Causal mediation analysis for a time-to-event endpoint in the presence of competing events
Yuhao Deng, Peter Gilbert, Youyi Fong, Bo Zhang
Fred Hutchinson Cancer Center
stat.ME
Submitted: 2026-11-09
Updated: 2026-11-09
Code: https://github.com/naiiife/MedCmprsk
Importance score: 92/100
The gist: Causal mediation analysis is employed to decompose a treatment's effect into distinct causal pathways, particularly in time-to-event settings complicated by competing events.
Terminology
Summary
Causal mediation analysis is employed to decompose a treatment's effect into distinct causal pathways, particularly in time-to-event settings complicated by competing events. The paper develops a unified treatment for this analysis, showing that the total effect on the cause-specific cumulative incidence decomposes into a direct effect, an indirect effect through the mediator (M), and an indirect effect through the competing events. p. 0 This decomposition is established under natural effects (NME, NDE, NCE), stochastic interventional effects (IME, IDE, ICE), and separable effects (SME, SDE, SCE) frameworks. p. 0 The authors derive efficient influence functions (EIFs) for counterfactual cumulative incidence curves and decomposed effects using multiply robust double-machine-learning estimators that achieve the semiparametric efficiency bound. p. 0 They also investigate the mediator–outcome relationship via kernel-smoothed estimands and study the dose–response relationship between the mediator and the counterfactual risk of the primary event in this context. p. 0
The methodology is applied to immune correlates analyses from COVID-19 vaccine and HIV-1 prevention trials. p. 0 Specifically, for a vaccine trial analyzing asymptomatic first infection, the total effect on cumulative incidence can be decomposed into components corresponding to: (i) NME1 (vaccine-induced changes in the immune marker), (ii) NDE1 (direct influence of the vaccine on asymptomatic first infection), and (iii) NCE1 (direct influence of the vaccine on symptomatic first infection). p. 3 This decomposition necessitates that the comparator-arm hazard be evaluated at marker values attained under the investigational vaccine, making positivity a critical condition in Assumption 1(iii). Furthermore, an alternative decomposition is proposed as NME + NCE' + NDE', where NCE' compares a = (1, 0, 1) with (1, 0, 0) and NDE' compares (1, 1, 1) with (1, 0, 1), which vanishes when the primary- and competing-event components act additively on the CIF.
Simulation studies using mock datasets mimicking vaccine trials confirm the robustness of the estimators; bias is generally small even at sample sizes as low as n=1000, and estimators perform well across various misspecifications, confirming multiple robustness. The asymptotic variance attains the nonparametric efficiency bound. p. 5 Sensitivity analyses examine unmeasured common causes across survival endpoints, showing that for instance, when m (related to the competing event incidence) is larger, the effect of increasing at-risk probability is weak. p. 5 The final result indicates that the EIF-based estimator is asymptotically normal and nonparametrically efficient. p. 4 In a case study involving immune correlates analysis, the estimated mediation effect was slightly negative but insignificant, suggesting that active treatment might lower asymptomatic first infection incidence through pathways not involving the cellular response or competing event (e. p. 8g., humoral immunity). p. 3
Green — the paper says it, and the page is printed. Red — no passage carries it, so treat that sentence as unverified. The same check runs on your own text: check a summary against its source, $5.
Improvements for AI systems
-
Improved Causal Mediation Analysis for Time-to-Event Endpoints under Competing Events: The system can decompose total effect into direct, indirect (through mediator), and indirect (through competing events) effects using three frameworks: natural effects, stochastic interventional effects, and separable effects. This allows the AI to compare the strengths and limitations of each framework in identifying causal pathways.
-
Development of Efficient Influence Functions for Counterfactual Cumulative Incidence Curves: The system can derive efficient influence functions (EIFs) for counterfactual cumulative incidence curves and decomposed effects, leading to multiply robust double-machine learning estimators that attain the semiparametric efficiency bound in a nonparametric model.
-
Implementation of Kernel-Smoothed Controlled Effects Estimation: The system can estimate the dose-response relationship between a mediator and the counterfactual risk by using kernel-smoothed estimands, allowing for interpretation of risks under different levels of the mediator.
-
Robust Inference via Cross-Fitting and Asymptotic Normality: The AI can utilize cross-fitting to estimate counterfactual CIFs and their decomposed effects, achieving asymptotic normality with an asymptotic variance that attains the nonparametric efficiency bound in the nonparametric model space.
-
Sensitivity Analysis for Unmeasured Common Causes: The system can perform sensitivity analysis by relaxing assumptions on unmeasured common causes between primary and competing events, allowing for a decomposition of direct and competing effects based on a sensitivity parameter Gamma.
Sources
- Debiasing hazard-based, time-varying vaccine effects using vaccine-irrelevant infections: An observational extension of a pivotal Phase 3 COVID-19 vaccine efficacy trial
- Inference for natural mediation effects under case-cohort sampling with applications in identifying COVID-19 vaccine correlates of protection
- Causal Effect Estimation after Propensity Score Trimming with Continuous Treatments
- Automatic Debiased Machine Learning via Riesz Regression
Related papers
- Learning Robust Treatment Rules for Censored Data
- Neural Estimation of Summary Statistics for Simulation-Based Inference with Non-Asymptotic Guarantees
- Spectral divide-and-conquer MCMC for long stationary time series
- Improving the Efficiency of Subgroup Analysis in Randomized Controlled Trials with TMLE
- Personalized Federated Vector Autoregression with Personalization Diversity
- Doubly Robust and Efficient Calibration of Prediction Sets for Right-Censored Time-to-Event Outcomes