Causal Inference -- 20/23 -- Staggered Adoption: The Bacon Decomposition

This series of online lectures covers the most important causal research designs in economics and other social sciences. This is the fourth of five videos on Difference-in-Differences (DiD). The video discusses the pitfalls of staggered adoption designs.
The course is based on two textbooks:
- Mostly Harmless Econometrics by Joshua Angrist and Steve Pischke (2009): press.princeton.edu/books/pap...
- Causal Inference -- the Mixtape by Scott Cunningham (2020): mixtape.scunning.com/
The slides for the course can be downloaded here: 8650ac82-9ca9-4f0f-936f-a74eb...
Course outline:
1 Basics of causal inference (2 videos)
2 Instrumental variables (5 videos)
3 Marginal treatment effects (4 videos)
4 Regression discontinuity and kink designs (4 videos)
5 Difference-in-differences (5 videos)
6 Synthetic control (2 videos)
Get in touch:
- follow me on twitter: / ben_elsner
- website: www.benjaminelsner.com
- profile on linkedin: / benjamin-elsner-b71b98bb

[@user-xc1ng2nv3p]

the BEST video for staggered DID on KZbin

[@rhyswilliams8141]

Very clear and helpful - thanks!

[@domfree3299]

thanks!

[@sangelstefan]

Hi Ben, thanks for your very insightful lecture. I was wondering if you could provide me with some thoughts or references on the following 2 questions: 1) Most staggered diff-in-diff examples refer to a situation where policies change in a stag-gered way at a regional level. But does the same argument also occur if we use individual level panel data and estimate 2-way FE models (individual & time fixed effects) for questions like "the effect of marriage on subjective well-being; effect of health deterioration on con-sumption spending etc.", i.e. where both the independent variable and the dependent var-iable are measured at the individual level? 2) Is the problem of the same relevance if e.g. the independent variable is metric at the mi-cro level (e.g the effect of your neighbors income on your own life satisfaction) or at the macro level (e.g. effect regional house price changes on the probability to move) Do you have some thoughts or references for that type of questions (let's put problems like reverse causality aside for the moment)? Thanks, Stefan

[@ben_elsner]

Hi Stefan, I haven't seen staggered adoption discussed that much when it comes to individual-level data, but the same mechanics apply there as well. Note, though, that many papers that use individual-level data and estimate models with person and time FE use event studies rather than the canonical TWFE framework. Here is an example of a paper that looks at changes in risk attitudes after the death of a close family member: www.aeaweb.org/articles?id=10.1257/app.20200164&&from=f Event studies are more suitable for these settings because typically only few observations are treated, and it makes sense to only consider variation within an event window. Event studies have similar problems as canonical TWFE models. Abraham & Sun (Journal of Econometrics, 2021) have a very insightful paper that highlights the methodological problems and proposes a solution. Scott Cunningham discusses their paper in his CodeChella videos (on KZbin). Another useful paper on event studies is by Schmidheiny & Siegloch: www.schmidheiny.name/research/docs/schmidheiny-siegloch_2020-11.pdf Re 2): I would say the problems for identification are the same. Inevitably, earlier treated are a control group for later treated and vice versa. BTW There is now a lot of useful work on the canonical TWFE model: Wooldridge's new Mundlak estimator, Callaway & Sant'Anna (JoE 2021) and a few papers by de Chaisemartin and d'Haultfoeille. I hope this helps. Ben

[@sangelstefan]

@@ben_elsner Thank you very much for your detailed and quick and helpful answer. I will have a look at these papers! The best way I guess is to stay up to date with the newest methods papers on these questions! Best wishes, Stefan