Dear all
I wonder if anyone has any references, and perhaps Stata applications, that can help me solve a problem like follows:
Code:
use http://fmwww.bc.edu/RePEc/bocode/o/oaxaca.dta, clear
* This is the benchmark
reg lnwage educ exper tenure female single
est sto m1
* now, i can create the following variables:
gen double female2=female*_b[female]
gen double educ2=educ*_b[educ]
gen double femaleeduc=female2+educ2
gen double lnwage2=lnwage-female2-educ2
* And estimate this model
reg lnwage femaleeduc exper tenure single
est sto m2
* or this other
reg lnwage2 exper tenure single
est sto m3

est tab m1 m2 m3, se
-----------------------------------------------------
    Variable |     m1           m2           m3      
-------------+---------------------------------------
        educ |  .08303358                            
             |  .00513458                            
       exper |  .00895939    .00895939    .00895939  
             |  .00156103    .00155551    .00153964  
      tenure |  .00613849    .00613849    .00613849  
             |  .00187615    .00186656    .00185977  
      female | -.09382159                            
             |  .02490175                            
      single | -.16080964   -.16080964   -.16080964  
             |  .02702903    .02694462     .0269209  
  femaleeduc |                       1               
             |               .05800787               
       _cons |  2.3408107    2.3408107    2.3408107  
             |  .07085517    .06106525    .02653445  
-----------------------------------------------------
                                         legend: b/se
All models provide the same point estimates, but different standard errors. The benchmark is column 1.
Column 2 combines the effect of female and educ, and adds them into the model. and column 3 simply extracts the effect of female and education from lnwage before estimating the model.
My question is, does anyone know how to correct the standard errors from model 3 or 2, and obtain the "correct" ones from model 1?
I know this could be done using bootstrap methods, but Im trying to see if it can be done in a different way.

For more details on why this. Im revising Robinson's Semiparametric estimator. (see reference below). Im aware about the userwriten command -semipar-. However, the application itself does not provide much detail on how to estimate the standard errors of the nonparametric section.
Looking through the code, it does it in a similar way as the results from column 3, but what we want are the results from column 1.
Thank you in advance.

Robinson, P. M. 1988. Root-n-consistent semiparametric regression. Econometrica 56: 931–954.