How does nominal differ from ordinal?

Ordinal categories have a meaningful order, such as low, medium and high priority. Nominal categories do not inherently have that ranking.

Can nominal variables predict numerical outcomes?

Yes. Category membership can be used as a predictor through an appropriate model representation.

Is a customer identifier nominal?

Its role is identification, not quantity. Whether it is useful as a predictor is a separate question.

AI Glossary / Statistics and probability / Advanced

Nominal scale

Definition

A nominal scale classifies observations into categories without assigning those categories an inherent order or numerical distance. Category names or codes identify differences in kind. For example, billing, access and delivery identify request types without ranking them.

What it means

A number stored in a dataset is not necessarily a quantity. If billing is coded 1 and access is coded 2, access is not twice as much of anything. The numbers are identifiers for categories.

The key information is which observations belong to the same category and which belong to different categories. Replacing the identifiers consistently should preserve that information. A meaningful analysis should not acquire a new conclusion merely because the category codes were renamed.

How it works

Define what each category means and how observations are assigned. For a single categorical variable, specify whether categories are mutually exclusive and how missing, unknown or unclassifiable cases are represented.

Summarize the variable with category counts, proportions or its most frequent category, where appropriate. When using it in a predictive model, choose a representation that matches the model's treatment of categories. One-hot encoding creates separate binary indicators; passing arbitrary integer codes directly to a numerical model can impose unsupported ordering or spacing.

Three practical examples

1. Ticket routing: billing, access and delivery are nominal request types. Their frequencies help allocate specialist staff, but their code average does not describe a typical request type.

2. Image classification: a model predicts bicycle, bus or train. These classes identify object types. A predicted class code of three is not numerically closer to class two than to class one in any inherently meaningful sense.

3. Data integration: two systems use different codes for the same delivery channels. Analysts map the codes using a shared dictionary before combining records, so that a code's local meaning is not mistaken for a universal definition.

Think of it this way

Think of colored folders. A blue folder and a green folder represent different categories, but blue is not two units above green. Replacing the colors with letters or numbers can make filing easier without creating a measurement scale. Counting the folders is meaningful; averaging their color labels is not.

Worked example

Consider twelve support requests: six billing, four access and two delivery. The proportions are 6/12 = 50 percent, 4/12 = 33.33 percent and 2/12 = 16.67 percent. Billing is the mode.

First encode billing as 1, access as 2 and delivery as 3. The average code is (6 × 1 + 4 × 2 + 2 × 3)/12 = 20/12, approximately 1.67. This arithmetic is possible, but the result has no defined interpretation as an average request type.

Now use billing = 30, access = 10 and delivery = 20. The same twelve requests produce an average code of (6 × 30 + 4 × 10 + 2 × 20)/12 = 260/12, approximately 21.67. Nothing about the requests changed. Their frequencies, proportions and modal category remain identical.

For one-hot encoding, use columns ordered billing, access, delivery. A billing request becomes (1,0,0), an access request (0,1,0), and a delivery request (0,0,1). The column averages are 0.50, 1/3 and 1/6. These averages are meaningful because each column explicitly measures membership in a named category. They are proportions, not averages of arbitrary labels.

The distinction also affects distance calculations. Under the first integer coding, billing and delivery are two units apart while billing and access are one unit apart. Under the second, those distances become ten and twenty. The apparent closeness reverses solely because of relabeling. With the stated full one-hot encoding, every pair of distinct category vectors has Euclidean distance sqrt(2), so the representation does not privilege one pair through arbitrary code spacing.

Illustrative case study

Fictional Cedar Support discovers that its clustering pipeline treats request-type codes as ordinary numbers. A revised integration changes the codes and unexpectedly changes the resulting clusters, even though the underlying requests are identical.

The analysts reproduce the distance reversal from the worked example. They then document request type as nominal and select a representation appropriate to their clustering method. They separately choose how much influence request type should have relative to numerical variables such as waiting time.

Cedar also adds distinct handling for unknown request types, rather than silently mapping them to billing. The team checks the revised pipeline on representative records before using its output. This fictional example does not establish that one encoding always improves clustering.

Common misconceptions and mistakes

“Numeric storage makes a variable quantitative.” A database can store category identifiers as integers without giving their numerical gaps meaning.

“Nominal categories cannot be ordered on a screen.” They can be sorted alphabetically or by frequency. That display order does not create an inherent measurement order.

“Means are never useful with categorical data.” The worked example's binary membership means are proportions. The problem is averaging arbitrary multi-category labels as though they measured quantity.

“One-hot encoding solves every modelling issue.” It avoids that particular artificial order, but does not decide feature weighting, category quality, missing-value handling or the suitability of a model.

Business use

Record measurement meaning in the data dictionary, not just the database type. For each nominal field, document allowed categories, definitions, code mappings and how changes are versioned. A dashboard should display meaningful names and denominators alongside percentages.

Keep distinctions between a missing value, an unfamiliar category and a valid “other” category explicit. If customers can select several request topics, represent those selections as multiple indicators or a suitable set-valued field; do not pretend the record contains one mutually exclusive category. Validate mappings when joining systems and preserve historical meaning when categories are merged or split.

Check your understanding

Why did the category-code average change?

Answer: The labels were arbitrarily reassigned, changing their arithmetic without changing the observations.

Why is the billing-indicator average meaningful?

Answer: It equals the fraction of requests classified as billing.

Does alphabetical sorting turn request types into ordinal data?

Answer: No; the display order supplies no substantive ranking.

Related terms

Categorical variable
Ordinal scale
Mode
One-hot encoding
Dummy variable
measurement scale
category frequency
data dictionary

Learn next

Recode a small categorical dataset using two different label schemes. Check that category summaries remain equivalent, then inspect whether your chosen model treats the codes as labels or as numerical values.

Sources

David C. Howell, Statistical Methods for Psychology glossary, “Nominal scale” and “Ordinal scale”: https://www.uvm.edu/~statdhtx/methods8/Glossary/Glossary.html

Scikit-learn documentation, Preprocessing data, Encoding categorical features: https://scikit-learn.org/stable/modules/preprocessing.html

Previous: Noisy text analytics Next: Nominal significance level