Survival with time-dependent covariates

The main page mentions starting an RFC for a new feature, but does not mention HOW to do do. Can someone clue me in?
I’m very interested in using counting process data with boosting. Also, I know how to do it efficiently (theory), can contribute the key slice of C code, and have applications. (I’m the author of the R survival package).

Terry Therneau
therneau@mayo.edu

Feel free to open a new GitHub issue for the proposed feature. Example: https://github.com/dmlc/xgboost/issues/7282. The more concrete proposal is, the better. You should also include papers and other references so that XGBoost maintainers can understand the proposed feature.

FYI, I wrote the AFT survival objective in XGBoost. It is not time-dependent. I’m quite curious as to how time-dependent AFT could be implemented. Looking forward to your GitHub issue. Thanks!

Thanks for pointing me in the right direction. I’ve added a request.

Time-dependent covariates for an AFT are a difficult issue. I get asked that question every few years wrt survival::survreg. For a Cox model, the underlying assumption is instantaneous risk: today’s covariates affect today’s hazard. But with an AFT your current position in time depends on the entire history of covariates: you spent two years of college partying hard and then reformed? That extra few months of “effective age” are still with you. This makes the bookkeeping for an AFT much harder, or I should rather say that I have not yet seen a way to make it simple.
Terry Therneau