Abstract: This article takes the general principles and application values of the discrete choice model system as a departure point and summarizes the classical model forms with respect to their basic theories and typical applications. Important latest developments are also introduced. Multinomial logit (MNL) model is the basis of the discrete choice model system, with the advantages of simplicity, reliability, and easy implementation. However, it also has some inherent theoretic defects, which led to the need for more refined models. Nested logit model is usually used to deal with problems of correlation among alternatives, no-choice alternative, and data enrichment. Its more general form is the generalized extreme value (GEV) model system; mixed logit model is suitable for handling random preference and some kinds of correlation problems, such as correlation among alternatives, panel data, random coefficients, and data for enrichment. A similar model form named latent class model is also widely used. Multinomial probit (MNP) model is highly flexible. However, its application is limited due to the complexity of model specification and very high computation demands. With regard to the new development of discrete choice model system, four important areas are introduced. These include complex new models derived from the combination of classical models; models suitable for dealing with revealed preference/stated preference (RP/SP), ordered, ranked, and multiple choice data; models based on bounded rationality choice which is more close to reality; and models considering the spatiotemporal background of choice.

Key words: discrete choice model, refining, applicability, new trends

摘要： 本文从离散选择模型(discrete choice model, DCM)体系的一般原理和应用价值出发,总结了各经典模型的基本理论和典型应用,并概括了近来年一些重要的研究新动向。多项Logit模型(multinomial logit model, MNL)是离散选择模型体系的基础,具有简洁、可靠、易实现等优点,但也存在固有的理论缺陷,由此产生了对更加精细化模型的需求。替代的精细化模型中,嵌套Logit模型(nested logit model, NL)常用于处理备选项相关、“都不选”备选项、数据合并等问题,一般极值模型(generalized extreme value model, GEV)体系是其更一般的形式;混合Logit模型(mixed logit model, MXL)可用于解决随机偏好问题和多种相关问题,包括备选项相关、面版数据相关、随机系数相关、数据合并等,与之类似的潜在类别模型也有着广泛应用;多项Probit模型(multinomial probit model, MNP)具有极高的灵活性,但其复杂的模型设定与庞大的运算量大大制约了其应用范围。本文在研究新动向上介绍了4个重要的研究关注点：由多种经典模型形式相结合而成的复杂模型;面向RP/SP数据、定序、排序、多选等不同数据类型的适宜模型;基于各种受限理性选择策略的更为真实的模型;以及考虑选择的时空背景的模型。

关键词: 离散选择模型, 精细化, 适用性, 新动向