Hi all,
The question I'm about to ask is an extension to previous thread I've created days ago:
https://www.statalist.org/forums/for...m-effect-model
While trying to understand what went wrong with the coefficient sign changing after adding the interaction term, I suddenly recognized that I've missed the most essential part.
To simply put it, I'm trying to deploy DID (Difference-in-Differences) approach, but I wasn't sure if I should trust result from the command
or random effects model.
1) So far, the main reason I was convinced to use 'random effects model' was that the control and the treatment groups are time-invariant.
And the
returns p-value of 1.
However, the Hausman test result is achieved without specifying the
the "robust" option.
I've conducted Hausman test without "robust" option, well because Hausman test didn't work with "robust" option included when saving the results using
.
Nevertheless, the p-value from the Hausman test is 1, which I assume I should run random effects model.
However, is random effects model appropriate for DID setting? (Although I know very little about the econometric setting behind the scene, doesn't DID operate based on fixed effects?)
2) Using the
which I believe was added in the new Stata 17 version, the estimated ATET is exactly the same when I run the fixed effect model with robust standard error option.
Which should to my understanding yield exactly the same result, as it did:
I've generated the "treat" variable as the following:
Since variable control and treatment groups over == 0 (as control) & over == 1 (as treatment) within the variable "treat" does not vary across time, I thought it would require me to use random effects model or hybrid model to observe the time-varying effects of treatment and control groups.
It sounded like to me that the treatment & control group specification is similar to that of gender (or any other demographical data) which can easily be found in the health organizational data, as they both are time-invariant.
So for example, if I were to run a regression with gender variable say
"sex" variable will be omitted but the interaction term will survive.
And from the previous posts that I've found in this forum seemed to indicate the interpretation of results in such case might differ from simple OLS result
.
Given the circumstances would it be appropriate for me to continue with the DID approach using
&
? or should I continue looking for solutions using the hybrid models to incorporate the issues with time-invariant predictors?
If xtdidreg is indeed way to go, should I include other control variables to control for unobserved heterogeneity? or does the panel fixed effect model takes care of them?
Any help or advices are appreciated in advance!
Kind regards,
Vincent
The question I'm about to ask is an extension to previous thread I've created days ago:
https://www.statalist.org/forums/for...m-effect-model
While trying to understand what went wrong with the coefficient sign changing after adding the interaction term, I suddenly recognized that I've missed the most essential part.
To simply put it, I'm trying to deploy DID (Difference-in-Differences) approach, but I wasn't sure if I should trust result from the command
Code:
xtdidreg, fe
1) So far, the main reason I was convinced to use 'random effects model' was that the control and the treatment groups are time-invariant.
And the
Code:
hausman fe re
However, the Hausman test result is achieved without specifying the
Code:
xtreg, fe robust
I've conducted Hausman test without "robust" option, well because Hausman test didn't work with "robust" option included when saving the results using
Code:
est store fe & est store fe & hausman fe re
Nevertheless, the p-value from the Hausman test is 1, which I assume I should run random effects model.
However, is random effects model appropriate for DID setting? (Although I know very little about the econometric setting behind the scene, doesn't DID operate based on fixed effects?)
2) Using the
Code:
xtdidreg
Which should to my understanding yield exactly the same result, as it did:
Code:
xtset id date
Code:
xtdidreg (OFF_AMT) (treat), group(id) time(date)
Code:
xtreg OFF_AMT treat i.date, fe robust
I've generated the "treat" variable as the following:
Code:
gen treat = (over == 1 & year == 1)
It sounded like to me that the treatment & control group specification is similar to that of gender (or any other demographical data) which can easily be found in the health organizational data, as they both are time-invariant.
So for example, if I were to run a regression with gender variable say
Code:
xtreg y i.sex##i.year, fe
And from the previous posts that I've found in this forum seemed to indicate the interpretation of results in such case might differ from simple OLS result
Code:
reg y i.sex##i.year
Given the circumstances would it be appropriate for me to continue with the DID approach using
Code:
xtreg, fe
Code:
xtdidreg
If xtdidreg is indeed way to go, should I include other control variables to control for unobserved heterogeneity? or does the panel fixed effect model takes care of them?
Any help or advices are appreciated in advance!
Kind regards,
Vincent
Attached Files
