Tree mortality models are important tools for simulating forest dynamic processes, and logistic regression is widely used for modeling tree mortality. However, most of the mortality models that have been developed generally ignore the hierarchical structure. In this study, Bayesian logistic multilevel mortality models were developed with the independent variables of initial the planting density, competition, site index and climate factors in Chinese fir plantations in southern China. The results showed that a Bayesian three-level model was best for describing tree mortality data with multiple sources of unobserved heterogeneity compared to fixed-effects and two-level models. The variance partition coefficient of tree mortality due to the tree level was much larger than that due to the plot level. The initial planting density and site index were positively correlated with mortality, and symmetric competition was negatively correlated. For climate variables, the mortality probability decreased with the increasing mean annual temperature and previous summer mean temperature. By contrast, the mortality probability increased with the increasing previous winter mean minimum temperature and annual heat-moisture index. Identifying different sources of variation in tree mortality will help further our understanding of the factors that drive tree mortality during climate change.