1

I am running a multinomial logistic regression model (with 3 possible outcomes) in R. I am trying to find the best way to assess the predictive power/accuracy of the model, and the best thing I've come up with is using a ROC curve.

For multi-class ROC analysis, I know that there is the one vs. one comparison or the one vs. all comparison. For the one vs. one comparison, would I need three separate ROC curves for each possible combination of outcome comparisons? If so, do I need to make a third model for comparing the two outcomes that were initially being compared to the baseline outcome?

For one vs. all comparison is the threshold for a r "random model" now 33% instead of 50%?

And finally, is there a better way to go about doing this/visualizing it?

EDIT: I know the pROC package has a multi class.roc function, but I don't totally get what it does.

user3451457
  • 11
  • 2
  • 1
    did you search through previous answers? have a look here [link](http://stats.stackexchange.com/questions/38541/creating-roc-curve-for-multi-level-logistic-regression-model-in-r?rq=1) – Jens Jul 29 '15 at 19:53

0 Answers0