Choice Sets, Utilities, And Multinomial Outcomes

From Outcomes to Choices

A multinomial outcome records one result from several possible categories. But a list of categories is not necessarily a list of options that a person could choose among. A response of “bus,” “train,” or “car,” for example, can represent a choice among travel modes; a response of “poor,” “fair,” or “good” instead records a rating on an ordered scale. The distinction matters because the interpretation of a model depends on what the observed categories mean.

A choice model starts by specifying which alternatives were available to each decision-maker. It then represents each available alternative by a utility: a latent measure of how desirable it is to that person in that situation. Utility is not usually observed directly. The observed choice can be modeled as the alternative with the greatest utility, while unobserved factors are represented by a random component.

Definition: For decision-maker \(i\), an alternative is one possible option, and the choice set \(C_i\) is the set of alternatives available to that person at the time of the decision. The observed choice \(y_i\) is the one alternative in \(C_i\) that the person selects. A multinomial choice outcome is unordered when its alternatives have no inherent ranking that defines their meaning.

The choice set is specific to the decision being studied. It may vary across people or occasions: one person may have access to a direct train while another does not. An alternative that was unavailable cannot be selected, even if it would have been attractive. Conversely, an outcome category should not be included as an available alternative merely because it appears in the data. A “none of these” option, for instance, belongs in the set only if it is genuinely available to the decision-maker.

Random Utility and the Choice Rule

A standard way to formalize choice is to write the utility that person \(i\) assigns to alternative \(j\) as the sum of a systematic component and an unobserved component. The systematic component may depend on observed characteristics of the person and the alternative; the random component collects relevant influences not represented by those observed variables.

U_{ij}=V_{ij}+\varepsilon_{ij}, \qquad j\in C_i

Here \(U_{ij}\) is total utility, \(V_{ij}\) is the modeled or systematic utility, and \(\varepsilon_{ij}\) is the unobserved utility component. The random-utility choice rule is that a person chooses an available alternative with the highest total utility:

y_i=\operatorname*{arg\,max}_{j\in C_i} U_{ij}

This is a behavioral assumption, not a direct observation of psychological quantities. Utility is latent, and its numerical scale is not intrinsically measured in dollars, minutes, or any other physical unit unless the model gives it a particular interpretation. In particular, the choice rule depends on comparisons between alternatives. Adding the same constant to every utility in a choice set leaves the maximizing alternative unchanged.

Implication of the choice rule: If alternative \(k\) is chosen, its utility must be at least as large as the utility of every other available alternative: \(U_{ik}\geq U_{ij}\) for all \(j\in C_i\). When utilities are distinct, the chosen alternative has strictly greater utility than each competitor.

This implication follows directly from maximization. If \(k\) is the maximizing alternative, then by definition \(U_{ik}\) is no smaller than any \(U_{ij}\) in the set. Substituting the random-utility decomposition gives a useful statement in terms of systematic and unobserved components:

V_{ik}+\varepsilon_{ik}\geq V_{ij}+\varepsilon_{ij} \quad\Longleftrightarrow\quad V_{ik}-V_{ij}\geq \varepsilon_{ij}-\varepsilon_{ik} \qquad (j\in C_i,\ j\ne k).

Thus, a choice does not establish that the chosen alternative has the largest systematic utility \(V_{ij}\) alone. Its unobserved component also matters. A model of choice probabilities requires assumptions about these unobserved components; the utility-maximization setup by itself does not supply a particular probability formula. The derivation of multinomial logit probabilities requires additional assumptions and is a separate step.

Worked Example: A Choice Set With Three Travel Modes

Suppose, hypothetically, that a commuter can choose among bus, train, and car. For this decision-maker, all three modes are available. A fitted systematic utility calculation gives \(V_{i,\text{bus}}=2.1\), \(V_{i,\text{train}}=2.5\), and \(V_{i,\text{car}}=1.8\). In this particular hypothetical realization, the unobserved components are \(\varepsilon_{i,\text{bus}}=0.3\), \(\varepsilon_{i,\text{train}}=-0.4\), and \(\varepsilon_{i,\text{car}}=0.6\). Find the choice implied by the random-utility rule.

Calculate total utility for each available alternative using \(U_{ij}=V_{ij}+\varepsilon_{ij}\):

U_{i,\text{bus}}=2.1+0.3=2.4,\qquad U_{i,\text{train}}=2.5-0.4=2.1,\qquad U_{i,\text{car}}=1.8+0.6=2.4.

The maximum total utility is \(2.4\), shared by bus and car. The deterministic utility is greatest for train, but its negative unobserved component makes its total utility lower in this realization. If ties are possible, the utility rule alone does not specify how to resolve this tie; a model may assume ties have probability zero or define a tie-breaking rule. For example, with a hypothetical car shock of \(0.7\), car utility would be \(1.8+0.7=2.5\), exceeding bus at \(2.4\) and train at \(2.1\), so car would be the unique choice. These calculations illustrate a realized choice rule, not a probability of choosing any mode.

Unordered Choice Versus Ordinal Outcomes

In an unordered choice problem, alternatives are distinct options, not points on a common scale. Choosing train rather than car does not mean choosing a “higher” or “lower” outcome. Relabeling the alternatives does not change their substantive meanings. The main modeling questions concern which options were available, their attributes, and how their utilities compare.

An ordinal outcome has a meaningful order, but the numerical spacing between its categories need not be meaningful. A rating of poor, fair, good, or excellent has a natural ranking. An ordinal response model typically represents an underlying propensity and category cutpoints, rather than treating each rating as a separate option in a choice set. A rating of “good” is not usually interpreted as the respondent choosing good instead of the available alternatives poor, fair, and excellent.

Worked Example: Choosing a Product or Reporting a Rating

Consider two hypothetical survey questions. One asks a customer to select a subscription plan: Basic, Standard, or Premium. The other asks the same customer to rate service as poor, fair, good, or excellent. Classify each response and explain what an observed category means.

For the plan question, suppose all three plans are available. The choice set is \(C_i=\{\text{Basic},\text{Standard},\text{Premium}\}\). If the customer selects Standard, the observed choice is \(y_i=\text{Standard}\). The outcome is unordered: Premium is not inherently a “larger” response than Basic in the sense of an ordinal measurement. In a random-utility formulation, the selection says that Standard’s utility is at least as high as the utilities of the other available plans.

For the rating question, the recorded response belongs to the ordered categories poor \(<\) fair \(<\) good \(<\) excellent. If the customer reports good, the observation places the response above fair and below excellent on the rating scale; it does not show that the customer selected “good” from a menu of competing rating options. The order is meaningful, but a numerical distance such as “excellent minus good equals good minus fair” is not implied. Treating this rating as unordered would discard its order; treating it as a choice among plans would give it the wrong behavioral interpretation.

Assumptions Needed for Interpretation

A multinomial choice model describes choices meaningfully only when the choice problem has been specified carefully. The following assumptions clarify what must be true for an interpretation based on random utility to be credible. They do not, by themselves, guarantee that a particular probability model is correctly specified or that an estimated effect is causal.

Alternatives are well defined and mutually exclusive. Each recorded decision corresponds to one alternative, rather than to a combination of categories or a ranking, unless the model explicitly represents combinations or rankings.
The choice set is correct and complete. It includes the alternatives genuinely available at the time of choice and excludes unavailable ones. If a meaningful “outside” option, such as not purchasing, was available but omitted, the modeled comparison does not represent the full decision.
The observed choice follows the stated decision rule, or the departures are modeled. The utility-maximization rule says the selected alternative has the greatest total utility. If decisions also reflect constraints, mistakes, habits, or a decision process not represented by utility maximization, those features may matter for interpretation.
The modeled systematic component is interpreted conditionally on the information used. Unobserved factors may be correlated with observed explanatory variables. A coefficient or contrast in systematic utility is not automatically a causal effect. Causal interpretation requires an appropriate design or additional assumptions that address confounding and selection.
The model’s assumptions about unobserved utility are stated separately. Different distributions or dependence patterns for \(\varepsilon_{ij}\) can produce different choice probabilities and substitution patterns. Random utility alone does not imply the multinomial logit formula, nor does it establish the independence-of-irrelevant-alternatives property.

Worked Example: Availability Changes the Meaning of a Choice

Suppose, hypothetically, a transport analyst records 20 commuters as choosing bus or train and models the choice set for every commuter as \(C_i=\{\text{bus},\text{train},\text{car}\}\). For one commuter, a planned car journey would have the greatest utility, but the commuter has no access to a car that day. What is wrong with the proposed choice set, and how should this commuter’s decision be represented?

The stated set includes car even though it is unavailable. Under the random-utility choice rule, the model would compare the utilities of bus, train, and car and could predict car as the selected alternative. But the actual feasible set for this commuter is \(C_i=\{\text{bus},\text{train}\}\), so car cannot be the observed choice. The comparison must be restricted to available options.

For a numerical illustration, suppose the commuter’s utilities are \(U_{i,\text{bus}}=3.2\), \(U_{i,\text{train}}=2.7\), and \(U_{i,\text{car}}=4.0\). If all three modes were available, the rule would select car because \(4.0>3.2>2.7\). With car unavailable, compare only bus and train: \(3.2>2.7\), so bus is the modeled choice. The observed bus choice is consistent with maximizing utility over the feasible set; it does not imply bus would be chosen over an unavailable car. If the dataset records only bus and train, it must still be established whether car was unavailable or whether a relevant alternative was omitted.

Why the Distinction Matters

The same set of category labels can describe different statistical objects depending on how they were produced. A response category may be an alternative selected from an available menu, an ordered rating, or a nominal classification assigned after an event. Those interpretations call for different models and support different conclusions. In particular, an unordered multinomial choice model does not automatically preserve a natural ranking, while an ordinal model does not automatically describe substitution among available alternatives.

Choice sets also shape the interpretation of comparisons. A model’s statements about choosing one alternative over another concern the alternatives available in the decision being analyzed. An omitted feasible option can change the behavioral question, and an incorrectly included unavailable option can make the modeled decision impossible. Finally, a random-utility representation is a framework for connecting observed choices to latent utilities; it is not direct measurement of preference and does not, without further assumptions, establish causation.

Common Mistake

Wrong reasoning: “There are four response categories, so this is automatically a four-alternative choice model. The selected category must have the greatest systematic utility.”
Correction: First determine whether the categories are genuinely available alternatives or ordered response levels. Then distinguish total utility \(U_{ij}\) from systematic utility \(V_{ij}\): a selected alternative maximizes total utility under the assumed choice rule, not necessarily the observed-variable component alone.

Key takeaway: A multinomial choice model begins with a well-defined set of available, mutually exclusive alternatives and an observed selection from that set. Random utility expresses the selection as maximization of total utility, while the unobserved components and the validity of the choice set govern what can be inferred from the observed outcome.

Check Your Understanding

Answer each question by identifying the decision or response being modeled and stating the relevant assumption.

A person chooses one of three available delivery services. What are the alternatives, the choice set, and the observed choice?
Explain why a poor-to-excellent service rating is ordinal rather than an unordered choice among four alternatives.
If an observed choice is alternative \(k\), write the utility inequality it must satisfy relative to another available alternative \(j\).
Why does random utility maximization not imply that the chosen option has the largest systematic utility?
A no-purchase option was available but omitted from a model of product choice. What part of the choice problem is misspecified?
Does a coefficient in systematic utility have a causal interpretation automatically? State one reason it may not.