Prediction Model Performance With Different Imputation Strategies: A Simulation Study Using a North American ICU...

Area of Science:

• Biostatistics
• Data Science
• Medical Informatics

Background:

• Missing data is a common challenge in large observational studies.
• Data are often not missing at random, complicating analysis.
• Traditional methods for handling missing data can introduce bias.

Purpose of the Study:

• To evaluate pragmatic imputation methods for estimating model coefficients.
• To compare the performance of different missing data handling strategies.
• To assess methods under varying degrees of data missingness.

Main Methods:

• Simulations using a pediatric intensive care unit (PICU) registry dataset.
• Two logistic regression models were developed with age-specific criteria.
• Missingness was introduced for WBC count and mentation, dependent on other variables.
• Evaluated methods included complete case analysis, assuming missing as normal, and multiple imputation by chained equations.

Main Results:

• Multiple imputation by chained equations significantly outperformed traditional approaches.
• Root mean square error for model coefficients was lower with multiple imputation by chained equations in 91% of simulations compared to complete case analysis.
• Multiple imputation by chained equations was superior to assuming missing data are normal (98% of simulations).
• Assuming missing data as abnormal yielded the worst performance.

Conclusions:

• Multiple imputation by chained equations is an effective strategy for handling missing data.
• This method significantly outperforms traditional approaches, even with naive implementation.
• Researchers should consider multiple imputation by chained equations for studies with any proportion of missing data.