Unequal Category Spacing Does Not Make An Outcome Continuous

Order Is Not Distance

A response scale can have a clear order without telling us how far apart its categories are. For example, a person may choose among “not at all,” “a little,” “moderately,” “very,” and “extremely” in response to a question about pain. These answers have a natural order. But the labels do not specify that the distance from “not at all” to “a little” equals the distance from “very” to “extremely”—or even provide a unit in which either distance can be measured.

This distinction matters when choosing an outcome representation. Assigning the codes 1 through 5 makes categories easier to store and display, but those codes record order; they do not establish a quantitative scale. A mixed model for an ordinal response can represent ordered category probabilities while retaining the response process. Treating the codes as continuous instead imposes additional assumptions about numerical distances and the meaning of arithmetic on the codes.

The question is not whether category labels look uneven, or whether someone can write a number beside each label. The question is what the measurement process tells us. An ordinal response identifies which ordered category contains a response. A continuous measurement, by contrast, has a quantitative scale on which differences have a defined interpretation in the measurement units.

Category Labels and Likert-Type Responses

A Likert-type item asks a respondent to select one ordered answer, such as a level of agreement or frequency. Its categories can be coded 1, 2, 3, 4, and 5 for data handling. The code 4 means that the fourth ordered category was selected; it does not establish that the category is twice as far from the first category as the category coded 2. Nor does it establish equal intervals between all five categories.

The same distinction applies if category labels themselves refer to amounts. “Never,” “rarely,” “sometimes,” “often,” and “always” are ordered frequency categories, but without defined time intervals or other quantitative boundaries, the words do not specify equal frequency differences. Conversely, if responses are recorded as actual minutes, counts, or another measured quantity, those recorded values may support quantitative modeling. The measurement process, not the choice of labels, determines which interpretation is warranted.

In an ordinal model, the order constrains the probabilities: categories higher in the order represent responses at least as high as categories below them. A model can estimate category probabilities or cumulative probabilities without asserting that the response is a point on an equally spaced numerical scale. The likelihood should be chosen in light of how the response categories are generated, as in the earlier material on choosing an ordinal likelihood from the response process.

Why Arithmetic on Category Codes Can Mislead

A useful test for a summary of ordinal data is whether it depends on arbitrary choices of numerical labels. Any strictly increasing recoding preserves the category order. If a conclusion changes under such a recoding, it is relying on distances that the ordinal scale did not specify.

For instance, the labels 1, 2, 3, 4, 5 and the labels 1, 2, 4, 7, 11 describe the same five-category order. The second set spaces some category codes farther apart than others. Neither coding is more faithful to the measurement process unless external information gives those numbers a quantitative interpretation.

Category Probabilities Without Equal Spacing

An ordinal probability model can work directly with ordered cutpoints on a model scale. The cutpoints are not measurements of distances between the observed category labels. For illustration, consider a cumulative-logit form with five categories and a fixed linear predictor value \(\eta=0\). Let \(\theta_1<\theta_2<\theta_3<\theta_4\) be ordered cutpoints, and let \(F(t)=1/(1+\exp(-t))\) be the logistic cumulative distribution function. The model specifies the probability of category \(k\) or below. In this example, the cutpoints are \(-1.2,-0.4,0.5,1.4\).

The model is conditional on the specified \(\eta\); in a mixed model, \(\eta\) can include covariates and a random effect. The example below sets \(\eta=0\) to keep the arithmetic clear. Its calculations use the logistic distribution as a model assumption, not because the observed category labels have equal intervals.

Justification: Why the Probabilities Sum to One

The subtraction rule follows directly from the fact that the cumulative events are nested. For categories \(1,\ldots,K\), define \(C_k=P(Y\leq k\mid\eta)\), set \(C_0=0\), and set \(C_K=1\). The event \(Y=k\) is the event \(Y\leq k\) with the event \(Y\leq k-1\) removed. Therefore its probability is the difference of those cumulative probabilities.

For these differences to be nonnegative, cumulative probabilities must be nondecreasing: \(C_k\geq C_{k-1}\). In the illustrative cumulative-logit model, the logistic function is increasing, so ordered cutpoints \(\theta_1<\cdots<\theta_{K-1}\) give ordered cumulative probabilities at a fixed \(\eta\). This is a basic probability condition, not an assumption that the observed categories are equally distant.

Assumptions, Diagnostics, and Alternatives

The classification as ordinal depends on the response process: categories must have a meaningful order, while the data do not supply a justified quantitative distance scale. If the model uses cumulative probabilities, it must also respect their ordering and yield valid category probabilities. Additional model assumptions depend on the likelihood and link chosen. For the illustrative cumulative-logit calculation, the logistic cumulative distribution and ordered cutpoints were specified; other ordinal likelihoods represent different response processes and should not be substituted merely because labels look uneven.

Useful checks begin with the response instrument and data coding. Confirm that category labels are in the intended order, that numerical codes have not been mistaken for measured values, and that the fitted or reported category probabilities are nonnegative and sum to one before rounding. Compare category and cumulative proportions across relevant covariate or group values. A category probability outside the interval from 0 to 1, decreasing cumulative probabilities, or category probabilities that fail to sum to 1 indicates a coding, calculation, or model problem.

A practical alternative is to analyze the ordinal categories as categorical without using their order. That avoids imposing equal distances but also discards ordering information. Treating category codes as continuous uses their numerical spacing and can be concise, but it changes the interpretation and requires a defensible quantitative scale. An ordinal likelihood retains category order without claiming that code differences are measured intervals. In a GLMM, adding random effects can represent dependence or heterogeneity; it does not convert the response scale into a continuous one.

This distinction has limits. If the data actually record a quantitative variable, or categories correspond to known measurement intervals and the scientific question concerns the underlying quantity, a quantitative or interval-based model may be appropriate. That is a different measurement situation from a Likert-type response whose labels only establish order. The classification should follow the information in the measurement, not a preference for one model family.

Common Mistake

A related mistake is to infer that unequal wording makes the response nominal. Unequal or unknown spacing does not remove the ordering. The categories remain ordinal when their order is meaningful; what remains unknown is the quantitative distance between them.