Hello. I am running a diff-in-diff regression with 2 periods. All treated units "idcomune" get treated at the same time, in period 2.
The "after" variable is built this way:
These are the results without time-varying covariates, using reghdfe:
These are the results without time-varying covariates, using drdid:
Same coefficients.
Now I add one unit-level time-varying covariate and one regional-level time varying covariate, using reghdfe:
and now using drdid:
Different coefficients.
Edit:
Same thing happens using:
All three specifications have the same coefficient before adding time-varying covariates and different coefficients after adding them.
Why does this happen? Thank you.
The "after" variable is built this way:
Code:
gen after=0 replace after=1 if period>=2
Code:
reghdfe vote_share ever_treated##after, absorb (idcomune period) cluster(idcomune)
(MWFE estimator converged in 2 iterations)
note: 1bn.ever_treated is probably collinear with the fixed effects (all partialled-out values are close to zero; tol = 1.0e-09)
note: 1bn.after is probably collinear with the fixed effects (all partialled-out values are close to zero; tol = 1.0e-09)
HDFE Linear regression Number of obs = 14,618
Absorbing 2 HDFE groups F( 1, 7308) = 5.78
Statistics robust to heteroskedasticity Prob > F = 0.0162
R-squared = 0.8938
Adj R-squared = 0.7875
Within R-sq. = 0.0005
Number of clusters (idcomune) = 7,309 Root MSE = 0.0313
(Std. err. adjusted for 7,309 clusters in idcomune)
------------------------------------------------------------------------------------
| Robust
vote_share | Coefficient std. err. t P>|t| [95% conf. interval]
-------------------+----------------------------------------------------------------
1.ever_treated | 0 (omitted)
1.after | 0 (omitted)
|
ever_treated#after |
1 1 | .0157764 .0065619 2.40 0.016 .0029132 .0286395
|
_cons | .2030246 .0000121 1.7e+04 0.000 .2030009 .2030484
------------------------------------------------------------------------------------
Absorbed degrees of freedom:
-----------------------------------------------------+
Absorbed FE | Categories - Redundant = Num. Coefs |
-------------+---------------------------------------|
idcomune | 7309 7309 0 *|
period | 2 1 1 |
-----------------------------------------------------+
* = FE nested within cluster; treated as redundant for DoF computation
Code:
drdid vote_share, ivar(idcomune) time(period) treatment(ever_treated) cluster(idcomune) reg
Doubly robust difference-in-differences Number of obs = 14,618
Outcome model : regression adjustment
Treatment model: none
(Std. err. adjusted for 7,309 clusters in idcomune)
------------------------------------------------------------------------------
| Coefficient Std. err. z P>|z| [95% conf. interval]
-------------+----------------------------------------------------------------
ATET |
ever_treated |
(1 vs 0) | .0157764 .006561 2.40 0.016 .0029171 .0286357
------------------------------------------------------------------------------
Now I add one unit-level time-varying covariate and one regional-level time varying covariate, using reghdfe:
Code:
reghdfe vote_share ever_treated##after tasso_disoccup_annoElez popComune_annoElez , absorb (idcomune period) cluster(idcomune)
(MWFE estimator converged in 2 iterations)
note: 1bn.ever_treated is probably collinear with the fixed effects (all partialled-out values are close to zero; tol = 1.0e-09)
note: 1bn.after is probably collinear with the fixed effects (all partialled-out values are close to zero; tol = 1.0e-09)
HDFE Linear regression Number of obs = 14,618
Absorbing 2 HDFE groups F( 3, 7308) = 27.95
Statistics robust to heteroskedasticity Prob > F = 0.0000
R-squared = 0.8949
Adj R-squared = 0.7897
Within R-sq. = 0.0114
Number of clusters (idcomune) = 7,309 Root MSE = 0.0311
(Std. err. adjusted for 7,309 clusters in idcomune)
-----------------------------------------------------------------------------------------
| Robust
vote_share | Coefficient std. err. t P>|t| [95% conf. interval]
------------------------+----------------------------------------------------------------
1.ever_treated | 0 (omitted)
1.after | 0 (omitted)
|
ever_treated#after |
1 1 | .018809 .0067149 2.80 0.005 .0056459 .0319722
|
region_unempl_year | .0070316 .0008053 8.73 0.000 .0054529 .0086102
city_pop_year | 5.75e-06 3.68e-06 1.56 0.119 -1.47e-06 .000013
_cons | .0852026 .0257956 3.30 0.001 .0346359 .1357693
-----------------------------------------------------------------------------------------
Absorbed degrees of freedom:
-----------------------------------------------------+
Absorbed FE | Categories - Redundant = Num. Coefs |
-------------+---------------------------------------|
idcomune | 7309 7309 0 *|
period | 2 1 1 |
-----------------------------------------------------+
* = FE nested within cluster; treated as redundant for DoF computation
Code:
drdid vote_share region_unempl_year city_pop_year, ivar(idcomune) time(period) treatment(ever_treated) cluster(idcomune) reg
Doubly robust difference-in-differences Number of obs = 14,618
Outcome model : regression adjustment
Treatment model: none
(Std. err. adjusted for 7,309 clusters in idcomune)
------------------------------------------------------------------------------
| Coefficient Std. err. z P>|z| [95% conf. interval]
-------------+----------------------------------------------------------------
ATET |
ever_treated |
(1 vs 0) | .0119468 .0060062 1.99 0.047 .0001748 .0237188
------------------------------------------------------------------------------
Edit:
Same thing happens using:
Code:
reg vote_share region_unempl_year city_pop_year ever_treated##after i.idcomune i.period, cluster(idcomune)
Why does this happen? Thank you.
Comment