Announcement

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

  • getting predicted probability larger than 1 using margins after xtologit

    Dear statalist:

    I have an individual-year panel, and I am estimating a fixed effect ordered logit model using xtologit. My dependent variable is happiness measured from 1 to 5. Here is my code, very simple:

    Code:
     xtologit happiness hhincome_per_log rd_hhper health , vce(cluster CID)
    To obtain the marginal effect at outcome(5), I run

    Code:
    margins, dydx(*) predict(pu0 outcome(5)) atmean post
    I get the following result:

    Code:
    . margins, dydx(*) predict(pu0 outcome(5)) atmean post
    
    Conditional marginal effects                      Number of obs   =       8644
    Model VCE    : Robust
    
    Expression   : Predicted mean (5.happiness), assuming u_i=0, predict(pu0 outcome(5))
    dy/dx w.r.t. : hhincome_per_log rd_hhper health
    at           : hhincome_p~g    =    9.837705 (mean)
                   rd_hhper        =    .0920682 (mean)
                   health          =    .6171911 (mean)
    
    ----------------------------------------------------------------------------------
                     |            Delta-method
                     |      dy/dx   Std. Err.      z    P>|z|     [95% Conf. Interval]
    -----------------+----------------------------------------------------------------
    hhincome_per_log |  -.0158093   .0054077    -2.92   0.003    -.0264082   -.0052103
            rd_hhper |  -1.386275   .1765556    -7.85   0.000    -1.732318   -1.040233
              health |   .1109692   .0088198    12.58   0.000     .0936827    .1282556
    ----------------------------------------------------------------------------------

    I don't understand why the absolute value of the coefficient of rd_hhper is larger than 1.

    I run dozen of similar regressions using the same dataset, most of the results look normal. As far as I understand, the coefficients here mean how the probability of having happiness being 5 will change with the change of the explanatory variable. But maybe I am wrong. Any help is appreciated!

  • #2
    Welcome to Statalist.

    The marginal effect given by dy/dx is the derivative of the predicted probability with respect to the rd_hhper. You think of it as how a one-unit change in rd_hhper would change the predicted probability of outcome 5, although in fact the probability is not linear in the independent variable.

    The question for you is, what is the range of rd_hhper and does a one-unit change in rd_hhper make sense? I suspect not, since the mean of rd_hhper is 0.092. So if a reasonable change in rd_hhper is 0.1, then the change in the probability would be something on the order of -.13.

    Comment


    • #3
      How is rd_hhper coded? What is its range, i.e. high and low values? You may wish to rescale it.
      -------------------------------------------
      Richard Williams, Notre Dame Dept of Sociology
      Stata Version: 17.0 MP (2 processor)

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

      Comment


      • #4
        Yes! That makes sense! I have scaled down the original value of rd_hhper by 1/100000 to avoid extremely small coefficients in the linear fixed effect estimation, so one unit change here means 100000 change in the original value. As you have correctly pointed out, the reasonable change here is 0.1. Thank you very much, you save me day.

        Comment

        Working...
        X