Huang Chen, Wei Kecheng, Yu Yongfu, Qin Guoyou
Department of Biostatistics, Key Laboratory for Health Technology Assessment, Key Laboratory of Public Health Safety of Ministry of Education, School of Public Health, National Commission of Health, Fudan University, Shanghai, China.
Shanghai Stomatological Hospital & School of Stomatology, Fudan University, Shanghai, China.
BMC Med Res Methodol. 2025 Jul 1;25(1):169. doi: 10.1186/s12874-025-02612-3.
In multi-site studies in clinical practice, the treatment effects may exhibit potential heterogeneity across different sites. Additionally, propensity score methods used to adjust for confounding may suffer from model misspecification, which can lead to biased estimates. Addressing both the potential heterogeneity of site-specific treatment effects and the issues of model misspecification in the context of multi-site survival data is a critical area that warrants further research.
We propose a novel partially heterogeneous weighted fusion learning method. This approach is designed to simultaneously identify and estimate potential heterogeneous treatment effects across sites, while also improving robustness against model misspecification in the estimation of survival causal effects across different sites. We evaluate the performance of this method through simulation studies and apply it to the real-world data from the Surveillance, Epidemiology, and End Result (SEER) database to assess whether the survival effect of surgery with adjuvant radiation therapy for breast cancer patients differs across sites in a multi-site study.
Simulation studies demonstrate that our proposed method accurately identifies the potential heterogeneous treatment effects, and when the candidate models include the correct model, our method performs comparably to methods based on correctly specified propensity score models in estimating the site-specific hazard ratios. The application of our method to the SEER database revealed two distinct survival effects of surgery with adjuvant radiation therapy for breast cancer patients across sites. However, compared to our method, traditional methods that pool all sites together failed to identify this heterogeneity, while analyzing each site individually led to a reduction in statistical power.
This study introduces a partially heterogeneous weighted fusion learning method for survival data that effectively identifies potential inter-site heterogeneity in treatment effects, while simultaneously addressing the issue of model misspecification.
在临床实践的多中心研究中，不同中心的治疗效果可能存在潜在的异质性。此外，用于调整混杂因素的倾向评分方法可能存在模型误设问题，这可能导致估计有偏差。在多中心生存数据的背景下，解决特定中心治疗效果的潜在异质性以及模型误设问题是一个关键领域，值得进一步研究。
我们提出了一种新颖的部分异质性加权融合学习方法。该方法旨在同时识别和估计不同中心潜在的异质性治疗效果，同时在估计不同中心的生存因果效应时提高对模型误设的稳健性。我们通过模拟研究评估该方法的性能，并将其应用于监测、流行病学和最终结果（SEER）数据库的真实世界数据，以评估在多中心研究中乳腺癌患者辅助放疗手术的生存效果在不同中心是否存在差异。
模拟研究表明，我们提出的方法能够准确识别潜在的异质性治疗效果，并且当候选模型包括正确模型时，我们的方法在估计特定中心的风险比方面与基于正确设定的倾向评分模型的方法表现相当。我们的方法应用于SEER数据库揭示了乳腺癌患者辅助放疗手术在不同中心有两种不同的生存效果。然而，与我们的方法相比，将所有中心合并在一起的传统方法未能识别这种异质性，而单独分析每个中心会导致统计功效降低。
本研究介绍了一种用于生存数据的部分异质性加权融合学习方法，该方法能有效识别治疗效果中潜在的中心间异质性，同时解决模型误设问题。
