Overview

BART-Survival is a high-performance Python package for discrete-time survival analysis using Bayesian Additive Regression Trees (BART). Combining the power and flexibility of PyMC probabilistic programming with specialized discrete-time hazard modeling, BART-Survival enables fully non-parametric time-to-event estimation, non-linear covariate interaction modeling, and causal effect estimation without requiring proportional hazards or linearity assumptions.

Key Highlights (2026 Release)

• Intuitive High-Level API (SurvBART): End-to-end scikit-learn compatible workflow with fit_survival(), predict_survival(), feature_importance(), and estimate_effect().
• Standardized Model Presets: Battle-tested presets (preset_standard(), preset_fast(), preset_high_precision()) calibrated for MCMC stability and rapid prototyping.
• Empirical Node Value Splitting: Proposes tree split cutpoints at exact empirical midpoints (up to 512 candidates per feature), capturing step-function risk dynamics with high fidelity.
• Support-Audited Causal G-Computation: Automated detection of conditional support and positivity violations (e.g., zero length-of-stay for outpatients) with conditional donor resampling to eliminate off-support extrapolation.
• Warm-Started DART Dirichlet Sparsity: Optional variable selection prior to identify true risk drivers in high-dimensional tabular cohorts.
• Fully Accessible Hyperparameters: Direct, open parameter exposure on SurvBART with property bindings and InferenceData serialization.
• Synthetic Survival Benchmarking: Built-in data generators (simulate_confounded_survival_data) for reproducible testing with known ground-truth hazard ratios.
• 100% Backward Compatible: Full compatibility with legacy BartSurvModel, get_surv_pre_train, and get_posterior_test pipelines.

---

Installation

Install from PyPI:

pip install bart-survival

Or install the latest development branch locally:

git clone https://github.com/CDCgov/BART-Survival.git
cd BART-Survival
pip install -e .

---

Quickstart

1. High-Level End-to-End Workflow

import numpy as np
import pandas as pd
from bart_survival import SurvBART, simulate_confounded_survival_data

# 1. Generate realistic synthetic cohort with known ground-truth treatment effect
df, meta = simulate_confounded_survival_data(
    N=400, 
    inpatient_hr=1.75, 
    censoring_rate=0.25, 
    seed=42
)
feature_cols = ["inpatient", "age", "stage", "los"]
X = df[feature_cols]

# 2. Initialize SurvBART with standard scientific preset
# (50 trees, 1000 draws, 250 tune, 4 chains, thin=4, empirical 512 cutpoints)
model = SurvBART.preset_standard()

# 3. Fit directly from tabular covariates and time-to-event outcomes
model.fit_survival(
    X=X,
    time=df["time"],
    status=df["status"],
    time_scale=5.0,     # Discretize continuous time into 5-day intervals
    progressbar=True,
    random_seed=42
)

# 4. Predict cumulative survival curves S(t) with 95% Credible Intervals
surv_summary = model.predict_survival(X, return_quantiles=True)
print("Survival curve shape (subjects x times):", surv_summary["mean"].shape)
print("Mean S(t) at final interval:", surv_summary["mean"][:, -1].mean())

# 5. Extract normalized posterior feature importance
print("\nFeature Split Importances:")
print(model.feature_importance())

# 6. Estimate Causal Treatment Effect (Support-Audited G-Computation)
# Audits conditional support for positivity violations (e.g. inpatient vs LOS)
causal_res = model.estimate_effect(
    X=X, 
    primary_col="inpatient", 
    values=[0, 1]
)
print("\nTotal Marginal HR (Policy Effect):     ", causal_res["total_marginal_hr"])
print("95% Credible Interval:                  ", causal_res["total_marginal_hr_ci"])
print("Total Conditional HR (Subject Effect):  ", causal_res["total_conditional_hr"])
print("95% Credible Interval:                  ", causal_res["total_conditional_hr_ci"])
print("Horizon Risk Difference:                ", causal_res["risk_difference_horizon_mean"])
print("Flagged Support Violations:             ", causal_res["flagged_support_violations"])

# 7. Continuous Shift Intervention (Modified Treatment Policy: Xi -> Xi + delta)
shift_res = model.estimate_effect(
    X=X,
    primary_col="age",
    delta=5.0  # Shift everyone's age by +5 years
)
print("\nAge +5yr Shift Marginal HR:    ", shift_res["total_marginal_hr"])
print("Age +5yr Shift Conditional HR: ", shift_res["total_conditional_hr"])
print("Age +5yr Risk Difference:      ", shift_res["risk_difference_horizon_mean"])

---

2. Standard Model Presets

SurvBART provides three pre-calibrated factory presets:

Preset	Trees	Draws	Tune	Chains	Thin	Update Fraction	Use Case
preset_standard() (Default)	50	1000	250	4	4 (1,000 final)	1.0 (100%)	Standard scientific benchmark models, publication-grade convergence
preset_fast()	25	500	250	4	5 (400 final)	1.0 (100%)	Fast exploration, interactive debugging, prototyping
preset_high_precision()	100	5000	2000	4	10	1.0 (100%)	High-dimensional data, complex multi-way interactions

# Launch fast exploration model
fast_model = SurvBART.preset_fast(split_method="empirical")

# Or instantiate with explicit custom parameters
custom_model = SurvBART(
    trees=50,
    draws=2000,
    tune=800,
    chains=4,
    thin=10,
    update_fraction=1.0,
    use_sparsity=False,       # Set True to enable DART Dirichlet selection
    split_method="empirical", # Proposes empirical midpoints up to max_cutpoints
    max_cutpoints=512,
    min_node_size=5
)

---

3. Step-by-Step Pooled Cross-Section API (Granular Control)

For researchers requiring explicit control over pooled person-time representations:

from bart_survival import (
    SurvBART,
    format_survival_data,
    format_prediction_data,
    compute_survival_curves,
    compute_causal_pdp,
    compute_hazard_ratio
)
from lifelines.datasets import load_rossi

# Load open Rossi recidivism benchmark cohort
rossi = load_rossi()
X_df = rossi[["fin", "age", "race", "wexp", "mar", "paro", "prio"]]
time = rossi["week"].values
status = rossi["arrest"].values

# 1. Format pooled person-time training cross-section
trn = format_survival_data(
    y_time=time,
    y_status=status,
    x=X_df.values,
    time_scale=4.0  # 4-week discrete intervals
)

# 2. Format out-of-sample prediction grid
post_test = format_prediction_data(
    y_time=time,
    y_status=status,
    x=X_df.values,
    time_scale=4.0
)

# 3. Fit SurvBART
model = SurvBART(trees=50, draws=1000, tune=500, chains=2, thin=5)
model.fit(
    y=trn["y"],
    X=trn["x"],
    coords=trn["coord"],
    progressbar=False,
    random_seed=42
)

# 4. Predict out-of-sample hazards and derive cumulative survival curves S(t)
post_preds = model.predict(post_test["post_x"], coords=post_test["coords"])
sv_curves = compute_survival_curves(post_preds, coords=post_test["coords"])

# S(t) shape: (n_draws, n_subjects, n_time_intervals)
print("Survival array shape:", sv_curves["sv"].shape)

---

---

Causal G-Computation and Counterfactual Treatment Effects

1. Support-Audited Binary G-Computation

In observational clinical survival studies, counterfactual interventions often suffer from positivity violations or structural zero dependencies. For example, assigning inpatient=0 (outpatient) while leaving length_of_stay=12 days creates non-existent clinical combinations that force tree ensembles into off-support extrapolation.

SurvBART addresses this via estimate_effect() and compute_causal_pdp():

1. Positivity & Support Audit: Automatically scans baseline covariates conditional on exposure levels to identify deterministic zeros, non-overlapping domains, or severe mass-ratio violations.
2. Conditional Donor Resampling: Re-samples dependent features conditional on exposure from observed donor subcohorts while preserving genuine baseline confounders.
3. Counterfactual Predictions: Evaluates both intervention branches through the posterior tree ensemble to yield unconfounded survival curves $S_1(t)$ and $S_0(t)$.

2. Continuous Shift Interventions (Modified Treatment Policies)

For continuous exposures (e.g., biomarker levels, blood pressure, age, radiation dose), arbitrary fixed point comparisons (e.g., $X=1$ vs $X=5$) frequently cause extreme positivity violations. SurvBART supports shift interventions (Modified Treatment Policies): $$\tilde{X}_i = X_i + \Delta \quad \text{vs.} \quad X_i$$ Evaluating a shift preserves empirical covariate correlations and natural population distributions while estimating realistic, policy-relevant counterfactual effects.

3. Hazard Ratios and Risk Differences

model.estimate_effect(...) and compute_hazard_and_risk_difference(...) compute comprehensive effect metrics (always formatted as "Intervention (1) over Reference (0)"):

• Event-Weighted Summary Hazard Ratio (hazard_ratio_weighted_mean): The standard summary HR, calculated as the geometric mean across follow-up intervals weighted by empirical event counts ($w_t = d_t / \sum d_t$). Matches the Cox proportional hazards estimand.
• Time-Varying Hazard Ratio Curve (hazard_ratio_curve_mean): Interval-specific hazard ratio computed via the complementary log-log link: $\mathrm{HR}(t) = \log(1 - \bar{h}_1(t)) / \log(1 - \bar{h}_0(t))$.
• Discrete Risk Ratio (discrete_risk_ratio_mean): Raw, interval-by-interval discrete risk ratio $\bar{h}_1(t) / \bar{h}_0(t)$ without temporal averaging.
• Risk Difference (risk_difference_mean, risk_difference_horizon_mean): Excess cumulative failure risk: $\mathrm{RD}(t) = (1 - S_1(t)) - (1 - S_0(t)) = S_0(t) - S_1(t)$.
• Restricted Mean Survival Time Difference (rmst_difference_mean): Area between the survival curves over the follow-up horizon.

---

Documentation

Comprehensive documentation, API references, and mathematical methodology:

• BART-Survival Documentation
• API Reference
• JOSS Paper

---

---

CDCgov General Disclaimers

This repository was created for use by CDC programs to collaborate on public health related projects in support of the CDC mission. GitHub is not hosted by the CDC, but is a third party website used by CDC and its partners to share information and collaborate on software. CDC use of GitHub does not imply an endorsement of any one particular service, product, or enterprise.

Related Documents

• Open Practices
• Rules of Behavior
• Disclaimer
• Contribution Notice
• Code of Conduct
• Licence

Public Domain Standard Notice

This repository constitutes a work of the United States Government and is not subject to domestic copyright protection under 17 USC § 105. This repository is in the public domain within the United States, and copyright and related rights in the work worldwide are waived through the CC0 1.0 Universal public domain dedication. All contributions to this repository will be released under the CC0 dedication. By submitting a pull request you are agreeing to comply with this waiver of copyright interest.

License Standard Notice

The repository utilizes code licensed under the terms of the Apache Software License and therefore is licensed under ASL v2 or later.

This source code in this repository is free: you can redistribute it and/or modify it under the terms of the Apache Software License version 2, or (at your option) any later version.

This source code in this repository is distributed in the hope that it will be useful, but WITHOUT ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the Apache Software License for more details.

You should have received a copy of the Apache Software License along with this program. If not, see http://www.apache.org/licenses/LICENSE-2.0.html

The source code forked from other open source projects will inherit its license.

Privacy Standard Notice

This repository contains only non-sensitive, publicly available data and information. All material and community participation is covered by the Disclaimer and Code of Conduct. For more information about CDC's privacy policy, please visit http://www.cdc.gov/other/privacy.html.

Contributing Standard Notice

Anyone is encouraged to contribute to the repository by forking and submitting a pull request. (If you are new to GitHub, you might start with a basic tutorial.) By contributing to this project, you grant a world-wide, royalty-free, perpetual, irrevocable, non-exclusive, transferable license to all users under the terms of the Apache Software License v2 or later.

All comments, messages, pull requests, and other submissions received through CDC including this GitHub page may be subject to applicable federal law, including but not limited to the Federal Records Act, and may be archived. Learn more at http://www.cdc.gov/other/privacy.html.

Records Management Standard Notice

This repository is not a source of government records, but is a copy to increase collaboration and collaborative potential. All government records will be published through the CDC web site.

Additional Standard Notices

Please refer to CDC's Template Repository for more information about contributing to this repository, public domain notices and disclaimers, and code of conduct.