Research map / Probabilistic and scientific modeling / Treatment effects
Heterogeneous Treatment Effect in Time-to-Event Outcomes: Harnessing Censored Data with Recursively Imputed Trees
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes MISTR, a non-parametric HTE estimator for right-censored survival data that replaces censoring-model estimation with recursively imputed survival trees (RIST) and multiple imputation over causal forests, extended to instrumental variables as MISTR-IV.
Problem
Estimating heterogeneous treatment effects on time-to-event outcomes under heavy right-censoring and possible unobserved confounding, where existing IPCW-based and causal survival forest methods degrade or cannot use instrumental variables.
Approach
Impute censored event times multiple times using recursively imputed survival trees (no explicit censoring model), fit a causal forest on each completed dataset, and pool estimates with Rubin's rule; the IV variant swaps in an instrumental forest.
Key results
- Under heavy censoring in MIMIC 1 (89.5%), MISTR achieves MSE 6.262 vs CSF 8.741 (x100).
- In IV simulation setting 204 (88% censoring), MISTR-IV achieves MAE 11.40 vs IPCW-IV 32.10 (x100).
- On ACTG 175 with added censoring, MISTR reduces MSE vs CSF (HIV-2: 1.209 vs 1.875).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use MISTR instead of CSF | HTE estimation on MIMIC semi-simulated data · MIMIC 1–4 | MSE: 8.741 -> 6.262; 6.532 -> 5.891; 5.632 -> 5.200; 5.349 -> 5.034 | Helped |
| Use MISTR instead of CSF | HTE estimation on MIMIC semi-simulated data · MIMIC 5 | MSE: 5.474 -> 5.494 | No clear effect |
| Use MISTR-IV instead of IPCW-IV | HTE estimation with unobserved confounding · IV settings 200, 200-a, 200-b, 201–204-b | MAE: 26.28 -> 14.33; 31.50 -> 16.32; 42.46 -> 21.12; 35.84 -> 18.24; 33.27 -> 17.50; 27.70 -> 12.82; 32.10 -> 11.40; 36.19 -> | Helped |
| Use MISTR instead of CSF under additional censoring | HTE estimation on ACTG 175 · HIV-1, HIV-2, HIV-3 additional censoring | MSE: 1.388 -> 1.207; 1.875 -> 1.209; 1.320 -> 0.777 | Helped |
| Increase censoring rate | MISTR versus CSF on MIMIC semi-simulated data · MIMIC 1–5 | MSE | Mixed |
| Increase the number of imputations | MISTR · MIMIC 1–5 | MSE | Helped |
Surprising findings
- On the low-censoring MIMIC 5 setting (69.9%), CSF slightly outperforms MISTR in MSE (5.474 vs 5.494).
Limitations
- Relies on RIST imputation hyperparameters and requires many imputations (A>=100) and bag-size tuning for variance.
- IV estimate represents HTE among compliers and requires IV assumptions that cannot be verified; no ground truth in real data.
- Performance advantage over CSF is marginal or absent at low censoring rates.
Experimental setup
Datasets: ACTG 175 MIMIC-IV 2.0 Illinois Unemployment Insurance Experiments synthetic simulation settings 1-10 and 200-204
Benchmarks: simulation settings 1-10 MIMIC semi-simulated settings IV simulation settings 200-204 ACTG 175 with additional censoring Illinois Unemployment Insurance
Metrics: MSE MAE SEM estimated SE
Baselines: Causal Survival Forest (CSF) IPCW IPCW-IV T-learner X-learner S-learner
Builds on: causal survival forest recursively imputed survival trees causal forest instrumental forest Rubin's rule
Closest papers by meaning
- Orthogonal Survival Learners for Estimating Heterogeneous Treatment Effects from Time-to-Event Data (NeurIPS 2025 · similarity 0.89)
- Assessing the robustness of heterogeneous treatment effects in survival analysis under informative censoring (NeurIPS 2026 · similarity 0.88)
- Doubly robust identification of treatment effects from multiple environments (ICLR 2025 · similarity 0.87)
- A Non-parametric Direct Learning Approach to Heterogeneous Treatment Effect Estimation under Unmeasured Confounding (NeurIPS 2024 · similarity 0.87)
- Doubly Protected Estimation for Survival Outcomes Utilizing External Controls for Randomized Clinical Trials (ICML 2025 · similarity 0.86)
- Estimating Heterogeneous Treatment Effects by Combining Weak Instruments and Observational Data (NeurIPS 2024 · similarity 0.86)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.