Figure 4

Misclassification rates of multivariate logistic models. All models were calculated using dead of disease at 10 years as a binary variable. Makers were assessed as a ratio of tumor expression to normal epithelial expression (Δ). All models included stage and grade. The ability of various models to predict outcome was assessed using LOOCV (a). 95% confidence intervals are shown in black (lower) and gray (upper). Analysis of stage and grade improves the misclassification of outcome from 26 (using a null model) to 19.6%. Adding VEGF or VEGF-Rs to this model does not significantly improve the misclassification rate. Two interesting interactions were noted: (b) shows that the ability of VEGF to predict outcome (y axis) improves on tumors with higher levels of VEGF-R1 (x axis). Without taking VEGF-R1 into account, the odds ratio for VEGF is 1.49 (dashed line). In contrast, (c) shows the ability of VEGF-R2 to predict outcome improves on tumors with lower VEGF-R1 levels. The dashed line shows the odds ratio of VEGF-R2 without the R1 interaction (0.71). The addition of the interactions outlined in panels b and c improves the misclassification rate to 17.4% (a).