Here is something I do not quite follow: Why do people almost always refer to two-way fixed effects (TWFE) models as synonymous to Difference-in-Differences (DID)??
TWFE is an *estimation method*, while DID is a *research design*. These are very different things!
TWFE is an *estimation method*, while DID is a *research design*. These are very different things!
IMO, what defines a DID analysis is the parallel trends assumption (PTA). There are different parallel trends assumption out there, sure, but this is the crucial part of a DID analysis. In general, the PTA shows how we can nonparametrically identify the causal effect of interest.
TWFE, on the other hand, does not depend on PTA! It is a estimation method (or model specification). You can use it regardless of the underlying causal assumption, though the interpretation of the results would of course change (your beta may not be an ATT anymore)
Perhaps it is easier to draw a parallel to the cross section setup many of us are more familiar with.
For example, unconfoundedness or selection on observables, would be the research design; it would provide the foundation for your *identification analysis*.
For example, unconfoundedness or selection on observables, would be the research design; it would provide the foundation for your *identification analysis*.
OTOH, inverse probability weighting, regression, or matching would be the *estimation method*.
Under unconf. ,these are estimators for the same causal quantity. However, they still can be used when unconf. does not hold, though the interpretation would of course change.
Under unconf. ,these are estimators for the same causal quantity. However, they still can be used when unconf. does not hold, though the interpretation would of course change.
Why do we acknowledge the difference between research design and estimation method in cross-section setups but not in panel/repeated cross-sections???
In fact, with DID, I would say that this distinction is even more important since different estimation methods lead to different causal estimates in setups with variation in treatment timing! TWFE and first difference regression estimators are not even comparable! =X
Perhaps it is time for us to emphasize more the differences between *estimation method* and *research design / identification analysis* in difference-in-differences setups.
As is evident here urlzs.com/6bRC6, I like the idea of building DID procedures that do not rely on the estimation method, i.e., to shift the attention towards the causal parameter of interest, and NOT the estimation method.
But this is something for another time!!!
But this is something for another time!!!
