A comparison of full information maximum likelihood and multiple imputation in structural equation modeling with...

Area of Science:

• Statistics
• Psychometrics
• Data Analysis

Background:

• Missing data is a common challenge in statistical analyses.
• Full Information Maximum Likelihood (FIML) and Multiple Imputation (MI) are two prominent methods for handling missing data.
• Understanding their performance under various conditions is crucial for accurate research.

Purpose of the Study:

• To compare the performance of FIML and MI missing data procedures.
• To investigate their relative accuracy against complete data analyses.
• To identify discrepancies when the statistical model is misspecified.

Main Methods:

• Monte Carlo simulation studies were employed.
• Researchers had access to original complete data for comparison.
• Analyses varied sample size, missingness percentage, and model misfit.

Main Results:

• FIML and MI produced equivalent results to each other and complete data when the model was correctly specified.
• When the model was misspecified, MI parameter estimates, CFI, and TLI were closer to complete data estimates.
• FIML chi-squares and RMSEA were closer to complete data estimates under model misspecification.

Conclusions:

• The choice between FIML and MI can impact results, particularly with model misspecification.
• Discrepancies arise from the interplay between imputation model parsimony and accuracy.
• Further research is needed to explore practical and methodological implications.