Model-based standardization using multiple imputation.

Area of Science:

• Biostatistics
• Epidemiology
• Clinical Research

Background:

• Parametric multivariable models adjust for covariates to estimate conditional treatment effects.
• Model-based standardization generates covariate-adjusted marginal treatment effects by averaging predictions.
• Standard methods use maximum-likelihood estimation and bootstrap for standardization.

Purpose of the Study:

• Introduce a novel, general-purpose model-based standardization method using multiple imputation.
• Develop an approach termed Multiple Imputation Marginalization (MIM) for generalized linear models.
• Enable principled propagation of uncertainty and incorporation of prior evidence within a Bayesian framework.

Main Methods:

• MIM involves generating synthetic datasets and subsequent analysis.
• Utilizes a Bayesian statistical framework for uncertainty propagation.
• Applicable to generalized linear models for outcome analysis.

Main Results:

• Simulation study benchmarks MIM's finite-sample performance.
• MIM demonstrates unbiased estimation and valid coverage rates.
• Statistical performance is comparable to standard model-based standardization.

Conclusions:

• Multiple imputation can effectively marginalize over covariate distributions.
• MIM provides appropriate inference with correctly specified parametric models.
• Offers comparable statistical performance to existing standardization approaches.