Estimating treatment effects with partially observed covariates using outcome regression with missing indicators.

Area of Science:

• Epidemiology
• Biostatistics
• Health Services Research

Background:

• Missing data in covariates is prevalent in observational health research.
• The missing indicator method is a simple approach to handle missing covariates.
• Previous criticisms questioned its validity in outcome regression due to potential bias.

Purpose of the Study:

• To identify and evaluate the assumptions under which the missing indicator approach provides valid inferences.
• To assess the conditions for unbiased estimation of the average treatment effect using outcome regression.
• To investigate the impact of violated assumptions on treatment effect estimates.

Main Methods:

• Theoretical analysis to derive conditions for unbiasedness.
• Simulation studies to quantify bias when assumptions are violated.
• Application to electronic health records data to illustrate findings.

Main Results:

• The missing indicator approach can yield unbiased average treatment effect estimates under specific assumptions.
• Key assumptions include no unmeasured confounding, conditional independence of missing covariate values, and correct outcome model specification.
• Violations of these assumptions can lead to biased treatment effect estimates.

Conclusions:

• The missing indicator approach is a potentially valid method for outcome regression in the presence of missing covariates.
• Its application requires careful evaluation of the plausibility of underlying assumptions.
• This method can provide valid inferences when assumptions are met, supported by simulation and real-world data.