Hi all,
I want to estimate the impact of occupation [occupation segregation] on the wage gap between two groups of individuals (native and migrants). There are 3 occupation types: occ1, occ2 and occ3.
As occupation choice is endogenous to individauls’ wages, I want to apply a user-written command –mtreatreg- to correct for occupational endogeneity. [Note: Everyone who has an occupation reports earnings; where in a typical Heckman problem, for instance, only those who work report earnings and are included in the earnings equation]
-Mtreatreg- estimates the multinominal logit model in the first stage, it then predicts the probability of individual's occupational choice and calculate/generate the correction terms (mills) for each person based on his/her occupation. In the second stage, a selection correction term for each occupation group are included in the wage equation and the coefficients for all the 3 selection correction terms and other explanatory variables are estimated.
I do not have any problem in estimating the earnings equations correcting for occupational endogeneity using –mtreatreg-.
However, as I am interested in the wage gap between native and migrant workers. After estimating the earning equation corrected for occupational endogeneity using –mtreatreg-, I want to apply Oaxaca-Blinder decomposition.
In order to carry out the decomposition using – Oaxaca – (Jann 2008. – Stata Journal. 8. No. 4) I need to store all the explanatory variables including the lambdas themselves [i.e. inverse mill ratio for each occupation group generated in the first stage using multinominal logit by –mtreatreg-].
For the official -heckman- command it has the options -mills(newvar)- or -nshazard(newvar)-, which create a new variable containing the inverse of the Mills ratio from the selection equation. Alternatively, I can calculate it manually by estimating a probit equation at the first stage:
predict phat, xb
gen mills = exp(-.5*phat2)/(sqrt(2*_pi)*normprob(phat)). [Though doing so, the standard errors might be slightly bias (Jann 2008: 474 – Stata Journal. 8. No. 4)]
Once the variable mills are generated, I can then apply the Oaxaca command in the context of Heckman.
However, according to -help mtreatreg-, and -help mtreatreg postestimation-, there are not analogous options such as those for -heckman-.
I am not sure how to proceed with the Oaxaca decomposition and whether I have run into a dead end. I would greatly appreciate any help/suggestions as to:
1) whether there are some ways to save the mills for each occupation groups;
2) if not, whether the mills for each occupation type can be calculated manually similar to the case of Heckman model.
Many thanks in advance.
Regards
Amy
I want to estimate the impact of occupation [occupation segregation] on the wage gap between two groups of individuals (native and migrants). There are 3 occupation types: occ1, occ2 and occ3.
As occupation choice is endogenous to individauls’ wages, I want to apply a user-written command –mtreatreg- to correct for occupational endogeneity. [Note: Everyone who has an occupation reports earnings; where in a typical Heckman problem, for instance, only those who work report earnings and are included in the earnings equation]
-Mtreatreg- estimates the multinominal logit model in the first stage, it then predicts the probability of individual's occupational choice and calculate/generate the correction terms (mills) for each person based on his/her occupation. In the second stage, a selection correction term for each occupation group are included in the wage equation and the coefficients for all the 3 selection correction terms and other explanatory variables are estimated.
I do not have any problem in estimating the earnings equations correcting for occupational endogeneity using –mtreatreg-.
However, as I am interested in the wage gap between native and migrant workers. After estimating the earning equation corrected for occupational endogeneity using –mtreatreg-, I want to apply Oaxaca-Blinder decomposition.
In order to carry out the decomposition using – Oaxaca – (Jann 2008. – Stata Journal. 8. No. 4) I need to store all the explanatory variables including the lambdas themselves [i.e. inverse mill ratio for each occupation group generated in the first stage using multinominal logit by –mtreatreg-].
For the official -heckman- command it has the options -mills(newvar)- or -nshazard(newvar)-, which create a new variable containing the inverse of the Mills ratio from the selection equation. Alternatively, I can calculate it manually by estimating a probit equation at the first stage:
predict phat, xb
gen mills = exp(-.5*phat2)/(sqrt(2*_pi)*normprob(phat)). [Though doing so, the standard errors might be slightly bias (Jann 2008: 474 – Stata Journal. 8. No. 4)]
Once the variable mills are generated, I can then apply the Oaxaca command in the context of Heckman.
However, according to -help mtreatreg-, and -help mtreatreg postestimation-, there are not analogous options such as those for -heckman-.
I am not sure how to proceed with the Oaxaca decomposition and whether I have run into a dead end. I would greatly appreciate any help/suggestions as to:
1) whether there are some ways to save the mills for each occupation groups;
2) if not, whether the mills for each occupation type can be calculated manually similar to the case of Heckman model.
Many thanks in advance.
Regards
Amy
Comment