0
$\begingroup$

I'm currently working with binary logistic regression in SAS to predict the probability of loan default and I have a problem with sensitivity and specificity.

I have a data sample of about N=3650 with dependent variable DEFAULT(=1) or NOT DEFAULT(=0) and 14 independent variables (both continuous and categoricals). I used stepwise selection with 0.1 significance level to select only the significant variables (finally 7 of them).

My problem is when I get the classification table with probability level 0.5, the percentages of sensitivity and specificity are 0% and 100% respectively. I have tried with different prob. levels (0.4 and 0.3) with no luck. Searching on the internet I found that one possible reason is the rare of events. In my data, the events(=1) are about 7% of the data and non events(=0) are about 93%.

What is the reason of that and finally how can I improve the classification?

Improve this question
$\endgroup$
2
  • $\begingroup$ How are you converting the predicted probabilities into predicted classes? It is common to call every case where the predicted probability is greater than some threshold a 1 & otherwise a 0. What is the threshold you are using? What does the distribution of pred probs look like? $\endgroup$
    – gung - Reinstate Monica
    Commented Aug 21, 2015 at 20:28
  • $\begingroup$ If the estimated event probability exceeds 0.5 the observation is predicted to be an event observation. Otherwise, it is predicted as a non event observation. $\endgroup$
    – andlak
    Commented Aug 21, 2015 at 21:17

0

Reset to default

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service and acknowledge you have read our privacy policy.