Announcement

Collapse
No announcement yet.
X
  • Filter
  • Time
  • Show
Clear All
new posts

  • Should I include gender dummy for the probit regression in Heckman?

    I'm estimating the wage equation for a selected sample of the self-employed, which contains both males and females, and afterwards which will be decomposed using the Oaxaca command in Stata, by gender.

    The wage equation for the Oaxaca-Blinder Decomposition by gender doesn't contain a gender dummy.

    However, I am also checking for sample selection bias, and so was wondering if the probit model which will be used to calculate the Inverse Mills Ratio (IMR) should include a gender dummy or not. This is because as mentioned earlier, the main equation doesn't have one.

    N.B. The IMR calculation I'm interested in is for participation into the sample for both genders.

    Code:
    svy: probit selfemp age agesq edu1-edu6 maritalstat1-maritalstat3 head childrenunder6 lnnethhincome hhasset1-hhasset3 landown
    
    predict xb if e(sample), xb
    
    generate mills=normalden (-xb) / (1-normal (-xb))
    
    svy: reg lnwage age agesq edu1-edu6 occ1-occ9 ind1-ind3 mills
    
    oaxaca lnwage age agesq edu1-edu6 occ1-occ9 ind1-ind3 mills, by (female) pooled svy relax adjust (mills) detail
    Now the problem is, if I include the female dunmy in svy: probit, the mills coefficient is significant in the svy: reg, otherwise not. Any suggestion regarding what is the appropriate approach here?

    Thank you in advance.


  • #2
    Bump, can anyone please kindly help? Thank you.

    Comment


    • #3
      Hi Alina,

      I'm not familiar with Oaxaca-Blinder Decomposition method, so this help may be worth what you pay for it. But apparently this is a common question about using the Oaxaca-Blinder Decomposition method because it has been previously posted twice on Statalist (see below) and it appears that neither of them got an answer:However, Ben Jann, the author of Stata's oaxaca command has an article explaining it in in the Stata Journal here . On p. 473-475 the author gives two ways to run the oaxaca with heckman (or, to do as you did and run the selection model outside of oaxaca and calculate the inverse mills ratio yourself.)

      Code:
      * Method #1
      . oaxaca lnwage educ exper tenure, by(female) model2(heckman, twostep
      > select(lfp = age agesq married divorced kids6 kids714))
      Code:
      * Method #2
      probit lfp age agesq married divorced kids6 kids714 if female==1
      *(output omitted )
      predict xb if e(sample), xb
      
      generate mills = normalden(-xb) / (1 - normal(-xb))
      
      replace mills = 0 if female==0
      
      oaxaca lnwage educ exper tenure mills, by(female) adjust(mills)
      So it doesn't look like he includes a gender dummy in the selection model, but runs it separately for women.

      You might also take at the following two Statalist posts that also discuss two decomposition models similar to the Blinder-Oxaca decomposition:Hope that can get you started in the right direction!

      Comment


      • #4
        Originally posted by David Benson View Post
        Hi Alina,

        I'm not familiar with Oaxaca-Blinder Decomposition method, so this help may be worth what you pay for it. But apparently this is a common question about using the Oaxaca-Blinder Decomposition method because it has been previously posted twice on Statalist (see below) and it appears that neither of them got an answer:However, Ben Jann, the author of Stata's oaxaca command has an article explaining it in in the Stata Journal here . On p. 473-475 the author gives two ways to run the oaxaca with heckman (or, to do as you did and run the selection model outside of oaxaca and calculate the inverse mills ratio yourself.)

        Code:
        * Method #1
        . oaxaca lnwage educ exper tenure, by(female) model2(heckman, twostep
        > select(lfp = age agesq married divorced kids6 kids714))
        Code:
        * Method #2
        probit lfp age agesq married divorced kids6 kids714 if female==1
        *(output omitted )
        predict xb if e(sample), xb
        
        generate mills = normalden(-xb) / (1 - normal(-xb))
        
        replace mills = 0 if female==0
        
        oaxaca lnwage educ exper tenure mills, by(female) adjust(mills)
        So it doesn't look like he includes a gender dummy in the selection model, but runs it separately for women.

        You might also take at the following two Statalist posts that also discuss two decomposition models similar to the Blinder-Oxaca decomposition:Hope that can get you started in the right direction!
        Dear Mr Benson,

        Thanks a lot for such a thorough response.

        So it appears it is an unresolved issue, and my problem is very similar to the first post you cited.

        I've already referred to the B. Jann article already, and as you correctly pointed out, it is for females only. So again it doesn't really answer my question.

        Really confused about what to do now, if I should run the first stage of Heckman for females only, or for both genders, and if yes, then should I include the gender dummy in the probit or not.

        Comment


        • #5
          Well, as an outsider it looks like you run the first stage of the probit for females only (because Jann does, and he then replaces mills=0 for the men.), and I assume that you include both genders in 2nd stage of the oaxaca, because he includes by(female)in the 2nd stage.

          A couple other ideas of how to proceed:

          1. You could email Ben Jann directly (his email is listed in both the journal article as well as if you type:
          Code:
           net describe oaxaca, from(http://fmwww.bc.edu/RePEc/bocode/o)

          2. Similarly, if you type help oaxaca there are several other relevant user-written commands, and you could email their authors as well. For example, the authors of the nldecompose command also list their emails. (I've listed the code to find it below). When people ask questions on StataList about user-written commands, the answer given is often "email the authors." So this is both common and expected.
          Code:
          net sj 9-2 st0152_1
          3. Finally you could also post on the other posts I listed in #3 (or see if you can contact them directly) to see if they ever got an answer.

          Also, once you do get an answer, I would ask that you post it here (or somewhere on Statalist), so that others coming along can find it it seems like a very common question, and it would be nice to get it

          Comment


          • #6
            Originally posted by David Benson View Post
            Well, as an outsider it looks like you run the first stage of the probit for females only (because Jann does, and he then replaces mills=0 for the men.), and I assume that you include both genders in 2nd stage of the oaxaca, because he includes by(female)in the 2nd stage.

            A couple other ideas of how to proceed:

            1. You could email Ben Jann directly (his email is listed in both the journal article as well as if you type:
            Code:
             net describe oaxaca, from(http://fmwww.bc.edu/RePEc/bocode/o)

            2. Similarly, if you type help oaxaca there are several other relevant user-written commands, and you could email their authors as well. For example, the authors of the nldecompose command also list their emails. (I've listed the code to find it below). When people ask questions on StataList about user-written commands, the answer given is often "email the authors." So this is both common and expected.
            Code:
            net sj 9-2 st0152_1
            3. Finally you could also post on the other posts I listed in #3 (or see if you can contact them directly) to see if they ever got an answer.

            Also, once you do get an answer, I would ask that you post it here (or somewhere on Statalist), so that others coming along can find it it seems like a very common question, and it would be nice to get it
            Thank you very much once again for responding!

            Well the problem is, authors tend to be unresponsive, which is understandable as I can presume the inflow for a public email address.

            I'll still try my luck I suppose, and will definitely post the answer here if the issue does get resolved.

            Thank you once again.

            Comment

            Working...
            X