We describe an efficient estimation method for large-scale tree logit models, using a novel change-of-variables transformation that allows us to express the negative log-likelihood as a difference of strictly convex functions. Exploiting this representation, we design a fast iterative method...