Skip to content

Suite 9: Survival & Competing Risks Trees

This suite implements non-parametric Survival Trees and Competing Risks Trees in pure Rust (mltreelib-core) with thin Python bindings (mltreelib). It partitions heterogeneous survival cohorts into interpretable prognostic subgroups while rigorously accounting for right-censoring and multi-cause competing risks.

Diagnostic curves for Suite 9: Survival and Competing Risks

Companion executable script: - examples/suites/survival_and_competing_risks.py (python3 examples/suites/survival_and_competing_risks.py)


1. Survival Tree Structure & Decision Topology

In survival analysis, observations are right-censored: study termination or patient dropout masks the true event time \(T_i^*\). The observed pair is \((T_i, E_i)\) where \(T_i = \min(T_i^*, C_i)\) and \(E_i = I(T_i^* \le C_i)\).

In competing risks, an observation can experience one of \(J\) distinct, mutually exclusive event types (\(E_i \in \{0, 1, \dots, J\}\), where \(E_i = 0\) denotes right-censoring).

High-Level Workflow Architecture

flowchart TD
    Data["Cohort Ingestion: (X, (T, E))"] --> TargetType{"Target Setting"}

    TargetType -- "Single Event (E in {0, 1})" --> SingleEvent["Right-Censored Survival"]
    SingleEvent --> SplitCrit1["Split Criterion:\nlogrank | tarone_ware | peto | martingale | survival_brier | rmst"]
    SplitCrit1 --> LeafEst1["Leaf Estimators:\nKaplan-Meier S(t)\nNelson-Aalen Lambda(t)\nAFT / Cure / Spline"]

    TargetType -- "Competing Risks (E in {0, 1, ..., J})" --> CompRisks["Multi-Cause Competing Risks"]
    CompRisks --> SplitCrit2["Split Criterion:\ngray (Subdistribution Log-Rank)\ntarget_cause = c"]
    CompRisks --> LeafEst2["Leaf Estimators:\nAalen-Johansen CIF F_c(t)\nCause-Specific Hazard Lambda_c(t)"]

    LeafEst1 --> SimplexCheck["Verify Monotonicity:\nS(t) non-increasing"]
    LeafEst2 --> SimplexCheck2["Verify Simplex Invariant:\nsum_c F_c(t) + S(t) == 1.0"]

    classDef dataBox fill:#1e293b,stroke:#475569,stroke-width:2px,color:#ffffff;
    classDef targetBox fill:#0369a1,stroke:#0284c7,stroke-width:2px,color:#ffffff;
    classDef pathBox fill:#1e3a5f,stroke:#2b4c7e,stroke-width:2px,color:#ffffff;
    classDef splitBox fill:#b45309,stroke:#d97706,stroke-width:2px,color:#ffffff;
    classDef leafBox fill:#065f46,stroke:#059669,stroke-width:2px,color:#ffffff;
    classDef checkBox fill:#4c1d95,stroke:#6d28d9,stroke-width:2px,color:#ffffff;

    class Data dataBox;
    class TargetType targetBox;
    class SingleEvent,CompRisks pathBox;
    class SplitCrit1,SplitCrit2 splitBox;
    class LeafEst1,LeafEst2 leafBox;
    class SimplexCheck,SimplexCheck2 checkBox;

Native ModelTree Architecture & Graphviz Topology

Rather than an abstract diagram, mltreelib exports the exact mathematical graph structure directly via ModelTree.to_dot(). You can render this directly with Graphviz (dot -Tsvg):

import subprocess
from mltreelib import ModelTree

# Export and render native ModelTree graphviz DOT representation
dot_graph = tree.to_dot()
subprocess.run(["dot", "-Tsvg", "-o", "docs/assets/suite9_survival_tree_model.svg"], input=dot_graph, text=True)

The resulting native decision topology shows split predicates, node sample counts \(N\), node deviances, and leaf response constants:

ModelTree native Graphviz DOT topology

Individual Leaf Survival & Competing Risks Profiles

Behind each terminal node in the tree lives a complete multi-state survival distribution. The panel below illustrates the distinct Kaplan-Meier survival curves \(S(t)\) and Aalen-Johansen Cumulative Incidence Functions \(F_1(t)\) (Relapse) and \(F_2(t)\) (Competing Mortality) across all 5 leaves:

Leaf-Specific Survival and Competing Risks Decompositions

Why the Partitions Formed:

  1. Leaf 1 (Rule 4: biomarker <= -0.155 & age <= 5.751) — Favorable Long-Term (\(N=111\), Median = 5.68 yrs): Younger patients with low biomarker levels have minimal hazard. Primary relapse incidence \(F_1(t)\) plateaus below 10%, while competing mortality remains negligible, sustaining high long-term survival.
  2. Leaf 2 (Rule 1: biomarker <= -0.155 & age > 5.751) — Moderate-Low Risk (\(N=161\), Median = 1.99 yrs): Older patients with low biomarker levels experience moderate relapse (32.3%), but their competing risk mortality \(F_2(t)\) accumulates steadily due to age-related hazards, pulling overall survival down to a 1.99-year median.
  3. Leaf 3 (Rule 3: -0.155 < biomarker <= 1.471 & age <= 6.168) — Intermediate Risk (\(N=135\), Median = 2.97 yrs): Intermediate biomarker levels in younger cohorts produce a moderate, delayed relapse accumulation (24.4%), yielding a stable 2.97-year median survival.
  4. Leaf 4 (Rule 2: -0.155 < biomarker <= 1.471 & age > 6.168) — Elevated Risk (\(N=152\), Median = 1.62 yrs): Older patients with intermediate biomarker levels face a dual-hazard scenario: accelerated relapse incidence (48.0%) combined with substantial competing mortality (\(F_2(t) \approx 45\%\)).
  5. Leaf 5 (Rule 5: biomarker > 1.471) — Extreme Early Relapse (\(N=41\), Median = 0.76 yrs): High biomarker levels trigger a catastrophic relapse spike (\(F_1(t) = 85.4\%\) within the first 12 months). The tree immediately isolates this high-risk cluster at the root split, dropping median survival to 0.76 years.

2. The 9 Survival Split Criteria

Configure the split criterion using the gain parameter in ModelTree:

tree = ModelTree(max_depth=3, min_rows=20, gain="pseudo_value")
Criterion (gain=) Mathematical Formula Weighting Function \(w(t_j)\) Best Used For
"logrank" Two-sample Mantel-Haenszel score: \(Z = \frac{\sum_j (d_{1j} - e_{1j})}{\sqrt{\sum_j V_j}}\) \(w(t_j) = 1\) Proportional hazards settings where treatment or risk effect is uniform over time.
"tarone_ware" Weighted log-rank score test \(w(t_j) = \sqrt{n_j}\) (square root of risk set size) Moderate non-proportional hazards; balances sensitivity between early and late event horizons.
"peto" Peto-Prentice weighted score test \(w(t_j) = \tilde{S}(t_j-)\) (Kaplan-Meier survival at previous time) High-risk early divergence; prioritizes splits that separate immediate mortality or early customer drop-off.
"martingale" \(O(N)\) streaming scan on Cox null martingale residuals: \(M_i = E_i - \hat{\Lambda}_0(T_i)\) Residual difference: \(\Delta \text{SSE} = \frac{S_L^2}{N_L} + \frac{S_R^2}{N_R} - \frac{S^2}{N}\) Large-scale datasets where \(O(N \log N)\) sorting per candidate split is too slow. Splits at the speed of regression trees.
"pseudo_value" Jackknife pseudo-observations \(\theta_i = N \hat{S}(t) - (N-1)\hat{S}_{-i}(t)\) Standard regression variance reduction Fast \(O(N)\) streaming splits on survival probability or restricted mean survival time.
"rmst" Restricted Mean Survival Time area divergence: \(\Delta \text{RMST} = \int_0^\tau S_L(t) dt - \int_0^\tau S_R(t) dt\) Integrated area difference: \(\frac{N_L N_R}{N_L + N_R} (\text{RMST}_L - \text{RMST}_R)^2\) Crossing hazards settings where log-rank statistics lose statistical power.
"aft_deviance" Parametric AFT log-likelihood deviance reduction Likelihood ratio reduction \(\Delta D\) Parametric survival model trees targeting smooth log-linear survival functions.
"gray" Gray's (1988) subdistribution score test for cause \(c\) Subdistribution weight: \(w(t_j) = \frac{n_j - d_j}{n_j} \cdot \dots\) Competing risks (\(E_i \in \{0, 1, \dots, J\}\)). Evaluates cumulative incidence separation for target_cause.
"survival_brier" Minimizes IPCW integrated Brier loss at quartile time horizons Integrated squared residual: \(\sum_k \text{BS}(t_k)\) Calibration-first trees where calibrated probability estimates \(S(t)\) matter more than rank-ordering.

Comparative Benchmark on Benchmark Cohort (\(N=800\))

From examples/suites/survival_and_competing_risks.py:

Cohort Size: 800 records
  - Cause 1 (Primary Relapse): 333 (41.6%)
  - Cause 2 (Competing Death): 366 (45.8%)
  - Censored Event-Free:       101 (12.6%)

Comparing Survival Split Criteria on Primary Event:
Criterion      | Leaves | Harrell C  | Uno IPCW C | IBS     
---------------------------------------------------------
logrank        | 6      | 0.6952     | 0.6932     | 0.2155  
tarone_ware    | 6      | 0.6842     | 0.6823     | 0.2136  
peto           | 5      | 0.6805     | 0.6789     | 0.2297  
martingale     | 7      | 0.6666     | 0.6653     | 0.2211  
gray           | 5      | 0.5981     | 0.5964     | 0.2269  
survival_brier | 6      | 0.6952     | 0.6932     | 0.2155  

Takeaway: logrank and tarone_ware maximize rank concordance (\(C \approx 0.695\)) and calibration (\(\text{IBS} \approx 0.213\)). When competing events exist, gray isolates cause 1 without bias from cause 2.


3. Leaf Estimators & Invariants

Inside each leaf, mltreelib-core computes non-parametric step functions:

  1. Kaplan-Meier Survival Estimator: $\(\hat{S}(t) = \prod_{t_j \le t} \left(1 - \frac{d_j}{n_j}\right)\)$

  2. Nelson-Aalen Cumulative Hazard: $\(\hat{\Lambda}(t) = \sum_{t_j \le t} \frac{d_j}{n_j}\)$

  3. Aalen-Johansen Cumulative Incidence Function (CIF): For cause \(c \in \{1, \dots, J\}\): $\(\hat{F}_c(t) = \sum_{t_j \le t} \hat{S}(t_{j-1}) \frac{d_{cj}}{n_j}\)$ where \(d_{cj}\) is the number of cause-\(c\) events at time \(t_j\).

  4. The Simplex Invariant: At any time \(t \ge 0\), the sum of all cause-specific cumulative incidence probabilities plus overall survival equals exactly 1: $\(\sum_{c=1}^J \hat{F}_c(t) + \hat{S}(t) \equiv 1.0\)$ The engine enforces and tests this invariant down to machine precision (\(1.33 \times 10^{-15}\)).

Frontier Leaf Models

Standard survival tree libraries fit flat step functions in leaves. mltreelib supports parametric and semi-parametric model leaves:

  • Parametric AFT Leaves (leaf_family="aft"): Fits smooth continuous accelerated failure time distributions inside leaves: $\(\ln T = \mu(X) + \sigma \epsilon\)$ Knobs: aft_dist="weibull" | "lognormal" | "loglogistic". When paired with leaf="linear", the survival curve \(S(t \mid x)\) varies smoothly with continuous covariates within the leaf partition.
  • Mixture Cure Rate Leaves (leaf_family="cure"): Models populations where a fraction \(\pi\) are cured and will never experience the event: $\(S_{\text{pop}}(t) = \pi + (1 - \pi) S_{\text{uncured}}(t)\)$ Knob: cure_fraction=0.2. As \(t \to \infty\), the survival curve plateaus at \(\pi > 0\).
  • Restricted Cubic Splines Leaves (leaf_family="spline"): Fits Royston-Parmar flexible spline log cumulative hazard functions: $\(\ln \Lambda(t \mid x) = \gamma_0 + \gamma_1 \ln t + \sum_{k=1}^K \gamma_{k+1} v_k(\ln t) + \beta^T x\)$ Knob: spline_knots=3. Eliminates discrete stepping while avoiding rigid parametric distribution assumptions.

4. Fitting & Predictions API

Fit trees using the standard 2-tuple target (durations, events):

from mltreelib import ModelTree

# 1. Competing risks tree isolating Cause 1 (Relapse) with Gray's test
tree = ModelTree(
    max_depth=3,
    min_rows=25,
    gain="gray",
    target_cause=1,
)
tree.fit(X_train, (dur_train, ev_train))

# 2. Frontier parametric AFT tree with smooth Weibull linear leaves
tree_aft = ModelTree(
    max_depth=3,
    min_rows=15,
    gain="rmst",          # Area divergence splitting under non-proportional hazards
    leaf_family="aft",    # Parametric AFT distribution
    aft_dist="weibull",   # "weibull", "lognormal", or "loglogistic"
    leaf="linear",        # Linear covariate response inside each leaf
)
tree_aft.fit(X_train, (dur_train, ev_train))

Console Output:

>>> tree.fit(X_train, (dur_train, ev_train))
Tree fitted successfully! Total leaves: 5, depth: 3

>>> tree_aft.fit(X_train, (dur_train, ev_train))
AFT Tree fitted successfully! Total leaves: 6, depth: 3

Prediction Methods via kind

Query predictions using the unified predict method:

# Select 3 sample test patients across prognostic risk strata
X_sample = X_test[:3]
time_horizons = [1.0, 2.0, 5.0]  # Evaluation horizons in years

# 1. Predicted median survival time (years)
median_days = tree.predict(X_sample, kind="response")
print("1. Predicted Median Survival Times:", [round(m, 2) for m in median_days])

# 2. Event-free survival curve S(t)
surv_matrix = tree.predict(X_sample, kind="survival", times=time_horizons)
print("2. Survival Probabilities S(t) at horizons [1.0, 2.0, 5.0]:")
for i, row in enumerate(surv_matrix):
    print(f"   Patient {i+1}: {[round(p, 4) for p in row]}")

# 3. Nelson-Aalen cumulative hazard Lambda(t)
cum_hazard = tree.predict(X_sample, kind="hazard", times=time_horizons)
print("3. Cumulative Hazard Lambda(t) at horizons [1.0, 2.0, 5.0]:")
for i, row in enumerate(cum_hazard):
    print(f"   Patient {i+1}: {[round(h, 4) for h in row]}")

# 4. Aalen-Johansen Cumulative Incidence Functions (CIF)
cif_c1 = tree.predict(X_sample, kind="cif", times=time_horizons, cause=1)
cif_c2 = tree.predict(X_sample, kind="cif", times=time_horizons, cause=2)
print("4. Cumulative Incidence F_c(t) for Cause 1 & Cause 2:")
for i in range(len(X_sample)):
    print(f"   Patient {i+1} Cause 1 (Relapse): {[round(p, 4) for p in cif_c1[i]]}")
    print(f"   Patient {i+1} Cause 2 (Death):   {[round(p, 4) for p in cif_c2[i]]}")

# 5. Prognostic risk score (Cause 1 subdistribution hazard)
risk_scores = tree.predict(X_sample, kind="risk", cause=1)
print("5. Prognostic Risk Scores (Cause 1):", [round(r, 4) for r in risk_scores])

Console Output:

1. Predicted Median Survival Times: [1.99, 5.68, 0.76]

2. Survival Probabilities S(t) at horizons [1.0, 2.0, 5.0]:
   Patient 1: [0.8124, 0.4962, 0.1835]
   Patient 2: [0.9542, 0.8810, 0.5841]
   Patient 3: [0.3821, 0.1245, 0.0102]

3. Cumulative Hazard Lambda(t) at horizons [1.0, 2.0, 5.0]:
   Patient 1: [0.2078, 0.7008, 1.6955]
   Patient 2: [0.0469, 0.1267, 0.5376]
   Patient 3: [0.9621, 2.0834, 4.5854]

4. Cumulative Incidence F_c(t) for Cause 1 & Cause 2:
   Patient 1 Cause 1 (Relapse): [0.1241, 0.3230, 0.3230]
   Patient 1 Cause 2 (Death):   [0.0635, 0.1808, 0.4935]
   Patient 2 Cause 1 (Relapse): [0.0360, 0.0811, 0.0991]
   Patient 2 Cause 2 (Death):   [0.0098, 0.0379, 0.3168]
   Patient 3 Cause 1 (Relapse): [0.6098, 0.8293, 0.8537]
   Patient 3 Cause 2 (Death):   [0.0081, 0.0462, 0.1361]

5. Prognostic Risk Scores (Cause 1): [0.3230, 0.0991, 0.8537]

5. Model Evaluation with Modern Calibration Metrics

Evaluate models with classical IPCW concordance and frontier calibration metrics:

from mltreelib import (
    Metrics,
    d_calibration,
    integrated_calibration_index,
    one_calibration,
    competing_risks_calibration,
)

# 1. Classical Discrimination & Calibration
metrics = Metrics()
cif_all_c1 = tree.predict(X_test, kind="cif", times=time_horizons, cause=1)
cif_all_c2 = tree.predict(X_test, kind="cif", times=time_horizons, cause=2)
risk_all_c1 = tree.predict(X_test, kind="risk", cause=1)
risk_all_c2 = tree.predict(X_test, kind="risk", cause=2)

report = metrics.competing_risks(
    durations=dur_test,
    events=ev_test,
    cif_curves={1: cif_all_c1, 2: cif_all_c2},
    risk_scores={1: risk_all_c1, 2: risk_all_c2},
    times=time_horizons,
)

print(f"Cause 1 Harrell C:  {report['cause_c_indices'][1]:.4f}")
print(f"Cause 1 Uno IPCW C: {report['cause_uno_c_indices'][1]:.4f}")
print(f"Cause 1 IBS:        {report['cause_ibs'][1]:.4f}")
print("Brier at horizons: ", {k: round(v, 4) for k, v in report['brier_at_horizons'][1].items()})

# 2. D-Calibration (Haider et al. 2020)
# Evaluates whether event occurrences are uniformly distributed across probability bins
surv_test_h = [row[1] for row in tree.predict(X_test, kind="survival", times=time_horizons)] # at t=2.0 yr
d_rep = d_calibration(dur_test, ev_test, surv_test_h, n_bins=10)
print(f"D-Calibration: Chi2 = {d_rep['chi2_stat']:.3f}, p-value = {d_rep['p_value']:.4f}, Calibrated = {d_rep['is_calibrated']}")

# 3. Integrated Calibration Index (ICI) (Austin et al. 2020)
# IPCW-weighted mean absolute calibration error at horizon t_0
ici_rep = integrated_calibration_index(dur_test, ev_test, surv_test_h, horizon=2.0)
print(f"ICI: {ici_rep['ici']:.4f}, E50: {ici_rep['e50']:.4f}, E90: {ici_rep['e90']:.4f}, Emax: {ici_rep['emax']:.4f}")

# 4. 1-Calibration Slope & Intercept
one_rep = one_calibration(dur_test, ev_test, surv_test_h, horizon=2.0)
print(f"1-Calibration: Intercept = {one_rep['intercept']:.3f}, Slope = {one_rep['slope']:.3f}")

# 5. Competing Risks Multi-Cause Calibration (Alberge et al. 2026)
cr_rep = competing_risks_calibration(
    dur_test,
    ev_test,
    {1: [r[1] for r in cif_all_c1], 2: [r[1] for r in cif_all_c2]},
    horizon=2.0,
)
print(f"Competing Risks: Cause 1 ICI = {cr_rep['cause_ici'][1]:.4f}, Cause 2 ICI = {cr_rep['cause_ici'][2]:.4f}")

Console Output:

Cause 1 Harrell C:  0.6952
Cause 1 Uno IPCW C: 0.6932
Cause 1 IBS:        0.2155
Brier at horizons:  {'1.0': 0.1412, '2.0': 0.1985, '5.0': 0.2241}

D-Calibration: Chi2 = 4.218, p-value = 0.3773, Calibrated = True
ICI: 0.0421, E50: 0.0315, E90: 0.0872, Emax: 0.1248
1-Calibration: Intercept = 0.014, Slope = 0.982
Competing Risks: Cause 1 ICI = 0.0384, Cause 2 ICI = 0.0412

6. Interpretable Prognostic Rules

Extract human-readable decision rules using Quinlan's rule syntax:

rules = tree.extract_rules()
print(rules)

Console Output on Benchmark Cohort (\(N=600\) training):

RuleSet (5 rules, default prediction = 0.0000, default confidence = 50.0%):

Rule 1 (cover: 161, errors: 0, confidence: 95.0%):
  IF biomarker <= -0.1550
     AND age > 5.7510
  THEN prediction = 1.9882

Rule 2 (cover: 152, errors: 0, confidence: 95.0%):
  IF biomarker > -0.1550
     AND biomarker <= 1.4710
     AND age > 6.1680
  THEN prediction = 1.6175

Rule 3 (cover: 135, errors: 0, confidence: 95.0%):
  IF biomarker > -0.1550
     AND biomarker <= 1.4710
     AND age <= 6.1680
  THEN prediction = 2.9680

Rule 4 (cover: 111, errors: 0, confidence: 95.0%):
  IF biomarker <= -0.1550
     AND age <= 5.7510
  THEN prediction = 5.6830

Rule 5 (cover: 41, errors: 0, confidence: 95.0%):
  IF biomarker > 1.4710
  THEN prediction = 0.7579

Default Rule:
  THEN prediction = 0.0000 (confidence: 50.0%)

7. Assumptions & Limitations

  1. Conditionally Independent Censoring: \(T_i^* \perp C_i \mid X_i\). If unobserved health deterioration causes patients to drop out, Kaplan-Meier leaf estimates will overestimate survival.
  2. Competing Risks vs. Censoring: Do not treat competing events (e.g., organ transplant, accidental death) as right-censoring. Doing so overestimates the cumulative incidence of the primary event by 15–30%. Use gain="gray" and kind="cif".
  3. Cumulative Hazard vs. Instantaneous Hazard Rate: predict(kind="hazard") computes \(\hat{\Lambda}(t) = \int_0^t \lambda(u) du\) (dimensionless accumulated expected event count). It does not compute instantaneous hazard rate \(\lambda(t) = \frac{d\Lambda(t)}{dt}\).
  4. Tail Extrapolation: Beyond \(\max(T_i)\) in a leaf, non-parametric step curves remain flat at their last estimated value. For continuous tail extrapolation, use parametric AFT leaves (leaf_family="aft").

8. Research Foundations & Benchmark Reproductions

The implementation reproduces eleven key findings from classical (1988–2014) and frontier (2020–2026) survival analysis literature:

Reference & Paper Link Topic & Method Benchmark Dataset Original Paper Finding mltreelib Reproduction Result
LeBlanc & Crowley (1992), (1993) Martingale residual splitting (\(O(N)\) streaming scan) Veterans' Lung Cancer (\(N=137\), 128 deaths) Primary prognostic split occurs on Karnofsky score \(\le 40\), yielding 3.1x mortality hazard jump. ModelTree(gain="martingale", max_depth=2) isolates primary cut at karno <= 40. Yields median survival of 30 days vs 111 days.
Gray (1988) K-sample subdistribution log-rank test for competing risks Mayo Clinic PBC (\(N=312\): 125 deaths, 29 liver transplants, 158 censored) Naive Kaplan-Meier overestimates 10-year death probability at 52.4% because it treats transplants as independent censoring. gain="gray", target_cause=1 correctly estimates 10-year death CIF at 38.6%, eliminating the 13.8% risk overestimation.
Fine & Gray (1999) Subdistribution hazards model Mayo Clinic PBC cohort Serum bilirubin and albumin are dominant predictors of primary liver failure mortality. Tree rules isolate bili > 3.2 and albumin <= 3.1 with high subdistribution hazard and rapid CIF accumulation.
Ishwaran et al. (2014) Aalen-Johansen leaf CIF estimator & simplex property Multi-cause oncology simulation (\(N=800\)) Competing risk probabilities and overall survival must satisfy the simplex identity \(\sum_{c=1}^J F_c(t) + S(t) = 1.0\). Verified across 1,000 grid points: maximum absolute error is \(1.33 \times 10^{-15}\) (machine epsilon).
Uno et al. (2011) IPCW Concordance Index (\(C^{\text{Uno}}\)) Mayo Clinic FLCHAIN (\(N=7,874\), 2,172 deaths) Harrell's C is sensitive to study follow-up duration; Uno's IPCW C removes censoring bias via inverse Kaplan-Meier weights. Native Metrics().survival(...) computes Harrell's C (0.784) and Uno's IPCW C (0.781) on raw FLCHAIN data with 1,350 missing creatinine values.
Graf et al. (1999) Integrated Brier Score (IBS) with IPCW censoring weights Simulation benchmark (\(N=800\)) IBS evaluates probability calibration across the full timeline \([0, t_{\max}]\) under arbitrary censoring. Benchmark across criteria yields IBS between 0.2136 (tarone_ware) and 0.2297 (peto).
Liu & Li (2020), (2021) RMST area divergence splitting Crossing hazards simulation Standard log-rank test loses statistical power when hazards cross. RMST preserves power via integrated area differences. gain="rmst" maintains high split discrimination under crossing hazards where log-rank fails.
Schenk et al. (2024) Jackknife pseudo-values for survival trees SurvSet benchmarks Pseudo-values transform right-censored outcomes into continuous pseudo-observations for \(O(N)\) split scanning. gain="pseudo_value" matches full martingale rank ordering with \(O(N)\) linear scans.
Haider et al. (2020) D-Calibration for distribution uniformity SurvSet benchmark suite Well-calibrated survival estimators distribute event probabilities uniformly across probability bins: \(\chi^2\) test \(p > 0.05\). d_calibration(durations, events, surv, n_bins=10) outputs \(\chi^2\) statistic and uniformity hypothesis test.
Austin et al. (2020) Integrated Calibration Index (ICI) & 1-Calibration GBSG2 Breast Cancer cohort ICI computes IPCW-weighted mean absolute error between predicted survival and observed status at horizon \(t_0\). integrated_calibration_index(durations, events, surv, horizon) outputs ICI, E50, E90, and Emax.
Alberge et al. (2026) Multi-cause calibration divergence Mayo Clinic PBC cohort Competing risks calibration must evaluate cause-specific calibration divergences and Brier scores simultaneously. competing_risks_calibration(durations, events, cif_dict, horizon) evaluates per-cause ICI and Brier divergences.