Evaluating FIML and multiple imputation in joint ordinal-continuous measurements models with missing data.

Area of Science:

• Statistics
• Psychometrics
• Data Analysis

Background:

• Missing data is prevalent in statistical analyses, particularly in confirmatory factor analysis (CFA).
• Existing research primarily addresses missing data in CFA with exclusively continuous or ordinal variables.
• Limited investigation exists for handling missing data in CFA models featuring a mix of continuous and ordinal observed variables.

Purpose of the Study:

• To evaluate the performance of four distinct methods for managing missing data within CFA models containing mixed observed variable types.
• To compare a joint ordinal-continuous full information maximum likelihood (FIML) approach against three multiple imputation techniques when used with the weighted least squares with mean and variance adjustment (WLSMV) estimator.

Main Methods:

• A Monte Carlo simulation study was conducted to assess the performance of the selected missing data handling techniques.
• Four approaches were investigated: joint ordinal-continuous FIML and three multiple imputation methods (fully conditional specification, latent variable formulation, expectation-maximization with bootstrapping).
• These imputation methods were combined with the weighted least squares with mean and variance adjustment (WLSMV) estimator.

Main Results:

• The joint ordinal-continuous FIML approach demonstrated unbiased estimation of factor loadings and standard errors across most simulated conditions.
• Fully conditional specification coupled with WLSMV performed well, yielding accurate estimates particularly with larger sample sizes.
• FIML exhibited minor non-convergence issues with skewed data, characterized by very low frequencies in certain ordinal categories.

Conclusions:

• The joint ordinal-continuous FIML method is generally recommended for handling missing data in CFA with mixed variable types due to its unbiased estimates.
• Fully conditional specification with WLSMV serves as a robust alternative when FIML is computationally infeasible or encounters convergence problems.
• Consider sample size and potential non-convergence issues when selecting a missing data strategy for mixed-type CFA models.