Estimation of structural equation modeling with missing data using full information maximum likelihood method: a simulation study

Structural equation modeling (SEM) is a powerful multivariate statistical approach for assessing complex relationships between latent variables in many human and behavioral sciences. A common challenge in estimating structural equation models, which is based on hypothesis testing, is the presence of missing data. Deleting subjects with missing values on each of items is the usual way of handling missing data, which leads to biased estimators and lose a considerable amount of sample information as the percentage of missing values increases. In estimating SEM with missing values, one can apply the full information maximum likelihood (FIML) approach that makes maximal use of all available data from every subject in the sample. In this paper, the performance of FIML is investigated under three missing value mechanisms, missing completely at random, missing at random, and missing not at random, in a simulation study. Two confirmatory factor analysis models are considered, where the data is generated under three mechanisms and the impact of two indexes, sample size (100, 500) and percentage of missing values (2%, 5%, 10%, 15%, 20%, 25%, 30%, 35%, 40%), are evaluated based on the root mean square error of approximation (RMSEA) index. Results show that the performance of SEM using FIML approach is generally better than the performance of SEM without using this approach in terms of some goodness of fit index.