Address a negative probability value when using the margins command

[DY Kim]

Hi
I ran a gologit2 model and then computed predicted probabilities for the ordinal dependent variable with 3 categories using the margins command. susr and offr are categorical variable with 4 categories. susw is a categorical variable with 3 categories. vio is a continuous variable.

After estimating the margins, however, I found a ‘implausible' negative probability and C.I values for the outcome(2). Probability should be positive. Is my understanding is right? I just wonder why it occurs and how to address this negative probability value? Any advice would be appreciated.

. margins, at(susr=3 offr=1 susw=3 vio=55.731)

Predictive margins Number of obs = 3,342
Model VCE : OIM

1._predict : Pr(forc==1), predict(pr outcome(1))
2._predict : Pr(forc==2), predict(pr outcome(2))
3._predict : Pr(forc==3), predict(pr outcome(3))
at : susr = 3
offr = 1
susw = 3
vio = 55.731

------------------------------------------------------------------------------
| Delta-method
| Margin Std. Err. z P>|z| [95% Conf. Interval]
-------------+----------------------------------------------------------------
_predict |
1 | .4373939 .0940638 4.65 0.000 .2530323 .6217555
2 | -.2668683 .063431 -4.21 0.000 -.3911907 -.1425458
3 | .8294743 .0555001 14.95 0.000 .7206961 .9382526
------------------------------------------------------------------------------

[DY Kim]

Please let me know if you need additional information or I didn't comply with any rules for posting. I greatly appreciate if anyone can answer my question.

[FernandoRios]

Hi DYKim,
I dont think there is much to be done with margins. You may want to use a different model specification.
What happens is that when you use a standard ologit, you are imposing parallel lines across the latent index for all order options. B
Since they are parallel, they do not cross (by definition), and the implied probabilities to belong to a group within those parallel lines is always defined between 0 and 1.

Now, when you use the generalized order logit, you are no longer imposing parallel lines. They will cross each other for some combinations of X's and Y's (most likely implausible combinations). And the implied probabilities will be, as you are observing, negative.

So, the only two things you could do.
1. change model specification, so it is less likely that the latent's cross each other.
2. Make sure that the point you are interested in analyzing is a plausible point (how many cases you observe that have a combination of X's (or similar) to at(susr=3 offr=1 susw=3 vio=55.731). If they are not very likely, then the chances of finding the negative predicted probabilities is high.

HTH
Fernando