Outcome-adaptive lasso: Variable selection for causal inference.

Area of Science:

• Epidemiology
• Biostatistics
• Observational Data Analysis

Background:

• Propensity score methods are vital for unbiased treatment effect estimation from observational data.
• Traditional covariate selection in propensity scores can lead to bias and reduced statistical efficiency.
• Including irrelevant covariates can inflate standard errors or miss important predictors.

Purpose of the Study:

• To propose and evaluate the outcome-adaptive lasso for selecting appropriate covariates in propensity score models.
• To improve confounding bias adjustment and maintain statistical efficiency.
• To perform variable selection effectively, even with numerous spurious covariates.

Main Methods:

• Development and application of the outcome-adaptive lasso method for covariate selection.
• Theoretical analysis and simulation studies to assess performance.
• Illustration using simulated data and a real-world dataset of patients on opioid therapy.

Main Results:

• The outcome-adaptive lasso effectively selects covariates that are true confounders and outcome predictors.
• The method successfully excludes covariates unrelated to exposure or outcome, improving efficiency.
• Simulations and real-world data confirm the method's ability to balance bias and precision.

Conclusions:

• The outcome-adaptive lasso offers a statistically efficient approach for covariate selection in propensity score models.
• This method enhances the reliability of treatment effect estimation from observational studies.
• It provides a robust alternative to traditional, less selective covariate selection strategies.