How does it differ from multinomial logit?

It uses joint normal latent errors and can specify their dependence, whereas the standard logit model has a different error structure and associated restrictions.

Why are probabilities harder to calculate?

They involve joint inequalities among correlated normal utility differences rather than the usual softmax formula.

Does it require ordered categories?

No. It is designed for unordered alternatives; ordered probit is a different model.

AI Glossary / Statistics and probability / Advanced

Multinomial probit

Definition

Multinomial probit is a statistical model for an outcome that takes one of several unordered categories, often interpreted as a choice among alternatives. It represents each alternative with an unobserved utility and selects the alternative with the largest utility. The random utility components follow a joint normal distribution. Their covariance structure can represent similarities or shared unobserved influences among alternatives. This distinguishes multinomial probit from the standard multinomial logit formulation. The additional flexibility comes with identification requirements and typically more demanding probability calculations and estimation.

What it means

The observed category reveals which alternative had the greatest utility, not the utilities themselves. A model links systematic utility to measured characteristics and uses random components to represent unexplained variation. If two alternatives share unobserved attributes, their utility errors may be correlated. Multinomial probit can incorporate that dependence rather than require the standard logit substitution pattern. However, a flexible covariance structure is still an assumption to estimate and evaluate. The model does not automatically recover true preferences or establish how choices would respond to an intervention.

How it works

Specify the available alternatives, explanatory variables and systematic utility functions. Define a joint normal distribution for the random components, with suitable covariance restrictions. Normalize location and scale because adding a common amount to all utilities or multiplying them by a positive constant leaves the chosen alternative unchanged. Calculate each choice probability as the probability that its utility exceeds every competitor's utility. This generally involves multivariate normal integration, often using numerical or simulation methods. Fit the model with an appropriate estimator, assess numerical stability and identification, and evaluate predicted probabilities on relevant data.

Three practical examples

Travel-choice research: represent similarities between transport alternatives that may share unobserved attributes.
Product analysis: study choices among exclusive options while allowing correlated unobserved utility components.
Service design: model selection among several service channels when a suitable choice dataset and substantive specification are available. Availability constraints and the way alternatives are presented must be represented rather than assumed identical for all observations.

Think of it this way

Think of several competitors receiving partly observed scores and partly uncertain scores. The chosen category is the competitor with the highest total. If two competitors share an uncertain influence, their scores can move together. The probability of winning depends on all comparisons jointly, not on treating each pairwise comparison as unrelated.

Worked example

Consider three artificial alternatives whose utilities are independent standard normal variables with equal means. By symmetry and the zero probability of a tie, each alternative has choice probability 1/3. Define D1 = U1 - U3 and D2 = U2 - U3. Each difference has variance 1 + 1 = 2, while their covariance is 1 because both contain the same negative U3 term. Their correlation is therefore 1/2. Alternative 3 wins when both differences are negative. Each individual comparison has probability 1/2, but multiplying them would incorrectly give 1/4 because the differences are dependent. The correct joint probability is 1/3. This original example shows why independent utility errors do not imply independent comparisons. Adding 7 to every realized utility leaves the winner unchanged, illustrating why absolute utility location cannot be identified from the chosen category alone.

Illustrative case study

In a fictional case, Aspen Mobility compares models of travel-mode choice. Analysts suspect that two public-transport alternatives share unobserved service conditions. They fit a multinomial probit specification with a documented covariance structure and compare its predictions with a simpler baseline. Reviewers check normalization and sensitivity to simulation settings before interpreting coefficients. The team also verifies which alternatives were actually available to each traveler, because a flexible error model cannot repair an incorrectly defined choice set.

Common misconceptions and mistakes

Multinomial probit is not several unrelated binary probit regressions. Winning against every competitor is a joint event, and pairwise probabilities cannot generally be multiplied. The normality assumption concerns latent random components, not a requirement that observed categorical outcomes or every predictor be normally distributed. Utility coefficients depend on the chosen normalization, so arbitrary raw comparisons across differently normalized models can mislead. Greater covariance flexibility does not guarantee better forecasts, and poor identification or simulation error can produce unstable estimates that appear more precise than the evidence supports.

Business use

Businesses should consider the model when the choice structure provides a substantive reason to represent correlated unobserved influences and the data can support the added complexity. Document alternatives, availability rules, covariates and normalization. Check predictive calibration and substitution behavior under realistic scenarios, while distinguishing model-based scenarios from established causal effects. Budget for numerical diagnostics and sensitivity analysis. A simpler model can remain preferable when its limitations are acceptable and the more flexible specification is weakly supported. The decision should follow evidence about the intended use rather than flexibility alone.

Check your understanding

Independent utilities produce two differences that share the same reference utility. Are those differences necessarily independent?

Answer: No. The shared reference term can create covariance, as in the example's correlation of one-half.

Related terms

Multivariate normal distribution
Covariance
Identification (information)
Discrete choice
latent utility
multinomial logit
probit
identification
simulated likelihood

Learn next

Review binary probit and the random-utility framework, then derive the covariance of utility differences. Study normalization and numerical probability estimation before interpreting a fitted multinomial probit model.

Covariance

Sources

Kenneth Train, Discrete Choice Methods with Simulation, second edition, Chapter 5: Probit: https://eml.berkeley.edu/books/choice2nd/Ch05_p97-133.pdf Kenneth Train, author-hosted book and chapter information: https://eml.berkeley.edu/books/choice2.html

Previous: Multinomial principal component analysis Next: Multinomial resampling