Weakening assumptions in the evaluation of treatment effects in longitudinal randomized trials with truncation by death or other intercurrent events

2025-09-12

AI Summary

In longitudinal randomized trials with intermediate events—such as death or treatment switching—conventional intention-to-treat (ITT) analyses suffer from interpretational ambiguity, while existing causal methods rely on untestable strong assumptions (e.g., time-varying no-unmeasured-confounding, positivity) and remain vulnerable to unobserved confounding. To address these limitations, we propose the Paired Last-Observed pre-Transition (PLOT) estimator: it matches units on baseline covariates and conducts pairwise comparisons at the last observed time point prior to the intermediate event, obviating modeling of time-varying confounders, dispensing with the positivity assumption, and avoiding reliance on surrogate endpoints or structural models. We establish its asymptotic unbiasedness under mild conditions. Simulation studies demonstrate PLOT’s robustness and efficiency across diverse truncation mechanisms, markedly reduced sensitivity to unmeasured confounding, and support for model-free, asymptotically valid hypothesis testing under the null.

Technology Category

• Reasoning under Uncertainty: Causality
• Machine Learning: Causal Learning
• Intelligent Robots: State Estimation

Application Category

• Graph Algorithms and Modeling for the Web: Algorithms and analysis for incomplete, noisy, or partially observed Web-related graphs
• Security and Privacy: Large-scale security measurements
• User Modeling, Personalization and Recommendation: Studies of user behavior, including longitudinal effects of personalized systems

Abstract

Intercurrent events, such as treatment switching, rescue medication, or truncation by death, can complicate the interpretation of intention-to-treat (ITT) analyses in randomized clinical trials. Recent advances in causal inference address these challenges by targeting alternative estimands, such as hypothetical estimands or principal stratum estimands (e.g., survivor average causal effects). However, such approaches often require strong, unverifiable assumptions, partly due to limited data on time-varying confounders and the difficulty of adjusting for them. Additionally, strict trial protocols frequently lead to (near) violations of the positivity assumption, resulting in limited information for identifying these estimands. In this paper, we propose a novel approach that sidesteps these difficulties by focusing on testing the null hypothesis of no treatment effect in the presence of arbitrary intercurrent events, including truncation by death, using longitudinal trial data. Our key idea is to compare treated and untreated individuals, matched on baseline covariates, at the most recent time point before either experiences an intercurrent event. We refer to such contrasts as Pairwise Last Observation Time (PLOT) estimands. These estimands can be identified in randomized clinical trials without requiring additional structural assumptions, and even in the presence of the aforementioned positivity violations. However, they may still be susceptible to a form of residual selection bias. We show that this bias vanishes under the conditions typically required by alternative methods, and find it to be more generally small in extensive simulation studies. Building on this, we develop asymptotically efficient, model-free tests using data-adaptive estimation of nuisance parameters. We evaluate the method's performance via simulation studies.

Problem

• Testing treatment effects with arbitrary intercurrent events
• Addressing strong assumptions in causal inference methods
• Handling truncation by death in longitudinal trials

Innovation

• Pairwise Last Observation Time estimands
• Baseline covariates matching strategy
• Model-free tests nuisance parameters