Built-in selection or confounder bias? Dynamic Landmarking in matched propensity score analyses

Area of Science:

• Epidemiology
• Biostatistics
• Health Services Research

Background:

• Propensity score matching (PSM) is widely used for causal inference in non-randomized studies.
• Estimating hazard ratios with PSM for time-to-event data is susceptible to bias from unobserved covariates.
• Researchers often lack knowledge of which unobserved factors may introduce bias.

Purpose of the Study:

• To adapt and evaluate Dynamic Landmarking for detecting bias in PSM for time-to-event outcomes.
• To provide a visual diagnostic tool for assessing the robustness of treatment effect estimates.

Main Methods:

• Extended Dynamic Landmarking, originally for randomized trials, to PSM settings.
• Employs successive deletion of sorted observations and univariable Cox models.
• Measures covariate balance, including for omitted variables, using sum of squared z-differences.

Main Results:

• Simulations demonstrate Dynamic Landmarking effectively detects and distinguishes selection and confounding bias.
• The method provides a visual tool to assess potential distortions in treatment effect estimates.
• Application to a cardiac surgery dataset illustrates practical interpretation and use.

Conclusions:

• Dynamic Landmarking serves as a valuable post-hoc diagnostic tool.
• It aids in visualizing potential confounding or selection bias affecting hazard ratio estimates.
• The method enhances the reliability of causal inference from observational studies.