Negative individual predicted probabilities were identified in the final model fit. You may want to

[Demographer]

Hi,

I want to run a ordered logistic model with unequalslopes. The predicted variables edunum has 3 categories: 0,1,2. The code is as follows:

proc logistic data=work.immigrants_from_lfs;
class sex pob_num(ref='0') period /param=ref;
model edunum(descending)= period sex|cohort2 pob_num|cohort2 /unequalslopes;
weight weight / norm;
score data=work.immigdataset out=work.scored_immig;
run;

I get this weird warning message:

WARNING: Negative individual predicted probabilities were identified in the final model fit. You may want to modify your

UNEQUALSLOPES specification.

WARNING: The LOGISTIC procedure continues in spite of the above warning. Results shown are based on the last maximum likelihood iteration. Validity of the model fit is questionable.

The message only appears when both sex|cohort2 pob_num|cohort2 are included in predictors. Sex has 2 categories (0,1), pob_num has 8 categories (0-7) and cohort 2 has 99 categories (0-98).

Parameters look however ok. I tried to remove the weight and score statement, but it does not fix it. Any idea why is this happening? I don't understand how a logistic model can predict negative probabilities.

[sbxkoenk]

Hello,

---

@Demographer wrote:

I don't understand how a logistic model can predict negative probabilities.

---

This paragraph (note) might give an insight into the answer to your question.

See the note on the bottom of this page :
SAS/STAT® 15.2 User's Guide
The LOGISTIC Procedure
Example 78.18 Partial Proportional Odds Model

https://go.documentation.sas.com/doc/en/statug/15.2/statug_logistic_examples22.htm

I am not sure if such a model can überhaupt be used.

I think @SAS_Rob can answer this.

Thanks,

Koen

[SAS_Rob]

I would not rely on the results when this occurs.  The problem is that the data does not support the model so that you will need to change the model specification in some way to get interpretable results.  The suggestion to modify the UNEQUALSLOPES option so that you fit a partial proportional odds model is a good place to start.

I will mention as well that although the estimates seem reasonable, clearly they are not, at least collectively, when they a negative predicted probability.