Estimation of average treatment effect based on a multi-index propensity score

Area of Science:

• Biostatistics
• Epidemiology
• Health Outcomes Research

Background:

• Observational studies are crucial for estimating treatment effects but susceptible to confounding bias due to unbalanced covariates.
• Estimating the average treatment effect (ATE) accurately is vital for medical research.
• Existing methods may not fully address confounding bias in complex observational data.

Purpose of the Study:

• To propose a novel estimator, the artificial neural network-based multi-index propensity score (ANN.MiPS), to correct confounding bias in ATE estimation.
• To enhance estimation consistency and robustness by combining information from multiple propensity score and outcome regression models.
• To evaluate the performance and practicability of the ANN.MiPS estimator through simulation and real-world data analysis.

Main Methods:

• Developed the ANN.MiPS estimator utilizing artificial neural networks to integrate information from multiple candidate models.
• Conducted a Monte Carlo simulation study to assess the estimator's performance across various scenarios (sample sizes, treatment rates, covariate types).
• Applied the ANN.MiPS estimator to real observational data to demonstrate its practical utility and stability.

Main Results:

• The ANN.MiPS estimator demonstrated very small bias and comparable standard errors to existing methods when at least one candidate model was correctly specified.
• ANN.MiPS generally yielded smaller standard errors than kernel function-based estimators when the correct model was included.
• Empirical application showed stable point estimation and bootstrap standard errors for ATE under diverse model specifications.

Conclusions:

• The ANN.MiPS estimator effectively combines information from multiple models, achieving multiply robust estimation for ATE.
• This novel approach offers enhanced efficiency compared to kernel-based methods for causal effect estimation in observational studies.
• ANN.MiPS provides a valuable new tool for addressing confounding in health research using observational data.