Diagnosing most cancers at an earlier stage improves survival because treatment is more effective and less invasive; however, except for cancer screening, no known factors are associated with stage at cancer diagnosis. Using a sequential, two-step approach, we employed different regularization penalties in binary logistic regression models to identify predictive factors and then evaluated them in ordinal regression models that varied in their proportionality assumptions. Using health, lifestyle and screening variables measured years before diagnosis from Alberta Tomorrow Project particpants, we identified sets of predictors that globally were associated with high or low cancer stage that were further evaluated for their association with more than two levels of stage. A simulation study will evaluate the large sample properties of this two-stage approach. This sequential approach enables both feature and model selection for ordinal response data using existing programs.