SurvHTE-Bench: A Benchmark for Heterogeneous Treatment Effect Estimation in Survival Analysis

Paper: ICLR 2026 — SurvHTE-Bench: A Benchmark for Heterogeneous Treatment Effect Estimation in Survival Analysis

GitHub: https://github.com/Shahriarnz14/SurvHTE-Bench

---

Overview

SurvHTE-Bench is a benchmark for heterogeneous treatment effect (HTE) estimation under right-censored survival outcomes.

The benchmark addresses an important gap at the intersection of causal inference and survival analysis. While heterogeneous treatment effect estimation has been widely studied in fully observed outcome settings, systematic evaluation in time-to-event data with censoring has been largely missing.

SurvHTE-Bench provides a unified framework for evaluating survival HTE estimators across:

• Synthetic datasets with known ground-truth treatment effects
• Semi-synthetic datasets combining real covariates with simulated treatments and outcomes
• Real-world datasets including a twin birth dataset (with ground-truth counterfactual outcomes) and an HIV clinical trial dataset

Across these datasets, the benchmark evaluates 53 estimator variants spanning three major methodological families:

• Outcome imputation approaches
• Direct survival causal methods
• Survival meta-learners

The benchmark focuses primarily on Conditional Average Treatment Effects (CATE) defined using Restricted Mean Survival Time (RMST) as the survival estimand.

---

Datasets

The benchmark includes five dataset groups spanning the full data-generation spectrum.

1. synthetic — Fully Synthetic

The synthetic benchmark consists of 40 datasets, constructed by crossing:

• 8 causal configurations (different treatment assignment mechanisms, confounding structures, positivity violations, and censoring mechanisms)
• 5 survival scenarios (different survival distributions and censoring regimes)

Each dataset contains up to 50,000 samples with:

• 5 covariates independently sampled from Uniform(0,1)
• binary treatment W
• observed time observed_time
• event indicator event
• potential survival times T0 and T1

Because both potential outcomes are generated, ground-truth individual treatment effects are available.

The causal configurations include randomized controlled trials and observational settings with violations such as unmeasured confounding, lack of positivity, and informative censoring.

2. actg_syn — Semi-Synthetic ACTG Dataset

Semi-synthetic datasets constructed from the ACTG 175 HIV clinical trial, which contains 2,139 patients.

• Covariates are real patient features from the trial.
• Treatment assignments and survival outcomes are simulated to generate known treatment effects.

This preserves realistic covariate distributions while enabling controlled evaluation.

3. twin — Twin Birth Dataset

A real-world dataset derived from the Twin Births dataset, containing 11,400 twin pairs.

The twin structure allows near-counterfactual evaluation: for each pair, one twin is treated and the other is untreated.

Treatment corresponds to being the heavier twin, and the outcome is time to mortality.

4. actgHC — ACTG High-Censoring Variant

A version of the ACTG dataset with high censoring rates, containing approximately 1,054–1,093 samples depending on the trial arm.

The dataset includes multiple time/event pairs (t0/e0 … t9/e9) representing repeated survival observations.

5. actgLC — ACTG Low-Censoring Variant

A lower-censoring version of the ACTG dataset.

The structure mirrors actgHC, but censoring rates are substantially lower.

6. mimic_syn — Semi-Synthetic MIMIC-IV Datasets

The benchmark also includes semi-synthetic datasets derived from covariates in the MIMIC-IV ICU database.

In the paper, we construct nine MIMIC-based semi-synthetic datasets (MIMIC-i – MIMIC-ix) using real patient covariates from MIMIC-IV while simulating treatment assignments and survival outcomes. These datasets are designed to capture realistic covariate structure while enabling controlled evaluation with known ground-truth treatment effects.

The datasets cover multiple regimes:

• MIMIC-i – MIMIC-v: varying censoring severity (approximately 53%–88%) under covariate-independent treatment assignment.
• MIMIC-vi – MIMIC-ix: covariate-dependent treatment assignment with more complex nonlinear outcome and censoring mechanisms.

Due to the MIMIC-IV data usage agreement, we cannot redistribute the original data or any datasets derived directly from it through this repository or the HuggingFace dataset.

Researchers must obtain access to MIMIC-IV through PhysioNet:

https://physionet.org/content/mimiciv/

After obtaining access, the semi-synthetic datasets used in our experiments can be reproduced using the notebook provided in the repository:

https://github.com/Shahriarnz14/SurvHTE-Bench/blob/main/data/semi-synthetic/generate_mimic_semi_synthetic.ipynb

---

HuggingFace Configuration Layout

Each dataset group is split into three HuggingFace configurations:

Config name	Split(s)	Contents
{name}	train	Full data with metadata
{name}_repeats	train	Random index permutations used for repeated splits
{name}_splits	train_0…train_9, val_0…val_9, test_0…test_9	Pre-computed splits for repeated experiments

So the full list of configs is: synthetic, synthetic_repeats, synthetic_splits, actg_syn, actg_syn_repeats, actg_syn_splits, twin, twin_repeats, twin_splits, actgHC, actgHC_repeats, actgHC_splits, actgLC, actgLC_repeats, actgLC_splits

---

Loading the Data

We provide a ready-to-use loader at
data_utils/hf_load.py
in the GitHub repository. Install dependencies first:

pip install datasets pandas numpy

Interface 1 — load_data: Full Dataset (mirrors local API)

Reconstructs experiment_setups and experiment_repeat_setups identically to the original local data loader.

from data_utils.hf_load import load_data

experiment_setups, experiment_repeat_setups = load_data(dataset_name="synthetic")

experiment_setups is a nested dict:

experiment_setups[setup_key][scenario] = {
    "dataset":  pd.DataFrame,   # all covariates + outcome columns
    "summary":  dict,           # summary statistics
    "metadata": dict,           # (synthetic only) DGP metadata
}

experiment_repeat_setups contains the pre-computed random index permutations used to generate reproducible train/val/test splits. For actgHC/actgLC it is a {setup_key: DataFrame} dict; for all other datasets it is a single shared DataFrame.

Supported dataset_name values: "synthetic", "actg_syn", "twin", "actgHC", "actgLC".

Interface 2 — load_splits: Pre-Split Arrays (drop-in for experiment loop)

Returns arrays already split into train/val/test for each configuration, scenario, and repeat index — ready to pass directly into model training.

from data_utils.hf_load import load_splits

split_dict = load_splits(dataset_name="synthetic")

The returned structure is:

split_dict[config_name][scenario_key][rand_idx]["train" | "val" | "test"]
    = (X, W, Y, cate_true)

where:

• X — covariate matrix (n, d) as np.ndarray
• W — treatment vector (n,) as np.ndarray
• Y — outcome matrix (n, 2) containing [observed_time, event] (or all t/e columns for actgHC)
• cate_true — ground-truth CATE (n,) (or proxy)

Example — accessing a specific split:

config_name  = "RCT-50"     # setup key
scenario_key = "Scenario_A" # scenario
rand_idx     = 0            # repeat index (0–9)

X_train, W_train, Y_train, cate_true_train = split_dict[config_name][scenario_key][rand_idx]["train"]
X_val,   W_val,   Y_val,   cate_true_val   = split_dict[config_name][scenario_key][rand_idx]["val"]
X_test,  W_test,  Y_test,  cate_true_test  = split_dict[config_name][scenario_key][rand_idx]["test"]

Example — iterating the full experiment loop:

results = load_splits(dataset_name="synthetic")

for config_name, scenarios in results.items():
    for scenario_key, repeats in scenarios.items():
        for rand_idx in range(10):
            X_tr, W_tr, Y_tr, cate_tr = repeats[rand_idx]["train"]
            X_te, W_te, Y_te, cate_te = repeats[rand_idx]["test"]
            # ... fit model, evaluate ...

---

Evaluation

The benchmark evaluates heterogeneous treatment effect estimators using metrics derived from the true Conditional Average Treatment Effect (CATE).

Primary evaluation metrics include:

• CATE Root Mean Square Error (RMSE)
  Measures the error between estimated and true individual treatment effects.

• ATE Bias
  Measures the deviation of the estimated average treatment effect from the true population ATE.

Additional auxiliary metrics are used to analyze component performance:

• Imputation accuracy (for methods using survival time imputation)
• Regression or survival model performance, such as MAE or time-dependent C-index

All experiments are averaged over 10 repeated train/validation/test splits.

---

Repository

Full code and experiment scripts are available at:

https://github.com/Shahriarnz14/SurvHTE-Bench

---

Citation

If you use SurvHTE-Bench in your research, please cite:

@inproceedings{noroozizadeh2026survhte,
  title={SurvHTE-Bench: A Benchmark for Heterogeneous Treatment Effect Estimation in Survival Analysis},
  author={Noroozizadeh, Shahriar and Shen, Xiaobin and Weiss, Jeremy and Chen, George H.},
  booktitle={International Conference on Learning Representations (ICLR)},
  year={2026}
}

---

License

This dataset is released under the Creative Commons Attribution 4.0 International (CC BY 4.0) license.