Why an Ordered Outcome Has a Different Structure
Some outcomes are categories with a meaningful ranking but no natural numerical distance between categories. A patient may report poor, fair, good, or excellent health; a regulator may classify a firm as low, medium, or high risk. The order matters, but it is not generally sensible to claim that the distance from poor to fair equals the distance from good to excellent.
An ordered response model represents this structure with an unobserved continuous index and cutpoints. Covariates shift the index, and the observed category records which interval contains it. This is different from treating each category as a separate choice with its own unrelated comparison against a baseline.
The Threshold Representation
For precision, define \(\kappa_0=-\infty\) and \(\kappa_J=+\infty\). The observation rule is
Here \(x_i\) is a vector of observed explanatory variables, \(\beta\) is a vector of coefficients, and \(\varepsilon_i\) is an unobserved disturbance. The thresholds divide the latent scale into consecutive regions. Category 1 occurs below the first threshold, category \(J\) above the last, and each intermediate category between two neighboring thresholds.
The cutpoints are not numerical scores assigned to the observed categories. For example, observing category 4 rather than category 3 means that the latent index crossed a particular boundary; it does not mean the outcome increased by a known one-unit amount. The model requires the order of categories, not equal spacing between them.
Worked Example: Translating a Rating Into a Latent Interval
Suppose a hypothetical five-level service rating runs from 1 (very poor) to 5 (excellent). For illustration, take thresholds \(\kappa_1=-1\), \(\kappa_2=0.5\), \(\kappa_3=1.5\), and \(\kappa_4=2.5\). A customer has latent index \(x_i'\beta=0.8\) and disturbance \(\varepsilon_i=-0.2\). Determine the model-implied category for this disturbance realization.
First calculate the latent variable:
Since \(0.5<0.6\leq1.5\), the observation rule assigns category 3. This calculation illustrates the threshold mapping only; in an estimated model the thresholds and coefficients are parameters, not values known in advance. A different disturbance realization could place the same customer's latent index in another interval.
What Ordered Logit and Ordered Probit Specify
The threshold rule defines the ordered-response structure, but does not by itself specify the distribution of the disturbance. Ordered logit and ordered probit use different distributions for \(\varepsilon_i\). In either case, the standard specification assumes that, conditional on the covariates, the disturbance follows the same distribution across observations and has no additional systematic dependence on \(x_i\). The chosen distribution determines how latent-index differences translate into the model's conditional category probabilities.
For either model, let \(F\) denote the disturbance cumulative distribution function. The threshold representation implies, for an internal category boundary \(j\),
This follows because \(Y_i\leq j\) is equivalent to \(Y_i^*\leq\kappa_j\), which in turn is equivalent to \(\varepsilon_i\leq\kappa_j-x_i'\beta\). For ordered probit \(F=\Phi\), the standard normal cumulative distribution function; for ordered logit \(F=\Lambda\). This cumulative expression is a direct implication of the model's foundations, not a claim that the observed categories are equally spaced.
The usual specifications normalize the latent disturbance scale: the probit disturbance has variance 1, while the standard logistic disturbance has variance \(\pi^2/3\). These conventions give the latent index a defined scale for estimation. The location of the latent index also has to be fixed relative to the thresholds; a separate intercept and a common shift of all thresholds cannot both be freely distinguished from the observed categories. Detailed parameter normalization is a separate identification issue.
Worked Example: Comparing Ordered Logit and Probit at One Cutpoint
Consider a hypothetical ordered outcome with a cutpoint for which \(\kappa_j-x_i'\beta=0.8\). Calculate the model-implied cumulative probability \(P(Y_i\leq j\mid x_i)\) under each standard specification.
For ordered logit, substitute \(0.8\) into the logistic cumulative distribution function:
For ordered probit, use the standard normal cumulative distribution function:
Thus, at this particular standardized index gap, the cumulative probability is about 0.6900 under logit and 0.7881 under probit. These values are model-based probabilities, not observed frequencies. The difference reflects the two assumed disturbance distributions; it does not change the fact that both models rank observations using the same single index and ordered cutpoints.
The Common Ranking Assumption
The defining restriction is not merely that the categories have an order. The ordered model assumes that covariates act through one common index \(x_i'\beta\), with the same coefficient vector at every threshold. To see the implication, consider two covariate profiles, \(x_A\) and \(x_B\), and suppose \(x_A'\beta>x_B'\beta\). At any cutpoint \(\kappa_j\), the quantity \(\kappa_j-x_A'\beta\) is smaller than \(\kappa_j-x_B'\beta\). Since a cumulative distribution function is nondecreasing,
The higher-index profile therefore has no greater conditional probability of being at or below any given category boundary. Its distribution is shifted toward higher categories in this cumulative sense. The direction of this ranking is shared across all cutpoints; the model does not permit one covariate profile to rank higher at one boundary but lower at another.
This conclusion concerns the model's conditional distribution, not a certainty about any individual's realized outcome. Disturbances still vary, and two observations with different indices can realize the same category or categories in an order that differs from their index ranking. The coefficient vector summarizes a systematic shift in latent propensity, not a deterministic rule for each observed response.
Worked Example: Reading a Shared Index Across Profiles
Suppose a hypothetical ordered assessment of program access has three categories: low, medium, and high. The index is \(x_i'\beta=0.4T_i-0.25W_i\), where \(T_i\) is training sessions and \(W_i\) is waiting time in weeks. Compare profile A, with \(T_A=1\) and \(W_A=2\), with profile B, with \(T_B=3\) and \(W_B=1\).
The indices are
Profile B's index exceeds profile A's by \(0.95-(-0.10)=1.05\). Consequently, for every category boundary \(j\), the model implies \(P(Y\leq j\mid x_B)\leq P(Y\leq j\mid x_A)\). The model places B's conditional response distribution higher in the ordered categories. It does not imply that B must receive a higher realized category, and the coefficient 0.4 is not a category-specific increase in the probability of “high.”
How This Differs From an Unordered Choice Model
An unordered choice model treats the alternatives as distinct outcomes without imposing a single natural ranking. Multinomial logit, for example, specifies comparisons between each nonbaseline outcome and a chosen baseline. The covariate coefficients in those comparisons can differ by alternative. This permits covariates to affect different alternatives in different ways.
An ordered model instead treats a move toward a higher latent index as a shift across the same ordered set of thresholds. The shared index and common coefficient vector are substantive restrictions. They can be useful when the categories represent increasing levels of one underlying trait, but they may be inappropriate when the labels are merely different options. “Low, medium, high” risk has an order; “bus, train, car” generally does not.
| Feature | Ordered logit or probit | Unordered multinomial choice model |
|---|---|---|
| Category structure | Categories have a meaningful order | Alternatives need not have a ranking |
| Covariate structure | One latent index and common coefficient vector across cutpoints | Alternative-specific comparisons can have different coefficients |
| Interpretation of a positive index shift | Shifts the conditional distribution toward higher categories | Can raise some alternative probabilities and lower others without a common ranking |
| Core modeling restriction | Observed category is determined by the interval containing the latent index | Choice is represented through comparisons among distinct alternatives |
Worked Example: Deciding Whether the Ranking Assumption Fits
A hypothetical survey records a respondent's preferred commute mode: bus, rail, or car. A researcher proposes assigning scores 1, 2, and 3 and fitting an ordered logit. Is the ordered structure justified by those labels?
No. The labels do not, by themselves, establish that car is “more” of the same underlying trait than rail, or that rail is “more” than bus. The numbers would impose a ranking that is not inherent in the outcome. An unordered choice model can instead compare each mode with a baseline and allow covariate relationships to differ across those comparisons.
If the outcome were instead a clearly ranked rating such as low, moderate, or high commuting burden, an ordered model could be substantively plausible. The key is not whether the categories can be numbered; it is whether they represent ordered regions of one underlying dimension and whether a common index is a reasonable restriction.
Interpretation and a Common Pitfall
A positive coefficient means that, holding the other covariates fixed, an increase in that covariate raises the latent index when the model is written as \(x_i'\beta\). Under the ordered model, this corresponds to a shift toward higher categories in the cumulative ordering described above. It does not mean that every category probability rises, nor does it directly give a change in the probability of any particular category. Those category-specific implications depend on the cutpoints and the covariate values.
The coefficient is not automatically causal. A causal interpretation requires an appropriate research design and assumptions beyond the ordered-response representation. In observational data, unmeasured factors associated with both a covariate and the response can undermine a causal claim, even when the ordered model is correctly specified as a conditional distribution.
- Wrong reasoning: “The outcome has three numbered categories, so an ordered model is appropriate.”
- Correction: Numbering is not enough. The categories must have a defensible substantive order, and a common latent index must be a plausible way to represent movement across them.
- Wrong reasoning: “A positive coefficient increases the probability of every higher category.”
- Correction: A positive coefficient shifts the cumulative distribution toward higher outcomes. The probability of a particular middle category can move either way, because it is bounded by two thresholds.
Check Your Understanding
Try these questions before moving on.
- In a four-category ordered response, write the latent-index interval that corresponds to category 3.
- What distribution does the standard ordered probit model assign to its latent disturbance? How does ordered logit differ?
- If one profile has a larger latent index than another, what inequality does the ordered model imply for their cumulative probabilities at each cutpoint?
- Why is a positive ordered-model coefficient not necessarily an increase in the probability of every higher category?
- Would preferred transport mode generally be a natural ordered response? Explain why or why not.