Propensity score analysis methods with balancing constraints: A Monte Carlo study.

Area of Science:

• Causal inference
• Statistical modeling
• Observational studies

Background:

• Inverse probability weighting (IPW) is a key method for estimating average treatment effects (ATE).
• Covariate balancing constraints can mitigate issues with large IPW weights and improve covariate balance.
• Alternative weighting methods exist that balance covariates without relying on propensity scores.

Purpose of the Study:

• To investigate if covariate balancing constraints eliminate the need for correct propensity score model specification.
• To determine if propensity score models enhance estimation performance when used with similar balancing constraints.
• To compare the performance of various weighting methods under different conditions.

Main Methods:

• Comprehensive Monte Carlo simulations were conducted.
• Methods compared included simple IPW, covariate balancing propensity score (CBPS), covariate balancing scoring rule, entropy balancing, and kernel balancing.
• The study evaluated the impact of propensity score model specification and the inclusion of propensity scores in methods with balancing constraints.

Main Results:

• Correct propensity score model specification remains critical for accurate ATE estimation, even when balancing constraints are effective.
• Weighting methods incorporating propensity scores showed improved estimation performance compared to those without, particularly when the number of covariates was large.
• Covariate balancing constraints alone did not fully circumvent the need for accurate propensity score modeling.

Conclusions:

• Propensity score model accuracy is essential for reliable causal effect estimation.
• Integrating propensity scores with covariate balancing constraints offers benefits, especially in high-dimensional settings.
• Developing flexible, data-driven propensity score models that inherently satisfy balancing conditions is recommended.