Compatibility in Missing Data Handling Across the Prediction Model Pipeline: A Simulation Study

Area of Science:

• Biostatistics
• Health Informatics
• Clinical Epidemiology

Background:

• Clinical prediction models require rigorous handling of missing data for reliable performance.
• Inconsistent methods for missing data can introduce bias during model validation and implementation.

Purpose of the Study:

• To evaluate bias in predictive performance estimation due to various missing data handling approaches.
• To identify compatible strategies for managing missing data throughout the clinical prediction model pipeline.

Main Methods:

• Assessed bias in predictive performance across different missing data handling techniques.
• Examined strategy compatibility between model validation and implementation phases.

Main Results:

• Identified four key strategies suitable for the entire model pipeline.
• Quantified the bias introduced by different missing data handling combinations.

Conclusions:

• Recommends specific strategies for handling missing data between model validation and implementation.
• Provides guidance based on different missingness mechanisms to ensure model robustness.