Ensemble learning of inverse probability weights for marginal structural modeling in large observational datasets

Area of Science:

• Biostatistics
• Machine Learning
• Epidemiology

Background:

• Inverse probability weights for marginal structural models are commonly estimated using logistic regression.
• Data-adaptive procedures may improve covariate information utilization.
• Ensemble learning combines predictions from multiple algorithms to reduce bias.

Purpose of the Study:

• To apply and compare two ensemble learning approaches (super learning and ensemble learner) for estimating stabilized weights.
• To evaluate the performance of ensemble methods against logistic regression in estimating marginal structural model parameters.
• To assess the impact of ensemble learning on bias reduction and confidence interval precision.

Main Methods:

• Application of super learning (SL) utilizing V-fold cross-validation.
• Implementation of an ensemble learner (EL) with a single training/validation data split.
• Analysis of longitudinal data from HIV-positive subjects in the CoRIS and CoRIS-MD studies.

Main Results:

• Both ensemble approaches yielded hazard ratio estimates further from the null compared to logistic regression.
• Ensemble methods resulted in tighter confidence intervals for hazard ratio estimates.
• The ensemble learner (EL) demonstrated less computation time than super learning (SL).

Conclusions:

• Ensemble learning provides a viable alternative to parametric modeling for inverse probability weights in marginal structural models.
• Ensemble methods, particularly EL, offer efficient bias reduction and improved precision for large datasets.
• Combining diverse algorithms in ensemble learning enhances the estimation of marginal structural model parameters.