Ordered Ratings Versus Event Counts

Why Integer Responses Can Mean Different Things

A response recorded as 0, 1, 2, 3, or 4 may look like a count. But the same codes could represent five ordered levels of pain, satisfaction, or severity. In one case, a value of 3 means three events occurred; in the other, it means the response belongs to the category labelled “3.” The entries have the same appearance, but they describe different quantities and support different likelihoods.

This distinction matters in a generalized linear mixed model (GLMM) because the response family describes the conditional distribution of the response given covariates and random effects. A random effect can represent dependence among observations in the same cluster; it cannot turn a ranked category into an event total. Classification therefore starts with the measurement and response process, not with a quick inspection of the numbers.

Definition: An event count records how many times a specified event occurred in a defined unit of observation. A value of 3 means three occurrences, and differences such as 3 versus 2 have a count interpretation. An ordered rating assigns each response to one of a set of ranked categories. A value of 3 identifies a category above category 2, but the labels alone do not imply that the categories are equally spaced or that 3 records three events.

The response process should make the distinction clear. Ask what was observed, what the possible responses mean, whether a value counts occurrences, and whether the distances between adjacent numbers have a defensible quantitative interpretation. For ordered ratings, the ordering is meaningful; equal spacing usually does not follow from the coding. For counts, integer increments represent additional events, although the distribution still depends on the event-generating process and any limits on possible values.

What the GLMM Response Family Must Represent

Let \(Y_{ij}\) denote the response for observation \(i\) in cluster \(j\), and let \(b_j\) denote a cluster-level random effect. A GLMM specifies a conditional distribution for \(Y_{ij}\) given its predictors and \(b_j\). The chosen family must assign probability to the responses that can actually occur and must represent what those responses mean.

For an event count, a Poisson model is a possible choice when the response is a nonnegative event total and the Poisson assumptions are reasonable. In a basic conditional form, \(Y_{ij}\mid b_j\) has a Poisson distribution with mean \(\lambda_{ij}\). Other count families may be appropriate under different features of the count process. If the response is the number of successes among a known number of opportunities, a binomial model may be appropriate. None of those choices follows simply because the values are integers.

For an ordered rating, the possible values represent categories with an order. A model for ordered categorical responses represents probabilities of those categories. Its structure should use the ordering without quietly treating the labels as event totals or assuming equal numerical gaps. The details of ordinal likelihoods are separate; the essential classification point here is that the response is a category, not a count.

A practical check is to imagine changing the labels while preserving their order. If a rating scale coded 1, 2, 3, 4, 5 were instead coded 0, 10, 20, 30, 40, the same response categories and ranking would remain. A category-based analysis can preserve that information. A count interpretation, or an analysis of the arithmetic mean as though code differences were measured units, would change substantially. By contrast, recoding an actual count from events to tens of events changes the unit of measurement and requires a meaningful conversion, not merely relabelling categories.

Worked Example: Identical Codes, Different Response Meaning

Suppose two hypothetical groups each have five recorded responses: 0, 1, 2, 3, and 4. In Group A, the values count the number of coughing episodes during a fixed observation period. In Group B, the values are ratings from “none” (0) to “very severe” (4). The numerical data are identical, but the response definitions differ.

For Group A, the total number of episodes is \(0+1+2+3+4=10\), and the sample mean count is \(10/5=2\) episodes per observation. If the Poisson model were being considered, its support would include all nonnegative integers, and its likelihood for these hypothetical observations at a common mean \(\lambda\) would be

L(\lambda)=\prod_{i=1}^{5}\frac{e^{-\lambda}\lambda^{y_i}}{y_i!} =\frac{e^{-5\lambda}\lambda^{10}}{0!\,1!\,2!\,3!\,4!}.

That calculation treats each response as an event total. It does not establish that Poisson assumptions hold; the observation design and conditional count variation would still need to be considered.

For Group B, the same arithmetic gives a mean code of \(2\), but that number is not automatically an average severity measured in equal-spaced units. The entries indicate which ordered categories were selected. A model should represent category membership and ordering rather than regard “4” as four episodes or assume that the change from rating 0 to 1 has the same measured size as the change from 3 to 4.

The example shows why examining the response definition changes the statistical interpretation even when every recorded integer is the same.

A Short Justification: Why Coding Alone Cannot Choose the Family

The distinction follows from what a likelihood assigns probability to. For counts \(y_1,\ldots,y_n\), the Poisson likelihood under a common mean \(\lambda\) is

L(\lambda)=\prod_{i=1}^{n}\frac{e^{-\lambda}\lambda^{y_i}}{y_i!}, \qquad y_i\in\{0,1,2,\ldots\}.

This expression uses the response values as event totals. For ordered categories \(k=1,\ldots,K\), write \(p_k\) for the probability of category \(k\), with \(\sum_{k=1}^{K}p_k=1\). If \(n_k\) responses fall in category \(k\), the category likelihood, apart from a factor that does not depend on the probabilities, is

L(p_1,\ldots,p_K)\propto\prod_{k=1}^{K}p_k^{n_k}.

Here the response is category membership; the integers name categories. The two likelihoods answer different questions. The Poisson likelihood assigns probability according to a count-valued process, while the category likelihood assigns probabilities to response levels. Replacing category codes with a different increasing sequence leaves the category identities and ordering intact, but it changes the numbers used as count values. Therefore, the printed integers cannot by themselves determine which likelihood describes the response.

State what one response records.
Describe the observation in words: an event total, a success count among opportunities, or a selected rating category.

List the possible values and their meanings.
Check whether values are restricted to a finite set of categories or arise as counts. A finite range alone does not settle the question; ratings can be finite, and counts can have known limits.

Match the family to the response process.
Use a count family only when the response represents a count and its distributional assumptions are plausible. Use an ordinal response formulation when the values are ordered categories.

Then address clustering.
Specify how observations within a cluster are related, for example through random effects, without changing the response’s basic meaning.

Worked Classification Examples

Worked Example: A Five-Level Symptom Rating

In a hypothetical clinical monitoring exercise, each participant reports symptom severity once per day on a scale coded 0, 1, 2, 3, or 4. The labels are none, mild, moderate, severe, and very severe. For one participant the daily codes are 1, 2, 2, and 4.

Classification. These are ordered ratings. The values identify severity categories; they do not count symptom episodes. The possible response set contains five categories, and the stated labels establish their order.

Why a count interpretation fails. The sum \(1+2+2+4=9\) does not mean that nine symptoms occurred. Likewise, the mean code \(9/4=2.25\) is an arithmetic summary of labels, not automatically an estimate of average severity on a scale with equal-sized units.

Model implication. If the goal is to model the recorded responses, the response family should treat them as ordered categories. A Poisson model would instead treat 4 as four occurrences and assign a count-process meaning that the measurement design does not support. Repeated days or participants may require a dependence structure, but that does not alter this classification.

Worked Example: Counting Events Rather Than Rating Them

In a hypothetical laboratory exercise, an observer records the number of equipment alarms in each of six equal-length sessions. The counts are 0, 0, 1, 1, 2, and 2 alarms. Each unit is one alarm during a session, so these values are event totals.

The total is \(0+0+1+1+2+2=6\), and the sample mean is \(6/6=1\) alarm per session. The sample variance is

s^2=\frac{(0-1)^2+(0-1)^2+(1-1)^2+(1-1)^2+(2-1)^2+(2-1)^2}{6-1} =\frac{4}{5}=0.800.

The mean and variance here are descriptive quantities in alarms per session and squared alarms per session, respectively. Their values alone do not prove that a Poisson model is suitable: a Poisson distribution imposes particular conditional mean-variance behavior, and the data-generating assumptions must also be plausible. But the response is genuinely a count, so count likelihoods are candidates in a way they were not for the severity ratings.

Worked Example: The Same Word “Score” Can Hide Two Measurements

Consider two hypothetical assessments, both reported on integer scores from 0 to 6. Assessment A records how many of six specified tasks were completed successfully. Assessment B asks a reviewer to assign one of seven ordered quality categories coded 0 through 6.

For Assessment A, a score of 4 means four successes among six opportunities. A binomial response is a natural candidate if the six opportunities and the assumptions behind the binomial model are appropriate. The score has a count interpretation tied to a known number of opportunities.

For Assessment B, a score of 4 means the reviewer selected category 4. It does not mean four quality events occurred. Its finite integer coding and the shared word “score” do not make it a binomial or Poisson response. The ordered categories should be modeled as categories.

The concrete measurement process, not the range from 0 to 6, distinguishes the two responses. Whether the two assessments use the same numerical storage format is irrelevant to that conclusion.

Assumptions, Diagnostics, and Failure Modes

Response-family choice should be made alongside the assumptions of the proposed GLMM. For example, a conditional Poisson model assumes that, given its predictors and random effects, each response is a Poisson event count with a specified conditional mean. The model also requires that the possible observed values fit its support. A category model requires that the recorded values correspond to its defined categories. At the GLMM level, the assumed conditional structure and the way clusters are represented must also be reasonable.

Useful checks begin with the data dictionary and collection protocol, not with a histogram. Confirm the operational definition of one response, the coding rules, whether zero is a real response or a special missing-value code, and whether the maximum is a physical or design limit. Tabulate observed values and compare them with the declared response set. Then check whether any proposed family assigns probability to impossible responses or excludes observed ones.

After fitting, residual patterns, predicted category or count frequencies, and unusual concentration at particular values can reveal model misfit. They cannot establish that the response was classified correctly in the first place. For instance, a Poisson model may return estimates for ratings coded as 0–4, but successful computation does not make those ratings event counts. Conversely, a finite set of observed counts does not prove the data are ordinal merely because the sample contains no large values.

A mismatch can also distort interpretation. A coefficient in a count model is tied to the expected event rate or count under the chosen link; applied to category codes, that interpretation is unsupported. Treating ratings as a numerical response can impose equal spacing and yield an effect in “code units” that the measurement scale may not justify. Cluster random effects address clustering under their model assumptions; they do not repair either problem.

Common Pitfalls and Alternatives

“It is an integer, so it is a count.” Integers are a storage format. Verify that each unit records one event or success, rather than a category label.
“It only ranges from 0 to 4, so it cannot be a count.” A count may be limited by observation design or opportunities. Ask how the value was generated; do not infer the response type from the range alone.
“The mean code is meaningful because the codes are ordered.” Ordering alone does not establish equal distances between categories. State what numerical differences mean before interpreting arithmetic summaries.
“A random intercept makes a count model suitable for ratings.” Random effects address dependence or heterogeneity represented by the model. They do not change the response from an ordered category to an event total.

The main alternative depends on the response process: an ordinal model for ordered ratings; a Poisson or another suitable count family for event totals; and a binomial model for successes out of known opportunities. These are not interchangeable methods selected by whichever produces the most convenient output. The model must correspond to the measurement, its support, and the scientific question. The introductory material on Poisson and binomial response families provides the relevant count-family background; here the key decision is whether the recorded integers are counts at all.

Key takeaway: Determine what each response means before choosing a GLMM family. Integer-valued ordered ratings are categories; event counts are totals. Their numerical labels may look alike, but their supports, likelihoods, and interpretations are not interchangeable.

Check Your Understanding

For each item, identify the response type and give the reason. Brief solutions follow the questions.

A nurse records each patient’s nausea as none, mild, moderate, or severe, coded 0 through 3. Is this an event count or an ordered rating?
A technician records the number of machine failures during each shift. The observed values happen to be only 0, 1, and 2. What feature determines the response type?
Two scales use codes 1–5: one counts successful steps out of five, and one records five ordered satisfaction categories. Why might their response families differ?
Why does adding a cluster random intercept not make an ordinal rating a count?

Solutions. (1) It is an ordered rating: the values name severity categories and do not count nausea events. (2) It is an event count because each value records the number of failures; the small observed range does not turn totals into categories. A suitable count family still requires its assumptions to be checked. (3) The first response counts successes among a fixed number of opportunities, making a binomial formulation a candidate if its assumptions are reasonable; the second records ordered categories and requires a categorical ordinal formulation. (4) A random intercept represents cluster-level dependence or heterogeneity. It does not change what an individual response measures, so the response family must still match the rating categories.