I find it more straight forward to just model the failure rate with the variables directly, and look metrics like AUC for out of sample data.