The critical thinking about the "staying forever" in Difference-in-Differences estimator of differential timing setting?

[Phuc Nguyen]

When dealing with the differential timing DiD setting, we may apply some modern approaches like Callaway, 2020, Borusyak, 2021 (I focus more on the imputation estimator of Borusyak because I am using it)

However, it comes to me a counterintuitive thought that: Why the treatment effect is assumed to start from the event date to the end of the sample period as indicated by (Borusyak, 2021). There should be some confounding events coming and change the pure examined effect. Why the effect do not just stay there just for 2,3,4 years, especially for accounting variables. And the effect for longer time further from the event date should be very messy.

For example, let us say a sample period lasts from 1990 to 2020, and US implement the law in 1993, is it fair to examine the effect of the laws on firms' asset growth by letting the treatment effect staying from 1993 to 2020 ?

[Clyde Schechter]

While I think the models that are used for teaching and explaining DID methodology often seek simplicity and assume that the treatment effect, once begun, is eternal, there is nothing inherent in the methodology that requires that assumption. If there is reason to believe that an effect is temporary, or that it gradually dissipates (say, in a linear, or piecewise linear, or exponential way) there is no reason not to build that into your model. Similarly, sometimes there is reason to believe that the onset of effect may be delayed after the onset of the treatment, or that the effect may gradually appear over some period of time rather than abruptly beginning at the onset of the treatment. All of these things can be included in DID modeling when there is reason to believe they are part of the real-world data-generating process, and they should be included if there is decent evidence to support and quantify those assumptions.

The problem with those more complicated models is that they require specific assumptions about how long the effect lasts, or how long it is delays, or how rapidly it dissipates, etc. And there may be little or no information to guide the model-building before seeing the data. If that is the case, one can be reduced to fitting multiple models that use different assumptions about the onset and decay of the treatment effect and trying to then select the best model. This is, needless to say, a dicey process, one which investigators (rightly or wrongly, depending on circumstances) try to avoid.

[Phuc Nguyen]

Clyde Schechter
It is a very helpful suggestion, I also confirmed with the code-write about that. Thanks a heap.