On evaluating how well a biomarker can predict treatment response with survival data