Extreme conditions that make current code fail

[CrystalXuR]

Examples are:

1. No one in treatment or control arm
2. No one is censored
3. PS is not the same across subjects (from an observational study)
4. Lengths of variables do not match, e.g., matrix dimensions
5. Types of variables are incorrect, e.g., numeric vs. factors
6. etc.

[erikcs]

Basic things like X,Y,Z etc are numeric and same lengths are easy to check.

For other things we need to think which invariants we wish to maintain given in input tuple (X,Y,W,D, times). For example if we require times > min(Y) and not all subjects censored then the following is always well defined https://github.com/som-shahlab/survmetabench/blob/master/r-package/survlearners/R/surv_fl_grf.R#L76

Some other things like X/T learners that split data into treated/control requires something more

[erikcs]

The requirement on n.min{W==0} and n.min{W==1} can maybe be made a function of n.folds and N to make things like https://github.com/som-shahlab/survmetabench/blob/master/r-package/survlearners/R/surv_xl_lasso.R#L39 always well defined

even if all D==0 for a given W things like https://github.com/som-shahlab/survmetabench/blob/master/r-package/survlearners/R/surv_xl_lasso.R#L42 are still well defined, the curve is just a constant, so ideally that doesn't need to be a requirement, though it's wierd input a user should have thought about