Technical Summary: MEC-Cox for ATT Marginal Hazard-Ratio Estimation

1. Problem Statement

The paper addresses the challenge of estimating causal treatment effects in externally controlled single-arm survival trials, a design increasingly used in oncology and rare diseases where concurrent randomized controls are infeasible. The specific target is the Average Treatment Effect on the Treated (ATT) marginal hazard ratio, which compares the marginal hazard under treatment against the counterfactual marginal hazard under control within the treated trial population.

The primary inferential difficulties identified are:

1. Model Misspecification Sensitivity: Standard Inverse Probability Weighted (IPW) Cox regression relies on the source propensity score. If the parametric model for the propensity score is misspecified (e.g., due to non-linear relationships), the resulting weighted Cox estimator suffers from severe bias.
2. Variance Estimation Complexity: Valid inference for IPW Cox estimators is technically challenging because the weighted partial likelihood depends on weights through both event contributions and risk-set averages.
   • Naive variance estimators ignore the weighting structure.
   • Robust sandwich estimators (e.g., Lin–Wei/Binder) treat estimated weights as fixed, ignoring the uncertainty from propensity score estimation.
   • Corrected sandwich estimators (e.g., Shu et al., 2021b) account for weight uncertainty but typically rely on correctly specified parametric propensity score models, making them sensitive to misspecification.
3. Integration of Machine Learning: While Targeted Maximum Likelihood Estimation (TMLE) and Double Machine Learning (DML) successfully incorporate flexible machine learning (ML) for mean-type estimands, extending these to Cox hazard-ratio estimation is non-trivial. It typically requires deriving an Efficient Influence Function (EIF) or Neyman-orthogonal scores, which are analytically complex for Cox models due to the non-linear dependence of the score on weights via risk-set averages.

2. Methodology: MEC-Cox

The authors propose MEC-Cox, a method that adapts the Machine-Learning-Assisted Generalized Entropy Calibration (MEC) framework (Lee and Kim, 2026) to IPW Cox regression. The method operates in two main stages:

A. Baseline Weight Construction

1. Source Propensity Score Estimation: The source propensity score $\pi(X) = P(A=1|X)$ is estimated using flexible ML methods (e.g., BART, Deep Learning) or parametric logistic regression via K-fold cross-fitting to avoid overfitting.
2. ATT Transport Weights: For external controls ($A=0$), baseline weights are constructed as normalized ATT odds: $d_i = n_1 \hat{q}(X_i) / \sum \hat{q}(X_j)$, where $\hat{q}(X) = \hat{\pi}(X)/(1-\hat{\pi}(X))$. These weights transport the external control distribution to the treated trial population.

B. Bregman Calibration Step

The core innovation is a calibration step that refines the baseline weights to balance prognostic summaries rather than just raw covariates.

1. Calibration Basis Construction: A cross-fitted prognostic score is generated for all subjects. The paper suggests using landmark control-survival probabilities $\hat{S}_0(t_k | X)$ estimated from the external control cohort using flexible survival learners (e.g., Random Survival Forests or Cox models). The basis vector is $h_i = (1, \hat{S}_0(t_1|X_i), \dots, \hat{S}_0(t_{p-1}|X_i))$.
2. Bregman Projection: The baseline weights $d_i$ are updated to new weights $w_i$ by minimizing a Bregman divergence $D_G(w_i \| d_i)$ subject to the constraint that the weighted average of the calibration basis in the external control group matches that of the treated trial group:
   $$\sum_{i \in I_0} w_i h_i = \sum_{i \in I_1} h_i$$
   The solution takes the form $w_i = g^{-1}(g(d_i) + \lambda^\top h_i)$, where $\lambda$ is the dual parameter solving the calibration equation.
3. Final Estimation: The updated weights $\tilde{\omega}_i = A_i + (1-A_i)w_i$ are used in the standard IPW Cox estimating equation to solve for the marginal log-hazard ratio $\theta_{ATT}$.

C. Variance Estimation

To account for the uncertainty introduced by the calibration step, the authors propose a stacked sandwich variance estimator.

• They stack the weighted Cox estimating equation for $\theta$ with the dual calibration equation for $\lambda$.
• The variance is derived from the influence function of this joint system, effectively treating the calibration parameter estimation as a source of variability similar to propensity score estimation in corrected sandwich estimators.
• This approach accounts for the first-order uncertainty of the calibration parameter $\lambda$ while treating the cross-fitted nuisance functions (propensity scores and prognostic models) as fixed in the sandwich calculation.

3. Key Contributions and Theoretical Results

• Consistency: Theorem 2 establishes that the MEC-Cox estimator is consistent for the ATT marginal log-hazard ratio under standard causal assumptions (consistency, survival transportability, positivity) and regularity conditions. Crucially, the consistency of the estimator does not require the calibration basis (prognostic model) to be correctly specified; it only requires the source propensity score to be consistently estimated and the calibration basis to be $L_2$-stochastically bounded.
• Efficiency Gain: Theorem 3 demonstrates that under a projection condition, the MEC-Cox stacked sandwich variance is no larger than the fixed-weight Lin–Wei robust variance. The calibration step effectively removes the component of the Cox score variation explained by the calibration basis, leading to potential efficiency gains.
• Oracle Basis: Proposition 4 shows that if an "oracle" control-prognostic score exists (a summary sufficient for the counterfactual control outcome), using it as the calibration basis minimizes the residual variation of the weight-normalized Cox contribution, theoretically maximizing efficiency.
• Flexibility: The method allows for the integration of flexible ML methods (BART, Random Survival Forests, Deep Learning) for both propensity score estimation and prognostic summary construction without requiring the derivation of complex EIFs.

4. Empirical Results

The paper evaluates MEC-Cox through simulation studies and a real-data application.

Simulation Studies

• Scenario 1 (Linear Models): In settings with correctly specified linear models, MEC-Cox (using logistic regression for propensity scores and Cox models for prognosis) achieved smaller bias and Root Mean Squared Error (RMSE) than standard ATT-IPW estimators, demonstrating efficiency gains even without model misspecification.
• Scenario 2 (Nonlinear Models): In settings with severe nonlinearity in both source selection and outcome models, standard IPW estimators using logistic regression exhibited significant bias due to propensity score misspecification. MEC-Cox, utilizing BART for propensity scores and flexible survival learners (RSF) for the calibration basis, successfully reduced bias and RMSE, maintaining coverage close to nominal levels.

Real-Data Application

The method was applied to compare hormonal therapy (GBSG trial) versus no hormonal therapy (Rotterdam tumor bank) for breast cancer recurrence-free survival.

• Imbalance: The unweighted comparison showed no significant difference, but raw covariate imbalance was severe (Max SMD = 1.158).
• Results: Standard ATT-IPW Cox regression suggested a significant benefit of hormonal therapy (HR = 0.569) but left residual imbalance (Max SMD = 0.270) and high weight variability.
• MEC-Cox Performance: MEC-Cox implementations (using both GLM/Cox and DL/RSF) produced similar qualitative conclusions (HR $\approx$ 0.61–0.65) but with improved weight stability (higher Effective Sample Size, lower Coefficient of Variation) and superior covariate balance (Max SMD < 0.15). The standard errors were also reduced compared to the standard IPW approach.

5. Significance and Claims

The paper claims that MEC-Cox bridges a gap in the literature regarding marginal hazard-ratio estimation in externally controlled trials.

• Robustness: It offers a solution to the sensitivity of existing methods to propensity score model misspecification by leveraging flexible ML.
• Inference: It provides a valid variance estimation framework (stacked sandwich) that accounts for the uncertainty of the calibration step, addressing a limitation of naive robust estimators.
• Efficiency: By balancing prognostic summaries rather than just raw covariates, the method improves finite-sample efficiency.
• Practicality: The dual Newton solver for the calibration step is computationally efficient, operating in the low-dimensional space of the calibration basis rather than the high-dimensional space of the weights.

The authors note that while MEC-Cox improves efficiency and robustness, it does not claim to achieve the semiparametric efficiency bound for the ATT marginal hazard ratio, nor does it establish a doubly robust property for weighted Cox regression in the same algebraic sense as for mean-type estimands. The method assumes source-specific independent censoring, and extensions to informative censoring are left for future work.