Why the Outcome’s Meaning Comes First
A response recorded as 1, 2, and 3 is not automatically an ordered outcome, and an outcome with words rather than numbers is not automatically unordered. The modelling decision depends on what the categories mean. “Bus, train, or car” names distinct alternatives; “low, moderate, or high satisfaction” gives categories with a meaningful order. Treating alternatives as ordered imposes a ranking that may not exist, while treating ordered categories as unrelated discards information about their order.
The distinction affects more than the model’s label. Multinomial logit assigns a separate systematic utility to each alternative relative to a reference alternative. An ordered response model instead represents categories as ranges of an underlying index separated by thresholds. Consequently, their probability restrictions and marginal-effect interpretations differ. The decision should begin with the outcome definition and the substantive question, not with which model produces a more appealing coefficient.
What the Two Models Assume
For an unordered outcome with \(J\) alternatives, multinomial logit can be motivated by a random-utility model. For individual \(i\) and alternative \(j\), write utility as \(U_{ij}=V_{ij}+\varepsilon_{ij}\), where \(V_{ij}\) is the systematic component and the individual chooses the alternative with the greatest utility. In the standard multinomial logit specification, the errors are independent and identically distributed Type I extreme-value variables across alternatives, conditional on the included information. This yields probabilities of the form below. A reference alternative is used to normalize its utility coefficients to zero.
This specification implies the independence of irrelevant alternatives (IIA): the odds of choosing alternative \(j\) rather than alternative \(k\) do not depend on the presence or characteristics of another alternative, conditional on the model’s covariates. IIA is a substantive restriction, not a consequence of categories merely being unordered. Earlier material on IIA testing and nested logit is relevant when substitution among alternatives makes that restriction doubtful.
An ordered logit or probit model uses a different structure. It supposes a latent continuous index \(Y_i^*=x_i'\beta+\varepsilon_i\), but observes only which interval the index falls into. With \(J\) ordered categories, there are \(J-1\) ordered thresholds. For ordered logit, the error has the standard logistic distribution with cumulative distribution function \(F\); for ordered probit, it is standard normal. The thresholds and error scale require a location and scale normalization for identification. In the conventional proportional-slope specification, the same \(\beta\) shifts the latent index at every threshold.
Here \(\tau_0=-\infty\), \(\tau_J=+\infty\), and the finite thresholds satisfy \(\tau_1<\cdots<\tau_{J-1}\). The category probabilities are differences between adjacent cumulative probabilities. This construction uses the order of the categories but does not assign numerical distances between them.
Worked Example: Choosing Among Unordered Commute Modes
Worked Example: Choosing Among Unordered Commute Modes
Suppose a hypothetical survey records each commuter’s chosen mode: bus, rail, or car. The research question concerns how an increase in a transit-cost index \(x\) changes mode choice. Bus, rail, and car are alternatives, not levels of a common quantity: there is no generally defensible ranking in which rail is “between” bus and car. A multinomial choice model is therefore a natural starting point, subject to its utility and substitution assumptions.
Take car as the reference alternative and specify \(V_{\text{car}}=0\), \(V_{\text{bus}}=\ln(0.5)+0.4x\), and \(V_{\text{rail}}=\ln(0.75)+0.1x\). At \(x=0\), the exponentiated utilities are \(0.5\), \(0.75\), and \(1\), so the common denominator is \(0.5+0.75+1=2.25\). The predicted probabilities are
The unrounded probabilities sum to 1; the displayed values also sum to 0.9999 because of rounding. For a different prediction, displayed probabilities of \(0.2142\), \(0.3357\), and \(0.4502\) sum to \(1.0001\). That small excess is likewise a rounding discrepancy: the unrounded probabilities sum to 1.
The coefficients \(0.4\) and \(0.1\) are changes in log utility relative to car per one-unit increase in \(x\); they are not probability changes. To obtain probability changes, differentiate the softmax formula. For an alternative \(j\), with slope \(\beta_j\), the derivative is \(p_j(\beta_j-\sum_k p_k\beta_k)\). At \(x=0\), the probability-weighted mean slope is \((2/9)(0.4)+(1/3)(0.1)+(4/9)(0)=0.1222\). Thus the derivatives for bus, rail, and car are approximately \(0.0617\), \(-0.0074\), and \(-0.0543\) per unit of \(x\), respectively. They sum to zero, as the probabilities must continue to sum to one. These are local changes in predicted probabilities, conditional on the specified model; they are not causal effects without additional causal assumptions.
Worked Example: Modelling Ordered Satisfaction
Worked Example: Modelling Ordered Satisfaction
Suppose a hypothetical survey records satisfaction as low, moderate, or high. The labels have a meaningful order, but the gaps between them are not measured. An ordered response model respects that order without pretending the categories are equally spaced. Consider an ordered logit model with \(Y_i^*=0.6x_i+\varepsilon_i\), thresholds \(\tau_1=-1\) and \(\tau_2=1\), and standard logistic errors. At \(x=1\), the latent index is \(0.6\). The cumulative probabilities are \(F(-1-0.6)=F(-1.6)\) and \(F(1-0.6)=F(0.4)\). Using \(F(z)=1/(1+\exp(-z))\), these are approximately \(0.1680\) and \(0.5987\).
The probabilities for low, moderate, and high satisfaction are therefore the first cumulative probability, the difference between cumulative probabilities, and one minus the second cumulative probability:
The unrounded probabilities sum to 1; the displayed values sum to 1.0000. The positive slope shifts the latent index upward, but it does not imply that every category’s probability rises. In fact, the probability of the middle category may rise or fall depending on the index and thresholds.
For a continuous covariate \(x\), let \(a_j=\tau_j-x\beta\) and let \(f\) be the logistic density. Since \(P(Y\leq j\mid x)=F(a_j)\), differentiation gives \(\partial P(Y\leq j\mid x)/\partial x=-\beta f(a_j)\). The low-category probability has derivative \(-\beta f(a_1)\); the middle-category probability, as the difference of two cumulative probabilities, has derivative \(\beta[f(a_1)-f(a_2)]\); and the high-category probability has derivative \(\beta f(a_2)\). At \(x=1\), \(f(-1.6)\approx0.1398\) and \(f(0.4)\approx0.2403\). With \(\beta=0.6\), the marginal effects are approximately \(-0.0839\), \(-0.0603\), and \(0.1442\) per unit of \(x\). Their unrounded sum is zero because the three category probabilities sum to one. Thus a positive latent-index slope increases the probability of high satisfaction here while decreasing the probabilities of low and moderate satisfaction.
Worked Example: Clinical Dispositions Are Not Necessarily Ordered
Worked Example: Clinical Dispositions Are Not Necessarily Ordered
Imagine a hypothetical clinic records one of three dispositions: discharge home, admit to the clinic’s ward, or transfer to another facility. Although these choices may be associated with severity, they are not necessarily successive levels of one ordered response. A transfer is not inherently “more” or “less” than admission; it may reflect available services, patient preference, or specialist needs. If the research question concerns which disposition is chosen, treating the labels as low, medium, and high would impose a questionable ranking. A multinomial model can instead represent them as distinct alternatives.
For illustration, suppose the fitted systematic utilities for a particular covariate profile are \(V_{\text{discharge}}=0\), \(V_{\text{admit}}=\ln(1.5)\), and \(V_{\text{transfer}}=\ln(0.5)\). The exponentiated utilities are \(1\), \(1.5\), and \(0.5\), with denominator \(3\). The corresponding probabilities are \(1/3=0.3333\) for discharge, \(1.5/3=0.5000\) for admission, and \(0.5/3=0.1667\) for transfer. Each probability describes the chance assigned to a named alternative, conditional on the model and covariates; none describes the chance of exceeding a severity threshold. The example does not establish that IIA is suitable: the model’s substitution assumptions still require substantive scrutiny and, where appropriate, diagnostics.
A Practical Choice and Interpretation Framework
The following sequence keeps the outcome’s meaning separate from the convenience of a particular estimator.
Write down what each recorded value means. Check whether the categories are mutually exclusive and whether their order is meaningful for the question being studied.
For distinct unordered alternatives, consider a multinomial choice model. For ranked categories with unknown spacing, consider an ordered logit or probit model.
For multinomial logit, examine whether IIA and the utility specification are plausible. For ordered response, examine threshold structure and the common-slope restriction; ordered-response diagnostics and proportional-odds material address these concerns.
Present category-specific predicted probabilities at meaningful covariate values and, when useful, marginal effects on each probability. State covariate units and the population or profile to which the predictions refer.
For both model families, predicted probabilities should be checked for sensible behavior across relevant covariate profiles. For an ordered model, it is also important to remember that the shared slope applies to the latent index or cumulative-logit structure, not as an identical probability change in every category. For multinomial logit, coefficients are relative to the chosen base alternative; changing the base changes coefficient presentation but not the fitted probabilities.
Model fit alone cannot decide whether categories are ordered. A model that fits the observed frequencies well may still encode an incoherent ranking or an unsuitable substitution structure. The outcome definition and estimand come first; diagnostics and predictive performance then help assess candidate specifications that make sense for that outcome.
Common Mistakes and a Key Distinction
- Ordering categories because they are coded with increasing numbers. The codes may only be labels. Use an ordered model only when the ranking has substantive meaning.
- Assuming ordered categories are equally spaced. Ordered logit and probit use thresholds; they do not require equal distances between adjacent observed categories.
- Reading an ordered-model slope as the effect on every category probability. The slope shifts the latent index. Category-specific marginal effects can have different signs, as the satisfaction example shows.
- Reading multinomial coefficients as probability changes. They describe changes in relative systematic utility or log odds against a reference alternative. Calculate marginal effects or predicted probabilities for probability-scale interpretation.
- Choosing multinomial logit simply because the outcome has several categories. The standard model has assumptions, including IIA. Distinct alternatives do not by themselves guarantee that IIA is plausible.
Check Your Understanding
For each question, explain the reasoning rather than relying only on the category labels.
- A survey records a person’s preferred payment method: cash, card, or mobile wallet. Is the outcome naturally ordered? Which model family might be a starting point, and what restriction should be considered?
- A rating is recorded as poor, fair, good, or excellent. Why might an ordered response model be appropriate even though the distances between ratings are not known?
- In multinomial logit, why is a coefficient relative to a reference alternative not itself a change in predicted probability?
- In an ordered logit model with a positive slope, must the probability of every higher category increase? Explain using the role of thresholds and category probabilities.
- Why is model fit insufficient, by itself, to decide whether a response should be treated as ordered or unordered?