Read on Twitter
Methods appendix 
https://abs.twimg.com/emoji/v2/... draggable="false" alt="π§΅" title="Thread" aria-label="Emoji: Thread">for the new version of our public library paper, as promised...
https://twitter.com/peternka/status/1381666459554869250?s=20
1/n">https://twitter.com/peternka/...
https://twitter.com/peternka/status/1381666459554869250?s=20
1/n">https://twitter.com/peternka/...
We use Callaway and @pedrohcgs (2020)βs staggered DiD method for the main results.
Thanks to @pedrohcgs @causalinf Sun/Abraham @jondr44 @agoodmanbacon @ChloeEast2 @arindube @Andrew___Baker @paulgp and everyone else for making the new DiD literature clear
Thanks to @pedrohcgs @causalinf Sun/Abraham @jondr44 @agoodmanbacon @ChloeEast2 @arindube @Andrew___Baker @paulgp and everyone else for making the new DiD literature clear
This is a perfectly timed tweet from @agoodmanbacon.
For us, the new methods are *not* a robustness check on a standard TWFE model. They fix "errant" pre-trends that we have been thinking about for years and could have torpedoed the project https://twitter.com/agoodmanbacon/status/1381013331058679810">https://twitter.com/agoodmanb...
For us, the new methods are *not* a robustness check on a standard TWFE model. They fix "errant" pre-trends that we have been thinking about for years and could have torpedoed the project https://twitter.com/agoodmanbacon/status/1381013331058679810">https://twitter.com/agoodmanb...
We have a long panel, staggered timing, and time-varying TEs. πΌ π§ππππ₯π ππ€π§ ππ¨π¨πͺππ¨.
Left, standard TWFE ES w/ outcome = ln child library circulation. Right, Callaway and @pedrohcgs results. Big difference, though in both cases it is clear something is going on
Left, standard TWFE ES w/ outcome = ln child library circulation. Right, Callaway and @pedrohcgs results. Big difference, though in both cases it is clear something is going on
I hope this is motivational: weird pre-trends might be due to TWFE problems! Not selection.
We tried to write our methods section in a way that is helpful for future CS adopters. Please read and let us know if anything is unclear :-)
Finally, can& #39;t thank Callaway, @pedrohcgs, and @jondr44 enough for their R packages. Incredible to use.
A few practitioner notes:
A few practitioner notes:
For now, the Stata implementations don& #39;t have the richness and options of the R version.
That richness is important. For example, we report multiple aggregations of the ES parameters.
The R package is easy to use and we are happy to share our code or discuss!
That richness is important. For example, we report multiple aggregations of the ES parameters.
The R package is easy to use and we are happy to share our code or discuss!
I also want to read/think more about the use of time-varying covariates in DiD papers.
As @causalinf said in today& #39;s substack: "My whole adult life, Iβd never until recently heard one person complain about including time varying covariates in a regression, let alone a DiD...."
As @causalinf said in today& #39;s substack: "My whole adult life, Iβd never until recently heard one person complain about including time varying covariates in a regression, let alone a DiD...."
The "standard way" of including time-varying covariates in ES models can be bad, but it is expected.
It seems important from a "current practice" point of view to fully explore the potential issues (the @agoodmanbacon paper also has some of this!)
It seems important from a "current practice" point of view to fully explore the potential issues (the @agoodmanbacon paper also has some of this!)
Our understanding is that the CS method is not built to handle "true" time-varying covariates
We put potential confounders on the LHS using CS and show that they did not change at the same time as our treatment. Things also are stable if we control for them in (biased) TWFE models
but more discussion on this point might be helpful!
but more discussion on this point might be helpful!
