The study utilized IRT Mixture Model to investigate latent heterogeneity of school learning and family relationships to differentiate latent classes and disturbance severity of middle school students. Four statistical models were examined,  including IRT, LCA, IRT Mixture （parameters constrained） and IRT Mixture （parameters non-constrained）. Results show that two-class two-factor IRT mixture model with constrained parameters provides the best fit of our data. Two factors are school learning and family relationships, and the two classes are high-risk and normal groups. Estimations of IRT Mixture Model are comparable to those of IRT and LCA. In conclusion, mixture models demonstrate more modeling flexibility compared to those of traditional statistical models, but they require larger sample size, longer computer running hours, and more difficulties in reaching algorithm convergence.