Announcement

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

  • Tiny Marginal effects after xtprobit

    I am running a random effects probit model with robust standard errors using the below code
    Code:
    xtprobit depvar indepvar, vce(cluster id)
    My problem is that when i try to get the average marginal effects using
    Code:
    margins, predict(pu0) dydx(*)
    I get very tiny numbers (between 6.07e-14 and 6.88e-08) that cannot be right as I get much higher marginal effects when i run the pooled probit model
    Code:
    probit depvar indepvar
    followed by
    Code:
    margins, dydx(*)
    as well as when I introduce dynamics (by including the lag of my dependent variable into my regression). Any insight as to what I am doing wrong here will be greatly appreciated

    In the attached file I paste the table with the random effects regression results, the output from the marginal effects when I use
    Code:
    margins, predict(pu0) dydx(*)
    and the output when I use
    Code:
    margins, dydx(*)
    Attached Files
    Last edited by Christina Chara; 27 Nov 2014, 07:37.

  • #2
    On the last one you are defaulting to predict(xb), whereas in the one before it you are using predict(pu0), So make sure you are comparing comparable things.

    Other than that I am afraid I can't tell much. You might try simplifying the model, e.g. just gender, and gradually making it more complex.
    -------------------------------------------
    Richard Williams, Notre Dame Dept of Sociology
    Stata Version: 17.0 MP (2 processor)

    EMAIL: [email protected]
    WWW: https://www3.nd.edu/~rwilliam

    Comment


    • #3
      thank you very much for this. Two more questions if I may: what is the difference in the interpretation of the two margins command? And which is the right one for the random effects probit model? the more I am reading about it the more it confuses me. Is the margins, dydx(*) command completely wrong as a command after xtprobit?
      Last edited by Christina Chara; 27 Nov 2014, 09:41.

      Comment


      • #4
        Personally I would use predict(pu0) with xtprobit. That way you are dealing with predicted probabilities rather than the linear prediction. Note that with xb, the marginal effects are almost the same as with the original xtprobit coefficients; the main differences are due to the age and age^2 term, whose effects get combined in the marginal effects.

        For categorical independent variables (e.g. gender, race) I am fine with either adjusted predictions or marginal effects. For continuous independent variables (e.g. income) I personally don't like marginal effects (I find them hard to interpret) although economists are more fond of them. I prefer to compute adjusted predictions across a range of values for the continuous variable.

        If none of that is making much sense, see these two handouts:

        http://www3.nd.edu/~rwilliam/stats/Margins01.pdf

        http://www3.nd.edu/~rwilliam/xsoc73994/Margins03.pdf

        Also consider looking at Vince Wiggins' excellent discussion of marginal effects and interactions:

        http://www.stata.com/statalist/archi.../msg00293.html

        I still find your tiny effects puzzling but maybe it will help if you simplify the model like I suggested earlier.
        -------------------------------------------
        Richard Williams, Notre Dame Dept of Sociology
        Stata Version: 17.0 MP (2 processor)

        EMAIL: [email protected]
        WWW: https://www3.nd.edu/~rwilliam

        Comment


        • #5
          Dear Richard,

          Than you very much for all the advice and for the indeed very useful references

          Comment

          Working...
          X