We propose a methodology to estimate missing information in mixed cell means models. This methodology improves on that Melo & Melo (2005), which is based on the methods of maximum likelihood estimation and covariate proposed by Bartlett (1937), and reduces the correlation between the observed and estimated information. Once the imputation of the missing information is done, we suggest a way to perform the analysis of variance in models without interaction, by generating a weighted test for the fixed and random effects involved in the model.