Maximum likelihood is an important method for covariance structure analysis in multilevel modeling. Existing algorithms for maximum likelihood estimation in two-level structural equation models require some undesirable restrictions on the sample design or on the model itself. An EM algorithm is developed to overcome some pitfalls of the existing algorithms. Asymptotic behavior of the maximum likelihood estimator and the chi-square test for model fit are studied. A simulartion study is performed to demonstrate the behavior of the EM algorithm. A comparison based on the analysis for a real data set by the proposed algorithm and by an existing software shows that the real EM algorithm can be used in application with confidence.