使用Cox proportion, AFT, Aalen Additive模型對於事件發生時間資料的中介效應估計的比較性研究

Abstract

本文探討事件發生時間(time-to-event)結果的中介效應分析，在模擬實驗 中考慮了四種不同的中介因子分佈，透過調整樣本數、設限比例、時間基線風 險，以相對偏差比較在參數模型間的差異，並使用拔靴法 (bootstrapping) 來估計 信賴區間以評估模型效果。當中介變量以廣義線性模型(generalized linear models, GLM)建模時，本文針對事件發生時間結果，採用了加速失效(accelerated failure time, AFT)模型、Cox 比例風險(Cox proportion hazard)模型和 Aalen 加性風 險(Aalen additive hazard)模型，考慮二元暴露的自然直接效應(Natural Direct Effect)並使用中介變量的自然間接效應(Natural Indirect Effect)之一般表達式。 本研究的重點在於中介分佈與模擬設定對模型估計效果的敏感性分析，以及方法 在實際數據(Melanoma 資料集 &HCC 資料集)中的應用。存活模型提供了在不 同尺度中的解釋，解決涉及醫學與可能涵蓋領域中的中因果中介分析問題，提供 相關政策制定者及未來研究建議與指引。

This study examines the mediation effect analysis of time-to-event outcomes. In sim- ulation experiments, we consider four different distributions of mediating variables and compare the differences among parametric models using relative bias by adjusting sample size, censoring proportion, and baseline hazard. Bootstrapping is used to estimate confi- dence intervals to evaluate model performance. When the mediating variable is modeled using generalized linear models (GLMs), this study employs the accelerated failure time (AFT) model, the Cox proportional hazards model, and the Aalen additive hazards model for time-to-event outcomes. We consider the natural direct effect of binary exposure and use the general expression of the natural indirect effect of the mediating variable. The focus of this study is on the sensitivity analysis of the effect estimates to the distribution of the mediators and simulation settings, as well as the application of these methods to real datasets (Melanoma and HCC datasets). Survival models provide an appropriate frame- work at different scales to address issues related to causal mediation analysis in medical research, offering recommendations and guidance for policymakers and future research.