Comparison

tausurv puts classical, tree-based and neural survival models behind one predictor interface, so every model is fitted, predicted and scored the same way, competing risks included. The table shows where that coverage overlaps with the established Python libraries and where it falls short. For penalized Cox, time-varying covariates, left truncation and hypothesis tests, lifelines and scikit-survival remain the better choice today.

Feature	tausurv	scikit-survival	lifelines	pycox	hazardous
Nonparametric
Kaplan-Meier	Yes	Yes	Yes	No	No
Nelson-Aalen	Yes	Yes	Yes	No	No
Aalen-Johansen	Yes	Yes	Yes	No	No
Left truncation (delayed entry)	No	Yes	Yes	No	No
Log-rank test	No	Yes	Yes	No	No
Parametric univariate fits	No	No	Yes	No	No
Regression
Cox proportional hazards	Yes	Yes	Yes	No	No
Penalized Cox (L1, L2, elastic net)	No	Yes	Yes	No	No
Time-varying covariates	No	No	Yes	No	No
Proportional-hazards test	No	No	Yes	No	No
AFT (Weibull, log-normal, log-logistic)	Yes	No	Yes	No	No
Aalen additive hazards	No	No	Yes	No	No
Fine-Gray	Yes	No	No	No	No
Trees and boosting
Survival tree	Yes	Yes	No	No	No
Random survival forest	Yes	Yes	No	No	No
Gradient boosting	Yes	Yes	No	No	Yes
Competing-risks gradient boosting	Yes	No	No	No	Yes
Survival support vector machine	No	Yes	No	No	No
Neural networks
DeepSurv	Yes	No	No	Yes	No
DeepHit, with competing risks	Yes	No	No	Yes	No
Logistic hazard	Yes	No	No	Yes	No
Cox-Time, MTLR, PC-Hazard	No	No	No	Yes	No
Deep survival machines	Yes	No	No	No	No
CopulaSurv (dependent censoring)	Yes	No	No	No	No
HACSurv (dependent competing risks)	Yes	No	No	No	No
Discrimination
Harrell's concordance	Yes	Yes	Yes	No	No
Uno's concordance (IPCW)	Yes	Yes	No	No	No
Antolini's time-dependent concordance	Yes	No	No	Yes	No
Time-dependent AUC	Yes	Yes	No	No	No
Prediction error and calibration
Brier score (IPCW)	Yes	Yes	No	Yes	Yes
Integrated Brier score	Yes	Yes	No	Yes	Yes
Censored negative log-likelihood ¹	Yes	No	No	No	No
CRPS	Yes	No	No	No	No
Calibration curve	Yes	No	Yes	No	Yes
D-calibration	Yes	No	No	No	No
Competing-risks metrics
Cause-specific concordance	Yes	No	No	No	Yes
Cause-specific Brier score	Yes	No	No	No	Yes
Cause-specific AUC	Yes	No	No	No	No
Cause-specific calibration	Yes	No	No	No	Yes
Workflow
Survival-aware cross-validation and tuning ²	Yes	Partly	No	No	No
Nested cross-validation	Yes	No	No	No	No
Save and load fitted models ³	Partly	Yes	Yes	Yes	Yes
Bundled datasets ⁴	Yes	Yes	Yes	Yes	No
Survival curves with at-risk tables	Yes	No	Yes	No	No
Time-dependent SHAP plots ⁵	Yes	No	No	No	No

pycox's nbll is the IPCW binomial log-likelihood, a Brier-type score, not the model likelihood.
scikit-survival ships scorers for scikit-learn's search tools, without survival-specific folds or nested cross-validation.
tausurv saves linear and neural models. SurvivalBoost pickles but has no save method; survival trees and forests can be neither saved nor pickled yet.
tausurv bundles 100 published cohorts, each verified against a pinned checksum.
tausurv plots SurvSHAP(t) values; computing them needs a SHAP library.

A cell is filled when the library ships a public class or function for exactly that feature. A subset (a ridge-only AFT), a workaround (Kaplan-Meier on the inverted event flag) or a neural stand-in for a classical model does not count. Checked on 2026-10-05 against scikit-survival 0.27.0, lifelines 0.30.3, pycox 0.3.0 and hazardous 0.2.0. If a cell is wrong, open an issue.