Flexible propensity score estimation strategies for clustered data in observational studies

Area of Science:

• Epidemiology
• Biostatistics
• Machine Learning

Background:

• Nonparametric machine learning often shows superior performance to logistic regression for propensity score estimation.
• The effectiveness of nonparametric methods in clustered settings, particularly with unmeasured cluster-level confounding, remains unclear.

Purpose of the Study:

• To compare the performance of logistic regression, Bayesian additive regression trees, and generalized boosted modeling for propensity score weighting in clustered settings.
• To evaluate the impact of sample size, cluster size, and unmeasured cluster-level confounding on these methods.

Main Methods:

• Simulated data from three hypothetical observational studies with varying sample and cluster sizes.
• Included cluster indicators or random intercepts to account for clustering.
• Generated confounders at individual and cluster levels, including an unobserved cluster-level confounder.
• Assessed performance based on covariate balance, bias reduction, and confidence interval coverage.
• Applied methods to the National Longitudinal Study of Adolescent to Adult Health data.

Main Results:

• Nonparametric methods demonstrated better covariate balance, bias reduction, and confidence interval coverage in large clustered settings, irrespective of model complexity.
• In small sample or cluster sizes, nonparametric methods showed increased vulnerability to unmeasured cluster-level confounding.
• Multilevel logistic regression may be a more robust alternative in smaller, clustered settings with potential unmeasured confounding.

Conclusions:

• Nonparametric propensity score estimation offers advantages in large clustered datasets but requires careful consideration in smaller clusters due to potential unmeasured confounding.
• The choice of method should account for sample size, cluster size, and the potential for unmeasured cluster-level confounders.