Overview
Research introduces a subsampling and reweighting strategy designed for longitudinal survival data. This approach addresses challenges associated with estimating the causal effect of time-varying treatments on survival outcomes in large observational studies, particularly when outcomes are rare. The method aims to reduce computational burden, maintain consistency, and enhance estimation stability in scenarios characterized by rare outcomes.
Research Context
Estimating causal effects of time-varying treatments on survival outcomes in large observational studies can be computationally intensive, especially when outcomes are rare. G-formula-based methods, such as the iterative conditional expectation (ICE) estimator, offer a principled framework for longitudinal causal inference. However, these methods can become computationally expensive, particularly when bootstrap-based variance estimation is required. Additionally, the rarity of outcomes at each time point can lead to severe class imbalance, which may result in instability and convergence issues when using models like logistic regression.
Approach
The proposed method involves a principled subsampling and reweighting strategy. This strategy is designed for application to longitudinal survival data and is compatible with a range of existing causal effect estimators in this setting, including the ICE estimator. The core objective of this approach is to mitigate computational demands and improve estimation stability.
Findings
- The proposed subsampling and reweighting strategy substantially reduces the computational burden associated with estimating causal effects of time-varying treatments on rare survival outcomes.
- The method preserves consistency in estimation.
- Estimation stability is improved, particularly in rare-outcome settings.
- The strategy addresses issues of severe class imbalance at each time point which can otherwise lead to instability and convergence problems in models such as logistic regression.
- Evaluation through simulations indicated the method's effectiveness.
- Validation using a large-scale Electronic Health Record (EHR) cohort study on social and behavioral determinants of health (SBDH) and suicide risk demonstrated its applicability for modeling rare outcomes in longitudinal data.
Why This Matters
This method offers a way to manage the computational demands and instability often encountered when analyzing rare outcomes in large longitudinal observational studies. By providing a strategy that reduces computational burden and improves stability while preserving consistency, it facilitates more efficient and reliable causal effect estimation in challenging data environments. The validation using an EHR cohort study on SBDH and suicide risk suggests its potential utility in addressing significant public health areas where rare outcomes are critical.