0
$\begingroup$

I'm running a logistic regression (N = 15000) and the percentage accuracy in classification (sensitivity) of the model with predictors included = 0 (the SAME as the constant model without predictors). Meanwhile the specificity remains unchanged from the constant model to the model with predictors (both 100%).

I start with a rather high overall success rate from the constant (95.2%). Maybe that is too high to start and suggests a problem?

But the output of the logistic regression seems to suggest that my added predictor variables (gender (dichotomous); age (continuous); country of origin (dichotomous dummy coded)) have no influence on membership to the group. Does this simply mean the model is without value and I shouldn't use any of the other outcomes from it?

Output

$\endgroup$
  • $\begingroup$ Can you edit your question (hitting the "edit" link in lower left) to share more details about your full model (i.e. what the predictors are, how they are coded, the sample size, etc.)? $\endgroup$ – Alexis Aug 1 '14 at 23:21
0
$\begingroup$

It seems that class imbalance may be a problem. You have 95% of observations in one class. Try to apply some random subsampling/oversampling to the training data to make the classes more even. In the testing data keep the proportions of the classes as is 95% - 5%.

Also, in such a high class imbalance overall classification accuracy (success) is not a good measure. Look at e.g. F1 score or area under ROC.

$\endgroup$

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Not the answer you're looking for? Browse other questions tagged or ask your own question.