Survival analysis model generation and performance evaluation

[#1]

I want to predict mortality over 15 years in a cohort of patients. Data is right censored, with a  mortality of 7% in 15 years of observation. The data set to derive a prediction model is about 3,500 subjects and 2,000 for external validation. 

I ran a Cox proportional hazard model, where I identified significant predictors and their Risk ratios, and a parametric survival analysis where significant predictors and probability estimation is available in the output.

However, 1. Why I can't generate a prediction probability formula with the fit proportional hazard report, 2.  how can 
I evaluate the performance (accuracy, AUROC, etc) of the prediction formula derived from the parametric survival platform in the same way I can do it with the nominal logistic platform ?

[#2 · Accepted solution]

In the Cox Proportional Hazards model, the dependent variable is continuous:  time to event.  So, there is no probability to estimate - the censored nature of the data does not really permit it.  Similarly, the measures you seek (AUC, etc.) do not arise for a continuous response variable.  You do get the survival curve, so you can use survival probabilities as a function of time and you do get information for confidence intervals.  You should be able to measure some performance characteristics using the saved prediction formula and your validation data (e.g., comparing the estimated survival percentage compared with the actual survival percentage as functions of the number of years).

[#3]

Thank you for your prompt and clear response