A Bayesian Two-Step Multiple Imputation Approach Based on Mixed Models for Missing EMA Data.

Area of Science:

• Psychological Science
• Biostatistics
• Data Science

Background:

• Ecological Momentary Assessments (EMA) yield rich longitudinal data but often suffer from significant missingness.
• Missing data can compromise the robustness of statistical analyses in EMA studies.
• Multiple imputation is a key technique for handling missing data in EMA.

Purpose of the Study:

• Introduce a novel two-step Bayesian multiple imputation framework for EMA data.
• Compare the performance of three mixed models within this framework: Random Intercept Linear Mixed, Mixed-effect Location Scale (MELS), and Shared Parameter MELS.
• Evaluate the effectiveness of handling simultaneous missing variables in EMA data.

Main Methods:

• Developed a two-step Bayesian multiple imputation framework utilizing mixed models.
• Compared three imputation models: Random Intercept Linear Mixed, MELS, and Shared Parameter MELS.
• Conducted a simulation study to assess imputation effectiveness for simultaneous missing variables in EMA data.

Main Results:

• Multiple imputation significantly outperforms single imputation for EMA data.
• The choice of imputation model critically impacts analysis outcomes.
• MELS models, particularly when accounting for within-subject variance and linking missingness to the response, show improved performance in specific scenarios.
• Applied the framework to the "Make Better Choices 1 (MBC1)" study, demonstrating differences in imputation results between models.

Conclusions:

• The proposed Bayesian multiple imputation framework effectively addresses missing data in EMA.
• Selecting appropriate mixed models, especially those capturing within-subject variance (MELS), is vital for robust EMA analysis.
• The findings emphasize the importance of considering the missingness mechanism and its relationship with the response variable.