Levels of Measurement: Nominal, Ordinal, Interval and Ratio Explained

This page covers every aspect of levels of measurement a student, researcher, or data analyst needs: definitions, extensive examples, the 3-question decision process for classifying any variable, a comprehensive 80+ variable lookup table, the Likert debate with citable references, statistical test selection, information loss when converting between levels, and implementation in SPSS, Python, and R.

For context, levels of measurement belong to the broader study of descriptive statistics and sit alongside types of data as two complementary frameworks for understanding what a variable can tell you.

The Origin — Stanley Smith Stevens (1946)

In 1946, American psychologist Stanley Smith Stevens published a landmark paper in Science titled "On the Theory of Scales of Measurement." Stevens proposed four hierarchical scales — nominal, ordinal, interval, and ratio — and argued that the level of measurement determines which statistical operations are permissible on a variable.

Stevens developed this framework while working in psychology, where researchers measured abstract phenomena like pain, satisfaction, and attitudes that resisted traditional physical measurement. His framework gave social scientists a principled way to decide which statistics were appropriate for their data. Before Stevens, this question had no systematic answer.

Stevens' Framework Is Widely Accepted but Not Universal

Stevens' four levels are the most widely used classification in statistics education and data analysis worldwide. They appear in virtually every introductory statistics course and research methods textbook. That said, the framework has academic critics:

• Critics like Duncan and Rozeboom argued that Stevens' nominal and ordinal categories do not fit the traditional mathematical definition of "measurement" at all — classifying objects into named groups or rank-ordering them is not the same as assigning quantities.
• The boundaries between levels are not always clear for psychological constructs. Whether a personality questionnaire scale has truly equal intervals is a theoretical assumption, not a provable fact.
• The framework originated in psychology and has been adopted and extended in some disciplines, and criticized or rejected by others.

Nominal Scale — Categories With No Order

The nominal scale is the simplest level of measurement. It classifies observations into named, mutually exclusive categories. The categories have no natural order: there is no sense in which one category is "more" or "higher" than another. Nominal comes from the Latin nomen, meaning name. Nominal scales are literally naming scales.

Nominal Scale Examples

Demographics: gender (male, female, non-binary), ethnicity, nationality, marital status, religion, political party affiliation.

Medical and biological: blood type (A, B, AB, O), species classification, hair color, eye color, cause of death (ICD codes), disease diagnosis category.

Research and surveys: experimental group assignment (treatment vs. control), yes/no questions, country codes, student ID numbers.

Binary and Dichotomous Variables as Nominal Data

Binary variables are a special case of nominal data with exactly two categories: yes/no, pass/fail, male/female, present/absent, treatment/control. These are also called indicator variables or dummy variables. In regression analysis, binary nominal variables are often coded as 0/1 to make them mathematically tractable, but the underlying variable is still nominal.

The "Numbers Are Just Labels" Trap

Valid Statistics for Nominal Data

You may use frequency counts and percentages, the mode (the most common category), frequency distribution tables, bar charts and pie charts, chi-square tests of independence, Fisher's exact test, and logistic regression (when nominal is the dependent variable). You may not use the mean, median, standard deviation, or Pearson correlation — none of these operations are meaningful when values are just category labels.

Ordinal Scale — Categories With a Meaningful Order

The ordinal scale adds a meaningful order or ranking to categories, but the differences between categories are not assumed to be equal. You know one value is "more" than another, but not by how much. Ordinal comes from Latin meaning "order" or "rank."

Ordinal Scale Examples

Education and achievement: education level (primary / secondary / tertiary / postgraduate), academic grades (A/B/C/D/F), class rank, military rank.

Health and wellness: pain scales (mild / moderate / severe), health status (poor / fair / good / excellent), cancer stage (I, II, III, IV), satisfaction ratings.

Competitive ranking: race finishing positions (1st, 2nd, 3rd), tournament seedings, Olympic medal standings.

Survey responses: Likert items (strongly disagree to strongly agree), Net Promoter Score categories (detractor / passive / promoter), frequency scales (never / rarely / sometimes / often / always).

Why the Intervals Between Ordinal Values Are Not Equal

Consider a foot race. The gap in time between 1st and 2nd place might be 0.01 seconds, while the gap between 2nd and 3rd could be 15 seconds. The ranks are meaningful, but the time differences behind them are not equal — and that is true of almost every ordinal variable. The same applies to a pain scale: the jump from "mild" to "moderate" is not the same experience as the jump from "moderate" to "severe" across different people or different types of pain.

Why the Median, Not the Mean, for Ordinal Data

The median uses only rank information: find the middle-ranked value. It does not assume that the distance between rank 1 and rank 2 equals the distance between rank 2 and rank 3. The mean, by contrast, calculates (sum ÷ n), which implicitly treats every step between adjacent ranks as exactly equal. Since ordinal data gives us no grounds for that assumption, the mean is technically inappropriate. For ordinal data, report the median and interquartile range.

Interval Scale — Equal Gaps, No True Zero

The interval scale has all the properties of ordinal data plus meaningful, equal intervals between values. The step from 10 to 20 is the same size as the step from 30 to 40. However, it has no true zero point — zero on an interval scale does not represent the complete absence of the quantity being measured.

Interval Scale Examples

Temperature: Celsius and Fahrenheit are the classic interval scales. The difference between 10°C and 20°C is the same as between 30°C and 40°C. But 0°C is the freezing point of water — an arbitrary reference, not "no temperature."

Standardized test scores: IQ scores (mean 100, SD 15), SAT scores (200–800), GRE scores, ACT scores. These are constructed to have equal intervals by design, but an IQ of zero would not mean "no intelligence."

Calendar years: The year 2000, 2010, and 2020 are interval. The difference of 10 years between 2000 and 2010 equals the difference between 2010 and 2020. But "year 0" is an arbitrary calendar convention.

No True Zero — What This Actually Means

The absence of a true zero means you cannot form meaningful ratio statements. Consider: saying "40°C is twice as hot as 20°C" sounds reasonable but is wrong. Convert both to Fahrenheit: 40°C = 104°F and 20°C = 68°F. Is 104°F twice as hot as 68°F? No — 104/68 ≈ 1.53. The ratio changed when you changed the unit. This only happens when zero is arbitrary.

Kelvin fixes this problem: 0 K = absolute zero = the complete absence of thermal energy. So 400 K does contain twice the thermal energy of 200 K, and that ratio holds regardless of unit conversion. Kelvin is a ratio scale.

Ratio Scale — Equal Gaps With a True Zero

The ratio scale is the highest level of measurement. It has all the properties of interval data plus a true zero point that represents the complete absence of the quantity. This makes ratio comparisons meaningful: a value of 80 is genuinely twice a value of 40.

Ratio Scale Examples

Physical measurements: height, weight, mass, length, distance, area, volume, speed, time duration, temperature in Kelvin.

Count data: number of children, number of correct answers on a test, number of customers, number of hospital visits, words in a document.

Financial: income in dollars (0 = no income), revenue, price, debt, profit.

Demographics and clinical: age in years (0 = birth), reaction time, blood pressure in mmHg, heart rate in bpm, BMI, drug dosage in mg.

The True Zero — What It Actually Means

Zero on a ratio scale means the complete absence of the measured quantity: 0 kg = no mass, $0 income = no income, 0 seconds = no time elapsed, 0 Kelvin = no thermal energy. Because of this, ratios between values are fixed and meaningful across any unit transformation. A person earning $80,000 earns exactly twice someone earning $40,000, and that relationship holds whether you convert to euros, pounds, or yen.

This is also why the coefficient of variation and geometric mean are only valid for ratio data — both depend on a meaningful zero and meaningful ratios.

The Four Levels Are Cumulative

Each level keeps all properties of the levels below it and adds exactly one new property. This is why ratio data can do everything nominal, ordinal, and interval data can do — plus ratio comparisons.

How to Identify the Level — 3 Questions

Every variable can be classified using three questions asked in order. Stop when you reach a "No" answer — that is your level.

What Level of Measurement Is It? — 80+ Variables Classified

Every variable below includes a "Why" explanation. The table is organized by domain. Use it as a lookup reference when classifying your own variables.

Is a Likert Scale Ordinal or Interval? The Complete Answer

This is the most searched controversy in all of levels-of-measurement content. Here is the full, citable, actionable resolution.

The Actionable Verdict

• Single Likert item → use ordinal methods: report the median, use Spearman's correlation, Mann-Whitney U, or Kruskal-Wallis for group comparisons.
• 4 or more Likert items summed or averaged into one scale → treating the composite score as interval is defensible. State this explicitly in your methods section and cite Boone & Boone (2012) or similar.
• In a thesis or dissertation: state whether you are treating your Likert data as ordinal or interval, why, and cite your justification. Reviewers will ask.

References to cite:

Boone, H. N., & Boone, D. A. (2012). Analyzing Likert data. Journal of Extension, 50(2).

Sullivan, G. M., & Artino, A. R. (2013). Analyzing and interpreting data from Likert-type scales. Journal of Graduate Medical Education, 5(4), 541–542.

Which Statistical Test for Each Level of Measurement?

Choosing the wrong test for your measurement level produces invalid results. This table maps each level to its correct descriptive and inferential tools. For deeper guidance on parametric vs. non-parametric choices, see parametric vs. non-parametric tests.

When to Use Non-Parametric Tests Even with Interval or Ratio Data

Even with interval or ratio data, switch to non-parametric tests when the distribution is severely skewed or non-normal, when the sample is very small (n < 20–30 as a rough guide), or when outliers are extreme and cannot be removed without losing meaningful data. The measures of central tendency section explains when the mean versus median better represents your data's center. For choosing between the Pearson correlation and Spearman rank correlation, the measurement level is one key factor — but data distribution matters too.

Converting Between Levels — A One-Way Street

You can always convert a variable down from a higher level to a lower level. You can never convert up. Once precision is discarded, it cannot be recovered.

Levels of Measurement in SPSS — Nominal, Ordinal, and Scale

SPSS combines Stevens' interval and ratio into a single category it calls Scale. SPSS therefore uses only three measurement levels:

• Nominal — same as Stevens' nominal
• Ordinal — same as Stevens' ordinal
• Scale — covers both interval and ratio data

To set this in SPSS: open your dataset and click the Variable View tab (the second tab in the Data Editor). The rightmost column is labeled Measure. Click any cell in that column and select Nominal, Ordinal, or Scale from the dropdown. SPSS uses this setting to determine which statistics appear in Frequencies output and which variables are eligible for certain chart types in Custom Tables. Setting the wrong level can exclude a variable from analysis options or include inappropriate statistics.

Levels of Measurement in Python and R

Python (Pandas)

import pandas as pd

# Create a sample dataset
df = pd.DataFrame({
    'blood_type': ['A', 'B', 'O', 'AB', 'A'],       # Nominal
    'education': ['High School', 'Bachelor', 'Master', 'PhD', 'Bachelor'],  # Ordinal
    'temperature_c': [22.5, 18.0, 25.3, 20.1, 19.8], # Interval
    'height_cm': [170, 165, 182, 178, 160]            # Ratio
})

# Nominal: unordered category dtype
df['blood_type'] = df['blood_type'].astype('category')
print(df['blood_type'].cat.ordered)   # False — unordered, nominal

# Ordinal: ordered Categorical
education_order = ['High School', 'Bachelor', 'Master', 'PhD']
df['education'] = pd.Categorical(
    df['education'],
    categories=education_order,
    ordered=True          # ordered=True marks it as ordinal
)
print(df['education'].cat.ordered)   # True — ordered, ordinal

# Interval / Ratio: standard float64 or int64
# pandas does not distinguish between them — use float64 for both

# Correct statistics by level
print(df['blood_type'].value_counts())       # Nominal: frequencies
print(df['blood_type'].mode()[0])            # Nominal: mode
print(df['education'].cat.codes.median())    # Ordinal: median by rank
print(df['temperature_c'].mean())            # Interval: mean valid
print(df['height_cm'].mean())               # Ratio: mean valid
print(df['height_cm'].std() / df['height_cm'].mean() * 100)  # Ratio only: CV

R (Ordered Factor)

# Create a sample dataset
df <- data.frame(
  blood_type = c("A", "B", "O", "AB", "A"),
  education = c("High School", "Bachelor", "Master", "PhD", "Bachelor"),
  temperature_c = c(22.5, 18.0, 25.3, 20.1, 19.8),
  height_cm = c(170, 165, 182, 178, 160),
  stringsAsFactors = FALSE
)

# Nominal: unordered factor
df$blood_type <- factor(df$blood_type)
is.ordered(df$blood_type)   # FALSE — nominal

# Ordinal: ordered factor
education_levels <- c("High School", "Bachelor", "Master", "PhD")
df$education <- factor(df$education,
                        levels = education_levels,
                        ordered = TRUE)   # ordered = TRUE marks as ordinal
is.ordered(df$education)   # TRUE

# Correct statistics by level
table(df$blood_type)               # Nominal: frequency table
which.max(table(df$blood_type))    # Nominal: mode
median(as.numeric(df$education))   # Ordinal: median by rank code
mean(df$temperature_c)             # Interval: mean valid
sd(df$height_cm)                   # Ratio: SD valid
sd(df$height_cm) / mean(df$height_cm) * 100  # Ratio only: CV

Tricky Variables — Edge Cases Explained

Temperature: Celsius, Fahrenheit, and Kelvin

Celsius and Fahrenheit are interval: their zero points are arbitrary conventions, not the absence of heat. The proof: 20°C = 68°F and 40°C = 104°F. The ratio 40/20 = 2, but 104/68 ≈ 1.53. The ratio changes with the unit — which only happens when zero is arbitrary. Kelvin is ratio: 0 K = absolute zero = no thermal energy. 400 K genuinely contains twice the thermal energy of 200 K, and this ratio holds regardless of conversion.

Age

Age in years is ratio — zero represents birth, a genuine starting point, so a 40-year-old has lived twice as long as a 20-year-old. Age grouped into brackets (under 18, 18–35, 36–55) is ordinal: you know rank order but lose the exact value. Birth year as a calendar year (e.g., 1985) is interval: the year 0 is an arbitrary calendar convention, and saying "year 2000 is twice year 1000" would be meaningless.

IQ Score

IQ is constructed to have a mean of 100 and a standard deviation of 15. The scale has no true zero — an IQ of 0 does not mean "no intelligence." Therefore IQ is interval. Most researchers treat IQ as interval data and apply parametric statistics, though some scholars note the equal-interval property rests on psychometric assumptions rather than proven fact.

Year: Duration vs. Calendar

Years of experience (e.g., "5 years in the field") is ratio — zero years means no experience. Calendar year (2000, 2010, 2020) is interval — year 0 is an arbitrary historical convention, and "year 2000 is twice year 1000" carries no meaning.

ZIP Code / Postal Code

ZIP codes are nominal despite being numbers. 90210 is not twice 45105. There is no meaningful order — ZIP codes are geographic region labels — and no equal intervals. The mean of ZIP codes is mathematically computable but geographically meaningless.

Race Finishing Position

Race positions (1st, 2nd, 3rd) are ordinal. The time gap between 1st and 2nd might be 0.001 seconds, while the gap between 2nd and 3rd is 2 minutes. The rank order is real; the gaps are not equal.

Shoe Size

Shoe sizes are treated as interval in most systems. Some systems (UK sizes) have a fixed unit of 1/3 inch per size, giving equal intervals. However, there is no true zero across systems — a size 0 does not represent "no foot" in any practical sense. Treat as interval.

Practical Rules for Researchers and Students

Frequently Asked Questions

Key Takeaways