Numerical Data

Numerical data is data expressed as numbers that measure a quantity — height, revenue, temperature, count of items sold — and the defining property that separates it from categorical data is that arithmetic on it is meaningful: averaging two heights produces a sensible third height, while averaging two zip codes produces nothing interpretable even though both are stored as numbers. This arithmetic validity is what unlocks the entire toolkit of classical statistics — means, variances, correlations, regression — none of which has a coherent definition on data that isn't numerical in this sense, which is why the numerical-versus-categorical distinction is usually the first branch point in deciding how to analyze any dataset.

Numerical data splits into discrete and continuous, and the split determines which distributions, charts, and tests are appropriate. Discrete numerical data takes distinct, separate, typically countable values — the number of items in a cart, the number of support tickets filed, where fractional values like 2.5 tickets are nonsensical. Continuous numerical data can in principle take any value within a range, limited only by measurement precision — weight, duration, temperature, where the gap between any two values can always be subdivided further. The practical consequence: discrete data is often modeled with distributions built for counts (Poisson, binomial), while continuous data is modeled with distributions built for measurements (normal, exponential), and applying a continuous-data technique to discrete data (or vice versa) tends to produce technically-running but statistically inappropriate results — treating a 0-to-5-star rating as continuous and computing a mean of 3.4 stars is common practice and usually fine in aggregate, but the same move on a true count variable with a heavily skewed distribution (most customers file 0 tickets, a few file dozens) can badly misrepresent the typical case.

Numerical data also carries a measurement-scale distinction that matters less often but can quietly break an analysis when ignored: interval versus ratio scale. An interval scale has equal spacing between values but no true zero point — temperature in Celsius is interval, because 0°C doesn't mean "no temperature," which is why saying 20°C is "twice as hot" as 10°C is statistically invalid even though the numbers suggest it. A ratio scale has equal spacing and a true zero that means "none of the quantity" — weight, income, and count data are ratio scale, where doubling the number does mean doubling the quantity. Multiplicative claims ("twice as much," "three times higher") are only valid on ratio-scale data; applying them to interval-scale data is a specific, recurring statistical error that looks reasonable on the surface because both scales look identical as plain numbers.

Numerical data is also the type most directly exposed to a class of errors that categorical data mostly avoids: unit inconsistency and precision loss. A dataset combining revenue figures recorded in different currencies, or distances recorded in a mix of miles and kilometers, produces numerically valid but substantively wrong aggregates unless every value is normalized to a common unit before combining — an error that's invisible in the output because nothing about a wrong sum looks different from a right one. Precision loss happens on the storage side: numerical data stored with insufficient floating-point precision, or truncated to fewer decimal places than the underlying measurement supports, silently discards information that can matter disproportionately in aggregate calculations across millions of rows, even when any single value's rounding error looks negligible.

Read numerical data as the category of data whose numbers mean something arithmetically, not just structurally: the discrete-versus-continuous split decides which distribution and chart family applies, the interval-versus-ratio split decides which claims about "how much more" are actually valid, and the recurring real-world failure modes — mismatched units, precision loss — are errors that only numerical data is exposed to precisely because it's the one data type where the numbers are trusted to carry quantitative meaning.