Prognostic score-based model averaging approach for propensity score estimation | JoVE Visualize

Area of Science:

Statistics
Epidemiology
Biostatistics

Background:

Traditional propensity score (PS) evaluation focuses on covariate balance, potentially overlooking predictive power for outcomes, leading to suboptimal bias reduction.
Prognostic scores, which incorporate covariate-outcome relationships, offer an alternative for evaluating covariate balance.
Standard PS model averaging methods may not optimize confounding adjustment; averaging based on prognostic score balance is proposed as a superior alternative.

Purpose of the Study:

To propose and evaluate a novel propensity score (PS) model averaging approach that utilizes prognostic score balance.
To demonstrate that averaging PS models based on prognostic score balance can reduce bias in treatment effect estimates.
To compare the performance of the proposed method against existing approaches through simulations and empirical data analysis.

Main Methods:

Conducted simulations and analyzed empirical observational data to compare treatment effect estimates.
Developed four candidate PS and prognostic score models using traditional and machine learning methods.
Employed model averaging for PS estimation, optimizing for treatment prediction accuracy, covariate balance, or prognostic score balance; used prognostic scores and model-averaged prognostic scores for balance assessment.

Main Results:

The proposed model-averaging approaches for PS estimation consistently demonstrated lower bias and less variability in treatment effect estimates compared to existing methods.
Utilizing optimally averaged prognostic scores as a balance metric significantly enhanced the robustness of weighted treatment effect estimates.
Both simulation and empirical analyses supported the superiority of the prognostic score-based model averaging approach.

Conclusions:

Prognostic score-based PS model averaging outperforms existing methods, yielding more robust treatment effect estimates.
The proposed method, using model-averaged prognostic scores as a balance metric, is recommended for various applications.
Application to complex, high-dimensional real-world data requires careful adjustments and complementary techniques to ensure accurate estimation.