Impact of Near-Positivity Violations on IPTW-Estimated Marginal Structural Survival Models With Time-Dependent...

Area of Science:

• Causal Inference
• Biostatistics
• Epidemiology

Background:

• Marginal structural models (MSMs) estimate causal effects in longitudinal studies with time-dependent confounding.
• Inverse probability of treatment weighting (IPTW) is commonly used to estimate MSMs.
• The positivity assumption is critical for valid IPTW-based causal inference but often overlooked.

Purpose of the Study:

• Investigate the impact of near-positivity violations on IPTW-based MSMs in survival analysis.
• Evaluate the performance of estimators under varying degrees of near-positivity and weight truncation strategies.

Main Methods:

• Developed two algorithms for simulating longitudinal data from hazard-MSMs.
• Simulated data with time-varying binary exposure, time-to-event outcome, and near-positivity violations.
• Analyzed scenarios with rare unexposed individuals within specific confounder levels.

Main Results:

• Near-positivity violations substantially destabilize IPTW-based estimators in longitudinal survival analyses.
• Bias and variance inflation were observed, particularly under aggressive weight truncation.
• Even minor violations significantly impact the reliability of causal effect estimates.

Conclusions:

• Overlooking the positivity assumption in longitudinal observational studies can lead to biased causal inference.
• Naive application of weight truncation strategies exacerbates issues caused by near-positivity violations.
• Emphasizes the critical need to assess and address positivity in causal effect estimation.