A comparison of imputation techniques for handling missing predictor values in a risk model with a binary outcome

Area of Science:

• Health Research
• Biostatistics
• Medical Informatics

Background:

• Risk models are crucial for predicting disease outcomes using routinely collected health data.
• Missing predictor values are common in health datasets, impacting model accuracy.
• Complete-case analysis, a frequent approach, can introduce bias and overfitting.

Purpose of the Study:

• To investigate the impact of various missing data imputation methods on risk model estimation.
• To evaluate the reliability of predictions generated from imputed datasets.
• To compare single and multiple imputation techniques in the context of risk modeling.

Main Methods:

• Simulated datasets were created from a large national cardiac surgery database.
• Methods evaluated included complete-case analysis, single imputation (e.g., conditional mean imputation), and multiple imputation (hot-decking, MICE).
• The effect on risk model estimation and prediction reliability was assessed.

Main Results:

• Complete-case analysis was found to produce unreliable risk predictions.
• Conditional mean imputation showed good performance but may be unsuitable with variable selection.
• Multiple Imputation by Chained Equations (MICE) demonstrated high-quality predictions, minimal bias, and good coverage.

Conclusions:

• Complete-case analysis should be avoided in risk modeling due to potential unreliability.
• Multiple imputation, particularly MICE, is a recommended method for handling missing data in risk model development.
• Accurate imputation is essential for reliable health risk predictions and unbiased estimates.