Methods for mediation analysis with missing data

Area of Science:

• Statistics
• Psychometrics
• Quantitative Psychology

Background:

• Mediation analysis is widely used but formally addressing missing data is uncommon.
• Existing methods for missing data in mediation analysis lack comprehensive comparison.

Purpose of the Study:

• To introduce and compare four methods for handling missing data in mediation analysis.
• To evaluate the performance of these methods under different missing data mechanisms.
• To provide practical tools for mediation analysis with missing data.

Main Methods:

• Comparison of listwise deletion, pairwise deletion, multiple imputation (MI), and two-stage maximum likelihood (TS-ML).
• Development of the R package 'bmem' for implementing these methods within structural equation modeling.
• Simulation studies under Missing Completely At Random (MCAR), Missing At Random (MAR), and Missing Not At Random (MNAR) conditions.

Main Results:

• Multiple imputation (MI) and two-stage maximum likelihood (TS-ML) demonstrated strong performance for MCAR and MAR data.
• MI and TS-ML were effective for Auxiliary Variable Missing Not At Random (AV-MNAR) data when auxiliary variables were included.
• Listwise and pairwise deletion showed low statistical power and significant parameter estimation bias in many scenarios.

Conclusions:

• Multiple imputation (MI) and two-stage maximum likelihood (TS-ML) are robust methods for mediation analysis with missing data.
• The R package 'bmem' facilitates the application of these advanced techniques.
• While less powerful, deletion methods can offer insights into missing data mechanisms.