Can the predicted count be 2.7?

Yes; an expected count can be fractional even though realized counts are integers.

Is a percentage count data?

A percentage is a ratio; retain its numerator and denominator when the modeling question depends on their sizes.

Are all zeros evidence of inactivity?

No; an active process can produce zero events by chance, while some zeros may arise from a distinct structural condition.

AI Glossary / Statistics and probability / Intermediate

Count data

Definition

Count data record how many events, objects, or observations occur within a defined unit, category, period, or amount of exposure. Ordinary counts are nonnegative integers, including zero. Examples include errors per batch, requests per hour, and respondents in each answer category. Their interpretation depends on what is counted and the opportunities for events to occur.

What it means

A count of twenty complaints means little without context if one service handled one thousand customers and another handled ten thousand. Counts and rates answer related but different questions: a count describes how many events occurred, while a rate relates those events to an exposure such as customer volume or operating time.
Not every integer-valued variable is a count. A category code, rank, or rounded temperature may use integers without representing a number of events. Counts can also be aggregated across categories or observed repeatedly for the same units, which affects the analysis.

How it works

Define the event and observational unit precisely. State whether repeat events from the same source count separately, which time window applies, and what constitutes exposure. Distinguish a genuine zero from a missing record, an unobserved period, or a unit that was never at risk of an event.
Inspect the distribution, zeros, large counts, exposure differences, and dependence. A Poisson model is a common starting point when its assumptions are appropriate; it has equal conditional mean and variance. A count outcome is not automatically Poisson merely because it is an integer.
In Poisson regression with a log link, the logarithm of the expected count is modeled using predictors. If counts scale proportionally with known exposure, the logarithm of exposure can enter as an offset with coefficient fixed at one. This models rates while retaining the count outcome. Zero exposure requires separate handling rather than taking its logarithm.
Extra conditional variation may motivate a negative binomial model, another variance specification, or changes to the mean model. Many zeros do not alone prove a separate zero-generating process; a standard count model with a low mean can already predict many zeros. Zero-inflated and hurdle models require substantive and statistical justification.
Counts of successes out of a fixed number of opportunities may be better modeled as binomial, while mutually exclusive category counts with a fixed total may call for a multinomial model. Match the model to the data-generating process and assess its fit and uncertainty.

Three practical examples

A service team analyzes the number of incoming requests per operating hour.
A manufacturer records defects per inspected area, distinguishing defect counts from the number of defective items.
An AI team tracks detected document errors while retaining document length or another relevant exposure measure.

Think of it this way

Think of a count as the tally marks in a ledger. To compare ledgers fairly, you also need to know how long each was open and what was eligible to be tallied. The same number can describe very different event frequencies when those opportunities differ.

Worked example

Suppose team A records twelve errors while reviewing three hundred documents, and team B records twenty errors while reviewing one thousand. If multiple errors per document are allowed, these are event counts with document exposure, not necessarily binomial counts of erroneous documents.
Team A's observed rate is 12/300 = 0.04 errors per document, or four per hundred. Team B's rate is 20/1,000 = 0.02, or two per hundred. Although B has more errors in total, A has twice the observed error rate.
Under an illustrative constant-rate Poisson model using A's rate as a fixed assumed parameter, the expected count for one hundred comparable documents is lambda = 100 x 0.04 = 4. The probability of no errors is exp(-4), approximately 0.018316. The probability of at least three errors is 1 - exp(-4)(1 + 4 + 4 squared/2), approximately 0.761897.
Those probabilities depend on the assumed model and treat the rate as known for the calculation. In practice, 0.04 was estimated from a finite sample, documents may differ, and errors within or across documents may be dependent. The descriptive rate ratio of two does not by itself establish a statistically reliable or causal team difference.

Illustrative case study

Fictional case: Copper Reports ranks departments by the number of flagged document errors. Larger departments appear worse because they process more material. Analysts add document volume and examine whether length and complexity also differ.
They preserve true zero counts and distinguish them from periods with no review activity. A count model is evaluated against the observed variability rather than assumed to fit automatically. The revised report shows both total workload burden and exposure-adjusted error frequency, allowing managers to ask whether they need more capacity, better quality, or both.

Common misconceptions and mistakes

Do not compare raw counts as rates when exposures differ. Do not replace missing observations with zeros without justification. An observed count must be an integer, but its expected value can be fractional. A sample variance exceeding the sample mean is not, by itself, proof of a particular alternative model, especially before accounting for predictors. Taking logarithms of counts also creates problems at zero and is not equivalent to a standard log-link count model.

Business use

Count analysis supports staffing, quality monitoring, incident review, and workload forecasting. Maintain event definitions and exposure data alongside the counts. Choose whether the decision concerns total demand, frequency per opportunity, or the chance of exceeding a capacity threshold. For AI-derived counts, evaluate detection errors as well as the statistical model, since missed or duplicated events can distort the apparent process.

Check your understanding

Which team has the higher observed rate in the example?

Answer: Team A, despite its smaller raw count.

Why is exposure needed?

Answer: It identifies the opportunities associated with the events.

What assumption underlies the displayed Poisson probabilities?

Answer: A constant rate of 0.04 per comparable document with the relevant Poisson event model, treated as known for illustration.

Related terms

Poisson distribution
negative binomial distribution
exposure
rate
offset
overdispersion
zero-inflated model
binomial data

Learn next

Compare counts, exposures, and rates in a small operational dataset. Then fit a suitable mean model, examine conditional dispersion and zeros, and assess how clustering or unequal exposure changes uncertainty.

Sources

David C. Howell, Statistical Methods for Psychology glossary, count data: https://www.uvm.edu/~statdhtx/methods8/Glossary/Glossary.html UCLA OARC, Poisson Regression, modeling counts and exposure: https://stats.oarc.ucla.edu/r/dae/poisson-regression/

Previous: Costmap Next: Count nouns