Propensity score analysis with partially observed covariates: How should multiple imputation be used?

Area of Science:

• Biostatistics
• Epidemiology
• Observational Studies

Background:

• Inverse probability of treatment weighting (IPTW) is crucial for estimating marginal treatment effects in observational studies, but is challenged by missing covariate data.
• Multiple imputation (MI) is a common strategy for handling missing data, yet its optimal implementation within IPTW frameworks remains unclear.

Purpose of the Study:

• To investigate and compare different multiple imputation strategies for propensity score analysis in the context of IPTW.
• To assess the impact of including the outcome in the imputation model and the application of Rubin's rules on treatment effect estimation.
• To evaluate the consistency, balancing properties, and empirical performance of various MI-IPTW methods.

Main Methods:

• Compared two primary MI approaches: MIte (combining treatment effect estimates) and MIps (combining propensity scores).
• Evaluated MIpar (combining propensity score parameters), complete case analysis, and missingness pattern analysis.
• Conducted a simulation study for a binary outcome, assessing bias, balancing properties, and variance estimation under a missing at random mechanism.

Main Results:

• Complete case and missingness pattern analyses yielded biased marginal treatment effect estimates.
• Multiple imputation approaches were approximately unbiased when the outcome was included in the imputation model.
• MIte demonstrated unbiasedness across all scenarios, with Rubin's rules providing accurate variance estimates and good covariate balancing properties.

Conclusions:

• For IPTW with missing covariate data, the MIte approach, incorporating the outcome in the imputation model, is the recommended strategy.
• This method ensures unbiased estimation of marginal treatment effects and reliable variance estimation, outperforming other MI and non-MI approaches.