DiD omitted treatment variable

[Clara Weber]

Hi everybody!

I want to run a difference in difference approach in Stata where I want to analyse whether gender-targeted IFC investment into the private sector of a country affects female employment in that country. My time dummy is Post15 since after 2015, IFC officially implemented the gender focus in their investment practices. My treatment group will be those countries that after 2015 saw a higher than average increase in gender-targeted investment. In stata, I create the time dummy Post15 which equals 1 if the year is higher than 2015. For the treatment dummy, does it need to equal 1 for all countries receiving higher than average gender-targeted investment only after 2015 or already before? For now, my treatment dummy equals 1 for treated countries in every year, i.e., also before 2015.

When I run the DiD using this command:

xtreg femployment i.Post2015##i.treatment i.year, fe vce(cluster CountryCode_num)

in my output table (using outreg2) the treatment dummy gets omitted due to collinearity. Is there a way to avoid the omitted dummy? When I exlcude the "fe", the interaction does not omitted, but what would it mean for my model to exlude the "fe"?

When I instead of the "fe" include i.country (country fixed effects), the coefficient for the treatment dummy gets unrealistically high.

Maybe anyone can help?

Thanks in advance!

[Clara Weber]

Edit: I assume that the treatment dummy gets omitted since it is time invariant (either always 0 or 1) and, thus, the fe already takes care of time invariant factors. But what is the difference between including "fe" or i.country?

[Jeff Wooldridge]

Clara: You are correct. You don't need to include the "eventually treated" dummy because it doesn't change over time. If you want to see the coefficient on it, just use reg or xtreg, re. With a balanced panel, you will get the same estimated treatment effect using all three methods if done properly.

Be very careful about i.county. Stata will drop one of the dummy variables essentially arbitrarily and you'll think you're estimating something meaningful. Do not use i.country because the resulting coefficient on "treatment" is meaningless.

[Clara Weber]

Dear Jeff,

thank you for your answer! That was very helpful!