Calibration of Cox model

calibration

#1

I am using Cox model for risk prediction, and I am kind of lost regarding calibration. I am reading on Nam D’Agostino test, but I am unable to grasp the time period to be selected for calculating predicted probabilities. The predicted probabilities vary depending on the time period I select for calculating the probabilities, which directly affects my observed vs predicted probabilities/rates across deciles. I was thinking of using the mean follow-up time in the data to calculate predicted probabilities but a post on statsexchange recommended against using it. I am not sure how to go about this.

(I am using administrative claims data with significant right censoring. Because the enrollment information is month-to-month, my predicted probabilities are monthly)

Thank you.


#2

It is not appropriate to use deciles or any other binning of continuous predicted values. My RMS book and course notes go into details related to obtaining smooth semi-parametric calibration curves for survival models, for a single time horizon. This is implemented in the R rms package calibrate.cph function. For examining the entire range of times and not selecting a single time point, there are residual-based methods you may want to take a look at, e.g., see the R rms package val.surv function.