What levels of measurement are and why they matter
Every variable you collect — an ORCID-linked demographic field, a Likert-scale survey item, a lab reading in degrees Celsius — carries with it a level of measurement (also called a scale of measurement): a classification of how much mathematical structure the numbers or categories actually represent. The framework, introduced by psychologist S. S. Stevens in a 1946 Science paper, defines four levels — nominal, ordinal, interval, and ratio — each one permitting a specific, cumulative set of mathematical operations and, in turn, a specific set of legitimate statistical tests.
Getting this wrong is not a cosmetic error. Averaging a nominal code, treating an ordinal Likert item as if its intervals were equal, or running a ratio-only test (like a coefficient of variation) on interval data all produce numbers that look valid but are not defensible. Reviewers, methodologists, and statistical consultants check this before anything else, because it determines which analyses in your plan are even mathematically coherent.
The four levels at a glance
| Level | Permissible operations | Permissible statistics | Examples | Common mistake |
|---|---|---|---|---|
| Nominal | Classify into categories; test equality (=, ≠) | Mode, frequency counts, chi-square test | Sex, blood type, country of origin, ORCID vs. no ORCID, treatment arm label | Computing a mean or standard deviation of numerically-coded categories (e.g., averaging “1 = control, 2 = treatment”) |
| Ordinal | Rank order (<, >), plus everything nominal allows | Median, percentiles, Spearman’s rank correlation, Mann-Whitney U, Kruskal-Wallis | Likert-scale agreement items, cancer stage (I-IV), education level, class rank, pain scale (mild/moderate/severe) | Treating the gaps between ranks as equal — assuming the distance from “agree” to “strongly agree” equals the distance from “neutral” to “agree,” then averaging the raw scores |
| Interval | Meaningful addition and subtraction, plus everything ordinal allows | Mean, standard deviation, Pearson correlation, t-tests, ANOVA | Calendar year, temperature in Celsius or Fahrenheit, IQ score, IQ-style standardized test scores | Computing a ratio (“40°C is twice as hot as 20°C”) — interval scales have no true zero, so ratios are meaningless |
| Ratio | Meaningful multiplication and division, plus everything interval allows | Geometric mean, coefficient of variation, all interval-level statistics, log transformations | Age, weight, height, reaction time, income, temperature in Kelvin, sample count, publication count | None inherent to the scale itself — but ratio data is sometimes needlessly collapsed into ordinal bins (e.g., binning exact age into “18-25, 26-35…”) before analysis, discarding statistical power |
How the four levels build on each other
The levels form a hierarchy, and each one inherits every permissible operation from the level below it:
- Nominal gives you category membership only — you can say two things are the same or different, nothing more.
- Ordinal adds order — you can say one thing is more or less than another, but not by how much.
- Interval adds equal, meaningful distances between values — you can say how much more, but “zero” is an arbitrary reference point (0°C is not an absence of temperature), so ratios are not meaningful.
- Ratio adds a true, non-arbitrary zero, meaning zero represents a genuine absence of the quantity — so ratios (“twice as much”) become meaningful, and the full range of mathematical and statistical operations applies.
This is why the four levels are sometimes grouped into two broader categories in applied statistics texts: categorical (nominal and ordinal) versus continuous (interval and ratio, also called metric or scale data in some software, notably SPSS’s variable-view labeling of “Nominal / Ordinal / Scale”).
Decision flowchart: classifying a variable
Work through these four questions in order for any variable in your dataset:
- Can the values be meaningfully ordered from low to high?
- No → the variable is nominal. Stop here.
- Yes → continue to question 2.
- Are the intervals between consecutive values equal and quantifiable? (Is the gap between values 1 and 2 the same size as the gap between values 3 and 4?)
- No → the variable is ordinal. Stop here.
- Yes → continue to question 3.
- Does the value zero represent a true, non-arbitrary absence of the quantity being measured?
- No (zero is an arbitrary reference point, e.g., 0°C, a calendar year “0,” or a standardized test score of 0) → the variable is interval. Stop here.
- Yes (zero genuinely means “none of this quantity exists,” e.g., 0 kg, 0 seconds, 0 publications) → the variable is ratio.
Applying this flowchart to a Likert-type survey item is the single most common place researchers get it wrong: the response options are ordered, which rules out nominal, but the psychological distance between “strongly disagree” and “disagree” is not established as equal to the distance between “agree” and “strongly agree” — so a single five-point Likert item is ordinal, not interval. In practice, many researchers treat summed or averaged multi-item Likert scales (not single items) as approximately interval for parametric analysis, a convention with methodological literature both defending and challenging it — but that is a deliberate, justified analytical choice, not a property of the raw scale itself, and it should be stated and defended explicitly in a methods section rather than assumed.
Why the classification determines which statistics are legitimate
Levels of measurement are the gatekeeper for statistical test selection because most common tests make assumptions about the arithmetic being performed on the data:
- A mean requires that equal numeric distances represent equal real-world distances — valid for interval and ratio data, not for ordinal or nominal data.
- A Pearson correlation coefficient or a t-test assumes interval- or ratio-level data (along with other assumptions like normality); for ordinal data, the non-parametric analogues — Spearman’s rank correlation, Mann-Whitney U, Wilcoxon signed-rank, Kruskal-Wallis — are the methodologically defensible choice instead.
- A chi-square test of independence is built for nominal (categorical) data and tests association between category frequencies, not correlation between magnitudes.
- Only ratio data supports statements like “twice as much,” geometric means, and coefficients of variation, because only ratio data has a true zero that makes multiplicative comparisons meaningful.
This is also why the level of measurement of your dependent variable is one of the first things a statistical consultant, methods reviewer, or pre-registration template asks about — it constrains the entire downstream analysis plan before sample size or study design are even considered.
Frequently asked questions
What is the difference between nominal and ordinal data?
Nominal data has categories with no inherent order — country of origin, blood type, treatment-arm label. Ordinal data has categories with a meaningful order but unequal or unknown distances between them — cancer stage, education level, a single Likert item. The practical test: if it makes sense to say one category is “more” or “higher” than another, the variable is at least ordinal, not nominal.
Is a Likert scale ordinal or interval?
A single Likert-type item (one question, one response scale) is ordinal: the response options are ordered, but the psychological distance between adjacent options is not established as equal. Many researchers treat multi-item Likert scales that are summed or averaged as approximately interval for the purposes of parametric analysis — a common and long-debated convention in psychometrics, not a property of a single raw item. State and justify this choice explicitly in the methods section rather than assuming it.
Why can’t you calculate a mean for ordinal data?
A mean assumes that equal numeric differences correspond to equal real differences in the underlying quantity — that the “distance” from 1 to 2 is the same as from 4 to 5. Ordinal data only guarantees rank order, not equal spacing, so a mean calculated on raw ordinal codes can be mathematically computed but does not have a defensible real-world interpretation. The median, which only requires rank order, is the appropriate measure of central tendency for ordinal data.
What is the difference between interval and ratio scales?
Both have equal, meaningful distances between values. The difference is the zero point: an interval scale’s zero is an arbitrary reference point (0°C does not mean “no temperature”), so ratio statements (“twice as hot”) are not meaningful. A ratio scale’s zero represents a true absence of the quantity (0 kg genuinely means no mass), so ratio statements (“twice as heavy”) are meaningful and operations like geometric means and coefficients of variation apply.
Is temperature interval or ratio data?
It depends on the unit. Celsius and Fahrenheit are interval scales — their zero points are arbitrary conventions, not a true absence of thermal energy. Kelvin is a ratio scale, because 0 K is absolute zero, a genuine absence of thermal energy, making statements like “200 K is twice as hot as 100 K” mathematically and physically meaningful.
Are nominal and ordinal data considered categorical, and interval and ratio considered continuous?
In applied statistics and most statistical software, yes — nominal and ordinal data are grouped as categorical (or “qualitative”), while interval and ratio data are grouped as continuous, quantitative, or “scale” data. This two-way grouping is a practical simplification used to decide broad analytical approach (e.g., chi-square family versus regression family), but it collapses distinctions — ordinal and ratio data, for instance, are treated very differently within their respective groups — that matter once you move past that first decision point.
Related concepts
Levels of measurement determine which reliability and validity checks are even applicable to an instrument. For consistency of measurement over time or between raters, see Intraclass Correlation Coefficient (ICC) and Test-Retest vs. Inter-Rater Reliability. For internal consistency of multi-item scales — which presumes the interval-like treatment discussed above — see Cronbach’s Alpha. For whether an instrument measures the construct it claims to, independent of scale level, see Types of Validity in Research. The level of measurement of your outcome variable also interacts with how you sampled it in the first place — see Sampling Methods and Simple Random Sampling.
This page is part of the CASRAI Research Methods hub.
