Comparing methods for estimating causal treatment effects of administrative health data: A plasmode simulation study

Area of Science:

• Health Services Research
• Biostatistics
• Health Economics

Background:

• Estimating causal effects of health policy interventions using real-world administrative data presents methodological challenges.
• Lack of clear guidance hinders accurate policy impact assessment.

Purpose of the Study:

• To evaluate and compare causal inference methods for health policy analysis using a plasmode simulation.
• To assess the performance of different methods in estimating treatment effects on healthcare utilization and costs.

Main Methods:

• Plasmode simulation using German administrative health care data from a deprived urban area.
• Comparison of propensity score matching, inverse probability of treatment weighting, and entropy balancing.
• Evaluation of augmented inverse probability weighting and targeted maximum likelihood estimation.
• Nuisance parameter estimation using regression models versus superlearner (ensemble learner).

Main Results:

• Superlearner demonstrated effectiveness in handling nuisance terms with large covariate sets, particularly when combined with doubly robust estimation.
• Regression-based nuisance parameter estimation performed best with small covariate sets, especially when paired with singly robust methods.
• The study identified varying strengths and weaknesses among the evaluated causal inference approaches.

Conclusions:

• Superlearner offers a robust approach for causal inference in health policy research with complex, high-dimensional data.
• The choice of nuisance parameter estimation method should consider the size of the covariate set and the robustness of the primary estimation method.
• Findings provide valuable methodological guidance for utilizing real-world health data to assess policy interventions.