Advanced variance-covariance structures are commonly used in genetic evaluation of crops to account for micro-site variability and achieve higher accuracy of predictions to increase selection efficiency. Various genetic variance-covariance structures were explored to predict best linear unbiased genetic merits of 453 loblolly pine (Pinus taeda L.) cloned progeny tested at 16 different locations in the southern U.S. Statistical models were compared using model fit statistics, variance components and genetic parameters. Among the models explored, spatial autoregressive error correlation with independent residual term for the R side with a factor analytic structure for the G side of the mixed model was superior. The model produced one of the smallest fit statistics (LogL equal to -2694), a small error variance (12.72), and the highest broad-sense heritability (0.45), compared with the default homogeneous error and genetic variance-covariance structure (statistical significance at P < 0.05). We concluded that the combination of specific structure for error and genetic design was effective to remove spatial-related variance, and to increase the accuracy of predictions of clonal genetic values, which could be used as analytical tool for increasing the selection efficiencies in forest genetic trials.