Preference heterogeneity is one of the central behavioral concepts in applied econometrics. Its centrality is particularly evident in the choice modeling literature, notably in its widespread application to environmental and health economics, marketing, and transport. Despite conceptual and empirical advances in modeling preference heterogeneity, the generalizability of preference heterogeneity to different decision contexts and different data generation processes remains an open question. The basic premise of this paper is that latent sources of preference heterogeneity can be decomposed into components general to decision contexts and others specific to them. We study the structure of preference heterogeneity in different data generation processes with the goal of reliably identifying common (presumably generalizable) and specific (presumably not generalizable) sources of preference heterogeneity. The contribution of the paper is both conceptual and methodological, leading to the testing of five rival model specifications which together elucidate the heterogeneity structure present in two preference data sources of the same choice behavior. In the empirical application, we find that the multitrait-multimethod model of preference heterogeneity has the best fit and most sensible interpretations, indicating that while each data source contributes uniquely to certain heterogeneity components, both data sources contribute also to common (generalizable) preference heterogeneity. Recognition of the separability of the common versus source-specific preference heterogeneity will lead to more reliable and accurate demand model forecasts and assessments of welfare impacts.

偏好异质性是应用计量经济学中的核心行为概念之一。它的中心性在选择建模文献中尤为明显，特别是在其广泛应用于环境和健康经济学、营销和运输方面。尽管在建模偏好异质性方面取得了概念和经验上的进步，但偏好异质性对不同决策背景和不同数据生成过程的普遍性仍然是一个悬而未决的问题。本文的基本前提是，偏好异质性的潜在来源可以分解为决策上下文通用的组件和特定于它们的其他组件。我们研究了不同数据生成过程中偏好异质性的结构，目的是可靠地识别偏好异质性的共同（可能可概括）和特定（可能不可概括）来源。该论文的贡献是概念性和方法性的，导致对五个竞争模型规范的测试，这些规范共同阐明了相同选择行为的两个偏好数据源中存在的异质性结构。在实证应用中，我们发现偏好异质性的多特征-多方法模型具有最佳拟合和最合理的解释，表明虽然每个数据源对某些异质性成分都有独特的贡献，但两个数据源也对共同的（可推广的）偏好异质性有贡献.