Scale Parameters And Random Utility Consistency In Nested Logit

Why the Nested-Logit Scale Parameter Matters

A central feature of nested logit is that alternatives within the same nest can share unobserved attributes. For example, two travel modes may be close substitutes because they share convenience or comfort characteristics that are not recorded in the data. The nested-logit dissimilarity parameter governs how strongly the model represents that within-nest dependence, and therefore how it allocates substitution across alternatives.

The parameter is not merely a numerical adjustment to the probabilities. Its range is restricted if the nested-logit probabilities are to come from a coherent random-utility model. Those restrictions depend on whether a nest contains multiple alternatives or only one. Understanding both the behavioral interpretation and the mathematical restriction is important before fitting or interpreting a nested-logit model.

Parameterization and Interpretation

Consider a choice set divided into mutually exclusive nests. Write \(V_{ij}\) for the systematic utility of alternative \(j\) for decision maker \(i\), and let \(m\) index nests. Alternative \(j\) belongs to nest \(m\), whose set of alternatives is \(B_m\). In the standard two-level nested logit, the top-level utility scale is normalized to one, and each nest has a positive dissimilarity parameter \(\lambda_m\).

The conditional probability of choosing \(j\), given that nest \(m\) is chosen, and the inclusive value for that nest are

P_{ij\mid m} = \frac{\exp(V_{ij}/\lambda_m)} {\sum_{k\in B_m}\exp(V_{ik}/\lambda_m)}, \qquad I_{im} = \log\left(\sum_{k\in B_m}\exp(V_{ik}/\lambda_m)\right).

The probability of choosing nest \(m\) is proportional to \(\exp(\lambda_m I_{im})\). Thus the unconditional probability of choosing \(j\in B_m\) is the product of the nest probability and the conditional probability:

P_{ij} = \frac{\exp(\lambda_m I_{im})} {\sum_n\exp(\lambda_n I_{in})} \cdot \frac{\exp(V_{ij}/\lambda_m)} {\sum_{k\in B_m}\exp(V_{ik}/\lambda_m)}.

Here \(\lambda_m\) is dimensionless: it is a scale parameter relative to the normalized top level. A smaller value means alternatives in that nest are represented as more similar in their unobserved utility components. It does not mean that the alternatives have lower systematic utilities. In a common normalized random-utility representation, the parameter is associated with within-nest error dependence; the precise correlation interpretation depends on the chosen error representation.

Definition: The dissimilarity parameter \(\lambda_m\) controls the scale of utility differences within nest \(m\), relative to the normalized top-level scale. In a multi-alternative nest, values nearer zero represent stronger within-nest dependence and more concentrated conditional probabilities. Random-utility consistency requires \(0<\lambda_m\leq1\). For a singleton nest, \(\lambda_m\) has no effect on the generating function or choice probabilities and is conventionally set to one.

How the Parameter Changes Choice Behavior

Within a nest, the conditional odds of two alternatives \(j\) and \(k\) have a particularly direct form:

\frac{P_{ij\mid m}}{P_{ik\mid m}} = \exp\left(\frac{V_{ij}-V_{ik}}{\lambda_m}\right).

A decrease in \(\lambda_m\) magnifies a given systematic-utility difference in these conditional odds. For a fixed set of utilities, the conditional probability distribution becomes more concentrated on the alternative with the largest \(V_{ij}\). At the same time, the model treats the alternatives as sharing more unobserved utility variation, so a change affecting one alternative tends to reallocate choice more strongly among members of its nest than among alternatives in different nests. These statements describe the model’s substitution structure; they do not establish a causal effect.

The parameter also enters the nest probability through \(\lambda_m I_{im}\). Consequently, its effect on unconditional probabilities is not captured by looking only at conditional probabilities. The conditional odds formula is useful for understanding within-nest behavior, while the full product probability is needed to assess overall choice probabilities.

Worked Example: Comparing Conditional Odds Within a Nest

Suppose a hypothetical nest contains alternatives A and B, with systematic utilities \(V_{iA}=1\) and \(V_{iB}=0\). Compare the conditional probabilities when \(\lambda_m=0.6\) and when \(\lambda_m=1\). Assume both alternatives belong to the same nest and use the standard nested-logit conditional probability.

For \(\lambda_m=0.6\), the conditional probability of A is

P_{iA\mid m} = \frac{\exp(1/0.6)}{\exp(1/0.6)+\exp(0/0.6)} = \frac{\exp(1/0.6)}{\exp(1/0.6)+1} \approx \frac{5.2945}{6.2945} \approx 0.8411.

Thus \(P_{iB\mid m}=1-0.8411=0.1589\). The conditional odds in favor of A are given directly by \(\exp((1-0)/0.6)\approx5.2945\).

When \(\lambda_m=1\), the corresponding calculation is

P_{iA\mid m} = \frac{\exp(1)}{\exp(1)+1} = \frac{2.7183}{3.7183} \approx 0.7311.

For this fixed utility difference, lowering the parameter from one to 0.6 raises A’s conditional probability from about 0.7311 to 0.8411. This illustrates the greater concentration of conditional probability when within-nest utility differences are scaled by a smaller \(\lambda_m\). It is not, by itself, a comparison of unconditional choice probabilities, because the probability of selecting the nest also matters.

Worked Example: Combining Conditional and Nest Probabilities

Consider a hypothetical choice set with nest \(m=\{A,B\}\) and a second, singleton nest containing C. Let \(V_{iA}=1\), \(V_{iB}=0\), and \(V_{iC}=0.5\). Set \(\lambda_m=0.6\) and use the conventional value \(\lambda=1\) for the singleton nest. These assumptions give positive parameters, an exhaustive choice set, and a multi-alternative nest within the consistency range.

For nest \(m\), the sum of exponentiated scaled utilities is \(\exp(1/0.6)+\exp(0/0.6)=5.2945+1=6.2945\). Its inclusive value is \(I_{im}=\log(6.2945)\approx1.8397\), so its nest-level weight is

\exp(\lambda_m I_{im}) = \exp(0.6\times1.8397) \approx \exp(1.1038) \approx 3.0156.

For the singleton nest, the weight is \(\exp(0.5)\approx1.6487\). Therefore,

P_{im} = \frac{3.0156}{3.0156+1.6487} \approx0.6465, \qquad P_{iC} = \frac{1.6487}{3.0156+1.6487} \approx0.3535.

Within nest \(m\), \(P_{iA\mid m}\approx0.8411\) and \(P_{iB\mid m}\approx0.1589\), as in the preceding calculation. Multiplication gives

P_{iA}\approx0.6465(0.8411)=0.5438, \qquad P_{iB}\approx0.6465(0.1589)=0.1027.

The three unconditional probabilities are approximately 0.5438, 0.1027, and 0.3535; they sum to 1.0000 after rounding. The calculation shows why a conditional probability alone does not describe overall choice: A’s probability is the probability of reaching its nest multiplied by its probability conditional on that nest.

Random-Utility Consistency Restrictions

The parameter restriction comes from the generalized extreme-value (GEV) generating function underlying the standard two-level nested logit. Write \(y_j=\exp(V_{ij})\), a positive quantity, and define

G(y) = \sum_m \left( \sum_{j\in B_m} y_j^{1/\lambda_m} \right)^{\lambda_m}.

A GEV generating function must be positive and homogeneous of degree one, and its mixed derivatives with respect to distinct alternatives must have alternating signs: a first derivative is nonnegative, a second mixed derivative is nonpositive, a third is nonnegative, and so on. These sign conditions are part of what makes the resulting choice probabilities consistent with a random-utility model.

For a nest containing at least two alternatives \(j\) and \(k\), let \(S_m=\sum_{r\in B_m}y_r^{1/\lambda_m}\). Differentiating the nest’s contribution twice with respect to the distinct alternatives gives

\frac{\partial^2 S_m^{\lambda_m}}{\partial y_j\,\partial y_k} = \frac{\lambda_m-1}{\lambda_m} S_m^{\lambda_m-2} y_j^{1/\lambda_m-1} y_k^{1/\lambda_m-1}.

All factors other than \((\lambda_m-1)/\lambda_m\) are positive. The required second mixed derivative is nonpositive, which, for a positive \(\lambda_m\), requires \(\lambda_m\leq1\). Positivity of the parameter is also required in the standard nested-logit scale parameterization. The bound is sufficient for the full nested-logit GEV sign conditions: for \(q\) distinct alternatives in one nest, the sign of the corresponding mixed derivative is the sign of \(\lambda_m(\lambda_m-1)\cdots(\lambda_m-q+1)\), which alternates as required when \(0<\lambda_m\leq1\). Mixed derivatives involving alternatives in different nests are zero because the generating function is a sum of separate nest terms.

Theorem: In the standard two-level nested-logit GEV model with top-level scale normalized to one, every nest containing at least two alternatives must have \(0<\lambda_m\leq1\) for random-utility consistency. A singleton nest contributes \(y_j\) to the generating function for every \(\lambda_m>0\), so its parameter is irrelevant and is conventionally set to one.

Worked Example: Checking the Sign Restriction and a Singleton Nest

Take a hypothetical nest with two alternatives, set \(y_j=y_k=1\), and consider \(\lambda_m=1.2\). Then \(S_m=2\), and the second mixed derivative is

\frac{\partial^2 S_m^{1.2}}{\partial y_j\,\partial y_k} = \frac{1.2-1}{1.2}\,2^{1.2-2} = \frac{1}{6}\,2^{-0.8} \approx0.09572.

This derivative is positive, contradicting the required nonpositive sign for distinct alternatives. Thus \(\lambda_m=1.2\) is inconsistent for this multi-alternative nest. In contrast, for \(\lambda_m=0.6\), the derivative is \((-0.4/0.6)2^{-1.4}\approx-0.2526\), which has the required second-derivative sign.

Now let a nest contain only alternative \(r\). Its generating-function contribution is

\left(y_r^{1/\lambda_m}\right)^{\lambda_m}=y_r \qquad\text{for every }\lambda_m>0.

Its conditional probability is one, its inclusive-value contribution to the nest weight is \(\exp(V_{ir})\), and its generating-function contribution is unchanged by \(\lambda_m\). The parameter therefore cannot affect the choice probabilities for a singleton nest; setting it to one is a convention, not a restriction imposed by the sign argument for multi-alternative nests.

What the Boundary Values Mean

At \(\lambda_m=1\), the conditional probabilities in nest \(m\) have the usual multinomial-logit form, and the nest’s contribution to the generating function is \(\sum_{j\in B_m}y_j\). If every nest has \(\lambda_m=1\), the model reduces to multinomial logit. In this sense, the nested structure adds no within-nest dependence beyond the multinomial-logit specification at that boundary.

As \(\lambda_m\) decreases toward zero, the model represents stronger within-nest dependence and a more concentrated conditional distribution for fixed utility differences. Zero itself is not an admissible parameter value: the expressions \(V_{ij}/\lambda_m\) are undefined there, and the standard model requires a positive scale. A value near zero can also make predictions very sensitive to utility differences, so it should not be described casually as just another ordinary value in the admissible range.

Common Mistakes

Applying the upper bound to singleton nests. The restriction \(\lambda_m\leq1\) follows from cross-derivatives between distinct alternatives in a nest. A singleton has no such pair, and its generating-function term is \(y_j\) for any positive \(\lambda_m\).
Calling a lower parameter a lower utility. The parameter rescales utility differences within a nest; it does not lower the systematic utility of every alternative in that nest.
Equating conditional and unconditional choice probabilities. The conditional probability is multiplied by the probability of choosing the nest. Omitting that factor can give an incorrect interpretation of a model prediction.
Claiming that nested logit proves alternatives are causally related. The parameter describes a model’s representation of unobserved dependence and substitution. It does not establish a causal relationship between choices.
Reporting a parameter without its normalization. The interpretation is relative to the top-level scale, which is normalized to one here. The value should not be treated as an absolute scale independent of the model parameterization.

Key takeaway: In a multi-alternative nest, \(0<\lambda_m\leq1\) is the random-utility-consistent range under the standard top-level normalization. Smaller values represent stronger within-nest dependence and affect both conditional choice odds and the nest-level probability. For singleton nests, \(\lambda_m\) is irrelevant and is conventionally fixed at one.

Check Your Understanding

Answer each question using the standard two-level nested-logit parameterization.

For two alternatives in the same nest, write their conditional odds in terms of their systematic utilities and \(\lambda_m\).
Why does a smaller \(\lambda_m\) represent stronger within-nest dependence, and why is that interpretation not the same as saying utilities are lower?
What sign must the second mixed derivative of a valid GEV generating function have for two distinct alternatives? How does this imply the upper bound for a multi-alternative nest?
Explain why a singleton nest’s \(\lambda_m\) does not affect its generating-function contribution or choice probability.
If a conditional choice probability is 0.8 and the probability of its nest is 0.5, what is the corresponding unconditional probability?