Probit average predicted probabilities

[Will Murphy]

Good afternoon,

I'm trying to generate average predicted probabilities of employment for those who are work-limiting disabled and the non-disabled to compare.

Whilst I have looked at numerous resources online and understood them, a paper's methodology I am trying to replicate uses the term 'average predicted probabilities'. I want to ensure the method I have found online is applicable to generate average predicted probabilities.

My dependent variable is: 'emp' (=1 if wages >0, =0 if otherwise) and my variable of interest is 'DISTYPE' ( 1 = work-limiting disabled; 2 = daily activity- limiting disabled; 4= non-disabled).

I was wondering whether any clarity could be provided regarding whether the following approach is valid:

 probit emp DISTYPE SEX ETH AGES1 URESMC1 HDPCH191 IND1 MARSTA1 HIQUAL81 REGWKR1 SKSBN911 FTPTWK1, nolog

 margins, at(DISTYPE=(1 4)) atmeans vsquish

Adjusted predictions                            Number of obs     =      4,830
Model VCE    : OIM

Expression   : Pr(emp), predict()
1._at        : DISTYPE         =           1
               SEX             =    1.538302 (mean)
               ETH             =    1.146791 (mean)
               AGES1           =    10.33685 (mean)
               URESMC1         =    10.54244 (mean)
               HDPCH191        =    .6463768 (mean)
               IND1            =    3.478054 (mean)
               MARSTA1         =    2.133747 (mean)
               HIQUAL81        =    26.86522 (mean)
               REGWKR1         =    6.403313 (mean)
               SKSBN911        =    .3763975 (mean)
               FTPTWK1         =   -.9333333 (mean)
2._at        : DISTYPE         =           4
               SEX             =    1.538302 (mean)
               ETH             =    1.146791 (mean)
               AGES1           =    10.33685 (mean)
               URESMC1         =    10.54244 (mean)
               HDPCH191        =    .6463768 (mean)
               IND1            =    3.478054 (mean)
               MARSTA1         =    2.133747 (mean)
               HIQUAL81        =    26.86522 (mean)
               REGWKR1         =    6.403313 (mean)
               SKSBN911        =    .3763975 (mean)
               FTPTWK1         =   -.9333333 (mean)

------------------------------------------------------------------------------
             |            Delta-method
             |     Margin   Std. Err.      z    P>|z|     [95% Conf. Interval]
-------------+----------------------------------------------------------------
         _at |
          1  |   .1761419   .0237056     7.43   0.000     .1296797     .222604
          2  |   .2090076   .0206316    10.13   0.000     .1685705    .2494448
------------------------------------------------------------------------------

Any help would be greatly appreciated.

[Maarten Buis]

What you computed are the probabilities at average values of the explanatory variables. This is not the same as the average probabilities. If that is what you want (and that is a good default) then you just have to remove the atmeans option (it is such a good default, that it is the default of margins).

[Will Murphy]

Perfect, thank you very much.

[Will Murphy]

What you computed are the probabilities at average values of the explanatory variables. This is not the same as the average probabilities. If that is what you want (and that is a good default) then you just have to remove the atmeans option (it is such a good default, that it is the default of margins).

Apologies, I have just been reading up on Blinder-Oaxaca decomposition as a possible means to assess a legislation's impact on wage differentials and discrimination between the non-disabled and work-limiting disabled. The results in #1 are strictly pre-legislation and utilise unlogged wages (I think removing the need for 2 step Heckman). Is the Blinder-Oaxaca decomposition a feasible method in which I can attempt to replicate the methodology in this literature (attached, pp 33- 34) for pre- and post-legislation to compare?

Any help would be greatly appreciated, thank you.