Missing data is a common occurrence in confirmatory factor analysis (CFA). Much work had evaluated the performance of different techniques when all observed variables were either continuous or ordinal. However, few have investigated these techniques when observed variables are a mix of continuous and ordinal variables. This study investigated the performance of four approaches to handling missing data in these models, a joint ordinal-continuous full information maximum likelihood (JOC-FIML) approach and three multiple imputation approaches (fully conditional specification, fully conditional specification with latent variable formulation, and expectation-maximization with bootstrapping) combined with the weighted least squares with mean and variance adjustment (WLSMV) estimator. In a Monte-Carlo simulation, the JOC-FIML approach produced unbiased estimations of factor loadings and standard errors in almost all conditions. Fully conditional specification combined with WLSMV was second best, producing accurate estimates if the sample size was large. We recommend JOC-FIML across most conditions, except when certain ordinal categories have extremely low frequencies as it was less likely to converge. If the sample is large, fully conditional specification combined with weighted-least-squares is recommended when the FIML approach is not feasible (e.g., non-convergence, variables that predict missingness are not of interest to the analysis).
Application of Multiple Imputation using the Two-Fold Fully Conditional Specification Algorithm in Longitudinal Clinical Data
