Modeling Versus Balancing Approaches to Addressing Instrumental Variables in Weighting: A Comparison of the...

Area of Science:

• Causal Inference
• Statistical Modeling
• Epidemiology

Background:

• Propensity score (PS) variable selection is critical for accurate causal effect estimation.
• Instrumental variables (IVs) can introduce bias and affect precision in PS-weighted estimators.
• Identifying and handling IVs is crucial for reliable causal inference.

Purpose of the Study:

• To compare the performance of outcome-adaptive lasso (OAL), stable balancing weighting (SBW), and stable confounder selection (SCS) in the presence of instrumental variables (IVs).
• To evaluate methods for variable selection in propensity score models when instrumental variables may violate positivity assumptions.

Main Methods:

• Outcome-adaptive lasso (OAL): A model-based approach to identify and exclude IVs from PS models.
• Stable Balancing Weighting (SBW): Directly estimates inverse probability weights while optimizing for weight variance and covariate balance.
• Stable Confounder Selection (SCS): Assesses the stability of model-based effect estimates.

Main Results:

• SBW demonstrated superior performance over OAL and SCS in simulation studies, particularly in reducing mean squared error when IVs were strong and covariates were highly correlated.
• SBW effectively handled limited overlap in an empirical application assessing the effect of abciximab treatment.
• Simulation studies generated extreme propensity scores using IVs and spurious variables to test method robustness.

Conclusions:

• Numerical results support the use of SBW in scenarios with instrumental variables or near-IVs that may lead to practical violations of positivity assumptions.
• SBW provides a robust method for causal inference in the presence of challenging instrumental variables.