Random Utility & Discrete Choice

Random utility theory says that every choice has a part we can explain and a part we can't. Think of it like an iceberg—the visible part is what we observe about a person's preferences, but there's a huge hidden portion of factors only they know about: their mood, memories, or private constraints. Two people can make different yet equally rational choices because they're combining the same observable options with different unobservable factors.

Random utility models start from a simple premise: we never fully know why people choose what they choose. What we observe (the choice) combines with what we don't (the random component). This 'random' part isn't mathematical chaos—it's everything the decision-maker knows that we don't. Maybe you chose the coffee shop because it's closer (observable), or maybe it's because the barista remembered your name yesterday (unobservable to me).

The framework revolutionized how we predict discrete choices—decisions among distinct alternatives like commuting by car versus train, voting for candidate A versus B, or choosing Product X versus Y. Before this approach, economists struggled with inconsistent choices: why would someone pick the bus today but drive tomorrow when nothing changed? Random utility explains this by treating preferences as having both a stable component and a situation-specific component that varies from moment to moment.

This matters because it makes prediction possible even with incomplete information. Instead of demanding perfect knowledge of every factor influencing a decision, we model the predictable patterns while acknowledging the residual variation. Companies use it to forecast market share for new products, transportation planners use it to predict route changes, and policymakers use it to simulate how regulations will shift behavior. The 'random' element actually makes models more accurate by not forcing false precision on inherently human decisions.

Critically, the approach reframes 'rationality.' You're rational if your choices reflect consistent preferences, not if you agree with others or maximize some objective standard. The burger-chooser and salad-chooser can both be acting rationally if each is optimizing given their own constraints and preferences—many of which are invisible to outsiders.

Random utility theory provides the mathematical foundation for discrete choice analysis, treating utility as a random variable composed of a systematic (observable) component and a random (unobservable) component. The probability that a decision-maker chooses a particular alternative depends on the probability that this alternative yields the highest utility among all available options. McFadden (1974) proved that if the random components follow an extreme-value distribution, the choice probabilities take a closed-form logit structure—making the model tractable for empirical work. [1]

Key research findings include:

• McFadden (1974): Derived the conditional logit model from random utility maximization, establishing that discrete choice models are consistent with utility maximization under specific distributional assumptions. [1]
• Train (2009): Showed that mixed logit models can approximate any random utility model, recovering both preference heterogeneity and flexible substitution patterns across alternatives. [2]
• Ben-Akiva and Lerman (1985): Demonstrated that random utility models enable welfare analysis through compensating variation calculations even when choices are probabilistic. [3]

Random utility models face several critiques. Distributional assumptions—typically the Gumbel (extreme value) distribution—determine substitution patterns; logit models impose the 'independence of irrelevant alternatives' property, which can be unrealistic when options share unobserved attributes. The 'red bus/blue bus problem' illustrates this: adding a nearly identical alternative shouldn't draw proportionate share from all options. Mixed logit and nested structures relax this but add computational complexity. The models also struggle with very large choice sets (hundreds of options) and with endogenous attributes—when prices or qualities respond predictably to demand. Finally, revealed preference only shows what people chose, not what they would have chosen under different conditions, limiting causal inference without experimental variation.