Using Super Learner Prediction Modeling to Improve High-dimensional Propensity Score Estimation.

Area of Science:

• Biostatistics
• Epidemiology
• Health Informatics

Background:

• Nonexperimental medical studies using electronic healthcare databases often struggle with confounding.
• The high-dimensional propensity score (hdPS) algorithm aids variable selection for confounding control but lacks clear guidance on optimal variable numbers.
• Overfitting the propensity score model can negatively impact effect estimates.

Purpose of the Study:

• To evaluate data-adaptive approaches for variable selection when using the hdPS algorithm.
• To assess methods for combining hdPS with prediction modeling to improve confounding control in large healthcare databases.
• To determine the optimal strategy for balancing bias and mean squared error (MSE) in effect estimates.

Main Methods:

• Utilized plasmode simulations based on empirical data.
• Compared strategies combining hdPS with Super Learner prediction modeling, collaborative targeted maximum-likelihood estimation, and penalized regression.
• Evaluated performance based on bias and MSE of effect estimates.

Main Results:

• The hdPS algorithm's performance is sensitive to the number of adjustment variables.
• Severe overfitting of the propensity score model can degrade the quality of effect estimates.
• Combining hdPS with Super Learner consistently reduced bias and MSE in effect estimates.

Conclusions:

• Data-adaptive methods are essential for optimizing variable selection with the hdPS algorithm.
• The combination of hdPS and Super Learner prediction modeling demonstrates robust performance in reducing bias and MSE.
• This combined approach shows promise for semiautomated, data-adaptive propensity score estimation in high-dimensional datasets.