Unclear output logistic regression

Dear All,

I performed a logistic regression with SPSS to determine which factors were associated with a binomial variable. Briefly, the binomial dependent variable is a journal (i.e. journal 1 vs journal 2), while the covariates are characteristics of the published papers.

I performed both "enter" and "forward conditional" methods.

I was expecting a normal logistic regression output, while all the covariates in the final model were not significant. Strangely, the Nagelkerke r square was .96; when I made the ROC curve it was like .99.

I never obtained this kind of "perfect" results with logistic regression.

I try to attach my output in here. In any case I had 11 covariates for 114 events (so I don't think the model is overfitted).

Can somebody assist me in the interpretation of this strange model results?

Thank you for assistance.



Omega Contributor
Yup, I have had a model like this. Surprised you didn't get a complete separation warning. I actually had one of my none eve ts with a 99.9996 probability of an event that was an asymptomatic patient. I dont see anything wrong with it, fits real well. A good program will pick up ifyou have an Independent variable that is a component of the dependent variable.


Omega Contributor
I use SAS and it will let me know, if it is detetable. I do not know your scenario, but you should know your variables well enough to be able to make an educated guess (e.g., having weight predict body mass index (BMI), etc.).
OK. Thank you.
My last concern is statistical significance of the covariates in the final model.

How is it possible that such "perfectly fitted" model do not have any covariate that is significant?

Maybe I am too cautious, but I've never see this kind of perfect fit in my field.


TS Contributor
I would say your effects are way too large, and I suspect that the results aren't stable. So I would not stick with this model. I would start with figuring out what is wrong with your variables.


Omega Contributor
maartenbuis, brings up a good point on the stability of the model. I had not realized that your covariates were not significant, since they were not a part of your output.

Are the effects big, but the standard errors even bigger (e.g., large or very small odds ratios with vary broad confidence intervals)