Does a decimal value imply an interval scale?

No; storage format does not determine measurement meaning.

Can a questionnaire score be treated as approximately interval?

That requires an argument or measurement model beyond assigning consecutive category numbers.

Are all properties equally represented after recoding?

No. A recoding may preserve categories or order while changing distances and ratios.

AI Glossary / Statistics and probability / Advanced

Scales of measurement

Definition

Scales of measurement describe the relationships represented by values assigned to observations. The familiar classification distinguishes nominal, ordinal, interval and ratio scales. These describe whether values represent categories, ordered categories, meaningful equal differences or meaningful ratios. A variable's appearance as a number in a spreadsheet does not establish which relationships that number represents.

What it means

A number can identify a customer, rank a response, express a temperature or measure elapsed time. Those uses support different interpretations. Customer 200 is not twice customer 100, while a duration of two minutes is twice a duration of one minute.

The useful question is what remains meaningful when the coding or unit changes. If a conclusion disappears after an equally legitimate relabeling, it may describe the chosen labels rather than the underlying observations. Measurement assumptions therefore belong in the explanation of an analysis, not just in its data dictionary.

How it works

Nominal values distinguish categories without an inherent order. A one-to-one relabeling preserves that information. Ordinal values additionally represent order; a strictly increasing recoding preserves it without necessarily preserving gaps.

Interval scales support comparisons of differences and permit changes of unit and origin, written x* = ax + b with a positive. Ratio scales also support comparisons by ratios and permit positive changes of unit, x* = ax, while retaining a meaningful zero.

These distinctions help interpret calculations, but they are not a complete recipe for selecting a statistical model. The question, data-generating process and additional modeling assumptions still matter. An arithmetic operation being possible does not by itself establish its substantive meaning.

Three practical examples

1. Support categories: codes 1, 2 and 3 denote billing, access and delivery. The codes identify groups; their numerical distances do not express how similar the issues are.

2. Review ratings: low, medium and high indicate an ordering. Coding them 1, 2 and 3 does not independently establish equally spaced levels of quality.

3. Response duration: sixty and one hundred twenty seconds represent elapsed times. Changing to minutes preserves the two-to-one ratio, so that comparison does not depend on the chosen time unit.

Think of it this way

Think of numbers as instructions: identify a category, establish order, quantify a gap or express a ratio. Check which instruction a measurement actually supports.

Worked example

Consider four fictional fields in an AI-service evaluation: request category, reviewer rating, room temperature and processing duration. Examine each before producing a dashboard summary.

First, five request categories are billing, billing, access, access and delivery. Coding these as 1, 1, 2, 2 and 3 produces a mean code of 9/5 = 1.8. Relabel billing as 30, access as 10 and delivery as 20. The mean code becomes (30 + 30 + 10 + 10 + 20)/5 = 20. The category frequencies remain two, two and one. The two numerical means do not describe a measured quantity of request type.

Second, three reviewers choose low, medium and high. Coding these as 1, 2 and 3 gives a mean of two. An order-preserving coding of 1, 2 and 100 gives a mean of 103/3, approximately 34.33. The ordering and middle category remain unchanged. A numerical mean can be reported under a specified scoring model, but equal spacing is an added assumption, not a consequence of the category order alone.

Third, two Celsius measurements are 10 and 20 degrees. Converting with F = 1.8C + 32 gives 50 and 68 degrees Fahrenheit. The Celsius numerical ratio is 20/10 = 2; the Fahrenheit ratio is 68/50 = 1.36. Calling the second measurement "twice as hot" based on the Celsius ratio is therefore not supported by that interval scale.

Their temperature difference is ten Celsius degrees or eighteen Fahrenheit degrees. Those are consistent descriptions of the same interval after accounting for unit conversion. The difference's number changes with the unit; its physical interpretation does not.

Finally, processing durations are sixty and one hundred twenty seconds, equivalent to one and two minutes. Both representations give a ratio of two. Here zero represents no elapsed duration, and changing the time unit preserves the ratio. Unlike the temperature example, no arbitrary origin shift is introduced.

Illustrative case study

Fictional Alder Reports initially treats every numeric column in the example as a continuous quantity. Its dashboard displays an average category code and labels the Celsius change a doubling.

The reviewer replaces the category-code average with counts and proportions, explains the temperature difference in stated units, and retains the twofold duration comparison. For the ratings, Alder records the response labels and makes its scoring assumption explicit before using an average.

The team also checks the AI model's input representation. A numeric category identifier could otherwise introduce an unintended order or distance. This fictional case illustrates measurement choices; it does not claim that one encoding guarantees better model performance.

Common misconceptions and mistakes

"An integer is automatically a count." Integers may be identifiers or category codes.

"Every scale containing zero is a ratio scale." A displayed zero can be an arbitrary reference point or a missing-value code.

"Ordinal values cannot be analysed statistically." They can; the analysis should respect or justify its assumptions about order and spacing.

"Changing units changes the underlying observation." A consistent conversion changes its representation, not the observed quantity.

Business use

For each field, document what values mean, their units, permitted categories, missing-value conventions and any scoring assumptions. Distinguish elapsed time from calendar time, identifiers from quantities and ordered labels from calibrated measurements.

When reviewing AI-generated insights, check ratio language such as "twice," differences described as improvements and averages of coded fields. A mathematically correct calculation can still support an incorrect interpretation. Preserve the original labels and units so that reviewers can determine what the computation actually compares.

Check your understanding

Why is the mean category code uninformative about request type?

Answer: Arbitrary relabeling changes it without changing the requests.

Why is the duration ratio stable?

Answer: A positive unit conversion multiplies both durations by the same factor.

What changes in the ordinal recoding?

Answer: Numerical gaps and the mean, while order remains intact.

Related terms

Learn next

Inspect a dataset's numeric columns and identify which are quantities, identifiers or ordered labels. Test whether a proposed conclusion survives a legitimate recoding.

Sources

David C. Howell, Statistical Methods for Psychology glossary, Scales of measurement: https://www.uvm.edu/~statdhtx/methods8/Glossary/Glossary.html

S. S. Stevens, On the Theory of Scales of Measurement, Science, 1946, original classification and transformations: https://static1.1.sqspcdn.com/static/f/221758/26986815/1461271683467/Stevens1946ScalesOfMeasurement.pdf

Previous: Scaled orthographic projection Next: Scaling