Quantile RSF

[YasserZ95]

I have a question about using Random Survival Forests (RSF) in scikit-survival to predict conditional quantiles.

Briefly, what I have done is the following. I fit an RSF model to right-censored data using RandomSurvivalForest. For prediction, I obtain the estimated survival function $S(t \mid x)$ via predict_survival_function and then compute the $\tau$-quantile (e.g., $\tau = 0.95$) as the first time point at which $S(t \mid x) \le 1 - \tau$.

I wanted to check whether this workflow is consistent with the intended use of RSF in the package, and whether there exists a different or more standard approach for quantile prediction under right censoring.

[sebp]

Do you want to predict the time point on an individual's survival function that corresponds to the $\tau=95\%$ percentile, i.e. find $t$ such that $\tau = S(t \mid x)$? If yes, your approach should work, but you have to consider that $t$ isn't defined when the survival function never reaches the $\tau$ percentile.

See also #190

[YasserZ95]

Yes! Thanks!