generalized difference-in-difference clarification

[Lucy Garcia]

Hello!

I come across a paper using a generalized difference-in-difference method, as shown in the screenshot, the authors use Treat × Post as the interaction, but, from my understanding, a generalized difference-in-difference can't have an interaction term, it only has a dummy indicating whether the event happens for firm i in year t, as in this post https://www.statalist.org/forums/for...in-differences, #6, Clyde said "In the generalized DID model, we have fixed effects for precincts and months and then a single variable which is 0 or 1 depending on whether treatment is "on" or "off" in that precinct in that month." So my question is, whether the method used in this paper is reasonable.

Source of this paper: Wu, Y., Zhang, Y., Li, G., & Li, F. (2022). Do property rights matter for bank loans? Evidence from China. Finance Research Letters, 102964.

[Clyde Schechter]

Yes, and notice that the way they have defined Treat and Post, Treat x Post is precisely the same as a variable that is 0 or 1 depending on whether the treatment is off or on in that precinct in that month.

Whether one chooses to call that an interaction term is a matter of taste, or how careful one wants to be in language usage. I generally do not call it that unless the constituent variables Treat and Post (or at least one of them) also appear in the model separately. But any variable that is defined as the product of two other variables could be called an interaction between those variables if you are using language a bit casually.

[Lucy Garcia]

Hi Clyde,

Thanks for your clarification! One more question, this generalized DID with Treat x Post can be applied when treatment is staggered, so Post is different for different treatment firms (rather than a specified point of time for all firms, such as after Oct 1, 2007 as above).

[Jared Greathouse]

You can do it like that, but I do not recommend this approach

[Lucy Garcia]

Hi Jared,

I see, thanks!