Social Sciences Brown Bag Seminar

Abstract: We characterize when discrete-choice datasets that involve aggregation, such as category-level items or an outside option, are consistent with a random utility model (RUM). The underlying alternatives that an aggregated category represents may differ across individuals and remain unobserved by the analyst. We characterize the observable implications of RUMs with unknown composition of aggregated categories and show that they consist of limited monotonicity of choice frequencies and standard RUM consistency on unaggregated menus. This characterization shows precisely how weak the RUM relative to the aggregated random utility model (ARUM) commonly assumed in empirical work. We identify two independent necessary and sufficient conditions that restore the implication of ARUM: non-overlapping preferences and menu-independent aggregation. Simulations show that violations of these conditions generate estimation bias, highlighting the practical importance of how data are aggregated.

With Yuexin Liao, and Alec Sandroni