Hi
#Econtwitter
, I'd like to share the slides from the PhD Applied Econometrics course I just had the privilege to teach at
@AreBerkeley
Regression & causality, selection on observables, panel data, IV, RDD --- usual topics but hopefully in a modern way
It's time to announce that this summer I'll be moving to Berkeley as an assistant professor
@AreBerkeley
and affiliated faculty at
@berkeleyecon
.
I'm thrilled for this new chapter but also sad to leave London and
@EconUCL
and very grateful for my four years here!
A plea to referees: don't push the authors of diff-in-diff papers to mindlessly implement all estimators.
Push them to be explicit and precise about their estimand and the underlying assumptions, and choose estimator(s) based on those primitives.
🥁 I’m thrilled to announce our paper w/
@XJaravel
and
@jannspiess
, “Revisiting Event Studies: Robust and Efficient Estimation” 🥁
It’s a fully revised version of our 2017 draft that the diff-in-diff loving audience may have seen
On the 1st anniversary of my twitter account, I'm happy to share w/the
#EconTwitter
community our
#stata
commands for event studies:
did_imputation: robust and efficient imputation estimator
event_plot: event study plots after various estimation methods
Very happy to report that "Non-Random Exposure to Exogenous Shocks," our beloved work w/
@instrumenthull
, has been accepted at
@ecmaEditors
A one-line summary? If you build your treatment or instrument by a formula, you should recenter it!
Link to draft:
This is Bartik-the-Cat.
And he should look happier than this, as our paper on shift-share ("Bartik") instruments with
@autoregress
and
@jaravel
has just been accepted at
@RevEconStud
. (My first pub, fwiw!)
Time to report that our work with
@XJaravel
and
@jannspiess
on diff-in-diff imputation has recently been accepted
@RevEconStudies
!
Final draft of "Revisiting Event Study Designs: Robust and Efficient Estimation" here:
New WP: “Non-Random Exposure to Exogenous Shocks” (w/
@autoregress
). Summary 🧵:
Papers often estimate causal effects by leveraging exogenous shocks that affect many observations jointly, to different extents
We show problems w/this & offer new solutions
📢 I’d like to share with
#TradeTwitter
a 🧵 on what
@XJaravel
and I have learned about the unequal effects of international trade through both cost-of-living and wages in the U.S.
For those of you who have seen my JMP, this is a much-revised draft
#Stata
commands for event study estimation and plotting based on my work with
@XJaravel
and
@jannspiess
are now available directly from SSC:
ssc install did_imputation
ssc install event_plot
Let me know if you notice any bugs!
On the 1st anniversary of my twitter account, I'm happy to share w/the
#EconTwitter
community our
#stata
commands for event studies:
did_imputation: robust and efficient imputation estimator
event_plot: event study plots after various estimation methods
📢
@XJaravel
,
@jannspiess
and I just posted a new draft of "Revisiting Event Study Designs: Robust and Efficient Estimation"
As before:
- Challenges of diff-in-diff estimation by OLS
- Robust and efficient imputation estimator
NEW:
- Application to MPC!
What did you always want to know about shift-share designs but were afraid to ask?
@instrumenthull
,
@XJaravel
, and I are writing a practical guide to shift-share IVs.
Let us know, and we may add some of your questions to the Q&A part of the paper!
I often receive emails with questions on did_imputation & other tools I've worked on
From now I will ask in return for an optional donation to Ukranians suffering in the war, e.g. via
@Mylovanov
's KSE or
@sguriev
's True Russia
Does $25 for students, $75 for faculty sound right?
🧵 Do shift-share IV regressions suffer from negative weight problems?
@CdeChaisemartin
and Lei recently posted a WP (Worrying Paper, one could say) arguing that way
@instrumenthull
and I are more optimistic and decided to share our view in a brief note
Мы считаем, что действия руководства России наносят огромный ущерб будущему России. Развязывая войну против Украины, руководители России действуют против интересов российских граждан. Мы требуем немедленного прекращения агрессии!
Stata commands for robust diff-in-diffs are in the news this week, so
@XJaravel
,
@jannspiess
, and I have our own:
A new version of
#did_imputation
available from SSC, with more convenient treatment effect heterogeneity analysis
More diff-in-diff news, anyone?
You know two-way FE regressions can have negative weights with staggered rollout
Perhaps you've heard the same can happen with simultaneous treatment but extra controls
But it is also possible without controls, just when the panel is incompete!
A tweet for everyone using robust diff-in-diff estimation
We have updated our
#did_imputation
command on SSC. Main changes:
- Leave-out SEs (recommended!)
- Coefficients on controls are now reported
- Bugs fixed
Hi Event-Study-with-Staggered-Adoption-
#EconTwitter
Could you please share with me (by DM/email/here) your code for generating the appropriate dummies, estimating the event study coefficients and plotting them?
And a reference to the corresponding figure in your paper. Thanks!
Jon is doing a good service to the profession by this clarification. But it's discouraging when empirical researchers, who should be looking for the best answers to their questions, don't bother to understand the methods they use - or even read the helpfile to the Stata packages!
#EconTwitter
Have you replaced your dynamic TWFE event-study with an event-study from one of the recent DiD methods for staggered timing?
If so, you may be interested in this short note on interpreting event-studies from these new methods. A short 🧵
I don't expect 2024 to be easy but at least we don't have to worry about negative weights in design-based OLS and IV specifications.
Come see
@instrumenthull
present our new paper at the AEA, or just read (it's only 5 pages!) if interested
Happy New Year! Kirill
@Borusyak
and I have a New (short) Paper on the infamous "negative weights" issue recently raised for TWFE and other popular OLS/IV specifications
Here's an (even shorter) summary thread
Very dark days in Russia, with
@navalny
arrested and many thousands detained for peaceful protest.
Among those detained was my friend Grigory
@Franguridi
. He's an econometrician and will be on the job market next year. Keep an eye out if you want to hire good people
If you used
#did_imputation
for treatment effect heterogeneity analysis via the project() option, please reinstall the command from SSC.
There was a bug in it when the sample includes always-treated units or periods. Many thanks to Zhihong Chen for reporting it!
A new WP & talks annoucement!
@autoregress
and I have just posted "Non-random Exposure to Natural Experiments", and I'll be presenting it on July 22 @ FREIT seminar and July 27
@osus_info
. All are welcome to tune in, details are below, and a summary thread will follow later
Very happy for our work w/
@XJaravel
to come out in the JIE!
We find that trade shocks primarily induce "horizontal inequality" in the US: welfare effects that mostly vary within, and not across, groups of initial earnings.
Why? 1/2
New: "Are trade wars class wars? The importance of trade-induced horizontal inequality" by Kirill Borusyak (
@borusyak
) and Xavier Jaravel (
@XJaravel
).
1/2
We wrote a review paper on identification with shift-share & other formula instruments.
One goal was to illustrate the general insight--recentering helps--in the context of linear shift-share IVs, giving a different perspective on them compared to our ReStud.
Comments welcome!
Kirill
@borusyak
, Xavier
@XJaravel
, and I have a new review article on shift-share instruments and other "formula" treatments/instruments
Check it out here! (comments welcome)
Now on the plotting command:
With event_plot, you can make event study plots for at least 5 methods:
- our did_imputation
- robust estimators of
@CdeChaisemartin
-D’Haultfoeuille, Callaway-
@pedrohcgs
, Sun-Abraham
- traditional OLS
And combine them with each other if you wish!
Anyone interested in my work with
@XJaravel
and
@jannspiess
on robust and efficient estimation in event study designs can now watch me present it on Youtube.
Thanks
@taylor_wright
for the invitation to the DiD reading group
Some are happy for the Peace prize today.
Others are angry that
@navalny
didn't get it.
I'm just sad that free journalism in my country deserves a Nobel prize.
In case anyone's interested, the video of my talk (Non-Random Exposure to Exogenous Shocks w/
@instrumenthull
) and Michal Kolesar's discussion is available from NBER for two weeks:
(from the start to 55min)
Slides are here:
Q to Econtwitter:
What's the simplest paper you know that uses a convincing shift-share design with exogenous shocks?
By this I mean:
👉 transparent shocks arising from a natural experiment (or RCT)
👉 simple OLS or IV specifications, without tons of arbitrary controls & FEs
Done with my first PhD lecture ever, in a topics trade class (today - on the spillover effects of the China shock).
Very enjoyable to teach things I've been thinking about, through the work of the giants + my own. Grateful for this in otherwise rather difficult times.
A question to anyone who have used our
#did_imputation
Stata command for diff-in-diff estimation:
We now have new options to conveniently analyze heterogeneous treatment effects. Would anyone be interested to beta-test them?
Please DM with your email or leave a comment!
For anyone worrying how to implement the new DiD techniques in Stata, you can contribute by testing our commands for event study estimation and plotting w/
@XJaravel
and
@jannspiess
in your application.
Just email or DM me, and let's catch all the remaining bugs!
Bitsy has a great idea. So here’s what I’ll propose. If someone will code up did in stata, the beautiful interface with the same options, then I will organize a fund me on Twitter to get donations to you. You will need to listen closely to authors of R package but we can do it
This popular thread looks well intentioned but misleading.
Design-based methods, which derive from a natural experiment in X, are great.
But diff-in-diff, as conventionally understood, is not one of them! [1/3]
Many papers I see as editor or referee still are not clear about the source of identification. My sense is many authors do not understand what is meant by this and how much it matters in modern empirical work.
To help level the playing field, I will try to explain in this thread
Call for Applications:
PhD Econ/Management at LSE
- Based at the Managerial Economics and Strategy Group
- First years joint with Econ Dpt
- Scholarships available
- Application deadline: 14 January 2022
#EconTwitter
More details here:
NBER Labor Studies SI has been fascinating so far. Tune in tomorrow for the Methods Session (livestreamed on Youtube)!
I'll present "Non-Random Exposure to Exogenous Shocks" w/
@instrumenthull
at noon ET; excited to see work by
@jannspiess
and others
Shift-share instruments are used to study robots --- not the other way round
(Stay tuned for the practical guide on shift-share IV that
@instrumenthull
@XJaravel
and I are writing without ChatGPT)
Had a very amusing discussion with chatGPT hitting on
@TimBartik
,
@lkatz42
, and a passing reference to
@paulgp
. I loved the first response to explaining "Bartik instruments". Full chat here
UCL (
@EconUCL
) is hiring, and we have several tenure-track positions!
#econtwitter
Applications are open in all fields, including financial econ. Please encourage candidates to apply by Nov 15. Interviews will be held at the European Job Market in Dec.
For
#econjobmarket
candidates:
@AreBerkeley
is hiring an AP in energy/enrivonmental/resource economics!
Our listing:
My colleagues and I will be happy to answer any questions
Here's my version of this very important point about negative weights:
You should consider the nature of treatment assignment and reasonable restrictions on treatment effects before worrying about negative weights, especially when you can’t avoid them by imputation
➡️🧵
It makes me sad when the results on event studies, and similarly shift-share IV, are viewed as "they came for..."
Both literatures are constructive, not destructive!
They give you appropriate language(s) & practical tools. Why not learn and adopt? The authors are here to help
First they came for event-study graphs in Diff-in-Diffs analysis. Now they're coming for Becker-style outcome tests for detecting bias. What will be left of my empirical training by the end of 2020?
A new, short, version of our non-random exposure paper w/
@instrumenthull
The paper studies treatments & instruments for which some determinants are (as-good-as-)random but their construction involves other variables, too
Kirill
@borusyak
and I just posted a revision of Non-Random Exposure to Exogenous Shocks
Version 2.0 is short (23 pages!) & focuses on the main OVB+recentering message. But fear not: we've updated the 125page working paper too :)
It's time for a reminder that the problems with DiD apply NOT ONLY with staggered treatment timing.
Similar issues arise in conventional DiDs with controls, especially w/unit-specific trends. Some of the solutions equally apply.
(I doubt that's what referees had in mind though)
referees who keep suggesting Calloway and Sant'Anna (2019) and Goodman-Bacon (2019) when the treatment happens to everyone in the same year:
plz stop it
For anyone interested, the recording of my talk at the NBER trade conference today, "Understanding migration responses to local shocks" (w/
@dix_rafael
and Brian Kovak), is currently available here:
If you are at the AEA this week, Peter
@instrumenthull
is presenting our early-stage work on structural estimation using natural experiments. We’d love any feedback, especially if you are an IO economist!
And if you are not at the AEAs, talk to us about it if this is of interest
First, on Friday morning I'll be presenting new work with
@borusyak
, on using natural experiments to estimate structural models like BLP
This builds on our recentered IV paper (new draft! ) and was the subject of this vaguetweet:
To add to what
@XJaravel
explained, shift-share designs involve lots of practical issues which can really matter. We explain what to do when shares do not add up to one, with panel data, with controls, etc. We are always happy to answer questions about your specific setup.
[1/n] Very happy to circulate a revised version of our paper on shift share design, with the brilliant
@borusyak
and
@autoregress
.
You can find it here:
A q to labor economists out here: what's the most common way to correct for top-coding of income, e.g. in the population census?
E.g. if I want to compute regional inequality measures or the skill premium? Or does everyone just ignore top-coding? Any good references?
For anyone interested, I'm presenting our work with
@instrumenthull
on design-based estimation of structural models, in particular mixed logit demand
Tomorrow (Saturday), 2:30pm at
#ASSA2024
, session with the great Vera Semenova, Tim Armstrong, and
@eric_auerbach
Thanks
@kylefbutts
for coding up a very nice R version of our ssaggregate Stata package for shift-share IV!
Now you can estimate SSIV regressions at the level of quasi-experimental shocks in R too! For the benefits of this see
Alright, let me try sharing 10 personal favorites in classical music, broadly defined (including modern, minimalist, and vocal or choral)
Ordered by how much I'd like everyone to listen to them
@analisapackham
I think, if anything, this new literature shows that TWFE is salvageable as a method of causal inference. We keep finding issues, sure, but really smart people keep coming up with solutions. I’m more encouraged than anything. Although it is kind of a pain to keep up with.
A question to structurally-minded economists:
What's your favorite simple example of a Roy model put to use?
I'd like a setting where, under reasonable assumptions that are rooted in economic theory and not inherently parametric, a Roy model identifies something that IV can't.
Hi
#Econtwitter
, do you know if I can pre-register an observational study, similarly to how the AEA holds the RCT register?
Any links or past examples would be appreciated.
Want to try randomization-based confidence intervals and specification tests?
PM/email me or
@autoregress
for our new
@stata
commands. We'd love user feedback!
Input: sets of counterfactual exogenous shocks (e.g. shock permutations)
Output: CI for OLS/IV + spec.test pvalues
A bonus to specifying the shock DGP? Randomization inference for free
This is useful when observations are “fuzzily clustered” by common exposure to the same observed and unobserved shocks
RI gives exact confidence intervals (under constant effects) and new specification tests
Anyone interested can now watch Brian Kovak's excellent talk at IZA on the paper he,
@dix_rafael
, and I have been working on:
"Understanding migration responses to local shocks"
with the great
@gordon_h_hanson
's talk right after. Comments most welcome!
Don't miss the comprehensive slides on Shift-Share IV prepared by
@autoregress
Check out to learn:
- the two alternative paths to identification (both formally and intuitively) and when they apply
- how to get standard errors
- practical issues, extensions, applications & more
By popular request, I've posted the slides from last week's lecture on Shift-Share IV methods here:
Thanks again to
@arindube
for the opportunity to develop these for his course. Questions/comments/other guest lecture opportunities are all very welcome!
Hi
#Econtwitter
, please share advice on how to make research assistance by a PhD student most productive!
What worked (or didn't work) for you would be great to know:
- Induction process
- Time management
- Types of tasks, etc.
If you are interested in international trade or organizational economics, please welcome to
#econtwitter
and follow
@FrankPisch
!
For some of you he needs no intro; for the others he's a
@CEP_LSE
PhD, now professor at TU Darmstadt, and my dear friend.
I look forward to reading this carefully soon and understanding how it relates to the results of our ReStud paper w/
@instrumenthull
@XJaravel
Our Appendix A.1 showed that the shift-share estimand is convexly weighted under pretty standard as-good-as-random assignment
I am excited to share this paper on panel Bartik regressions. Headline result: without assuming constant effects, one cannot conclude that Chinese imports reduced US manufacturing employment from the data used by Autor el al in their China shock paper.
Stacked DiD is robust & transparent but too flexible
Consequences? It's arbitrary and inefficient
A model of Y(0) formalizes which comparisons are valid & yields efficient estimators for any given estimand
Stay tuned for our implementation of this idea w/
@XJaravel
@jannspiess
Stacked DID is very flexible. It depends on whatever YOU think is admissible for your control group! You can only allow never-treated. Or treated more than a 10y ago. Or those who share certain observables. Up to you!
What it does is to be transparent about choice of controls.
Honored and excited to present my work with
@autoregress
to the development econ community at the "BREAD et al." conference, surrounded by very distinguished speakers
"Non-random exposure to exogenous shocks" @ noon ET on Friday Oct 2. Attendance open to all
On 1-3 October, join BREAD et al's virtual Conference on Development Economics which is open to all who are interested.
See the full programme and signup here:
#EconTwitter
#DevEcon
A nice compilation of recent applied micro methods!
I'd also include the papers by
@ArkhangelskyD
and
@guido_imbens
on double robustness in DiD and by
@Susan_Athey
et al. on matrix completion
Excited to try the R implementation of our
#did_imputation
estimator, along with the alternative methods, that
@kylefbutts
has generously prepared!
(Although I wouldn't label the new event study methods "robustness checks" 🙃)
Conference on Urban and Regional Economics is happening this week. Looking forward to its unusual format: 10min talk + 50min discussion.
I'll present "Non-Random Exposure to Exogenous Shocks" (with
@autoregress
) on Sat at noon ET; registation is free
But what about the other estimators then? Is there ever a reason to use them???
My paper (pt 1):
YES! If treatment adoption is non-staggered and the outcome has “persistent shocks” (Random Walk errors), then in fact csdid/did_multiplegt are THE best unbiased estimators.
6/N
This Wednesday on Zoom: a webinar on recent advancements in applied IV methods
In 2 hours you can see four papers! I'll talk about non-random exposure to exogenous shocks, based on the work with
@instrumenthull
, who'll also be the harsh moderator
Very excited for Wednesday's
@YoungStatS2
webinar on Recent Advancements in Applied IV Methods, featuring
@borusyak
,
@paulgp
,
@stephencoussens
, and Alex Torgovitsky!
We have 600+ attendees so far but still have plenty of room in the zoom! Register here:
Q to development folks out here: could you please recommend papers that:
➡️ Estimate spillovers by regressing the outcome on the % of treated neighbors, and
➡️ Randomize treatment at the unit level, so % of treated friends is not directly randomly assigned (as in a two-tier RCT)
🚨New working paper tweet🚨
Many questions in economics involve the causal effects of treatments which are computed from multiple sources of variation, according to a known formula. Consider three examples. First, when estimating spillovers from a randomized intervention, one
Excited to see our
@voxeu
piece out, on the effect of schools on the covid spread (w/
@claravobi
and Uta Schönberg). Thanks to
@BaldwinRE
for the interest and
@econromesh
for the help in the process.
Talks in Europe: you don't need masks, tests, forms. Just come, we trust you!
A talk in the US: you have to present on Zoom. We can't have you on campus: you only got two Sputniks, two Modernas, and had covid recently, so you don't count as boosted! And an exemption was denied
Jan Bakker is our great job market candidate in trade and urban
@EconUCL
Check out his JMP on why large cities specialize in exporting, and how agglomeration forces shape the unequal gains from international trade across space
More papers are on his website. h/t
@TradeDiversion
Find out about our fantastic Job market candidates.
Economics Postdoc Jan David Bakker’s research interests include International Trade, Urban Economics & Development Economics. Learn more about Jan & our Candidates at
Context: this plot (from my work w/
@claravobi
and Uta Schönberg) is generated by just 2 lines of Stata code: 1 for estimation (by the method from my work w/
@XJaravel
and
@jannspiess
) & 1 for plotting
But I’d like to make the plotting command compatible w/other estimation methods
The great multi-year effort by
@alterelim
et al. on the mental health survey of European econ departments is complete.
Time to review the detailed data and think how to improve PhD student (and junior faculty) life
@EconUCL
Today,
@ClaraSievert
,
@vbolotnyy
, Paul Barreira, and I shared results of our study of mental health in European Economics Departments with the 14 participating departments.
A thread about what we did and learned.
The executive summary is here:
1/N
To recap:
1) Event studies are fine but be clear about the assumptions
2) Don’t use old-school OLS
3) The imputation estimator is robust to treatment effect heterogeneity and to pre-testing, efficient w/homoskedasticity & transparent
Comments welcome!
Very glad to hear our event study estimation command is fast
But there is still room for improvement. If anyone who knows Stata programming could volunteer to help us speed it up, it'd be a great public service
While everyone else is invited to try the command as is!
To estimate causal effects with care,
Recentering can help us prepare,
Simulating what could have been,
In partially-randomized scene,
The true effect we can then declare.
My new favorite genre of poetry is ChatGPT-generated poems about econometric methods:
There once was a method quite fine
They called it regression kink design
It's a graph with a bend
That helps us comprehend
The causal effects that we need to find
Thanks everyone for interesting suggestions on the Roy model.
What I was looking for was Adão (2016): a transparent analysis of both selection and outcomes, both non-parametric and cleanly parameterized.
A question to structurally-minded economists:
What's your favorite simple example of a Roy model put to use?
I'd like a setting where, under reasonable assumptions that are rooted in economic theory and not inherently parametric, a Roy model identifies something that IV can't.
@toniwhited
Let me join those who say that pushing one button OBVIOUSLY must reproduce all the results in any published work
Easier implementation (and ways to deal with proprietary data replications) are most welcome, but it's pretty lexicographic to me, regardless of disparate impacts
An event for those of you interested in event study designs.
A new draft of our paper with major revisions (and with the great
@jannspiess
on the team) is also coming soon
#econtwitter
I'm setting up another discussion on event study models!
This time we'll discuss great
@borusyak
and
@XJaravel
paper
If you already DMed for last one, I'll include you in invite, if not, DM & I'll add you!
In about 2 weeks; all welcome!
Great to see this paper finally published!
Besides the empirical findings, it enriched the modelling toolbox by Non-Homothetic CES preferences which I've become a big fan of (some new insights about using them in the trade context to follow)
Why is economic growth accompanied w changes in the sectoral structure of economies? Micro (~ US & India) & macro data (~ 40 countries) suggest main driver is Engel curves: systematic relations between household income and sectoral composition of demand