Double Variable Importance Matching to Estimate Distinct Causal Effects on Event Probability and Timing

Yuqi Li, Quinn Lanners, Matthew M. EngelhardarXiv 2026 · DOI · arXiv

Abstract

In many clinical contexts, estimating effects of treatment in time-to-event data is complicated not only by confounding, censoring, and heterogeneity, but also by the presence of a cured subpopulation in which the event of interest never occurs. In such settings, treatment may have distinct effects on (1) the probability of being cured and (2) the event timing among non-cured individuals. Standard survival analysis and causal inference methods typically do not separate cured from non-cured individuals, obscuring distinct treatment mechanisms on cure probability and event timing. To address these challenges, we propose a matching-based framework that constructs distinct match groups to estimate heterogeneous treatment effects (HTE) on cure probability and event timing, respectively. We use mixture cure models to identify feature importance for both estimands, which in turn informs weighted distance metrics for matching in high-dimensional spaces. Within matched groups, Kaplan-Meier estimators provide estimates of cure probability and expected time to event, from which individual-level treatment effects are derived. We provide theoretical guarantees for estimator consistency and distance metric optimality under an equal-scale constraint. We further decompose estimation error into contributions from censoring, model fitting, and irreducible noise. Simulations and real-world data analyses demonstrate that our approach delivers interpretable and robust HTE estimates in time-to-event settings.

Introduction: The "If" vs. "When" Problem

Standard survival analysis often operates under the assumption that every subject will eventually experience the event of interest (e.g., disease recurrence) if followed for a sufficient duration. However, in many clinical scenarios—such as long-term cancer remission—a "cured" subpopulation exists for whom the event will never occur.

Traditional causal inference tools typically conflate two fundamentally different treatment mechanisms:

  1. The "If": Does the treatment increase the probability of being cured?
  2. The "When": For those not cured, does the treatment delay the event?
Distinct treatment effects on cure probability (vertical shift) and event timing (horizontal shift).

Neglecting this distinction can lead to a paradoxical understanding of treatment. For instance, a drug might successfully cure more people but, for those it fails to cure, it might actually hasten disease progression. Standard models average these effects, potentially leading to sub-optimal clinical decisions.

Methodology: Double Variable Importance Matching

To address this, researchers at Duke University developed Double Variable Importance Matching (MCM KNN). This framework disentangles the two effects by learning separate feature importance metrics for each outcome. The workflow consists of three stages:

1. Auxiliary Mixture Cure Modeling

The process begins by fitting a parametric "mixture cure model" to the data. This model comprises two parts: a logistic component to estimate the probability of being cured and an Accelerated Failure Time (AFT) component to model the timing of the event for the non-cured.

2. Learning Distance Metrics

From these models, the algorithm extracts feature importance (coefficients) to create two distinct "weighted Euclidean" distance metrics. One metric prioritizes variables relevant to being cured, while the other prioritizes variables relevant to event timing. This is a critical "aha" moment: the variables predicting if someone is cured are often different from those predicting when a relapse occurs.

3. Double Matching and Estimation

For every individual, the algorithm performs two separate matching procedures. It identifies "nearest neighbor" groups in the treatment and control arms using the cure-specific metric to estimate the cure effect, and a different set of neighbors using the timing-specific metric to estimate the timing effect. Finally, non-parametric Kaplan-Meier estimators are applied within these matched groups to calculate the final treatment effects.

Experimental Evaluation

The framework was evaluated using both simulated and real-world datasets:

  • Simulated Data: 20,000 samples with 20 covariates across four scenarios (effects on cure only, timing only, disjoint variables, and overlapping variables).
  • Real-World Clinical Data: A multi-center cohort of 1,157 patients with acute lymphoblastic leukemia (ALL) who received stem cell transplants (Haplo-SCT vs. MSDT), tracking leukemia-free survival (LFS) over 10 years.

Key Results and Insights

  • Superior Accuracy: In simulations, MCM KNN consistently achieved lower Mean Absolute Error (MAE) than standard Euclidean, Propensity Score, and Prognostic Score matching. In complex scenarios where effects overlapped, MCM KNN reduced the MAE for cure probability by more than half compared to Propensity Score matching.
  • Clinical Impact: In the leukemia study, the method estimated that Haplo-SCT improves leukemia-free survival rates by approximately 5% on average compared to MSDT.
  • Theoretical Robustness: The authors provided mathematical proof that their weighting scheme is optimal for minimizing estimation error. By using parametric models for weights but non-parametric estimators for final effects, the method remains robust to model misspecification.

Limitations and Future Directions

The method requires a pre-specified time horizon ($H$) to define a "cure," and the accuracy of the initial distance metrics depends on the auxiliary mixture cure model identifying the correct important variables. Future work may involve integrating "neural" mixture cure models to capture complex, non-linear feature importances in higher-dimensional datasets.

Summary updated

Abstract

In many clinical contexts, estimating effects of treatment in time-to-event data is complicated not only by confounding, censoring, and heterogeneity, but also by the presence of a cured subpopulation in which the event of interest never occurs. In such settings, treatment may have distinct effects on (1) the probability of being cured and (2) the event timing among non-cured individuals. Standard survival analysis and causal inference methods typically do not separate cured from non-cured individuals, obscuring distinct treatment mechanisms on cure probability and event timing. To address these challenges, we propose a matching-based framework that constructs distinct match groups to estimate heterogeneous treatment effects (HTE) on cure probability and event timing, respectively. We use mixture cure models to identify feature importance for both estimands, which in turn informs weighted distance metrics for matching in high-dimensional spaces. Within matched groups, Kaplan-Meier estimators provide estimates of cure probability and expected time to event, from which individual-level treatment effects are derived. We provide theoretical guarantees for estimator consistency and distance metric optimality under an equal-scale constraint. We further decompose estimation error into contributions from censoring, model fitting, and irreducible noise. Simulations and real-world data analyses demonstrate that our approach delivers interpretable and robust HTE estimates in time-to-event settings.
BibTeX
@misc{li2026double,
  author = {Yuqi Li and Quinn Lanners and Matthew M. Engelhard},
  title = {{Double Variable Importance Matching to Estimate Distinct Causal Effects on Event Probability and Timing}},
  year = {2026},
  eprint = {2602.05022},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2602.05022},
}

Loading related content...

Figure