Discrete Variable

A discrete variable is a numerical variable that can take only separate, countable values — typically the whole numbers 0, 1, 2, 3 — with gaps between neighboring values in which nothing can occur. The number of defective chips in a batch, goals in a match, children in a household, or emails received in an hour are discrete: there is no household with 2.4 children, and the possible outcomes can be listed one by one even if the list never ends.

The mathematical signature of a discrete variable is that probability sits on individual points, described by a probability mass function that assigns a positive probability to each possible value and sums to exactly 1. Because each value has its own probability, questions such as "what is the chance of exactly 3 defects?" have direct answers, and the probability of a range is found by adding up the point probabilities. The cumulative distribution function climbs in steps rather than smoothly, jumping at each possible value. This is the basic contrast with a continuous variable, whose probability is spread over intervals; the two kinds of variable need different tools (sums versus integrals), even though most of the intuition about center and spread carries over.

A handful of named distributions cover most discrete variables in practice, each tied to a story about how the counts arise. The binomial distribution counts successes in a fixed number of independent yes-or-no trials and traces back to Jacob Bernoulli's work published in 1713. The Poisson distribution, introduced by Siméon Denis Poisson in 1837, counts events in a fixed window of time or space when they occur independently at a constant average rate, and its mean equals its variance; a famous illustration is Ladislaus Bortkiewicz's 1898 analysis of Prussian soldiers killed by horse kicks, which fit that model closely. The geometric and negative binomial distributions count trials or failures until a success threshold, and the hypergeometric covers sampling without replacement.

Discrete outcomes call for models built for counts, which is where applications in data science and science concentrate. Poisson and negative binomial regression model counts of claims, visits, or mutations as functions of predictors; logistic regression handles the binary case; classification in machine learning predicts a discrete label, and metrics like the number of misclassified examples are discrete too. In queueing, reliability, and genetics, discrete models describe arrivals, failures, and allele counts. A plain linear regression on a small count can produce negative or fractional predictions, which is the practical reason for these specialized models.

Two pitfalls recur: assuming that discrete means few values, and assuming that whole-number storage means discreteness. A count can run into the millions and, when large, is routinely approximated by a continuous normal curve, as in the de Moivre-Laplace theorem of the 1730s and 1810s for binomial counts, usually with a continuity correction. Conversely, a measurement of height rounded to the nearest centimeter is stored as integers but is a continuous quantity recorded coarsely. A further trap is the Poisson assumption itself: real event counts are often clustered, with variance exceeding the mean, so a model assuming equality will report standard errors that are too small.

Read a discrete variable as something that comes in countable steps: every possible outcome can be named, each carries its own probability, and the modeling work is mostly about choosing the counting mechanism — fixed trials, steady arrivals, waiting for a success — that actually generated the numbers.