A Sequential Foundation for Two-Level Nested Logit
Listen
Radio episode about this paper
Transcript
Introduction to the show: ident: Economics. The day in economics: theory, mechanisms and the evidence beneath them.
Ruth: I'm Ruth.
Warren: And I'm Warren, here with Iris.
Ruth: We're looking at the paper: "A Sequential Foundation for Two-Level Nested Logit".
Warren: Here's the gist of the paper. The two-level nested logit model arises exactly from a sequential choice process in which group, subgroup, and product shocks are revealed in stages.
Ruth: This result extends to any number of levels and holds for any nesting parameters below one, without the bounds and ordering that random utility maximization requires.
Warren: Sequential Choice Process Products are partitioned into subgroups within groups, and the utility of a consumer from product j is defined by equation (1).
Ruth: The process involves three stages where the consumer compares nests through their inclusive values, expected utility, net of shocks already observed, and optimal choices within each nest once taste shocks are revealed.
Title and authors: Ruth: So this paper is called "A Sequential Foundation for Two-Level Nested Logit". It sounds like it's trying to figure out the basic structure behind how we model choices in markets segmented along two dimensions.
Warren: Yeah, that’s what the title says. It suggests they are building a foundation for that nested logit model by looking at it through a different lens.
Ruth: Exactly. The authors are essentially showing that this whole structure of comparing groups, then subgroups, then products isn't just some arbitrary setup; it comes directly from a sequential choice process where you observe different shocks in stages.
Iris: So the core idea is to map the static nested logit model onto a sequence of choices where information—like product tastes—is revealed piece by piece as we move through groups and subgroups.
Ruth: Right, and they make it work for any number of levels and any nesting parameters below one. That’s a big deal because normally, when you use the standard random utility maximization framework, you have these strict rules about how those nesting parameters have to be ordered, like sigma two being smaller than sigma one.
Warren: Because if you don't follow that order, your estimates get discarded as just noise. This paper seems to suggest that we can ignore those strict ordering rules and still get consistent results for any parameter values below one.
Iris: It really is about showing that the sequential process doesn't impose those same bounds on how the parameters behave or are ordered.
The paper's summary: Ruth: So, in terms of what they actually did, this paper shows that the two-level nested logit model isn't just an abstract math construct; it arises naturally from a consumer making choices sequentially.
Warren: They describe a process where you first look at group shocks, then subgroup shocks, and finally product shocks as you decide what to buy. That’s the sequence they are following.
Iris: And at each stage of that sequence, the consumer compares options using three things: their inclusive values, their expected utility adjusted by all the shocks they've seen so far, and then figuring out the best choice within any nest once those specific product shocks are revealed.
Ruth: That’s a very concrete way to think about it. It moves away from just plugging numbers into a static equation and shows the mechanism that generates that equation in the first place.
Warren: The key result they establish is that this sequential process works for any number of levels, not just two, and it holds even when those nesting parameters aren't ordered in the traditional way we see them in simpler models.
Iris: And they show how the share of a product within a subgroup is determined by an exponential function related to those parameters. Specifically, the share of product j in subgroup h is given by that expression involving sigma one and sigma two, which they call s jhg = exp (delta j over one minus sigma one) exp Ihg over one minus sigma one.
Ruth: That formula shows how much influence the group level has on the subgroup choice, and how the product itself fits in once you’re down to that level. It links everything back to that sequential structure they described earlier.
The paper's improvements: Warren: So what are the actual improvements they propose over what we already had? They're focusing on showing how this sequential foundation fixes the problems with the old model setup.
Ruth: Well, one of the main things is that it validates parameter estimates that might have been discarded before because they violated those standard ordering constraints. They’re showing those estimates are actually consistent with a sequential choice process even if they don't fit the neat zero at most sigma two at most sigma one < one rule <ref:2610.38098#pg2>.
Iris: And they show that the linear inversion and the welfare formula hold up under these conditions, which is important because those formulas are how we actually go about estimating the parameters in practice.
Warren: That means we don't have to throw away whole sets of estimates just because a parameter doesn't fit the textbook ordering. It opens up more flexibility in model selection for people doing demand estimation.
Ruth: And they also look at what happens when you push those parameters outside that restricted region, like when sigma two is actually bigger than sigma one. They show that these values can still be consistent with random utility maximization at the data level, even if they aren't globally consistent across all possible data points.
Iris: They also point out a specific condition for cross effects within a subgroup to be positive. It depends on whether the share of product h in group g is small enough relative to some ratio involving sigma two and sigma one, which is R equals sigma two minus sigma one over one minus sigma one.
Ruth: So they’re not just showing that it works more broadly; they are giving us a condition—that R has to be big enough relative to the share of the group—for things like cross effects within a subgroup to show positive demand relationships.
Conclusion: Warren: To wrap up, this paper, "A Sequential Foundation for Two-Level Nested Logit", is basically showing that the two-level nested logit model arises exactly from a sequential choice process where shocks are revealed in stages.
Ruth: So the main implication is that we don't need to worry about those strict ordering rules for nesting parameters anymore if we follow this sequential structure. It lets us use a much wider range of parameter values below one for our analysis.
Iris: And they confirm that the linear inversion and the welfare formula remain valid under these conditions, which means our standard tools for estimating demand estimates are still reliable even when we relax those constraints.
Warren: They also give us a condition about cross effects within a subgroup, showing when you get positive results in those interactions based on how the group share compares to that ratio involving sigma two and sigma one.
Ruth: So, if you’re looking at demand estimation, this paper suggests that the way shocks are revealed matters more than forcing the parameters into a specific order to get a consistent result. It's about modeling the process itself.
Iris: It moves the focus from just finding a single "best" ordering to understanding how different sequential information flows through the model structure.
Warren: Yeah, this paper provides a more robust way to think about how these complex demand structures actually behave when you introduce multiple levels of segmentation and shocks.
Florida State University
econ.TH
Submitted: 2026-10-28
Updated: 2026-11-06
License: http://creativecommons.org/licenses/by/4.0/
Importance score: 79/100
The gist: The two-level nested logit model arises exactly from a sequential choice process in which group, subgroup, and product shocks are revealed in stages.
Key concepts
- Sequential Choice Process
- This describes a consumer's decision-making path where choices are made in stages. Shocks related to larger groups, then smaller subgroups, and finally individual products are observed sequentially, leading to the nested logit structure.
- Nested Logit Shares
- These are the probabilities that a consumer chooses a specific product within a particular subgroup. The model derives these shares based on assumptions about how consumers compare options across different levels of grouping.
- Generalization to K Levels
- The model can be extended to any number of product levels (K) by using specific formulas for shares and expected utility. This contrasts with static models that require a strict ordering of nesting parameters.
- Cross Effects within a Subgroup
- This refers to how the choice probability of one product changes when the utility of another product in the same subgroup is altered. The paper provides conditions, based on R, to determine when these cross effects are positive.
Terminology
Summary
The two-level nested logit model arises exactly from a sequential choice process in which group, subgroup, and product shocks are revealed in stages. p. 0 A Sequential Foundation for Two-Level Nested Logit
Davide Luparello
Department of Economics, Florida State University, 278 Bellamy Building,
113 Collegiate Loop, Tallahassee, FL 32306, USA
November 2, 2026
Abstract
I show that the two-level nested logit model arises exactly from a sequential choice
process in which group, subgroup, and product shocks are revealed in stages. The
result extends to any number of levels and holds for any nesting parameters below
one, without the bounds and ordering that random utility maximization requires. A Sequential Foundation for Two-Level Nested Logit
Davide Luparello
Department of Economics, Florida State University, 278 Bellamy Building,
113 Collegiate Loop, Tallahassee, FL 32306, USA
November 2, 2026
Abstract
I show that the two-level nested logit model arises exactly from a sequential choice
process in which group, subgroup, and product shocks are revealed in stages. The
result extends to any number of levels and holds for any nesting parameters below
one, without the bounds and ordering that random utility maximization requires.Where the paper says it
Where the paper says it
Sequential Choice Process
Products are partitioned into subgroups within groups, and the utility of a consumer from product j is defined by equation (1). The process involves three stages where the consumer compares nests through their inclusive values, expected utility, net of shocks already observed, and optimal choices within each nest once taste shocks are revealed. p. 1 1
In this paper, I show that the two-level nested logit share system and its linear inversion
arise exactly from a sequential choice process in which group, subgroup, and product
taste shocks are independent extreme value draws, revealed one stage at a time. At each
stage, the consumer compares nests through their inclusive values, the expected utility,
net of the shocks already observed, of choosing optimally within each nest once the taste
shocks within it are revealed. ASSUMPTION 2. The consumer observes the shocks {ζ
0
i
, ζ
g
i
}g∈G and chooses between the outside
option and a group g; then observes the subgroup shocks {ζ
h
i
}h∈Hg
and chooses a subgroup h;
then observes the product shocks {ϵij}
j∈Jhg
and chooses a product.Where the paper says it
Where the paper says it
Derivation of Nested Logit Shares
Proposition 1 establishes the two-level nested logit shares based on Assumptions 1 and 2 for any σ1, σ2 < 1. The share of product j in subgroup h is given by s jhg = exp (δj/(1 − σ1)) exp Ihg/(1 − σ1). p. 3 PROPOSITION 1. Under Assumptions 1 and 2, for any σ1
, σ2 < 1, the probability that consumer
i chooses product j ∈ Jhg is sj
= s
j|hg sh|g
sg, with
(3) s
j|hg =
exp
δj
/(1 − σ1
)
exp
Ihg/(1 − σ1
)
, sh|g
=
exp
Ihg/(1 − σ2)
exp Ig/(1 − σ2)
, sg =
exp(Ig)
1 + ∑g
′∈G exp(Ig
′)
. These are the two-level nested logit shares. Consequently,
(4) log s
j − log s0 = δj
+ σ1
log s
j|hg + σ2 log sh|g
,
and the consumer’s ex ante expected utility is W = log 1 + ∑g∈G exp(Ig)
.Where the paper says it
Where the paper says it
Generalization to K Levels
Proposition 2 shows that the choice probability of product j can be expressed for any number of levels K, with the share being s j = (Y+1)/K sum k=1 snk−1(j)nk(j). The ex ante expected utility is W = InK+1(j). p. 4 These are the K-level nested logit shares. Consequently,
(8) log s
j − log s0 = δj
+
X
K
k=1
σk
log snk−1
(j)|nk
(j)
,
and the consumer’s ex ante expected utility is W = InK+1(j)
. Applied level by level, the construction of Cardell (1997) requires 0 ≤ σK ≤ · · · ≤ σ1 < 1. The sequential process allows any ordering at any depth.Where the paper says it
Where the paper says it
Estimation and Interpretation
Existing estimates of σ1 and σ2 can be read as estimates of the sequential process even if they violate the standard ordering constraints. The condition for cross effects within a subgroup to be positive is given by Proposition 3, which depends on whether shg is small enough relative to R = (σ2 − σ1)/(1 − σ1). p. 5 Cross effects within a subgroup. Let β = (1 − σ1
)/(1 − σ2). Kovach and Tserenjigmid
(2022) read negative nesting parameters as complementarity among products, and Fosgerau, Monardo, and de Palma (2024) obtain complementarity in demand from groupings
that are not nested, with parameters inside the restricted parameter region. The sequential process generates positive cross effects within a subgroup with nested groupings and
single-unit demand.Where the paper says it
Where the paper says it
Green — the paper says it, and the page is printed. Red — no passage carries it, so treat that sentence as unverified. The same check runs on your own text: check a summary against its source, $5.
Improvements for AI systems
- Sequential Choice Modeling for Demand Estimation
The model can estimate parameters for any number of levels and any nesting parameters below one, without requiring them to satisfy a specific ordering like 0 <= sigma2 <= sigma1 < 1, as it arises exactly from a sequential choice process where shocks are revealed in stages.
- Welfare Formula Consistency
The system can utilize the standard welfare formula because "ex ante expected utility equals the nested logit inclusive value, so the standard welfare formula holds for any σ1, σ2 < 1."
- Interpretation of Parameter Estimates
The AI can interpret parameter estimates outside the restricted parameter region as consistent with sequential choice, and it can analyze cross effects within a subgroup using Proposition 3 to determine when raising one product's utility raises another's share.
Abstract
I show that the two-level nested logit model arises exactly from a sequential choice process in which group, subgroup, and product shocks are revealed in stages. The result extends to any number of levels and holds for any nesting parameters below one, without the bounds and ordering that random utility maximization requires. Parameter estimates that applied work discards for violating the standard ordering or the lower bound are consistent with sequential choice. The linear inversion and welfare formula hold. When the standard ordering fails, improving one product can raise a rival's market share in its subgroup, even with single-unit demand.