Evaluating Missing Data Imputation Strategies for Environmental Mixture Models: A Simulation Study and Applied...

Area of Science:

• Environmental Epidemiology
• Statistical Modeling
• Data Science

Background:

• Missing data are prevalent in environmental mixture studies, potentially biasing results.
• Variable selection in mixture models is sensitive to how missing data are handled.

Purpose of the Study:

• To evaluate the impact of different imputation strategies on variable selection performance across six mixture modeling frameworks.
• To compare single imputation (SI) and multiple imputation (MI) methods under various missing data scenarios (MAR, MNAR).

Main Methods:

• Monte Carlo simulations generated data with missing exposure and outcome variables under MAR and MNAR mechanisms.
• Six mixture modeling frameworks were assessed: WQS, BWQS, Q-gcomp, BKMR, Elastic Net, and LASSO.
• Imputation methods included mean, median, KNN (SI), MICE, and Amelia (MI). Performance was evaluated using sensitivity, specificity, and FDR.

Main Results:

• MI consistently outperformed SI and listwise deletion under MAR, improving variable selection accuracy.
• All methods showed reduced performance under MNAR, with flexible models being more unstable.
• Quantile g-computation (Q-gcomp) offered a robust balance of performance metrics and resilience to MAR assumption violations.
• Proper imputation model specification, including relevant covariates, enhanced stability, especially for flexible models and binary outcomes.

Conclusions:

• The choice and specification of imputation strategy significantly impact variable selection in environmental mixture models.
• Multiple imputation is recommended for handling missing data in these studies, particularly under MAR.
• Quantile g-computation and Bayesian WQS demonstrated robust performance in simulation and real-world data analysis (NHANES).