all the ppl linked above for their input
4. I now follow a couple ppl on #epitwitter
5. There's a thing called tough love and econs are really good at it..
/FIN
1. I'll read much more on this
2. #AcademicTwitter is the best Twitter
3. I'm massively indebted to all the ppl linked above for their input
4. I now follow a couple ppl on #epitwitter
5. There's a thing called tough love and econs are really good at it..
/FIN
