The impact of nonnormality on full information maximum-likelihood estimation for structural equation models with...

Area of Science:

• Statistics
• Quantitative Psychology
• Econometrics

Background:

• Structural equation models (SEMs) often encounter nonnormal indicator variables and missing data.
• Full information maximum-likelihood estimation (FIML) is a common method for handling missing data in SEMs.
• Ad hoc missing data techniques may introduce bias and reduce efficiency.

Purpose of the Study:

• To evaluate the performance of FIML estimation in SEMs with nonnormal indicator variables.
• To compare FIML with ad hoc missing data techniques under various conditions.
• To assess the impact of missing data patterns, rates, sample size, and distribution shape on estimation.

Main Methods:

• A Monte Carlo simulation was employed to examine FIML estimation.
• Independent variables included missing data algorithm, missing data rate, sample size, and distribution shape.
• Outcome measures comprised parameter estimate bias, parameter estimate efficiency, standard error coverage, and model rejection rates.

Main Results:

• FIML parameter estimates demonstrated less bias and greater efficiency compared to ad hoc methods for missing completely at random and missing at random data.
• Standard errors were negatively biased, and model rejection rates were inflated, similar to complete-data maximum-likelihood estimation.
• Recently developed correctives for missing data showed potential in mitigating issues arising from nonnormal data.

Conclusions:

• FIML offers advantages in parameter estimation accuracy and efficiency for SEMs with nonnormal data and missing values.
• Inflation in standard errors and model rejection rates remain concerns with FIML.
• Corrective methods like rescaled statistics and bootstrapping show promise for improving FIML performance in challenging data conditions.