Nested Logit Choice Trees And Correlated Substitution

Why Put Alternatives in a Tree?

A standard multinomial logit model treats the unobserved utility shocks of the available alternatives as independent in the particular way that yields the independence of irrelevant alternatives property. That structure can be restrictive when some alternatives are close substitutes. For example, removing one bus service might plausibly shift more demand to another bus service than to a private car. A nested logit model represents such structure by arranging alternatives into groups called nests.

The tree is not merely a convenient diagram. It encodes a claim about unobserved utility: alternatives in the same nest may have dependent utility shocks, while alternatives in different nests have a different dependence relationship. This structure can make substitution more concentrated among alternatives that share a nest. It does not, by itself, prove that the alternatives are similar or that the grouping is empirically appropriate.

Definition: A nested-logit choice tree is a hierarchical grouping of the alternatives in a choice set. In the basic two-level form, the root represents the choice occasion, its branches are nests, and each nest contains one or more alternatives. The nests form a partition: every available alternative belongs to exactly one nest. A nested logit model combines a systematic utility for each alternative with a joint distribution of utility shocks that permits structured dependence among alternatives in the same nest.

Utility and the Tree's Assumptions

Let \(C\) denote the available alternatives for a decision maker on a choice occasion, and let \(U_j\) be the utility of alternative \(j\). The random-utility specification is

U_j = V_j + \varepsilon_j, \qquad j \in C,

where \(V_j\) is the systematic component, often a function of observed attributes, and \(\varepsilon_j\) is the unobserved component. The decision maker is assumed to choose an available alternative with the highest utility:

j \text{ is chosen if } U_j \geq U_k \text{ for every } k \in C.

In nested logit, the shocks \(\varepsilon_j\) are not treated as independent across all alternatives. The model uses a generalized extreme-value error structure organized by the tree. Alternatives sharing a nest have a dependence structure that can represent common unobserved appeal; alternatives in different nests do not share that same within-nest structure. This is a structured probability model, not permission to specify any arbitrary set of pairwise correlations.

The phrase “correlated utility shocks” is useful shorthand, but the essential assumption concerns the joint distribution of the shocks. It is not generally equivalent to adding a normally distributed random intercept shared by the alternatives in a nest. That construction would specify a different random-utility model and would not, on its own, establish the nested-logit probabilities.

For a proposed tree to be interpretable, the choice set and the systematic utilities must also be defined coherently. An alternative that is unavailable on a choice occasion cannot be chosen there. Attributes entering \(V_j\) should be measured and coded for the alternatives to which they apply. The tree then describes how the unobserved components relate across the available alternatives; it does not replace the systematic utility specification.

Build the Tree From a Choice Set

A useful way to construct a tree is to begin with the actual alternatives, then ask which groups have a defensible behavioral interpretation. The grouping should reflect plausible similarities in unobserved utility and likely substitution, not merely labels that happen to be in the data. Each alternative must appear in one and only one nest in a standard nonoverlapping nested-logit tree.

A simple tree has two levels: a root and its nests. More elaborate hierarchical trees can have additional levels, but every branch must still lead unambiguously to an alternative. A singleton nest contains only one alternative. It is permitted in a choice tree, although it does not express within-nest substitution among multiple alternatives.

Worked Example: Grouping Commuting Alternatives

Suppose, hypothetically, a travel model offers five alternatives: local bus, express bus, metro, commuter rail, and car. The analyst argues that local and express bus share unobserved features, metro and commuter rail share other features, and car is distinct. A defensible initial tree is:

Root: Bus {local bus, express bus}; Rail {metro, commuter rail}; Car {car}.

This is a partition: each of the five alternatives appears once. For illustration, let the systematic utility be \(V_j=-0.3\,\text{fare}_j+0.5\,\text{comfort}_j\), where fare is in dollars and comfort is a hypothetical score. For a local bus with fare \(2\) dollars and comfort score \(2\),

V_{\text{local bus}} = -0.3(2) + 0.5(2) = -0.6 + 1.0 = 0.4.

For an express bus with fare \(3\) dollars and comfort score \(3\),

V_{\text{express bus}} = -0.3(3) + 0.5(3) = -0.9 + 1.5 = 0.6.

These calculations specify the observed-attribute part of utility; they do not determine the nests. The proposed bus grouping is a separate assumption about the relationship among unobserved utility shocks. Its behavioral implication is that the model can allow bus alternatives to be closer substitutes for one another than either is for car, subject to the complete model specification.

How Dependence Changes the Substitution Story

The source of the substitution distinction can be seen directly from the choice rule. The event that alternative \(j\) is chosen is

\{j \text{ chosen}\} = \{U_j \geq U_k \text{ for every } k \in C\}.

For alternatives in the same nest, these comparisons involve shocks whose joint distribution allows dependence. A change to one alternative can therefore affect the relative attractiveness of its nest in a way that is not captured by treating every alternative's shock as independent of every other. In practical terms, removing or making one option less attractive can lead to a relatively stronger shift toward its nest-mates than toward alternatives in other nests.

This is a qualitative implication, not a fixed numerical rule that says all lost choices must go to the nearest alternative. The exact choice probabilities and the size of any substitution response depend on the full model, including systematic utilities and the parameters governing the error structure. The tree determines where the model allows structured dependence; it does not dictate that all alternatives within a nest are equally attractive or equally substitutable.

This organization addresses one limitation associated with the independence of irrelevant alternatives property in multinomial logit. It does not remove every restriction. In particular, nested logit imposes a particular form of dependence based on the tree. If two alternatives are placed in different nests, the model does not treat their unobserved utilities as sharing the same within-nest relationship, even if the analyst believes they are close substitutes.

Worked Example: A Tree for Book Formats

Consider a hypothetical choice among an e-book, an audiobook, a paperback, and a hardcover. A researcher proposes two nests: Digital {e-book, audiobook} and Print {paperback, hardcover}. Each option belongs to exactly one nest, and the grouping has a clear behavioral interpretation: format-specific unobserved features may make choices within a format more alike.

Suppose the researcher uses the illustrative systematic utility \(V_j=-0.2\,\text{price}_j+0.4\,\text{convenience}_j\), with price in dollars and convenience on a hypothetical scale. For an e-book priced at \(10\) dollars with convenience score \(4\),

V_{\text{e-book}} = -0.2(10) + 0.4(4) = -2.0 + 1.6 = -0.4.

For an audiobook priced at \(12\) dollars with convenience score \(5\),

V_{\text{audiobook}} = -0.2(12) + 0.4(5) = -2.4 + 2.0 = -0.4.

The equal systematic utilities do not mean the two alternatives have identical total utility: their shocks may differ on a choice occasion. Nor does equal \(V_j\) establish that the alternatives should be in the same nest. The nest assignment rests on the behavioral claim about their unobserved features. If format is not a plausible source of shared unobserved appeal, the tree would need another justification.

Check Whether a Proposed Tree Is Coherent

A tree is a substantive specification, so it should be checked against the decision process and available alternatives. First, confirm that the alternatives are mutually exclusive outcomes of one choice occasion. Then verify that each alternative is assigned once, that the nests have behavioral meaning, and that the grouping is compatible with the substitution pattern the analyst wishes to represent.

Worked Example: Correcting an Overlapping Nest Assignment

A hypothetical food-delivery model includes four alternatives: restaurant pickup, restaurant delivery, grocery pickup, and grocery delivery. An analyst proposes a “restaurant” nest containing restaurant pickup and restaurant delivery, and a “delivery” nest containing restaurant delivery and grocery delivery. Restaurant delivery appears in both nests, so this is not a partition and does not define the standard nonoverlapping tree described here.

Suppose the analyst instead chooses the tree Restaurant {restaurant pickup, restaurant delivery} and Grocery {grocery pickup, grocery delivery}. The alternatives now appear once each, but this grouping asserts that shared restaurant-versus-grocery features are central to unobserved utility dependence. If the main hypothesized similarity is pickup versus delivery, a different partition—Pickup {restaurant pickup, grocery pickup} and Delivery {restaurant delivery, grocery delivery}—would represent that claim instead.

The two candidate trees answer different behavioral questions. The data description alone does not decide which one is right. The analyst must justify the grouping using the decision context and the anticipated substitution pattern; model comparison and diagnostics can then assess the specification. A tree cannot represent both overlapping classifications at the same level of a standard nested-logit partition.

Why the Tree Matters

Nesting matters because substitution is often an economic or behavioral claim, not just a mathematical feature. If two bus routes, two brands, or two service formats share important unobserved attributes, a model that accounts for that dependence may offer a more credible account of how choices respond when one alternative changes. A carefully defined tree makes the claim visible and interpretable.

At the same time, a tree can create false confidence if its branches are chosen for convenience. Assigning alternatives to the same nest does not demonstrate that they have correlated shocks. It encodes that assumption in the model. The usefulness of the structure depends on whether the proposed nests correspond to meaningful similarities and whether the imposed dependence is a reasonable approximation.

Key takeaway: A nested-logit tree partitions alternatives into behaviorally motivated nests. The model permits structured dependence among utility shocks within a nest, which can concentrate substitution among nest-mates relative to alternatives in other nests. The tree is a substantive assumption, not a data-free guarantee of realistic substitution.

Common Mistakes

Putting an alternative in multiple nests. A standard tree requires a partition: each alternative belongs to exactly one nest.
Treating the tree as a list of observed categories. A nest is a model assumption about utility-shock dependence and substitution, not simply a descriptive label.
Assuming any within-nest correlation can be specified. Nested logit imposes a particular structured joint error distribution; it is not a general model for arbitrary correlations.
Claiming that all substitution stays within a nest. Nesting can make within-nest substitution relatively stronger, but choices across nests remain possible and exact responses depend on the full model.
Confusing systematic utility with total utility. Observed attributes determine \(V_j\); the unobserved shock \(\varepsilon_j\) also contributes to \(U_j\).

Check Your Understanding

Use the definitions and assumptions above to answer the following questions.

What does it mean for nests to form a partition of the available alternatives?
Why does placing two alternatives in the same nest change the model's assumptions about their unobserved utilities?
A proposed tree groups two close substitutes in different nests. What behavioral feature of the model might this fail to represent?
Why do equal systematic utilities not imply that two alternatives have equal total utility on a choice occasion?
Can a standard nonoverlapping nested-logit tree put an alternative in both a product-type nest and a purchase-channel nest? Explain.