Covariate association eliminating weights: a unified weighting framework for causal effect estimation

Area of Science:

• Epidemiology
• Biostatistics
• Causal Inference

Background:

• Weighting methods are used to estimate causal effects in observational studies.
• Maximum likelihood estimation of weights can lead to bias and variance if the treatment assignment model is misspecified.

Purpose of the Study:

• To propose a unified framework for constructing weights that eliminate covariate-treatment assignment associations.
• To extend covariate balancing methods to longitudinal settings.

Main Methods:

• Developed a framework for weight estimation by deriving conditions to eliminate covariate-treatment associations.
• Demonstrated that existing covariate balancing methods are special cases of this framework.
• Extended the framework to handle longitudinal data.

Main Results:

• Simulations show the proposed method yields lower bias and variance compared to maximum likelihood estimation under model misspecification.
• The framework successfully eliminates associations between measured pretreatment covariates and treatment assignment after weighting.

Conclusions:

• The proposed unified framework offers a robust approach to causal effect estimation in observational studies, particularly when treatment assignment models are misspecified.
• This method improves the reliability of causal inference in both cross-sectional and longitudinal settings.