Comparing multiple imputation methods for systematically missing subject-level data

Area of Science:

• Biostatistics
• Longitudinal Data Analysis
• Research Synthesis

Background:

• Research synthesis often involves combining studies with varying measured variables, leading to missing data.
• Missing data in longitudinal studies can occur at the observation-level (time-varying) or subject-level (non-time-varying).
• Existing methods primarily address missing observation-level data, leaving subject-level missingness under-addressed.

Purpose of the Study:

• To compare two multiple imputation approaches for handling missing subject-level variables in longitudinal research synthesis.
• To evaluate the performance of joint modeling versus sequential conditional modeling for imputing missing subject-level data.

Main Methods:

• The study compares a joint modeling approach with a sequential conditional modeling approach for multiple imputation.
• The focus is specifically on missing subject-level variables within longitudinal datasets for research synthesis.

Main Results:

• The joint modeling approach is generally preferable to the sequential conditional approach for handling missing subject-level data.
• The sequential conditional method can lead to attenuated and less efficient regression coefficient estimates compared to the joint method.
• In some cases, the sequential conditional method yields less efficient estimates than a complete case analysis, indicating a loss of efficiency in research synthesis.

Conclusions:

• The joint modeling approach is recommended for handling missing subject-level variables in longitudinal research synthesis, offering better efficiency and accuracy.
• The sequential conditional approach may be suitable only under specific conditions of homogenous variance and exchangeable correlation.
• Failure to use appropriate imputation methods can diminish the statistical power and reliability of findings from research syntheses.