Date of Award


Document Type


Degree Name

Doctor of Philosophy (PhD)


Researchers using the LISREL (linear structural relations) program developed by Joreskog and Sorbom (1984) for the purpose of estimating parameters of structural equation models must make important assumptions. The researcher must assume large sample size (asymptotic theory) and a multivariate normal distribution for the indicator variables. In addition, theoretical assumptions as to the appropriateness of the model specification must also be made. The Monte Carlo study examines the robustness of maximum likelihood parameter estimates under varying degrees of model misspecification. The effects of errors of omission, errors of inclusion, and simultaneous errors of omission and inclusion are studied for sample sizes of 100 and 200. A true population model containing four latent variables and eight continuous indicator variables was developed. From this true model, a population covariance matrix was derived, and sample matrices were generated by use of a FORTRAN program. The true model was modified and tested for each type of misspecification considered. Parameter estimates, standard errors, and the (chi)('2) statistic were averaged over a minimum of 300 replications. Results indicate that certain types of specification errors are more serious in terms of parameter bias and/or model fit. Simultaneous errors are in general more problemmatic than single errors. Nonconvergence problems are related to the type of error made in relation to the overall pattern of the model. Sample sizes of 200 (in contrast to N = 100) particularly affect the estimates of parameter standard errors. The chance of rejecting a misspecified model improves when the sample size is increased. However, the ability in general of the LISREL program to detect misspecifications is limited.