A generalization of the multinomial logit (MNL) model is developed in which discrete alternatives are ordered so as to induce stochastic correlation among alternatives in close proximity. It is designed for situations where the alternatives are ordered, but is more flexible than previous ordered models. The model belongs to the Generalized Extreme Value class introduced by McFadden,and is therefore consistent with random utility maximization. A straightforward extension can handle cases where observations have been selected on the basis of a truncated choice set. A two-stage procedure using MNL computer software provides a specification test for MNL against either of these alternative models. The models' properties are investigated through two empirical applications whose rather unsatisfactory results are very briefly described.