Evaluating treatment effects on longitudinal outcomes with attrition due to death: Methods for a two-dimentional estimand with a case study in Quality of Life

2025-09-29

AI Summary

In populations with non-negligible mortality risk, death induces missingness in longitudinal outcomes, complicating causal effect estimation. Conventional approaches either merge survival and longitudinal outcomes or overlook the distinct mechanisms governing them, compromising valid inference for either process. This paper proposes a two-dimensional “alive-only” causal framework that decouples survival status from longitudinal health outcomes (e.g., quality of life) observed exclusively among survivors, and defines identifiable two-dimensional causal estimands. Methodologically, it jointly models survival and longitudinal trajectories, integrating regression standardization with inverse probability weighting to simultaneously address confounding bias and right-censoring. Simulation studies demonstrate robust finite-sample performance, and real-world analysis of oncology patient data confirms its interpretability and practical utility. To our knowledge, this is the first framework to enable principled, separate causal inference for survival and longitudinal outcomes—avoiding estimation bias inherent in traditional one-dimensional composite approaches.

Technology Category

• Reasoning under Uncertainty: Causality
• Machine Learning: Causal Learning
• Search and Optimization: Mixed Discrete/Continuous Search

Application Category

• User Modeling, Personalization and Recommendation: Studies of user behavior, including longitudinal effects of personalized systems
• Search and Retrieval-Augmented AI: Web learning to rank, online learning, and counterfactual learning for ranking
• Graph Algorithms and Modeling for the Web: Algorithms and analysis for incomplete, noisy, or partially observed Web-related graphs

Abstract

When longitudinal outcomes are evaluated in mortal populations, their non-existence after death complicates the analysis and its causal interpretation. Where popular methods often merge longitudinal outcome and survival into one scale or otherwise try to circumvent the problem of mortality, some highly relevant questions require survival to be acknowledged as a unique condition. " extit{What are my chances of survival}" and " extit{What can I expect for my condition while still alive}" reflect the intrinsically two-dimensional outcome of survival and longitudinal outcome while-alive. We define a two-dimensional causal while-alive estimand for a point exposure and compare two methods for estimation in an observational setting. Regression-Standardization models survival and the observed longitudinal outcome before standardizing the latter to a target population weighted by its estimated survival. Alternatively, Inverse Probability of Treatment and Censoring Weighting weights the observed outcomes twice, to account for censoring and differences in baseline-case-mix. Both approaches rely on the same causal identification assumptions, but require different models to be correctly specified. With its potential to extrapolate, Regression-Standardization is more efficient when all assumptions are met. We show finite sample performance in a simulation study and apply the methods to a case study on quality of life in oncology.

Problem

• Evaluating treatment effects with mortality-related attrition
• Defining two-dimensional causal estimands for survival outcomes
• Comparing standardization and weighting methods for longitudinal data

Innovation

• Two-dimensional causal estimand combining survival and outcomes
• Regression-Standardization method models survival before standardization
• Inverse Probability Weighting accounts for censoring and case-mix