Comparison of propensity score methods for pre-specified subgroup analysis with survival data

Area of Science:

• Epidemiology
• Biostatistics
• Public Health

Background:

• Evaluating medical product safety across diverse populations is crucial for public health.
• Observational studies can inform subgroup-specific risks if confounding is adequately controlled.
• Limited guidelines exist for simultaneous confounding control and subgroup analysis.

Purpose of the Study:

• To evaluate propensity score methods for confounding control in subgroup analyses.
• To assess bias, efficiency, and coverage of different methods across various simulation scenarios.
• To provide guidance on best practices for subgroup safety analyses in cohort studies.

Main Methods:

• Simulation study evaluating six propensity score methods (matching and weighting).
• Estimation of subgroup-specific hazard ratios for average treatment effect in the treated using Cox regression.
• Scenarios varied by subgroup size, subgroup-exposure/outcome associations, and outcome incidence.

Main Results:

• Subsetting data by the subgrouping variable to estimate propensity scores and hazard ratios resulted in the lowest bias.
• Subsetting outperformed methods that borrow information across subgroups, despite potential precision trade-offs.
• Weighting methods showed higher bias than matching methods when propensity score models were misspecified and subgroups were strong confounders.

Conclusions:

• Subsetting data is a robust strategy for minimizing bias in propensity score-based subgroup analyses.
• Careful consideration of propensity score model specification is vital, especially with strong confounders.
• Findings inform the design and analysis of safety studies for medical products in specific populations.