The potential mechanisms that drive the behavior of subducted oceanic plates at the 660-km discontinuity are subject to debate. Here we conduct logistic regression analysis and model selection to determine the key subduction-zone parameters in natural subduction zones that discriminate the plate behavior along the discontinuity. We select the key variables based on three information criteria: leave-one-out cross-validation score (LOO), Akaike Information Criterion (AIC), and Bayesian Information Criterion (BIC). Among the 17 subduction-zone parameters analyzed, only the trench velocity, convergence rate, and trench width are selected in the simplest model that minimizes BIC. The thermal parameter and several other variables are also selected to minimize AIC and LOO. Our results suggest that a stagnant slab occurs along the 660-km discontinuity when there is a narrow oceanic plate and a retreating trench in natural subduction zones, which has also been modeled in previous numerical simulations. Neither the stress nor the deformation rate of the upper-plate margin is selected in the three optimal models, which suggests that back-arc spreading in natural subduction zones does not globally characterize plate behavior at the 660-km discontinuity, although back-arc spreading and a stagnant slab coincide in some numerical simulations. The combination of subduction-zone data analysis and numerical simulations will therefore provide deep insights into the dynamics of Earth’s deep interior.