Ratio Scale

The ratio scale is the highest level in Stanley Smith Stevens's 1946 typology: values are ordered, equally spaced, and anchored to a true zero that means the complete absence of the quantity, so ratios between values are meaningful. Mass, length, duration, counts of objects, electrical current, and temperature in kelvins are ratio-scaled: 40 kilograms really is twice 20 kilograms, and a stopwatch reading of 0 seconds means that no time has elapsed.

The only permissible transformation on a ratio scale is multiplication by a positive constant, y = c × x, which changes the unit but never the origin — and that is exactly why ratio statements survive every change of units. Stevens called this the similarity group. Converting kilograms to pounds, meters to feet, or minutes to seconds rescales every value by the same factor, so "twice as heavy" stays true while the numbers change. Compared with the interval scale, the fixed zero is the whole difference: temperature in Celsius allows addition of shifts but not ratios, while kelvin temperature, with its zero at absolute zero (about minus 273.15 degrees Celsius), supports a statement like "a gas at 600 kelvin has twice the thermal energy per molecule of one at 300 kelvin."

Everything computable on the lower scales remains valid, and the ratio scale additionally licenses the multiplicative statistics: ratios, percentage change, the geometric and harmonic means, and the coefficient of variation. The coefficient of variation, the standard deviation divided by the mean, lets spread be compared across variables in different units, such as the variability of body mass in mice against that in elephants, and it is meaningful only because the zero is genuine. Fold change and log ratio, ubiquitous in genomics and finance, depend on this scale, as do growth rates, elasticities in economics, and the log transformation of positively skewed ratio data such as incomes or concentrations, which turns multiplicative effects into additive ones.

Ratio data is the norm in the physical sciences, engineering, medicine, economics, and business analytics. Dose, concentration, body weight, distance, price, revenue, and counts of events or items are all ratio-scaled, so models can use multiplicative structure freely: log-log regressions estimate elasticities, count models use rate ratios, and relative risk compares an event probability in two groups. Many machine learning features are ratio-scaled and highly right-skewed, which is why log or Box-Cox transforms are common preprocessing, and why relative error measures such as mean absolute percentage error are interpretable (while blowing up near a true zero, a practical hazard in demand forecasting).

The practical difficulty is recognizing a false zero and avoiding both over-promotion and needless demotion of scales. A zero on a Likert item or on a Celsius thermometer is a label or a convention, so computing ratios there is unsound. Conversely, some quantities that look awkward are still ratio-scaled: a count of children with a legitimate zero qualifies, even though it is discrete. Another pitfall is that a ratio scale does not guarantee a normal distribution or a linear relationship, since ratio data is frequently skewed and bounded below by zero. Stevens's four-level scheme, though influential, was challenged by Frederic Lord in 1953 and by Velleman and Wilkinson in 1993 for being too rigid, so the choice of statistic should also follow the question and the distribution, not the scale label alone.

Read the ratio scale as measurement with a genuine nothing at its foot: because zero truly means none, the numbers can be multiplied, divided, compared as percentages, and rescaled by any positive factor without changing what they say, which makes it the level where nearly every statistical operation is permitted.