Sparse Ordinal Categories And Response Coding

Why Sparse Ordinal Levels Need Careful Coding

Ordinal responses often have a fixed set of ordered levels, such as none, mild, moderate, and severe. Some levels may occur rarely in a particular sample. That is a warning to inspect how much information the data provide about probabilities involving those levels; it is not, by itself, a reason to discard a level or combine it with another. The categories still describe distinct responses, and an analysis should preserve those distinctions unless there is a substantive reason to change them.

In a generalized linear mixed model, observations may also be clustered. Let \(Y_{ij}\) be the response for observation \(i\) in cluster \(j\), with ordered categories \(1,\ldots,K\). Let \(x_{ij}\) denote covariates and \(b_j\) a cluster-level random effect. An ordinal response model specifies probabilities for the ordered categories conditional on covariates and random effects. The data must therefore communicate both the category order and which observed response belongs to which category.

The response should ordinarily be stored as an ordered categorical variable, with the intended levels and their order explicitly declared. Integer labels can be convenient internal codes, but their numerical distances do not represent distances between responses. Coding mild, moderate, and severe as 1, 2, and 3 does not make the difference from mild to moderate equal to the difference from moderate to severe.

What Sparse Levels Change in an Ordinal Model

For a cumulative-link representation, define \(F_k(x_{ij},b_j)=P(Y_{ij}\leq k\mid x_{ij},b_j)\), for \(k=1,\ldots,K-1\). The probability of an interior category \(k\) is \(F_k-F_{k-1}\), with \(F_0=0\) and \(F_K=1\). The probability of the top category is \(1-F_{K-1}\). Thus every observation contributes to the likelihood through the probability assigned to its actual category, while the category boundaries determine which cumulative probabilities are relevant.

A rare interior category can leave its probability, or the thresholds that determine its boundaries, weakly informed. A rare endpoint may particularly limit information about the corresponding tail probability. The exact consequences depend on the chosen ordinal likelihood and the covariate pattern: a level may be rare overall but well represented across covariate values, or common overall but absent in an important subgroup. In a mixed model, it also matters whether responses occur across many clusters or are concentrated in just a few.

A useful intuition comes from a single cumulative split. Suppose a split defines a binary indicator \(I=1\) for a response on one side of a boundary, with probability \(p\). For \(n\) independent observations with a common \(p\), the Bernoulli log-likelihood is \(I_{\mathrm{sum}}\log(p)+(n-I_{\mathrm{sum}})\log(1-p)\). Its expected Fisher information for the logit of \(p\) is \(np(1-p)\). If \(p\) is very small, this information is small because few observations are expected on that side of the split. This is an intuition, not a formula for the variance of every ordinal-model estimate; covariates, random effects, and the likelihood representation also matter.

The derivation follows by differentiating with respect to \(\eta\). Since \(dp/d\eta=p(1-p)\), the score is \(I_{\mathrm{sum}}-np\), and the negative expected second derivative is \(np(1-p)\). When observed outcomes perfectly separate a covariate pattern, estimates can become extremely large or fail to be finite; sparse categories can make such patterns more likely. This is one reason to inspect the data before trusting a convergence message or a large coefficient.

Response Coding and Representation

Coding begins with the meaning of the response, not with whichever coding makes software run most easily. Preserve the scientifically defined order, retain all possible levels in the declared response scale, and check that the recorded values map to those levels correctly. A numerical score model imposes assumptions about spacing and linear effects that an ordinal-category model does not require; assigning integers to categories is not a justification for treating the response as quantitative.

The response process also guides the likelihood choice. Cumulative-link models describe probabilities relative to ordered cut points; adjacent-category models compare neighboring levels; continuation-ratio models describe progression through sequential stages. These are different representations of the response probabilities, and their assumptions and interpretations differ. Sparse counts may expose weak support or convergence problems under any of them. Choosing a representation solely because it seems to avoid a sparse level is not a defensible substitute for matching the model to how the response was generated.

Distinguishing Sample Sparsity From Structural Absence

Before changing the response scale, determine why a level is empty or rare. A level may be possible but unobserved by chance, especially in a small sample or a small subgroup. It may be impossible by design for some observations, such as a response option not offered under a particular protocol. Or the apparent absence may result from a data-processing error: a misspelled label, an incorrect mapping, or an omitted level. These situations have different implications. Verify the questionnaire or measurement process and the data coding before interpreting a zero count.

Diagnostics, Alternatives, and Limits

Sparse-category assessment is not a single numerical cutoff. The implications depend on total sample size, the number of clusters, the covariate structure, random-effect structure, and the selected likelihood. Useful checks include:

• Tabulate counts for every declared category, overall and by important predictors and clusters.
• Confirm the response is ordered correctly and that no values are missing, miscoded, or silently dropped.
• Inspect fitted category probabilities and uncertainty, particularly for rare endpoints and sparse covariate patterns.
• Check convergence warnings, extreme estimates, large standard errors, and sensitivity to starting values or influential clusters.
• Compare scientifically plausible ordinal representations and assess whether the substantive conclusions are stable.

A model that converges is not necessarily well supported: convergence only indicates that a numerical algorithm met its stopping criteria. Conversely, a rare category does not automatically invalidate a model. Estimates may remain useful when the likelihood is identifiable, the uncertainty is honestly represented, and results are not driven by a few observations. The limitations should be stated when intervals are wide, probabilities are imprecise, or estimates change substantially under reasonable specifications.

Possible responses to a serious support problem include obtaining more data, simplifying an over-ambitious covariate or random-effect structure, or using a carefully justified alternative model. Merging categories is another possible change to the response definition, but it changes what the outcome distinguishes and can change the estimand. It should not be an automatic software repair or a rule triggered by a particular count. The rationale must be substantive and transparent, and the altered coding must be used consistently in interpretation.