A staged-choice explanation for nested logit’s unusual estimates

Economists use demand models to describe how consumers divide their purchases among competing products. In nested logit, products sit inside subgroups, which sit inside broader groups. The hierarchy lets the model represent different patterns of competition within and between these categories. But estimated parameters sometimes violate the restrictions required by its conventional interpretation.

In A Sequential Foundation for Two-Level Nested Logit, Davide Luparello shows that the same choice probabilities can arise from a different decision process: consumers learn their preferences one stage at a time. This construction accommodates a wider range of parameters while preserving familiar estimation and expected-utility formulas. The paper is an arXiv preprint; the version examined, 2610.38098v3, was revised on November 6, 2026.

Choosing before everything is known

The conventional static interpretation treats consumers as choosing with their relevant preference shocks already known. A shock represents an individual taste component beyond a product’s mean utility. Global consistency with this random-utility interpretation requires the two nesting parameters to be nonnegative and ordered: the group parameter cannot exceed the subgroup parameter, and both must remain below one.

Luparello instead specifies three successive decisions. First, consumers observe group-level shocks and choose a group or the outside option. Within their chosen group, they observe subgroup shocks and choose a subgroup. Only then do they observe product-specific shocks and select a product.

At earlier stages, consumers do not know which later shocks they will receive. They compare the expected utility of their future choices, given the information available. The paper calls these expected values “inclusive values”: measures of how attractive a collection of options is before the remaining preferences are revealed.

The assumptions are specific. Shocks are mutually independent draws from a mean-zero extreme-value distribution with unit scale. Product and subgroup shocks are multiplied by positive scale factors determined by the nesting parameters. Each parameter must therefore remain below one, but neither needs to be nonnegative or follow the conventional ordering.

A proof, not a market experiment

The method is backward induction. Luparello first solves the final product choice, then uses its expected value to solve the subgroup decision, and finally solves the group decision. A standard mathematical property of extreme-value shocks supplies both the choice probabilities and expected maximum utility at each step.

Multiplying the three conditional choice probabilities gives exactly the two-level nested logit market shares. The familiar linear inversion also survives: it relates a product’s log share relative to the outside option to mean utility and the relevant conditional log shares. Researchers need not change that estimating equation to interpret estimates through the sequential model.

The proof extends to any finite number of nesting levels. Expected utility is calculated recursively from smaller nests to larger ones, while consumers make decisions in the opposite direction as information arrives.

The usual inclusive-value expression for ex ante expected utility also holds throughout the expanded parameter region. That preserves the model’s welfare formula under the sequential assumptions; it does not establish that those assumptions describe a particular market. The paper reports no empirical dataset or market-level test.

When an improvement helps a rival

One implication concerns two distinct products in the same subgroup. Improving one product increases that subgroup’s expected attractiveness, drawing consumers toward it before they know their product-specific tastes. Some newcomers subsequently discover that they prefer the rival.

This creates two opposing effects: the improved product diverts choices away from its rival within the subgroup, but the subgroup attracts additional consumers. The rival’s total market share rises if the inflow outweighs the diversion.

When the group nesting parameter exceeds the subgroup parameter, the paper proves that a positive effect occurs if the subgroup’s conditional share is sufficiently small, provided the group contains at least two subgroups. Positive effects can also occur under certain negative parameter values. They do not occur within the conventional restricted parameter region. Consumers still select only one option: the result is not evidence that they buy the products together.

What the interpretation can establish

Within the conventional parameter region, static and sequential models generate identical final choice probabilities. Final purchases alone therefore cannot distinguish their explanations; evidence about when consumers learn their preferences could help.

Outside that region, the contribution is a mathematically coherent interpretation, not automatic validation of an estimate. Its relevance depends on the assumed information timing, shock distributions and decision hierarchy.

This report uses the supplied exact-revision PDF text extraction. Extraction may omit figures and damage mathematical layout; it is not a complete visual inspection of the PDF.