odd ratios ologit regressions

[Ralf Gier]

Hi everyone! I have run some ologit regressions for my ordinal categorical variable which can take on a value of -1, 0 or 1. my ologit results are like this:

     

Ordered logistic regression                       Number of obs   =      15410
                                                  Wald chi2(8)    =     104.18
                                                  Prob > chi2     =     0.0000
Log pseudolikelihood = -16743.615                 Pseudo R2       =     0.0019

                                  (Std. Err. adjusted for 21 clusters in year)
------------------------------------------------------------------------------
             |               Robust
     UPGRADE | Odds Ratio   Std. Err.      z    P>|z|     [95% Conf. Interval]
-------------+----------------------------------------------------------------
        CASH |   .9213927   .0336132    -2.24   0.025     .8578123    .9896857
       VALUE |   1.044413   .0127019     3.57   0.000     1.019812    1.069607
        SIZE |   1.033685   .0182644     1.88   0.061     .9985004     1.07011
  EXPERIENCE |   .9717727   .0277187    -1.00   0.315     .9189358    1.027648
        DAYS |   .9997179   .0002154    -1.31   0.190     .9992958     1.00014
   FOLLOWING |   .9983839   .0009214    -1.75   0.080     .9965797    1.000191
     AFF_ACQ |   1.102509   .1018432     1.06   0.291     .9199266    1.321329
     AFF_TAR |   .9554293   .0950913    -0.46   0.647     .7861054    1.161225
-------------+----------------------------------------------------------------
       /cut1 |  -.1086019   .1233542                     -.3503718    .1331679
       /cut2 |   1.002035   .1175542                      .7716332    1.232437
------------------------------------------------------------------------------

I would like to have the odd ratio's of these independent variables for dependent variable =1. therefore i have used the following code and got the following results:

mfx, predict(outcome(1))

Marginal effects after ologit
      y  = Pr(UPGRADE==1) (predict, outcome(1))
         =  .35006879
------------------------------------------------------------------------------
variable |      dy/dx    Std. Err.     z    P>|z|  [    95% C.I.   ]      X
---------+--------------------------------------------------------------------
    CASH*|  -.0186356      .00827   -2.25   0.024  -.034854 -.002418   .523297
   VALUE |   .0098869      .00276    3.58   0.000   .004473  .015301   5.55294
    SIZE |   .0075378      .00402    1.88   0.061  -.000335  .015411   8.37543
EXPERI~E |  -.0065147      .00646   -1.01   0.313  -.019182  .006153   1.21899
    DAYS |  -.0000642      .00005   -1.31   0.190   -.00016  .000032   95.7663
FOLLOW~G |   -.000368      .00021   -1.75   0.079  -.000779  .000043   19.8554
 AFF_ACQ*|   .0225002      .02152    1.05   0.296   -.01967   .06467   .024205
 AFF_TAR*|  -.0103046      .02235   -0.46   0.645  -.054114  .033505   .021285
------------------------------------------------------------------------------
(*) dy/dx is for discrete change of dummy variable from 0 to 1

can I interpret the dy/dx as the odd-ratios? or do I need to use another code.

[Clyde Schechter]

The dydx outputs are most definitely not the odds ratios. For one thing, there is no such thing as a negative odds ratio!

The dydx outputs represent the change in the probability of outcome = 1 associated with a unit increase in the corresponding independent variables. And they are in the probability metric, not the odds metric.

I would like to have the odd ratio's of these independent variables for dependent variable =1.

I don't know what this means. The -ologit- model relies on the proportional odds assumption, which stipulates that the odds ratio for being above vs below any off the levels of the outcome is the same. So in this model there is no odds ratio specific to outcome = 1. The output of -ologit- itself gives you your estimate of the odds ratio that contrasts outcomes -1,0 with outcome 1. It is, within this kind of model, necessarily the same as the odds ratio that contrasts -1 with 0, 1.

[Ralf Gier]

Thanks for your fast reply Clyde.

I think there is a fault on my side. I understood from several papers that when you run ordered logistic regressions and multinominal regressions, you report the odd ratio of your independent variables. but because i'm working with categorical dependent variables, it might make more sense to report the marginal effects.

As I understand from your post, the dxdy is the marginal effect, am I correct?

kind regards

[Clyde Schechter]

Yes, dydx is the marginal effect. As for which to report, marginal effects or odds ratios, it is a matter of taste. Or, more pragmatically, it is a matter of what is customary in your discipline.

[Marcos Almeida]

Clyde fully clarified the dydx "issue" to Ralf.

Below, I just wish to share an excerpt from the manual (http://www.stata.com/manuals13/rmargins.pdf), where it approaches the dydx as well as the (related) elasticities in a nutshell:

[...]the interpretations are
dydx(): change in y for a change in x
eyex(): proportional change in y for a proportional change in x
eydx(): proportional change in y for a change in x
dyex(): change in y for a proportional change in x

I liked the visual presentation of a set of possibilities on a par with the interpretation.

Best,

Marcos