Fig. 5

Bootstrap resampling provides confidence intervals for anticipated performance on independent CRC data sets. Bootstrap resampling of 5-fold cross-validation of the subC-LP signature was used to generate distributions of hazard ratios for sensitive and relatively-resistant groups (a), number of patients in each group (b), and area between curves Abc (c). Predictive risk for the signature is found to be pRisk = P (Abc < 0) = 0.018. The histograms in (d) show in more detail that the distributions of hR(S) and hR(R) barely overlap. Figures e-h display the corresponding statistics generated from a randomized data set, showing that the distributions of hR(S) and hR(R) (e, h) are almost completely overlapping, as expected for this negative control. The predictive risk for the randomized model is 0.44 (g)