I've recently built a survival model with time-varying covariates (assuming equal time periods) using R and I am now in the process of putting this into Oracle for production using just the coefficients from the model. The goal is to score new data that comes in each day in order to produce a cumulative hazard estimate; however, I just realized that the production system doesn't contain any of the training data.

Am I right to assume that I can only calculate the instantaneous hazard as opposed to the cumulative hazard without historical data?



It depends on what kind of survival model you use and what you mean by "using just the coefficients from the model". You may need to integrate the instantaneous hazard to get the cumulative hazard, but that should be straightforward given that your time periods are equal and I assume you know the value of covariates for the all relevant times (if you do not have a full history for the individual you are trying to predict, then I cannot see how you would calculate cumulative hazards). So assuming you have the full history for the individual in question, the question becomes if you know the instantaneous hazard for all times.

If you have a parametric survival model (e.g. a Weibull proportional hazards model), having the parameters for the baseline (e.g. scale, shape) and the covariates (regression coefficients) is sufficient to calculate everything you want. Since you said that you have time-varying covariates, I suspect that you have a semi-parametric model, such as a Cox proportional hazards model, thus getting the baseline hazard out is slightly more work, but perfectly doable. It is my understanding that using the standard Cox model with time-varying covariates allows you to continue to use all the standard methods with it, so there should be no problem that methods do not work as expected. There is some discussion of getting the hazard out of a Cox model here: Cox baseline hazard

If you can give me a reproducible example of what kind of model you want, I can check and attempt to give you a code example.


Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Not the answer you're looking for? Browse other questions tagged or ask your own question.