Multiple imputation approaches for handling incomplete three-level data with time-varying cluster-memberships

Area of Science:

• Biostatistics
• Longitudinal Data Analysis
• Medical Research Methodology

Background:

• Three-level, cross-classified data structures are prevalent in medical research, particularly with repeated measures on individuals nested within changing clusters over time.
• Multiple imputation (MI) is a common technique for handling missing data, but its application in complex cross-classified settings requires careful consideration of the imputation model.
• The performance of various MI approaches in three-level, cross-classified data remains underexplored, necessitating simulation studies to guide best practices.

Purpose of the Study:

• To compare the performance of different multiple imputation (MI) methods for analyzing three-level, cross-classified data with missing values.
• To evaluate MI strategies within joint modeling (JM) and fully conditional specification (FCS) frameworks in the context of longitudinal medical research.
• To identify robust imputation methods that accurately account for time-varying cluster memberships in cross-classified random effects models.

Main Methods:

• Conducted simulation studies based on a longitudinal cohort of students nested within schools to mimic a real-world cross-classified data scenario.
• Evaluated methods including ignoring time-varying clusters, JM extensions using dummy indicators (DI) or wide-format imputation, and a specialized three-level FCS MI approach.
• Assessed imputation performance based on bias and precision in the context of an acute-effects cross-classified random effects substantive model.

Main Results:

• Fully conditional specification (FCS) implementations of multiple imputation (MI) demonstrated superior performance, exhibiting minimal bias and good precision.
• Joint modeling (JM) approaches performed poorly, indicating potential issues when applied to this complex data structure.
• The use of dummy indicator (DI) extensions within both JM and FCS frameworks warrants caution, especially when dealing with sparse data.

Conclusions:

• Specialized three-level FCS MI methods are recommended for handling missing data in three-level, cross-classified longitudinal studies.
• Researchers should exercise caution when employing dummy indicator strategies for imputation in sparse cross-classified datasets.
• The findings provide crucial guidance for biostatisticians and medical researchers on selecting appropriate MI techniques for complex hierarchical data structures.