Are differences meaningful?

Yes, under the interval-scale interpretation.

Can the origin be changed?

Yes, while preserving the scale's equal-difference structure.

AI Glossary / Statistics and probability / Advanced

Interval scale

Definition

An interval scale is a measurement scale on which equal numerical differences represent equal differences in the measured attribute. Its zero point is conventional rather than an intrinsic absence of the attribute. Differences are meaningful, but ratios of the recorded levels are not generally meaningful. Celsius temperature is a familiar example.

What it means

Moving from 10 to 20 degrees Celsius represents the same temperature difference as moving from 20 to 30 degrees Celsius. However, 20 degrees Celsius is not twice the thermodynamic temperature of 10 degrees Celsius. The zero used by this scale does not mean the absence of temperature.
An interval scale therefore contains more information than ordering alone. If values merely rank options from worse to better, equal-looking numerical steps do not establish equal changes in the underlying attribute. Writing ordered categories as 1, 2, and 3 does not supply that missing evidence.

How it works

The characteristic change of representation is y = ax + b, where a is positive. Multiplying by a changes the unit; adding b changes the origin. Such a transformation preserves order and equality of differences. Differences are multiplied by a, while the added constant cancels.
This gives a useful interpretation check: a substantive conclusion should survive an admissible change of units and origin. A claim based on the ratio of two levels can fail that check. Ratios of two nonzero differences, in contrast, remain unchanged because both differences acquire the same multiplier.
The measurement interpretation does not by itself establish a sampling distribution, independence, or the suitability of a particular statistical model. Those questions require separate consideration.

Three practical examples

A facilities team compares temperature changes around computing equipment. It reports differences in degrees and avoids percentage changes calculated directly from Celsius levels.
An analyst compares event dates measured as elapsed days from a chosen reference date. Changing the reference changes the recorded levels but leaves elapsed durations unchanged.
A research team evaluates whether a constructed assessment score supports equal-difference interpretations. The appearance of numerical labels alone is insufficient evidence that a ten-point change has the same meaning everywhere.

Think of it this way

Imagine a ruler whose unit is consistent but whose starting label can be shifted. The distance between two marks remains meaningful after relabeling the origin. A claim that one mark's label is twice another's need not survive that relabeling.

Worked example

Consider three invented room-temperature readings: 10, 20, and 30 degrees Celsius. The first gap is 20 minus 10 = 10 degrees, and the second is 30 minus 20 = 10 degrees. The two differences are equal.
Convert them to Fahrenheit using F = 1.8C + 32. The readings become 50, 68, and 86 degrees Fahrenheit. Their gaps are 18 and 18 degrees. Equality of differences survives, even though the numerical size of each gap changes with the unit.
Now examine ratios of levels. On the Celsius scale, the ratio of the second reading to the first is 20/10 = 2. On the Fahrenheit scale it is 68/50 = 1.36. The claim that the second room is twice as hot cannot be justified by these ratios, because changing the permitted scale representation changes the ratio.
Compare differences instead. The gap from the first reading to the third is 20 Celsius degrees, twice the ten-degree gap from the first to the second. In Fahrenheit, the corresponding gaps are 36 and 18 degrees, also a ratio of two. A ratio of temperature differences can be meaningful even though the ratio of Celsius readings is not.
The average of the three Celsius readings is 20. Converting that average gives 68 Fahrenheit. Averaging the converted readings also gives (50 plus 68 plus 86)/3 = 68. The sample standard deviation is 10 Celsius degrees and 18 Fahrenheit degrees. It changes by the unit multiplier, without acquiring the origin shift of 32.
Now calculate the coefficient of variation, standard deviation divided by mean. It is 10/20 = 0.5 in Celsius but 18/68, approximately 0.265, in Fahrenheit. This inconsistency illustrates why that relative-spread interpretation is inappropriate when the origin is arbitrary.
For a separate illustration of ordering versus equal spacing, suppose three service-quality categories are labeled 1, 2, and 3. Relabel them 1, 2, and 10 while retaining the same order. The gaps change from 1 and 1 to 1 and 8. If only the ordering was justified, both labelings preserve the information available, but conclusions about equal differences do not survive.
Finally, consider dates encoded as days 100 and 130 from a reference point. Shifting the reference so the values become 10 and 40 preserves the thirty-day duration. The ratios of the date labels change from 1.3 to 4. This is another reason to distinguish an elapsed duration from a location on an interval scale.
These comparisons do not establish causality or validate a measurement construct.

Illustrative case study

Fictional company Birch Facilities reviews an AI-generated report claiming that a room's temperature doubled from 10 to 20 degrees Celsius. The analyst replaces the claim with a ten-degree increase and retains the original unit.
The team also checks other automated summaries for percentages calculated from arbitrary score origins. It documents which measures support differences, which support ratios, and which are only ordered categories.

Common misconceptions and mistakes

An interval scale is not an interval estimate. One describes measurement structure; the other expresses estimation uncertainty.
Having a value labeled zero does not establish an absolute zero.
Averages can transform consistently on an interval scale, but this does not establish that every average answers a useful substantive question.
A numerical survey response is not automatically an interval measurement.
Negative recorded values do not invalidate an interval scale; they can arise from the chosen origin.

Business use

Document units, reference points, and the comparisons a measure supports. Review generated claims such as twice as much, percentage increase, and relative variability before publishing them. For assessment scores, distinguish assumed equal spacing from evidence supporting that assumption, especially when comparing changes across groups or time.

Check your understanding

Why do the Fahrenheit readings preserve equal gaps?

Answer: Both gaps are multiplied by 1.8.

Why does 20/10 fail as a temperature-ratio claim?

Answer: It depends on the arbitrary origin.

What remains unchanged when the date reference shifts?

Answer: The thirty-day difference.

Related terms

Ordinal scale
Ratio scale
Affine transformation
Coefficient of variation
Measurement scale
origin
unit
measurement validity

Learn next

Compare admissible transformations across measurement scales and examine how measurement assumptions affect the interpretation of statistical summaries.

Sources

Howell, Statistical Methods for Psychology glossary: https://www.uvm.edu/~statdhtx/methods8/Glossary/Glossary.html . Stevens, On the Theory of Scales of Measurement, 1946: https://faculty.ucmerced.edu/jvevea/classes/290_21/readings/week%208/Stevens%201946.pdf . Numerical comparisons and company scenario are original illustrations.

Previous: Interval predictor model Next: Interval tree