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

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

Problem Statement
In clinical time-to-event analyses, standard survival methods often fail to distinguish between two distinct treatment mechanisms: (1) the probability of being "cured" (i.e., the event never occurring within a relevant timeframe) and (2) the timing of the event among those who are not cured. Traditional approaches, such as standard survival analysis or causal inference methods that treat all censored individuals as potential event experiencers, conflate these effects. This conflation obscures the specific mechanisms of treatment, leading to imprecise prognostic information and suboptimal clinical decision-making. The challenge is further compounded by high-dimensional covariates, confounding, and censoring, which make it difficult to identify which features drive cure probability versus event timing.

Methodology
The authors propose a Double Variable Importance Matching (DVIM) framework designed to separately estimate Heterogeneous Treatment Effects (HTE) on cure probability and conditional mean event time (CMET). The methodology proceeds through the following steps:

1. Data Splitting and Mixture Cure Modeling: The dataset is split into training and estimation sets. On the training set, separate mixture cure models are fitted for the treatment and control arms. These models consist of two components:

   • A logistic regression model for the probability of being cured ($P(\text{cure}|X)$).
   • An Accelerated Failure Time (AFT) model (specifically log-normal) for the event time distribution among non-cured individuals.
   • Parameters are estimated via maximum likelihood.

2. Learning Distance Metrics: The absolute values of the estimated coefficients from the mixture cure models are used to quantify feature importance. Two distinct weighted Euclidean distance metrics are constructed:

   • $d_{\text{cure}}$: Weighted by coefficients from the cure probability component.
   • $d_{\text{time}}$: Weighted by coefficients from the event time component.
   • These weights approximate the optimal covariate weighting for each specific estimand, upweighting features most relevant to the respective outcome.

3. Double Matching Procedure: For each individual in the estimation set, two separate matched groups are formed:

   • One group matched using $d_{\text{cure}}$ to estimate the HTE on cure probability.
   • One group matched using $d_{\text{time}}$ to estimate the HTE on CMET.
   • Nearest neighbors are identified from the respective treatment and control cohorts.

4. HTE Estimation: Within the matched groups, non-parametric Kaplan-Meier estimators are applied:

   • Cure Probability HTE: Estimated as the difference in survival probabilities at a pre-specified time horizon $H$ ($S_1(H) - S_0(H)$).
   • CMET HTE: Estimated as the difference in the expected event time conditional on the event occurring before $H$, derived from the area under the Kaplan-Meier curve.

Key Contributions
The paper outlines six primary contributions:

1. Novel Framework: A matching-based framework that separately estimates HTE on cure probability and event timing, addressing the limitation of conflating these distinct clinical effects.
2. Tailored Distance Metrics: The construction of cure-specific and timing-specific distance metrics derived from covariate importances in an auxiliary mixture cure model.
3. Theoretical Optimality: Proof that these learned metrics are optimal for HTE estimation under an equal-scale constraint, minimizing the worst-case discrepancy (Lipschitz constant) under local perturbations.
4. Consistency: Demonstration that the proposed estimators are consistent for both causal effects under standard assumptions (consistency, unconfoundedness, positivity, and non-informative censoring).
5. Error Decomposition: A theoretical and empirical decomposition of estimation error into contributions from censoring, model fitting, and irreducible noise.
6. Validation: Comprehensive validation across diverse simulation settings and a real-world case study.

Results

• Simulations: In four simulation settings (varying by whether treatment effects modify cure, time, both, or overlapping covariates), the proposed method consistently achieved lower Mean Absolute Error (MAE) compared to standard Euclidean matching, propensity score matching, prognostic score matching, and Cox models. The method's performance closely approached that of "oracle" estimators (which know the true data-generating parameters), suggesting the learned distance metrics effectively approximate optimal matching structures.
• Real-World Application: The method was applied to a multi-center cohort of 1,157 acute lymphoblastic leukemia patients comparing haploidentical versus matched sibling donor transplants. The proposed method estimated a more conservative and concentrated HTE on leukemia-free survival (approx. 5% improvement) compared to other matching methods, which produced more dispersed or bimodal estimates. The authors attribute this to the method's ability to disentangle cure probability from short-term survival timing.

Significance and Claims
The authors claim that their approach offers a robust and interpretable solution for causal inference in time-to-event data with a cured subpopulation. By explicitly separating the estimation of cure probability and event timing, the method provides clearer insights into treatment mechanisms, which is critical for high-stakes clinical decisions. The paper emphasizes that while the method is theoretically grounded and performs well in simulations, it relies on the correct specification of the mixture cure model for the distance metric learning step. The authors note that the method requires further validation on diverse real-world datasets and remains sensitive to model misspecification, despite combining parametric (for metric learning) and non-parametric (for final estimation) elements. The work does not claim to solve all causal inference challenges but specifically targets the under-explored area of jointly addressing cured subpopulations and distinct treatment mechanisms via matching.