Imputation and Missing Indicators for Handling Missing Longitudinal Data: Data Simulation Analysis Based on...

Area of Science:

• Statistics
• Biostatistics
• Health Informatics

Background:

• Missing data in electronic health records (EHRs) presents analytical challenges, including bias and reduced statistical power.
• The missing indicator method, treating missingness as a category, is a simple approach for handling unknown covariate values.
• Its utility in longitudinal data, where repeated measures can be leveraged, remains unclear compared to cross-sectional analyses.

Purpose of the Study:

• To evaluate the missing indicator method's impact on model performance and imputation accuracy for longitudinal data.
• To assess its effectiveness in a simulation study mimicking EHR-based clinical prediction models for falls in older adults.

Main Methods:

• Simulated longitudinal binary outcomes using mixed-effects logistic regression.
• Incorporated time-invariant and dynamic predictors, inducing missing data under random and non-random scenarios.
• Evaluated model performance using Area Under the Receiver Operating Characteristic Curve (AUROC) and imputation quality via normalized root-mean-square error and percent falsely classified.

Main Results:

• Model performance (AUROC) and imputation accuracy were similar with or without missing indicators, irrespective of the missing data mechanism (random or non-random).
• Inclusion of missing indicators did not affect performance or imputation quality, even when the outcome was related to missingness.
• Imputation methods showed comparable AUROC means to complete case analysis, though complete case analysis had greater variability.

Conclusions:

• The missing indicator method neither enhances nor diminishes performance or imputation accuracy in longitudinal data modeling.
• Further research is warranted to explore its utility in prediction modeling with longitudinal data, particularly in high-dimensional settings.