Missing data and prediction: the pattern submodel.

Area of Science:

• Statistics
• Machine Learning
• Data Science

Background:

• Missing data present significant challenges in developing and deploying predictive models.
• Existing methods for handling missing data often involve trade-offs between computational efficiency and predictive accuracy.

Purpose of the Study:

• To introduce and evaluate Pattern Submodels (PS) as a computationally efficient strategy for addressing missing data in prediction algorithms.
• To demonstrate the superior predictive performance of PS compared to traditional missing data handling techniques.

Main Methods:

• Developed Pattern Submodels (PS), where a unique submodel is fitted for each distinct missing data pattern.
• Conducted simulations and re-analyzed the SUPPORT study data to compare PS with imputation methods like zero-imputation, mean-imputation, complete-case analysis, and multiple imputation (MI).

Main Results:

• PS maintain predictive accuracy even when data are not missing at random (MAR).
• PS generally outperform other standard missing data strategies, including multiple imputation (MI), in terms of predictive accuracy.
• The performance improvement of PS is influenced by the missingness mechanism and the effect size of missing predictors.

Conclusions:

• Pattern Submodels (PS) provide a computationally efficient and highly predictive approach for handling missing data in prediction algorithms.
• PS offer a robust alternative to traditional imputation methods, particularly when dealing with complex missing data patterns and non-random missingness.