Abstract: Based on the branch analysis data of 2 598 branches from 40 sample trees of Chinese fir plantation in Jiangle National Forest Farm in Fujian Province. The linear mixed effects(LME)models was used build primary branch size(diameter and length)LME models with the trees effect as the random effect by R language. The results showed that the LME models provide better performance than that of the traditional multiple linear regression models for the branch size prediction of Chinese fir plantation. The LME models with different combinations of the random effects parameters had different fitting precisions, among which the LME model with three random effects parameters provides the best performance. However, the LME models with more than three random effects parameters showed no convergence. The LME models including variance structures could effectively remove the heteroscedasticity in the data, among which the LME model with the power function as the variance structure has better fitting precisions. Model validation confirms that the LME models with the random effect and heteroscedasticity structure could significantly improve the precision of prediction, compared with the traditional regression models.