Propensity Score Modeling in Electronic Health Records with Time-to-Event Endpoints: Application to Kidney...

Area of Science:

• Biostatistics
• Epidemiology
• Health Services Research

Background:

• Propensity scores (PS) are crucial for bias reduction in observational studies.
• Existing PS methods lack consensus, especially for multiple treatments and time-to-event outcomes with high censoring and clustering.
• Kidney transplantation data presents unique challenges including confounding, censoring, and clustered survival outcomes.

Purpose of the Study:

• To investigate and compare Generalized Boosted Model (GBM) and Covariate-Balancing Propensity Score (CBPS) procedures.
• To evaluate these PS methods for estimating causal effects in a complex, right-censored kidney transplant dataset.
• To address challenges like multiple treatment groups, high censoring, and data clustering in time-to-event analysis.

Main Methods:

• A two-step estimation procedure was employed using a kidney transplantation dataset.
• Multinomial PS models were fitted using GBM and CBPS to adjust for numerous confounders across multiple treatment subgroups.
• Estimated PS were incorporated into a semi-parametric cure rate Cox proportional hazard frailty model using inverse probability of treatment weighting, adjusted for multi-center clustering and excess censoring.

Main Results:

• The study compared GBM and CBPS for propensity score estimation in a complex observational dataset.
• The analysis revealed that the full model, incorporating various data complexities, provided more informative and superior treatment effect estimation.
• The proposed two-step procedure effectively addressed challenges observed in the United Network of Organ Sharing (UNOS) database.

Conclusions:

• The developed two-step procedure using GBM and CBPS offers a robust framework for causal inference in complex time-to-event data.
• The full model demonstrated superior performance in estimating treatment effects compared to simplified models.
• This approach enhances the reliability of treatment effect estimation in large observational studies with challenging data characteristics.