The t linear mixed model with AR(p) dependence structure is proposed for the analysis of longitudinal data in which the underlying repeated measures contain thick tails and serial correlations simultaneously. For parameter estimation, I develop a hybrid maximization scheme that combines the stability of the Expectation Conditional Maximization Either (ECME) algorithm with the rapid convergence property of the scoring method. Empirical Bayes estimation of random effects and prediction of future values for the proposed model are also considered. The proposed methodologies are applied to a real example from a tumor growth study on twenty-two mice. Numerical comparisons indicate that the proposed model outperforms the normal model from both inferential and predictive perspectives.