A Boosting Algorithm for Estimating Generalized Propensity Scores with Continuous Treatments.

Area of Science:

• Statistics
• Biostatistics
• Epidemiology

Background:

• Causal inference with continuous treatments is challenging due to high-dimensional covariates.
• Traditional propensity score methods struggle with the curse of dimensionality.
• Existing two-step procedures often involve modeling the mean function.

Purpose of the Study:

• To propose a novel boosting algorithm for estimating the mean function in continuous treatment causal inference.
• To introduce the average absolute correlation coefficient (AACC) for optimal tuning of the boosting algorithm.
• To evaluate the performance of the proposed method against existing approaches.

Main Methods:

• Utilizing a boosting algorithm to estimate the mean function of the treatment given covariates.
• Defining the generalized propensity score as the conditional density of treatment levels.
• Employing inverse probability weighting with estimated propensity scores for dose-response function estimation.
• Proposing the average absolute correlation coefficient (AACC) to select the optimal number of trees in boosting.

Main Results:

• The proposed boosting approach demonstrates superior performance compared to linear approximation and L2 boosting in simulations.
• The AACC criterion effectively determines the optimal number of trees, balancing bias and variance.
• The methodology is successfully applied to the Early Dieting in Girls study.

Conclusions:

• The developed boosting algorithm offers an effective solution for causal inference with continuous treatments, especially in high-dimensional settings.
• The AACC criterion provides a reliable method for tuning the boosting algorithm.
• The study highlights the influence of maternal weight concern on adolescent girls' dieting behaviors.