Hypothesis-Testing In Multivariate Linear-Models With Randomly Missing Data

A common problem in multivariate general linear models is partially missing response data. The usual analysis method in the presence of missing data is listwise deletion. An approach is presented which allows hypothesis testing using all data which are observed. An EM algorithm was used for parameter estimation. Rao's F approximation for Wilks’ A with adjusted error degrees of freedom was evaluated using a Monte Carlo simulation. The resulting test statistic consistently yielded slightly conservative test sizes and substantially greater test powers than listwise deletion.