Abstract
Breast cancer patients may experience relapse or death after surgery during the follow-up period, leading to dependent censoring of relapse. This phenomenon, known as semi-competing risk, imposes challenges in analyzing treatment effects on breast cancer and necessitates advanced statistical tools for unbiased analysis. Despite progress in estimation and inference within semi-competing risks regression, its application to causal inference is still in its early stages. This article aims to propose a frequentist and semi-parametric framework based on copula models that can facilitate valid causal inference, net quantity estimation and interpretation, and sensitivity analysis for unmeasured factors under right-censored semi-competing risks data. We also propose novel procedures to enhance parameter estimation and its applicability in real practice. After that, we apply the proposed framework to a breast cancer study and detect the time-varying causal effects of hormone- and radio-treatments on patients' relapse-free survival and overall survival. Moreover, extensive numerical evaluations demonstrate the method's feasibility, highlighting minimal estimation bias and reliable statistical inference.
Introduction
In clinical oncology, particularly in breast cancer research, investigators typically monitor two primary endpoints: Relapse-Free Survival (RFS) and Overall Survival (OS). Analyzing these outcomes simultaneously presents a significant statistical challenge known as "semi-competing risks." In this scenario, while a relapse event does not preclude the subsequent occurrence of death, the occurrence of death prevents the observation of a relapse.
This structure induces "dependent censoring." If researchers ignore the fact that patients who die may have been at a higher latent risk for relapse, standard survival estimates become biased. Traditional methods often fail to capture the "net" effect of a treatment—the hypothetical effect on relapse if death were removed as a competing risk. This paper addresses these gaps by proposing a frequentist, semi-parametric framework designed to untangle these dependencies and provide rigorous causal interpretations.
Methodology
The proposed framework integrates three sophisticated statistical pillars to ensure robust causal inference:
1. Copula-Based Joint Modeling
Rather than analyzing RFS and OS as independent variables, the authors employ Archimedean copulas (specifically the Frank copula) to model the joint distribution of the two events. This approach allows for the quantification of "net" treatment effects while explicitly accounting for the inherent association between the timing of relapse and the timing of death.
2. Principal Stratification and SACE
To achieve a true "apples-to-apples" comparison in an observational setting, the framework utilizes Principal Stratification. The authors focus on the Survivor Average Causal Effect (SACE), which looks at treatment effects within specific latent strata, such as the "Always-Diseased" (those who would relapse regardless of treatment assignment) and "Never-Diseased." This provides a more granular and clinically relevant interpretation of how treatments perform across different patient risk profiles.
3. Advanced Estimation and Sensitivity Analysis
The study enhances standard Nonparametric Maximum Likelihood Estimation (NPMLE) by implementing refined initial value selection to avoid local optima. Furthermore, to address potential unmeasured confounding, the authors introduce a modified Monte Carlo Expectation-Maximization (MCEM) algorithm. This allows for sensitivity analysis regarding "frailty" variables—latent factors that influence both relapse and survival—enabling researchers to quantify how robust their conclusions are to hidden biases.
Data and Evaluation
The framework was validated through both real-world application and rigorous simulation:
- METABRIC Cohort: The authors analyzed 1,114 patients from the Molecular Taxonomy of Breast Cancer International Consortium. The analysis focused on two primary comparisons: hormone therapy versus no treatment, and combined hormone and radiotherapy versus hormone therapy alone.
- Numerical Simulations: Four distinct scenarios (Ex1–Ex4) were tested to evaluate the model's performance under varying levels of censoring, different covariate counts, and the presence of unmeasured confounding factors.
Key Results
Clinical Insights
In the METABRIC study, the model revealed that hormone therapy alone significantly improved survival outcomes for patients with a lower baseline risk of relapse. For patients in the higher-risk "Always-Diseased" stratum, the addition of radiotherapy to hormone therapy showed a detectable, albeit weaker, benefit for overall survival compared to hormone therapy alone. These findings suggest that treatment benefits are highly dependent on the patient's underlying risk strata.
Statistical Performance
Simulation results confirmed that the proposed estimators exhibit minimal bias. The 95% confidence intervals maintained high accuracy, with coverage probabilities consistently near 0.95, even in environments with high levels of data censoring.
Comparison to Traditional Methods
A critical finding was the discrepancy between this copula-based approach and traditional Kaplan-Meier (KM) analysis. KM analysis, which assumes independent censoring, produced survival curves that occasionally contradicted clinical expectations. The copula-based framework yielded more conservative and reliable estimates by properly accounting for the dependency between relapse and death.
Discussion and Conclusion
This research provides a powerful new tool for oncology researchers to evaluate treatment efficacy in the presence of semi-competing risks. By shifting the focus to "net" quantities and employing principal stratification, the framework offers a "purified" view of how interventions influence disease progression.
Future work may expand this model to include time-dependent or covariate-dependent copula associations, as the current model assumes a homogeneous association. Additionally, extending the framework to handle multi-arm trials or continuous dosages would further enhance its utility in complex clinical decision-making. Ultimately, this methodology moves the field closer to a more honest and rigorous assessment of life-saving treatments in observational cancer data.
Summary updated