Time to celebrate! “DID estimators of intertemporal effects” accepted at @restatjournal, and very very fast did_multiplegt_dyn Stata & R commands available from SSC & CRAN, thanks to Mélitine Malézieux @fe_knau
@diegociccia1 & @DSDoulo! Celebratory thread! papers.ssrn.com/sol3/papers.cf…
The paper proposes DID event-study estimators, for complex designs where treatment might be binary, discrete, or continuous, groups may experience multiple changes of their treatment, and outcome may be affected by current and lagged treatments. Widely applicable!

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.
https://t.co/IRP4QZ1WTlpapers.ssrn.com/sol3/papers.cf…
We start by showing that panel Bartik coeff theta^b not robust to heterogeneous effects.

With T=2, theta^b=sum_g Delta Y_g (Delta Z_g-Delta Z_.)/sum_g Delta D_g (Delta Z_g-Delta Z_.).

But if Y_gt=Y_gt(0)+alpha_gt D_gt, Delta Y_g= Delta Y_g(0)+alpha_g2 D_g2-alpha_g1 D_g1, so:

New version of our survey on TWFE and DID, with an application: we revisit @JustinWolfers's AER paper on unilateral divorce laws, using several new DID estimators. papers.ssrn.com/sol3/papers.cf…
Dofile may be useful for those trying to implement those estimators: drive.google.com/file/d/156Fu73… Conclusions obtained by @JustinWolfers using TWFE regression stand, perhaps because TWFE fairly robust to hetero effects in this application (no negative weights, few contamination weights). Due to simple design: binary and staggered treatment, many never treated states.

Sorry for other instance of self-promotion, but I wanted to flag new paper with Xavier, Félix, and Gonzalo. Proposes DID estimators for treatments continuously distributed at every time period as is often the case of trade tariffs, or precipitations 1/n papers.ssrn.com/sol3/papers.cf… Main insight can be conveyed with 2 time periods. We assume there are movers: units whose treatment changes from period 1 to 2, and stayers: units whose treatment does not change. Our estimators compare outcome evolutions of movers and stayers with same period-one treatment 2/n

You feel a bit lost in the recent whirlwind of TWFE/DID papers? Our new, short survey of this literature may help you! papers.ssrn.com/sol3/papers.cf…
Here is a short thread on what is/isn't in the survey 1/n We start by reviewing papers that have:
i) shown that TWFE regressions may not be robust to heterogeneous treatment effects,
ii) explained why/when TWFE regressions may not be robust,
iii) proposed diagnostic tools researchers may use to assess their TWFE’s robustness.

2/n

As @essobecker reminded me, did_multiplegt turned one year! I had forgotten, what a bad parent… It has grown so much! E.g. it can now estimate dynamic effects even if treatment not binary or design not staggered, following new paper with Xavier: papers.ssrn.com/sol3/papers.cf… 1/n Syntax: did_multiplegt y g t d
y=outcome, g=group (e.g. county), t=time, d= treatment. With robust_dynamic option, estimators in our new paper are computed; without it, estimators in our previous paper are computed: aeaweb.org/articles?id=10… So, what’s the difference? 2/n

5 years ago, a referee for aeaweb.org/articles?id=10… spotted issue with our econometrics. His/her comment gave rise to papers.ssrn.com/sol3/papers.cf…, now accepted at Econometrica! We complain, often for good reasons, about review process, but can also be fruitful. Thread on paper. 1 Oversubscribed treatments often allocated by randomized waitlists. Applicants are ranked randomly. If, say, 3 seats available, first 3 get initial offers. However, some of those 3 may be non takers who decline treatment => next applicants get offer, until all seats filled.

For all pair of time periods t-1 & t, command computes DID comparing outcome evolution among switchers, groups whose treatment changes from t-1 to t, & groups whose treatment doesn’t change. DIDM = average of DIDs across all pairs of periods. 3/n Under parallel trends assumption, DIDM consistent for switchers treatment effect. DIDM robust to heterogeneous treatment effects, contrary to two-way FE estimator (to diagnose two-way FE: twowayfeweights Stata command) 4/n

#econtwitter: Want to produce event-study graph like this one, using estimators robust to heterogeneous treatment effects, across units or over time? You may want to check our new did_multipleGT #Stata package. "ssc install did_multipleGT" to install package and help file. 1/n did_multipleGT can be used in DID designs with multiple groups and periods, where all units in same group and period have same treatment (sharp designs). E.g.: treatment is county- or state-level variable. Treatment does not have to be binary. 2/n

So here is a thread to explain why/when the fuzzydid @Stata package may be useful. Many papers use regressions with group and time fixed effects to estimate effect of treatment on an outcome. In papers.ssrn.com/sol3/papers.cf… , we conduct a lit review and find that 20% of applied papers published in the AER between 2010 and 2012 estimate these regressions.