Mean and Covariance Structure Analysis with Missing Data

Most of the existing methods for missing data analysis are density based imputations. A serious drawback of these methods is that, when the observed data do not obey the assumed density, consequent inferences may not be reliable. In the context of mean and covariance structure analysis with missing data, use of a pseudo-maximum likelihood method has been proposed, but its properties have not been established for missing data. In this paper, a new method for mean and covariance structure analysis is developed to handle missing data under minimal assumptions. Under very modest assumptions, the estimator is asymptotically e cient. A test statistic is given for the overall t of a model; the test is asymptotically distribution free. Several nite sample versions of such a statistic are investigated through some limited simulation studies. As in the complete data case, the performance of these statistics varies for small to medium sample sizes. Unless sample size is very large, only the corrected statistics and the residual based statistics are recommended for use to evaluate the adequacy of a model.