Hi,
I am interested in understanding whether the effect of a policy differs by some dimension (e.g. whether the person is white). I ran the DID separately for the two groups, and I want to report p-test of difference between the estimated effect of the policy for the two groups, but I am unsure how to do this. The criteria for treated is a different one, in my case it is whether the person has two or more children vs just one child.
The separate regressions look like this, where the coefficient on treatedpost is the estimated treatment effect (it is just treated*post) . c.year#i.state estimates state specific time trends.
reghdfe y treatedpost treated `controls' c.year#i.state, absorb(year state) cluster(state)
I tried to get the p-test result in two ways:
1. reghdfe y treatedpost##white treated `controls' c.year#i.state, absorb(year state) cluster(state)
and
2. reghdfe y treatedpost09 treated `controls' c.year#i.hh_state (c.(controls_a c.year#i.hh_state)i.(treatedpost09 treated controls_b))#race, absorb(year hh_state) cluster(hh_state)
I think 1 seems to be working correctly and reporting sensible reports. The idea of 2 is trying to estimate the other coefficients all differently based on race. But I don't think it is currently giving me what I want. I have an estimate for treatedpost09 which I think is for the the base level (0,0), and an estimate for (0,1) and (1,0) each, and (1,1) is omitted (first category is for treated-post and second category for race), with corresponding p-values. I am not sure why 1,1 should be omitted. Also I think what I am interested in is the difference between (1,0)-(0,0) and (1,1)-(0,1), and the associated p-values, not any of these coefficients separately. It would be great if you can point me to the right way to do this. I read that a triple difference is probably not what I want right now.
Thanks!
Best,
Angela
I am interested in understanding whether the effect of a policy differs by some dimension (e.g. whether the person is white). I ran the DID separately for the two groups, and I want to report p-test of difference between the estimated effect of the policy for the two groups, but I am unsure how to do this. The criteria for treated is a different one, in my case it is whether the person has two or more children vs just one child.
The separate regressions look like this, where the coefficient on treatedpost is the estimated treatment effect (it is just treated*post) . c.year#i.state estimates state specific time trends.
reghdfe y treatedpost treated `controls' c.year#i.state, absorb(year state) cluster(state)
I tried to get the p-test result in two ways:
1. reghdfe y treatedpost##white treated `controls' c.year#i.state, absorb(year state) cluster(state)
and
2. reghdfe y treatedpost09 treated `controls' c.year#i.hh_state (c.(controls_a c.year#i.hh_state)i.(treatedpost09 treated controls_b))#race, absorb(year hh_state) cluster(hh_state)
I think 1 seems to be working correctly and reporting sensible reports. The idea of 2 is trying to estimate the other coefficients all differently based on race. But I don't think it is currently giving me what I want. I have an estimate for treatedpost09 which I think is for the the base level (0,0), and an estimate for (0,1) and (1,0) each, and (1,1) is omitted (first category is for treated-post and second category for race), with corresponding p-values. I am not sure why 1,1 should be omitted. Also I think what I am interested in is the difference between (1,0)-(0,0) and (1,1)-(0,1), and the associated p-values, not any of these coefficients separately. It would be great if you can point me to the right way to do this. I read that a triple difference is probably not what I want right now.
Thanks!
Best,
Angela
Comment