Why Check the Response Support First?
Before choosing a response family for a generalized linear mixed model, establish what values the response can actually take. A number stored in a data column is not necessarily a count: it might be a measured quantity, a proportion, or a numeric code for an ordered category. These responses have different supports and meanings, so treating them as interchangeable can make a model inappropriate even if software accepts the data.
A count records how many times a defined event occurred for an observational unit during a specified period, area, or other scope. The word how many matters: the recorded value is a whole-number tally of occurrences, not a measurement on a continuous scale and not a fraction of a total. In a GLMM, identifying this structure is an initial classification step. It does not by itself establish which count distribution is suitable.
Support Is a Property of the Response, Not Its Data Type
For a discrete response, its support is the set of values that can have positive probability under the model. A count support contains zero and nonnegative integers, although a particular study design may restrict which of those integers are possible. A continuous measurement, by contrast, can in principle take fractional values on a measurement scale. A proportion lies between zero and one and represents a numerator relative to a denominator. An ordered response has categories with a ranking, but the numerical labels assigned to those categories do not create a counting scale.
This distinction can be stated as a simple validity check. If a model assigns zero probability to a value that was observed, then that model cannot give a positive likelihood to the data. For example, a discrete count model assigns probability only to the values in its count support. If the recorded response is 1.7 milligrams, it is not made into a count by placing it in a numeric column; it is a measurement that a count likelihood does not describe.
Here, \(\mathcal{S}\) is the support specified by the response model. In a mixed model, this support condition applies to the response distribution conditional on the random effects. If \(P(Y=y\mid b)=0\) for every possible value of the random effects \(b\), then averaging over those effects cannot make the probability positive:
This is why checking support comes before considering the effects structure or estimating coefficients. Random effects can represent dependence or heterogeneity, but they cannot turn an incompatible response into a valid count.
Worked Example: Identifying a Count From Event Records
Suppose, in a hypothetical service log, five sites record the number of equipment failures during the same one-month period. Their recorded values are \(0, 2, 1, 4, 3\). Determine whether these are counts and calculate their sample mean.
Each value is a tally of occurrences of the same defined event, equipment failure, for one site over a specified period. All observed values are nonnegative integers, including zero, so the response is a count. The values are not measurements of failure duration or a fraction of equipment that failed.
The sample mean is:
Thus the mean observed count is 2.0 failures per site during the one-month period. This calculation summarizes the records; it does not establish that a particular probability distribution is appropriate. In particular, being a count does not, by itself, justify a Poisson model or any other particular count family.
Counts, Measurements, Proportions, and Ordered Labels
The response’s meaning is as important as its appearance. Several types of responses can be written as numbers, yet differ in what those numbers represent. The following comparison is a practical first screen.
| Response type | What the values represent | Support or example | Count? |
|---|---|---|---|
| Event count | Number of occurrences in a defined scope | 0, 1, 2, ... failures in a month | Yes |
| Continuous measurement | Amount measured on a scale | Temperature in degrees or concentration in milligrams per litre | No |
| Proportion | Part divided by a total or denominator | 0.25 or 25 percent | No; it is a ratio |
| Ordered category | One ranked label among a set of categories | 1 = poor, 2 = fair, 3 = good | No; the labels encode rank |
A proportion may be calculated from a count, but it is not itself a count. For instance, three successes among twelve eligible trials produce the proportion \(3/12=0.25\). The numerator is a count; the proportion is a ratio. Retaining the denominator matters because a proportion of 0.5 from two trials and a proportion of 0.5 from two hundred trials do not convey the same amount of information. The earlier material on grouped binomial data covers the trial-count representation when it applies.
Likewise, category labels can be stored as integers without becoming counts. If the values 1, 2, and 3 mean “low,” “medium,” and “high,” the difference between 1 and 2 need not represent the same amount as the difference between 2 and 3. Their order is meaningful; arithmetic distances generally are not.
Worked Example: Separating Counts From Proportions
In a hypothetical audit, two wards have 3 and 5 incomplete records among 12 and 20 eligible records, respectively. A report lists both the number incomplete and the fraction incomplete. Classify each response and calculate the two fractions.
The numbers of incomplete records, 3 and 5, are counts: each is a tally of occurrences among a defined set of records. The fractions are:
Each fraction is a proportion, not a count. The equal proportions do not make the underlying data identical: the first is based on 12 eligible records and the second on 20. If only the two proportions were retained, the denominators would no longer be visible in those values. Therefore, determine whether the analysis concerns the number of events or the fraction of eligible units with the event before selecting a response representation.
Worked Example: Ordered Scores Are Not Event Counts
Suppose five hypothetical respondents choose a service rating coded from 1 (“poor”) to 4 (“excellent”), giving \(1, 2, 2, 3, 4\). A separate instrument records temperature in degrees Celsius as \(18.4, 19.1, 20.0, 20.6, 21.3\). Classify both responses and examine what their numerical values mean.
The ratings are ordered category labels. Their mean can be computed arithmetically:
But 2.4 is not an event count, and its interpretation depends on treating the gaps between rating categories as meaningful and comparable. The coding alone does not establish that assumption. By contrast, the temperature values are measurements; decimal values are possible and meaningful on the Celsius scale. Neither response is a tally of occurrences, so neither should be classified as a count merely because the recorded values are numbers.
A Support-First Classification Process
Use the following sequence before choosing a response family. It keeps the question focused on the observations and their possible values rather than on the software format or the appearance of a summary table.
State what one response describes and over what period, area, or set of opportunities it is recorded.
For a count, one unit must represent one occurrence of a clearly defined event. Identify what qualifies as an occurrence and check that it is counted consistently.
Ask whether zero and nonnegative integers are the relevant values, and whether the design imposes a maximum. Fractional measurements, ratios, and rank codes indicate different response structures.
Once the response is identified as a count, assess the study design and model assumptions before selecting a particular count distribution.
A fixed number of eligible opportunities can impose an upper bound on a tally. That fact is important: a count with a known maximum is not automatically interchangeable with a count that has no fixed maximum. It remains a tally, but the opportunity structure can affect which family is suitable. The count-support classification is therefore a necessary first step, not a complete model-selection rule.
Diagnostics, Alternatives, and Common Errors
Several checks can reveal that the response has been misclassified or recorded inconsistently. Inspect the raw values, not just a mean or a software variable type. For a proposed count, look for negative values, fractional values, and values that exceed the number of opportunities available. Each requires investigation: it may indicate a data-entry problem, an unclear event definition, a different response scale, or a mismatch between the stated design and the data.
Also check for inconsistent units of observation or scope. A count of events per day and a count of events per month do not describe the same exposure window. Likewise, one row might count events for a person while another counts them for a site. Such differences can make a column look like a single response when its values do not have a consistent interpretation. These checks diagnose the response definition; they do not substitute for assessing the assumptions of a fitted model.
The main alternatives depend on what the response means. For a continuous measurement, use a model suited to a measurement scale rather than forcing the values onto integer support. For a proportion formed from eligible trials, preserve the numerator and denominator when the design calls for a trial-based representation. For ordered categories, use a representation that respects category order instead of assuming that numeric codes are equally spaced measurements. These are different response structures, not competing ways to label the same count.
Support screening is not a complete diagnostic assessment. A valid count response can still be paired with an unsuitable count distribution, an incorrect dependence structure, or an inadequate description of how observation opportunities differ. Conversely, fractional values do not automatically indicate bad data: they may correctly be proportions or measurements. The goal is not to force every response into a count representation, but to select a model whose support and interpretation match the recorded outcome.
Check Your Understanding
For each question, identify the response structure and explain the key reason.
- A hypothetical laboratory records the number of contamination events in each sample during a fixed inspection period. The values are 0, 0, 1, and 3. Is this a count?
- A report records 0.40 as the fraction of eligible items that failed inspection. Is 0.40 a count? What information would help identify the underlying tally?
- A survey records pain as 1 = mild, 2 = moderate, and 3 = severe. Why is this not a count even though all recorded values are integers?
- A device records a chemical concentration of 2.6 milligrams per litre. Should a count model be chosen because the value is positive?
- For a proposed count response, the data include a value larger than the number of opportunities available to that observational unit. What should be checked?
Solutions. (1) Yes. Each value is a nonnegative-integer tally of defined events for a sample over a stated period. (2) No. It is a proportion. The numerator—the number of failed items—and the denominator—the number of eligible items—are needed to recover the underlying count and its opportunities. (3) The numbers label ordered categories; they do not count occurrences, and their spacing is not established by the coding. (4) No. Concentration is a measurement, and the decimal value is meaningful on that measurement scale. (5) Check the event definition, denominator or opportunity count, observational unit, and data entry. A value exceeding the stated opportunities is inconsistent with that design and should be resolved before selecting a response family.