Measurement Model Specification Error in LISREL Structural Equation Models

Abstract

This LISREL study examines the robustness of the maximum likelihood estimates under varying degrees of measurement model misspecification. A true model containing five latent variables (two endogenous and three exogenous) and two indicator variables per latent variable was used. Measurement model misspecification considered included errors of omission, errors of inclusion, and simultaneous errors of omission and inclusion. A sample size of 200 was used for all replications. LISREL programs were written for each of six models to be tested. The PC version of LISREL VII was used to estimate the models. Assessments of results were based on average parameter estimates for each model across replications, average chi-square value across replications, modification indices for errors of omission, and t-values for errors of inclusion. Results indicate that minor misspecifications in LISREL measurement models in the LY matrix are the most problematic, particularly for compound errors. In general, the ability of the LISREL program to detect measurement model misspecification is quite good. One figure illustrating the true population model and six data tables are included. (TJH)