Dropping the previous quarter just before the intervention for illustrating the parallel trends

[James Sonela]

Hello,

This is not directly about STATA but a conceptual question.

I am running a DiD model with FE in STATA and I conduct my analysis without dropping any observation and results are significant. However, just for the sake of presenting the effects (and the story) in a cleaner way, I am dropping the previous quarters (t-1) observations while depicting my parallel trend graph. It is highly likely that my t-1 has anticipation effect and also includes some small portion of the post-treatment data. Thus, there is a bit unusual behavior just before the intervention date. That's why I would like to get rid of it only in my parallel trends graph (I am not dropping the observations in my analysis or whatsoever).

Do you think it makes sense? Have you read any similar paper that does this (so I can justify this selection)?

[Mike Murphy]

Not a Stata answer, but my reaction would that it doesn't make sense to do this. The best way to tell the story is with all of the data- presenting it with the preceding quarter gives you the opportunity to highlight these issues and tee up how you will handle them in your subsequent analysis. Purely a personal take, but I generally expect that a dataset will have at least one quirky feature and tend to take it as a negative signal when a presenter doesn't note any kind of wrinkle in the DGP.

[James Sonela]

Not a Stata answer, but my reaction would that it doesn't make sense to do this. The best way to tell the story is with all of the data- presenting it with the preceding quarter gives you the opportunity to highlight these issues and tee up how you will handle them in your subsequent analysis. Purely a personal take, but I generally expect that a dataset will have at least one quirky feature and tend to take it as a negative signal when a presenter doesn't note any kind of wrinkle in the DGP.

Thanks for your answer.