Hello
I am trying to use switch_probit (Stata Journal, volume 11, number 3: st0233) but am puzzled by the results it gives. Specifically, by how the estimated marginal treatment effect relates to the estimated average treatment effect.
I can illustrate this by running the code from switch_probit_example.do that comes with the installation:
use switch_probit_example
switch_probit works age age2 wedu_2-wedu_5 hhsize hhsize2 reg_*, select(migrant age age2 wedu_2-wedu_5 hhsize hhsize2 reg_* pmigrants)
predict tt, tt
predict mte, mte
I can then summarise to get the ATT
su tt
Variable | Obs Mean Std. Dev. Min Max
-------------+--------------------------------------------------------
tt | 1694 .1160357 .0705859 .0046564 .4448425
Equation 8 of Aakvik, Heckman and Vytlacil (2005) shows that the ATT is a weighted sum of the MTEs. My confusion arises from the fact that the maximum value of the MTE (0.093 - see below) is smaller than the ATT (0.116), suggesting that the ATT is not a weighted average of the MTEs.
su mte
Variable | Obs Mean Std. Dev. Min Max
-------------+--------------------------------------------------------
mte | 100 .0282171 .047762 -.0681092 .0929032
Am I missing something? I'm using this type of model for the first time so I am very keen to understand this fundamental point.
I wondered whether the X variables were somehow to blame. But if I estimate the model with no regressors other than the instrument I get similarly-puzzling results - now the ATT is smaller than the smallest MTE:
switch_probit works , select(migrant pmigrants)
predict tt, tt
predict mte, mte
su tt mte
Variable | Obs Mean Std. Dev. Min Max
-------------+--------------------------------------------------------
tt | 1694 -.352801 .015625 -.36504 -.2810359
mte | 100 -.1768098 .0319067 -.2333341 -.1250045
Thanks for any guidance/clarification.
Richard
I am trying to use switch_probit (Stata Journal, volume 11, number 3: st0233) but am puzzled by the results it gives. Specifically, by how the estimated marginal treatment effect relates to the estimated average treatment effect.
I can illustrate this by running the code from switch_probit_example.do that comes with the installation:
use switch_probit_example
switch_probit works age age2 wedu_2-wedu_5 hhsize hhsize2 reg_*, select(migrant age age2 wedu_2-wedu_5 hhsize hhsize2 reg_* pmigrants)
predict tt, tt
predict mte, mte
I can then summarise to get the ATT
su tt
Variable | Obs Mean Std. Dev. Min Max
-------------+--------------------------------------------------------
tt | 1694 .1160357 .0705859 .0046564 .4448425
Equation 8 of Aakvik, Heckman and Vytlacil (2005) shows that the ATT is a weighted sum of the MTEs. My confusion arises from the fact that the maximum value of the MTE (0.093 - see below) is smaller than the ATT (0.116), suggesting that the ATT is not a weighted average of the MTEs.
su mte
Variable | Obs Mean Std. Dev. Min Max
-------------+--------------------------------------------------------
mte | 100 .0282171 .047762 -.0681092 .0929032
Am I missing something? I'm using this type of model for the first time so I am very keen to understand this fundamental point.
I wondered whether the X variables were somehow to blame. But if I estimate the model with no regressors other than the instrument I get similarly-puzzling results - now the ATT is smaller than the smallest MTE:
switch_probit works , select(migrant pmigrants)
predict tt, tt
predict mte, mte
su tt mte
Variable | Obs Mean Std. Dev. Min Max
-------------+--------------------------------------------------------
tt | 1694 -.352801 .015625 -.36504 -.2810359
mte | 100 -.1768098 .0319067 -.2333341 -.1250045
Thanks for any guidance/clarification.
Richard
Comment