Discrete Data

Discrete data is numerical data restricted to distinct, separate values with no meaningful values in between — the number of defects on a production line, the number of emails in an inbox, the number of goals scored in a match. The defining property is countability: discrete data can always in principle be listed one by one (0, 1, 2, 3...), and a value like 2.5 defects is not just unusual but nonsensical, which is a stronger constraint than continuous data carries, where any value between two points is a legitimate possibility.

Discrete data is what count-based probability distributions were built for, and using the right one changes what an analysis can honestly claim. The Poisson distribution models the number of events occurring in a fixed interval of time or space when those events happen independently and at a constant average rate — customer arrivals per hour, typos per page, server errors per minute. The binomial distribution models the number of successes across a fixed number of independent yes/no trials — defective units in a batch of 100, conversions out of 500 visitors. Fitting a normal distribution (built for continuous data) to a discrete count variable is a common shortcut that works reasonably well when counts are large, but breaks down visibly for low counts — a Poisson-distributed variable with a mean of 0.3 events is heavily right-skewed and mostly zeros, nothing like the symmetric bell curve a normal-distribution assumption would impose on it.

Discrete data also includes a subtype worth separating out: binary or Boolean data, where the only two possible values are typically coded 0 and 1 — did the customer convert, did the part fail, did the email get opened. Binary discrete data is common enough and statistically distinct enough to have its own dedicated modeling family, most notably logistic regression, which predicts the probability of the 1-outcome rather than attempting to predict 0 or 1 directly as if it were a continuous quantity — a distinction that matters because treating a binary outcome as continuous and fitting ordinary linear regression to it can produce predicted "probabilities" below 0 or above 1, values that are mathematically possible under linear regression but meaningless for a probability.

Aggregating discrete data correctly requires care that aggregating continuous data mostly doesn't: summing counts across groups is always valid, but averaging counts across groups of very different sizes without normalizing by exposure (the time window, population size, or number of trials each count was observed over) produces comparisons that look like apples-to-apples but aren't. Comparing "12 errors in store A" to "9 errors in store B" is meaningless without knowing store A processed three times the transaction volume of store B — the correct comparison is a rate (errors per thousand transactions), not the raw discrete count, a normalization step that's easy to skip because the raw counts look directly comparable at a glance.

Read discrete data as the subset of numerical data defined by what it can't be — fractional, continuous, infinitely subdividable — and that single constraint is what routes it to its own family of distributions (Poisson, binomial), its own regression method for the binary case (logistic regression), and its own aggregation discipline, where a raw count is almost never directly comparable across groups without first converting it into a rate.