Why Response Scale Matters
A response recorded as 1, 2, 3, or 4 may look quantitative in a data file. But those numbers can be labels for ordered categories rather than measurements. For example, a rating of 4 may mean “severe” while a rating of 3 means “moderate”; the numbers establish which category is higher, but do not by themselves say how much higher it is. Coding a response as continuous changes the statistical question: it treats differences between scores as meaningful quantities.
This decision is important in a generalized linear mixed model (GLMM). A random effect can represent dependence among observations in the same cluster, such as patients within clinics. It does not determine whether the response is ordinal or continuous. That choice must follow the way the outcome was generated and measured.
Write \(Y_{ij}\) for the response from observation \(i\) in cluster \(j\). An ordinal response with \(K\) categories can be represented as \(Y_{ij}\in\{1,\ldots,K\}\), where the labels indicate order only. A continuous response is instead a measured value, such as \(Z_{ij}=2.4\) hours. The notation does not make a response continuous: its meaning and measurement process do.
What Each Coding Choice Assumes
If the response is treated as continuous in a Gaussian mixed model, a common conditional specification is
Here \(x_{ij}\) and \(z_{ij}\) are covariate vectors, \(\beta\) is a vector of fixed-effect coefficients, and \(b_j\) is a cluster-specific random effect. This model assumes that, conditional on the random effects and covariates, the response varies around a numerical conditional mean with an approximately Gaussian error distribution and common residual variance \(\sigma^2\). It also assumes conditional independence of residuals unless further dependence is explicitly modeled. In this formulation, changes in the conditional mean are expressed in the response’s measurement units.
By contrast, an ordinal model treats the outcome as a set of ordered categories and models their probabilities. Its predictions must correspond to valid probabilities for the categories, and the category order is part of the response structure. The choice of ordinal model and its likelihood is a separate modeling decision; the essential point here is that category membership, not a measured distance between labels, is the outcome. The introductory material on ordinal responses establishes why ordered category probabilities differ from an unordered set of category probabilities.
A useful check is to ask what an estimated one-unit change would mean. If it means one additional hour, one additional degree, or another defined amount, continuous modeling may be sensible. If it means moving from “mild” to “moderate,” with no defensible interpretation of that step as a fixed amount, a continuous interpretation is not automatic.
Worked Example: Choosing a Response Scale for Symptom Ratings
Suppose a hypothetical study records each patient’s pain interference at each visit as none, mild, moderate, or severe. Patients are observed repeatedly within clinics. Which response coding is supported by this description?
The observed outcome has four ordered categories, so a suitable representation is \(Y_{ij}\in\{1,2,3,4\}\), with larger labels assigned to more interference. The labels record rank. Nothing in the description says that the difference between none and mild is the same amount of pain interference as the difference between moderate and severe.
The repeated visits and clinic membership may require dependence to be represented in the model, for example through patient- or clinic-level random effects if appropriate to the design. Those effects address clustering; they do not turn the four response categories into measurements. The response coding supported by the information given is ordinal.
If the study instead recorded the number of hours per day that pain interfered with activity, that value would have a direct unit of hours. A continuous-response model could then be considered, subject to checking its distributional and residual assumptions. The measurement process, not the presence of digits in a data file, separates the two cases.
Why Category Scores Do Not Automatically Have Measurable Distances
The problem with treating ordinal labels as measurements can be shown directly. Let \(c_k\) be the numerical score assigned to category \(k\), and let \(p_k\) be the probability of that category. The mean of the scores is
The order of categories restricts the scores to increase with category rank, but it does not uniquely specify their numerical values. A different set of increasing scores can therefore give a different mean and a different contrast between groups. If the score mean changes under a relabeling that preserves all the observed ordering, it cannot be interpreted as a quantity determined by the ordering alone.
For each group, record the proportion in every ordered category. These proportions describe the observed distribution without assigning distances to categories.
Use \(E(C)=\sum_k c_kp_k\) to find the mean for each group and their contrast.
If the contrast changes, its magnitude depends on the chosen scores and is not identified by the ordinal categories alone.
Worked Example: A Score Contrast Depends on the Labels
In a hypothetical sample, group A has four ratings: two in category 1 and two in category 2. Group B has four ratings: two in category 2 and two in category 3. The categories are ordered from lower to higher, but their labels are not measurements.
With scores \(1,2,3\), the group means are
The score-mean difference is \(2.5-1.5=1.0\). Now assign the same ordered categories scores \(1,2,10\). This preserves every category ranking, but gives
The contrast is now \(6.0-1.5=4.5\), even though no response changed category and the ordering is unchanged. The direction is the same in this example, but the numerical size is not determined by the ordinal information. Calling either contrast a measured amount requires justification beyond the category order.
Worked Comparison in a Clustered Setting
The following example shows how to describe an ordinal outcome without treating its category codes as measured values. It also illustrates why the cluster structure should be kept conceptually separate from response scale.
Worked Example: Summarizing Ratings Across Clinics
In a hypothetical study, patients in several clinics rate treatment burden in four ordered categories, coded 1 through 4 from low to high. Pooling the records for a descriptive comparison gives 20 ratings in each of two groups. Group A has category counts \((2,6,8,4)\), and group B has counts \((0,3,7,10)\).
For group A, the proportions are \(2/20=0.10\), \(6/20=0.30\), \(8/20=0.40\), and \(4/20=0.20\). For group B, they are \(0/20=0\), \(3/20=0.15\), \(7/20=0.35\), and \(10/20=0.50\). These category proportions show that group B has a greater share in the highest category: \(0.50\) compared with \(0.20\).
If the codes are averaged, group A’s score mean is
Group B’s coded mean is
The coded difference is \(3.35-2.70=0.65\) score units. This is arithmetic on the chosen labels, not automatically a difference of 0.65 units of treatment burden. A model for the ordinal outcome should respect the ordered categories and account for the clinic clustering when the design and inferential target require it. The pooled calculations above are descriptive only: they do not account for clustering and are not a fitted mixed model.
Diagnostics, Alternatives, and Limits
Before choosing continuous-response coding, check both the measurement definition and the consequences of the proposed model. Ask whether the response has a meaningful unit, whether differences in the response carry the same substantive meaning across its range, and whether the intended coefficient interpretation can be stated in that unit. Then assess the conditional residual behavior and fitted values implied by the Gaussian mixed model. Warnings include strong skewness, residual spread that changes substantially with the fitted mean, extreme ceiling or floor accumulation, and predictions outside the possible response range. These checks can reveal a poor continuous approximation, although no single diagnostic can establish that ordinal labels are measurements.
For ordinal data, inspect category frequencies within relevant covariate groups and clusters. Very sparse categories can make an ordinal analysis difficult, while nearly empty categories may make model estimates unstable. A few clusters or little information within clusters can also limit reliable estimation of random effects. These are limitations to address in model design and analysis; they are not reasons by themselves to reinterpret categories as continuous.
The main alternatives serve different purposes. An ordinal model uses the order while retaining categories as the response. A nominal multicategory model does not use an order and may be appropriate when categories have no meaningful ranking. A continuous-response model uses a numerical scale and supports mean-based interpretations in its units when its assumptions are reasonable. Choosing among them is not a contest in which one model is always best: it is a decision about which features of the outcome are supported and relevant.
A continuous approximation can sometimes be useful as a pragmatic simplification, but it needs an explicit rationale. It may be more plausible when the response has many ordered levels and the score has an accepted measurement interpretation, for example a validated scale designed to summarize multiple items. Even then, analysts should assess whether the chosen model answers the scientific question and whether its conclusions are sensitive to reasonable alternatives. A category’s numeric storage format, a familiar coding convention, or convenience alone is not sufficient evidence.
Common Mistake
Correction: Numeric labels can simply mark ordered categories. A Gaussian mixed model interprets differences as numerical quantities and makes assumptions about conditional means and residual variation. First establish that the response scale supports that interpretation. If it does not, retain the ordinal response structure and separately represent clustering as required.
Check Your Understanding
Answer each question, then compare with the brief solutions below.
- A hospital records pain as none, mild, moderate, or severe. What feature makes this response ordinal, and what information is not supplied by the labels?
- A device records temperature to the nearest tenth of a degree Celsius. Is this necessarily an ordinal rating? Explain the relevant distinction.
- Two groups have the same ordered-category distributions, but an analyst changes category scores from \(1,2,3\) to \(1,2,8\). What should be checked before interpreting a change in the score means?
- In repeated clinic observations, what does a clinic random effect address, and what does it not establish about the response scale?
Solutions. (1) The categories have a meaningful order. The labels do not specify measured distances between adjacent categories. (2) No. The values represent a temperature measurement with a stated unit; rounding does not make it a mere category label. (3) Whether the numerical score differences have a defensible measurement interpretation, because preserving order alone does not preserve the mean contrast. (4) A clinic random effect can represent between-clinic variation and dependence induced by shared clinic membership; it does not make ordinal labels continuous or validate arithmetic on them.