Omit Main Effect (Treat*Post) and keep interactions (Dummy*Post)

[Wendy Wang]

I have a difference-in-differences model for outcome "y" after multiple groups undergo a treatment. I want to drop the main interaction term (treat*post) in the following model, so that I can keep all group dummy interactions:

xtset id year
xtreg log_y i.treat##i.post##i.group_dummy i.year, fe

Otherwise, in the above specification, STATA uses one of my groups as the baseline to measure the treat*group_dummy interaction, but I would like to have coefficients for the treatment effect for all of the treat*post*group_dummy interactions (without falling into the dummy variable trap).

Thanks--in advance--for your advice!

[Maarten Buis]

You may want to look at http://www.maartenbuis.nl/publications/ref_cat.html

[Wendy Wang]

Thanks Maarten! Different solution--also interesting.

I wonder how that affects the interpretation of the result, though. Either:

(1) I omit the Treat*Post interaction, then the Treat*Post*Group_Dummy coefficient will be the treatment effect for each group (still unclear in STATA how to do this).

(2) I omit the constant and get a main treatment affect, as well as interaction effects. The interaction affects will be difficult to interpret, so if I go with (2), perhaps I should calculate the margins for treat#post#group_dummy to get a more interpretable comparison?