event study after propensity score matching with exact match?

[Maurice McCourt]

I create a propensity-score matched Firm-Year panel sample as follows:

kmatch ps T X1 X2 (Y), ematch(X3 Year) att wgenerate(W) replace

where T is the treatment variable, Y is the outcome variable, Year is the year, and X1, X2 and X3 are firm characteristics that vary over time. The matching is exact for X3 and Year.



Now I want to run a staggered event study regression using the matched sample.

My question then is, as Year is already included in the propensity score matching (as an exact match), do I need to include it again (as a fixed effect) in the event study regression?

In other words, does the event study regression look like this:

reghdfe Y T [pweight=W], absorb(Firm Year)

or like this:

reghdfe Y T [pweight=W], absorb(Firm)

Note: kmatch and reghdfe are SSC contributed commands.


Thanks

[Nick Cox]

This appears to repeat https://www.statalist.org/forums/for...th-exact-match

If interested, please follow that thread.

[Maurice McCourt]

hi Nick

Thanks for the repsonse. I didnt think this question was a repeat of the previous question, given the event study context I envisaged it more as a follow-on to the previous quesion.

But your response suggests that maybe it is the same question (or at least the answer is the same): variables used in the propensity score matching are not used in the follow-on regressions to estimate treatment effects with the matched sample. In this case the Year fixed-effects could be omitted.

[Maurice McCourt]

As a follow-on to the previous question, consider these two scenarios which control for Year effects when estimating an event study model with a matched sample. The objective is to estimate the effect of treatment T on the outcome variable Y for the subset of firms that receive the treatment. Treatment is staggered, that is, firms receive the treatment in different years.


Scenario 1 (same as in previous message, with Year variable included in the matching step as an exact match and excluded in the follow-on event study regression:

 kmatch ps T X1 X2 (Y), ematch(X3 Year) att wgenerate(W) replace  
 reghdfe Y T [pweight=W], absorb(Firm)

Scenario 2 (with Year variable excluded friom the matching step but included as a fixed-effect in the follow-on regression:

 kmatch ps T X1 X2 (Y), ematch(X3) att wgenerate(W) replace  
 reghdfe Y T [pweight=W], absorb(Firm Year)

Both these scenarios should control for Year effects, but in my data, they give completely different coefficients on the treatment variable (T) in the event study regression.


So two questions:

- Why would the two approaches give completely different results?

- Which approach is "correct" (or how to choose between the two approaches)?




Thanks