Logistic Regression Diagnostics (zero model sensitivity)

[Hakuna_101]

greetings,
I built a model using univariate analysis then Walt statistic then LR test and arrived at a model that looks good.
There were no collinearity issues and the GOF tests all look good.

Then when looking at deviance residuals and the standarized Pearsons residuals I realized there was something wrong. But I am new to Stata and I can not tell what the problem is.

Also, the classification of the model looks really bad (zero sensitivity and zero positive predictive power) !

Can someone please steer me in the right direction ? am I missing something huge here ?

I should note that the outcome frequency in this sample (positives) is around 0.03

Somebody advised me that I am expecting too much of the model to leave the cuttoff point at 0.5.

I attached some pictures for clarity.

Thank you in advance

[Richard Williams]

When one of the outcomes is very rare, as in this case, my experience is that the classification table usually predicts every case to be the more common outcome. Personally I don't find the table very useful in such cases.

[Hakuna_101]

Thank you Richard,

That sounds about right. However, I am not sure that the graphs look ok. Because of the very low probabilities (and therefore the low cuttoff point that should be chosen), the deviance residuals and the standarized Pearson's are too big. Am I correct in that assumption?

Is there a way to adjust for rare outcomes ?

[Richard Williams]

The user-written firthlogit (available from SSC) is often used for rare events. See http://www.statalist.org/forums/foru...git-estimation for suggested readings.