Coefficient signs flip after replacing Post with time indicators

[Jacob Helium]

Dear all,

Hope you are doing well! I just have a quick question. Initially, I run the following regression to estimate the simple pre-post difference. The Post coefficient is negative, suggesting that y decreases after the event.

reghdfe y post $controls , absorb(industry) vce(cluster industry)

I wanted to assess how the effect persists into three years following the event. Thus, I replace Post with a series of time indicators. Post1 (Post2, Post3) is a dummy variable for the first (second, third) year after the event. There are three years in total after the event in the sample.

reghdfe y post1 post2 post3 $controls , absorb(industry) vce(cluster industry)

However, coefficients of Post1, Post2, and Post3 become all positive! Is this possible?

Thank you for reading my question.

Best,
Jacob

[Clyde Schechter]

You do not show example data, and you do not show the detailed output of the commands you ran. With so much left to the imagination, it is hard to give confident advice.

That said, I will speculate that the way post, post1, post2, and post3 are defined, it is likely that they are all colinear with the year fixed (absorbed) effects. This would imply that the results of both regressions are invalid for purposes of interpreting the post or post* variables. But example data, the code for creating post and post1-post3, and the full -reghdfe- output would be needed in order to verify this.

Added: A more benign explanation, possibly only if you have more than 3 years of post-event data, is that the response is, in fact, positive during the first 3 years post event but then turns negative later, and when all is smashed together into a single post period, the net response is negative.

[Jacob Helium]

You do not show example data, and you do not show the detailed output of the commands you ran. With so much left to the imagination, it is hard to give confident advice.

That said, I will speculate that the way post, post1, post2, and post3 are defined, it is likely that they are all colinear with the year fixed (absorbed) effects. This would imply that the results of both regressions are invalid for purposes of interpreting the post or post* variables. But example data, the code for creating post and post1-post3, and the full -reghdfe- output would be needed in order to verify this.

Added: A more benign explanation, possibly only if you have more than 3 years of post-event data, is that the response is, in fact, positive during the first 3 years post event but then turns negative later, and when all is smashed together into a single post period, the net response is negative.

Thank you for your response, Clyde. I have been examining my data and haven't been able to produce an example data set. I will share example data and outputs with the forum.

Though, I actually did not include year indicators. I wish to just compare 3 years before to three years after the event.