Dear Statalister,
I am running the following triangular model using Stata 18.
(hc + excluded variables) ->ret3 (+excluded variable)-> hirecando) -> + ret3 on fsk3.
The triangular model has been implemented using the user written routine CMP in the following way:
exclusion restrictions
the equation for ret3 has the follwing variables entering only this equation: hc_class wpsupp eidelay
the equation for hirecando has the follwing variables entering only this equation: itprodimp itperfmon skillch trski trflex supchek neverint
except for the exclusion restrictions all the equations contain the same set of contorls
In the last equation fsk is regressed on ret3 hirecando and the controls.
All the three equations belonging to the triangular system have been modelled as ordinal probits, and they have been jointly estimated.
The same triangular system has also been implemented using the extended regression suite of Stata
the dendent variables are the same, the controls are the same, the exclusion restrictions are the same, the specification is the same (ordinal probit), yet the results are different. CMP uses a somewhat larger sample, 18,928 observations, while eoprobit uses 18,093 observations. I guess this difference could be traced back to the FIML (all usable observations for each equation) approach used in CMP. But it is a guess.
For example, for the equation concerning the variable fsk I get the following results:
CMP
eoprobit
The first two coefficients, those on ret3 and hirecando, are very different in the two specifications.
The difference in sample size between the two commands may have caused some change.
I assumed that both CMP and eoprobit would have implmented the same triangular system. I am missing something.
Thank in advance for your help with this matter
Kind regards
Giovanni
I am running the following triangular model using Stata 18.
(hc + excluded variables) ->ret3 (+excluded variable)-> hirecando) -> + ret3 on fsk3.
The triangular model has been implemented using the user written routine CMP in the following way:
Code:
cmp (fsk3 = ret3 hirecando $reg) (ret3 = i.hc_class wpsupp eidelay $reg) (hirecando = ret3 itprodimp itperfmon skillch trski trflex supchek neverint $reg) [pweight= s5_wgt_final], ind( 5, 5, 5 ) vce(cluster country) difficult
the equation for ret3 has the follwing variables entering only this equation: hc_class wpsupp eidelay
the equation for hirecando has the follwing variables entering only this equation: itprodimp itperfmon skillch trski trflex supchek neverint
except for the exclusion restrictions all the equations contain the same set of contorls
In the last equation fsk is regressed on ret3 hirecando and the controls.
All the three equations belonging to the triangular system have been modelled as ordinal probits, and they have been jointly estimated.
The same triangular system has also been implemented using the extended regression suite of Stata
Code:
eoprobit fsk3 ret3 hirecando $reg [pweight= s5_wgt_final], endogenous(ret3 = i.hc_class wpsupp eidelay $reg, oprobit nomain) endogenous(hirecando = ret3 itprodimp itperfmon skillch trski trflex supchek neverint $reg, oprobit nomain) vce(cluster country) difficult
For example, for the equation concerning the variable fsk I get the following results:
CMP
| Robust | ||||||
| Coefficient | std. err. | z | P>z | [95% conf. | interval] | |
| fsk3 | ||||||
| ret3 | .7808615 | .1765825 | 4.42 | 0.000 | .4347662 | 1.126957 |
| hirecando | -.2332199 | .1152787 | -2.02 | 0.043 | -.459162 | -.0072779 |
| comorg_1 | .0074937 | .0123713 | 0.61 | 0.545 | -.0167536 | .0317411 |
| compprobs_1 | .0245087 | .0143221 | 1.71 | 0.087 | -.0035621 | .0525795 |
| contr_1 | .053524 | .0150753 | 3.55 | 0.000 | .023977 | .0830709 |
eoprobit
| Robust | ||||||
| Coefficient | std. err. | z | P>z | [95% conf. | interval] | |
| fsk3 | ||||||
| ret3 | 1.046546 | .2508465 | 4.17 | 0.000 | .5548954 | 1.538196 |
| hirecando | -.0236865 | .3363055 | -0.07 | 0.944 | -.6828332 | .6354602 |
| comorg_1 | .0110178 | .0102237 | 1.08 | 0.281 | -.0090203 | .031056 |
| compprobs_1 | .0259533 | .0144384 | 1.80 | 0.072 | -.0023454 | .0542521 |
| contr_1 | .0374643 | .0256565 | 1.46 | 0.144 | -.0128216 | .0877501 |
The first two coefficients, those on ret3 and hirecando, are very different in the two specifications.
The difference in sample size between the two commands may have caused some change.
I assumed that both CMP and eoprobit would have implmented the same triangular system. I am missing something.
Thank in advance for your help with this matter
Kind regards
Giovanni
