Variables and Measurement Scales

A variable is any measured or recorded characteristic that differs among the individuals, samples, or observations in a study — see Types of Biological Data for the broader typology of what gets recorded in biology. Measurement scale refers to the mathematical structure imposed on the possible values of a variable: what relationships between values are defined, and consequently what arithmetic can be legitimately performed on them. Stevens (1946) proposed that every variable falls into one of four scales, each nested inside the next in terms of the operations it permits. The four scales, in order of increasing mathematical structure, are nominal, ordinal, interval, and ratio. Each scale is defined by which of four properties its values possess: (1) identity — values can be distinguished from each other; (2) order — values can be ranked from least to most; (3) equal intervals — the numerical distance between any two adjacent values represents the same real magnitude of difference; and (4) a true (absolute) zero — the value zero represents a genuine absence of the quantity, not merely an arbitrary reference point. Nominal scales have only identity. Ordinal scales add order. Interval scales add equal intervals. Ratio scales add a true zero. Because each scale inherits the properties of the ones before it, the scales form a strict hierarchy — you can always demote a ratio-scale variable to interval, ordinal, or nominal by discarding information (e.g., converting exact masses into "light/medium/heavy" categories), but you can never promote a nominal variable into a ratio-scale one without additional real information. The practical payoff of the classification is a hard constraint: only certain arithmetic operations are defined at each scale, and a statistic is only meaningful if it uses operations that are defined for that scale. Addition and subtraction require equal intervals; multiplication, division, and ratios additionally require a true zero. This is why you can compute a mean body mass (ratio scale — arithmetic works) but not a meaningful mean of tumor grades I, II, III (ordinal — the "distance" between grade I and II is not guaranteed equal to the distance between II and III), and why a temperature of 20°C is not "twice as hot" as 10°C (interval scale — the zero is arbitrary, fixed by the freezing point of water rather than by the absence of thermal energy).

Step 1 — Nominal scale: names only. A nominal variable assigns each observation to one of several unordered categories that differ in kind, not amount. Examples: blood type (A, B, AB, O), species, sex, genotype at a biallelic locus (AA, Aa, aa treated purely as labels), habitat type (forest, grassland, wetland). The only legitimate operations are checking equality (is this observation the same category as that one?) and counting. You can tally frequencies, compute proportions, and report the mode (the most frequent category) — but you cannot rank the categories, subtract them, or average them. Assigning numeric codes (1 = forest, 2 = grassland, 3 = wetland) for data-entry convenience does not change the underlying scale; computing "the mean habitat" as 2.3 is meaningless, because the numbers are arbitrary labels, not measured quantities.

Step 2 — Ordinal scale: order without equal spacing. An ordinal variable's categories have a natural rank, but the size of the gap between adjacent ranks is unknown and typically unequal. Examples: Braun-Blanquet plant cover classes, tumor stage (I–IV), a 5-point Likert agreement scale, dominance rank in a primate troop, pain score (0–10). Because order is defined, you can now compute medians and percentiles, and you can use rank-based statistics (Spearman correlation, Mann–Whitney U, Wilcoxon tests). But you still cannot legitimately compute a mean or a standard deviation, because those operations assume the numerical distance between category 1 and 2 equals the distance between category 3 and 4 — an assumption ordinal data explicitly do not satisfy. The jump from "mild" to "moderate" pain need not represent the same physiological change as "moderate" to "severe."

Step 3 — Interval scale: equal units, arbitrary zero. An interval variable has units of equal, known size, so subtraction is meaningful — a difference of 5 units means the same magnitude of change anywhere on the scale. But the zero point is a convention, not an absence of the quantity. Examples: temperature in Celsius or Fahrenheit, pH, calendar year, standardized test scores (e.g., IQ). You can compute means, standard deviations, and correlations, and differences are interpretable ("this reaction is 10°C warmer than that one" is a valid statement). But ratios are not: 20°C is not twice as hot as 10°C, because 0°C does not mean "no heat" — it is simply the freezing point of water, an arbitrary reference. Multiplying or dividing interval-scale values, or computing a coefficient of variation on them, produces numbers that depend on the arbitrary choice of zero and are therefore not biologically meaningful.

Step 4 — Ratio scale: equal units, true zero. A ratio variable has equal units and a zero that genuinely represents the absence of the measured quantity. Examples: mass, length, age, enzyme reaction rate, number of offspring, heart rate, absolute (Kelvin) temperature, concentration. All arithmetic operations are valid: addition, subtraction, multiplication, division, geometric means, coefficients of variation, and log transformations. "This mouse weighs twice as much as that one" and "metabolic rate dropped to a quarter of baseline" are both meaningful statements, because zero mass and zero metabolic rate are real, physically meaningful states.

Choosing statistics follows directly from this hierarchy. Each scale permits every operation valid at the scales below it, plus one more:

Step 5 — Discrete versus continuous within ratio and interval scales. Ratio and interval variables can be further split into discrete (countable, integer-valued — number of offspring, number of alleles) and continuous (any value within a range — mass, enzyme activity, height). This distinction affects which probability distributions and graphical summaries are appropriate (e.g., Poisson vs. normal), a topic developed further under Frequency Distributions, but it does not change the scale classification itself: discrete ratio data still support all ratio-scale arithmetic.

Step 6 — Why the scale, not the number of digits, decides validity. A common trap is assuming that if a variable "looks numeric" (e.g., pain scored 0–10) it behaves like a ratio-scale measurement. It does not. The diagnostic question is always: does a one-unit difference mean the same thing everywhere on the scale, and does zero mean "none of it"? If either answer is no, the variable is ordinal or interval, not ratio, regardless of how the data are recorded.

Measurement scale theory was not derived from a single wet-lab experiment; it is a conceptual and mathematical framework, formalized in a landmark 1946 Science paper, "On the Theory of Scales of Measurement," by the Harvard psychophysicist Stanley Smith Stevens. The "evidence" for the framework comes from Stevens's own research program in psychophysics — the quantitative study of the relationship between physical stimuli and perceived sensation — which had spent the preceding decade demonstrating that different sensory judgments obey fundamentally different mathematical rules, and that treating them identically produced contradictions. Stevens and collaborators ran magnitude-estimation experiments: subjects were presented with physical stimuli (tones of varying loudness, lights of varying brightness, weights of varying mass) and asked either to rank them, to judge whether the difference between two stimuli was equal to the difference between two others, or to estimate a stimulus as a numerical multiple of a reference stimulus ("this tone sounds three times as loud as the reference"). The key empirical finding was that for stimuli like brightness and loudness, subjects could reliably perform ratio judgments — reporting consistent multiples across repeated trials and across different reference stimuli — while for other perceptual judgments (e.g., subjective "goodness" of a color, or apparent complexity), subjects could reliably rank stimuli but could not consistently agree on equal-interval or ratio judgments. This showed empirically that not all measurements are created equal: some support full arithmetic, others only support ordering, and forcing the same statistical treatment onto both produces judgments that fail to replicate. Stevens's 1946 paper crystallized these empirical patterns into the four-scale taxonomy still taught today, explicitly tying each scale to a permissible group of mathematical transformations — a formalism borrowed from the mathematician's notion of invariance. A nominal scale is invariant under any one-to-one relabeling of categories (calling species "A" instead of "1" changes nothing). An ordinal scale is invariant under any monotonic (order-preserving) transformation. An interval scale is invariant only under linear transformations (y = a + bx, such as converting Celsius to Fahrenheit). A ratio scale is invariant only under multiplication by a positive constant (changing units of mass from grams to kilograms, but not shifting the zero). This transformation-invariance criterion is the rigorous version of the "what operations are meaningful" heuristic: a statistic is meaningful only if it gives the same substantive conclusion regardless of which permissible transformation was applied to the raw data. The mean of Celsius temperatures converts correctly to the mean of the corresponding Fahrenheit temperatures (both are legitimate, linearly related interval statistics) — but the ratio of two Celsius temperatures does not correspond to the same ratio in Fahrenheit, proving empirically and mathematically that "ratio" statements are illegitimate on an interval scale. Subsequent statisticians (notably Frederick Lord, in a widely cited 1953 rebuttal "On the Statistical Treatment of Football Numbers") challenged Stevens's strict prohibitions, pointing out that arithmetic operations are computable regardless of scale and that the real question is whether the resulting inference is valid given the data-generating process — a debate that persists in statistical methodology. In practice, biology has settled on a pragmatic middle ground: Stevens's four-scale framework is retained as the standard classification for choosing appropriate descriptive statistics and nonparametric versus parametric tests, while researchers are encouraged to think about the substantive meaning of a statistic (not just its formal permissibility) — a matter taken up in detail when choosing the right summary statistic for each data type.

Example 1 — A field ecology dataset with all four scales at once. A stream ecologist samples 40 sites and records, for each site: (a) dominant substrate type (bedrock, cobble, sand, silt) — nominal; (b) a visual water clarity rank on a 5-point scale (1 = very turbid to 5 = very clear) — ordinal; (c) water temperature in °C — interval; (d) dissolved oxygen concentration in mg/L — ratio.

Suppose the ecologist wants to know whether clearer water is associated with higher dissolved oxygen. Because clarity is ordinal, the correct approach is a Spearman rank correlation between clarity rank and DO concentration — not a Pearson correlation, which would implicitly assume the 1-unit gap between clarity ranks 2 and 3 equals the gap between ranks 4 and 5, an assumption the visual rating scale cannot support. If the ecologist further wants to summarize temperature across sites, the mean is valid (interval scale supports subtraction and averaging), but she cannot say "site A's water is 1.5 times as warm as site B's" even if the mean at A is 18°C and at B is 12°C, because 0°C is not "no heat." Converting to Kelvin (291 K vs. 285 K) would make a ratio statement almost meaningless too, since the relative difference (291/285 ≈ 1.02) barely resembles the naive Celsius ratio (18/12 = 1.5) — a direct demonstration that the "1.5×" claim was an artifact of an arbitrary zero, not a real physical relationship.

Example 2 — Clinical trial: ordinal outcome mistakenly analyzed as interval. A pain-management trial randomizes 120 post-surgical patients (asymmetric groups: 80 to a new analgesic, 40 to standard care) and records pain on an 11-point numeric rating scale (0 = no pain, 10 = worst imaginable) at 24 hours. Raw group means are:

• New analgesic (n = 80): mean pain score = 3.1, SD = 1.8
• Standard care (n = 40): mean pain score = 4.4, SD = 2.0

A naive analysis runs a two-sample t-test on these means and reports p = 0.002, concluding the new drug reduces pain by "1.3 points," and — worse — a downstream reviewer claims patients on standard care felt "42% more pain" (4.4/3.1 ≈ 1.42). Scale analysis exposes two errors. First, the pain scale is ordinal: the physiological/psychological difference between "0" and "1" is not guaranteed to equal the difference between "8" and "9" (ceiling effects are common near the top of such scales), so computing a mean and SD, while numerically possible, assumes equal intervals the instrument was never validated to provide. A more defensible approach reports medians and interquartile ranges and tests the group difference with a Mann–Whitney U test (equivalently, Wilcoxon rank-sum), which uses only the rank information the scale actually supports. Second, the "42% more pain" claim compounds the error by treating the scale as ratio — implying "0" means zero pain sensation in some absolute physical sense and that 8 is twice as painful as 4, claims the 11-point scale was never designed to license. In practice, many pain researchers treat 0–10 numeric rating scales as approximately interval for large samples because parametric tests are often robust to this violation and results converge with nonparametric alternatives — a pragmatic compromise, but the study should report which choice was made and ideally show both analyses agree, rather than silently asserting ratio-level claims the data cannot support.

Clinical medicine. Every clinical trial reporting a Likert-type quality-of-life score, a tumor response grade (complete/partial/stable/progressive disease under RECIST criteria), or an Apgar score must decide — and justify — whether to treat that ordinal outcome with rank-based nonparametric statistics or with parametric methods under an approximate-interval assumption; regulatory agencies increasingly require this justification in trial protocols.

Genomics and molecular biology. Genotype calls (homozygous reference, heterozygous, homozygous alternate) are nominal, while gene expression levels (ratio-scale, log-transformed counts) support full parametric statistics — mixing these scales incorrectly in a single regression model is a common source of statistical error in genome-wide association studies.

Ecology and conservation. Vegetation surveys routinely use ordinal cover-abundance scales (Braun-Blanquet, Domin), and species-richness counts are ratio-scale; conservation status categories (IUCN Red List: Least Concern through Extinct) are ordinal, so computing a "mean conservation status" across a region is a scale violation that nonetheless appears in some poorly reviewed reports.

Agriculture and animal science. Body condition scoring in livestock (typically a 1–5 or 1–9 ordinal scale) is routinely — and controversially — averaged across herds; researchers who instead report medians and use ordinal logistic regression produce more defensible management recommendations.

Public health surveillance. Disease severity staging, socioeconomic categories, and self-reported health status (a very commonly used ordinal instrument) all require scale-appropriate summary statistics before being fed into models that inform resource allocation, illustrating why this seemingly basic classification step underlies decisions with real population-level consequences.

Ask the tutor: If I have a 7-point Likert scale with roughly symmetric responses across 500 participants, is it defensible to report means and use a t-test instead of a Mann–Whitney U test?