My data is a balanced panel.
12 years X 600 ids/year X 12 months = 864,000 observations.
For about 150 of my 600 ids, at a certain point in this 12 year period a treatment is applied switching them to the "treated" or "on" state. A dummy variable AFTER indicates when this is true.
My outcome is a count, positive only and quite skewed, looks like a power law.
I'm interested in estimating the impact of treatment on outcome.
I've tried these 3 specifications in STATA and I dont seem to understand why they don't accord well with each other.
1. OLS ---> reg y AFTER i.id i.year, robust cluster(id)
2. Logged --> reg ln(y) AFTER i.id i.year, robust cluster(id)
3. Poisson ---> xtpoisson y AFTER i.year, fe vce(robust)
4. Negative Binomial ---> nbreg y AFTER i.id i.year, vce(robust)
Here i.id and i.year are fixed effects for id and year respectively.
Estimated coeffs are as follows:
1. OLS: 403922 (223946) *
2. Log : -0.664 (0.0252) ***
3. Poisson : -.0188 (0.019)
Sum stats for my outcome variable are:
range: [0,3.025e+08] unique values: 59191
mean: 1.3e+06
std. dev: 1.1e+07
percentiles: 10% 25% 50% 75% 90%
0 13480 74690 322982 1.3e+06
12 years X 600 ids/year X 12 months = 864,000 observations.
For about 150 of my 600 ids, at a certain point in this 12 year period a treatment is applied switching them to the "treated" or "on" state. A dummy variable AFTER indicates when this is true.
My outcome is a count, positive only and quite skewed, looks like a power law.
I'm interested in estimating the impact of treatment on outcome.
I've tried these 3 specifications in STATA and I dont seem to understand why they don't accord well with each other.
1. OLS ---> reg y AFTER i.id i.year, robust cluster(id)
2. Logged --> reg ln(y) AFTER i.id i.year, robust cluster(id)
3. Poisson ---> xtpoisson y AFTER i.year, fe vce(robust)
4. Negative Binomial ---> nbreg y AFTER i.id i.year, vce(robust)
Here i.id and i.year are fixed effects for id and year respectively.
Estimated coeffs are as follows:
1. OLS: 403922 (223946) *
2. Log : -0.664 (0.0252) ***
3. Poisson : -.0188 (0.019)
Sum stats for my outcome variable are:
range: [0,3.025e+08] unique values: 59191
mean: 1.3e+06
std. dev: 1.1e+07
percentiles: 10% 25% 50% 75% 90%
0 13480 74690 322982 1.3e+06