Undeniably, calibration curves for survival probabilities are the preferred method. The idea of risk groups as proposed from Roystol et al was used as a technical device to indicate how well the model fits and predicts in a validation sample, for example by graphing or tabulating performance. In order to create a mean survival curve for a KM survival curve, is So(t)^ exp(mean(LP)) a valid way?
Could “β being far from 1.0”, mean we had inadequate sample size (~300) or was something wrong with our usage of R in estimating it? If not, what can be done next?
As always, thank you for your insight and time.