Polytomous Variable – Definition, Purpose and Examples

A polytomous variable is a categorical variable with three or more possible categories. Its categories may be unordered, such as blood groups, or ordered, such as low, medium, and high satisfaction. The variable may serve as a predictor, an outcome, or a scored item, and its role and measurement level determine the appropriate analysis.

Introduction

Researchers frequently collect information that cannot be represented adequately by a simple yes-or-no variable. A respondent may select one of several transport modes, report one of five satisfaction levels, belong to one of four blood groups, or earn a score from 0 to 3 on a test item.

Each of these may be represented by a polytomous variable. However, knowing that a variable has several categories is only the first step. Researchers must also determine whether the categories are nominal or ordinal, whether the variable is a predictor or an outcome, and which coding and statistical methods preserve its meaning.

This guide explains how to identify, design, code, analyze, interpret, and report polytomous variables in academic research.

Key Takeaways

• A polytomous variable has three or more categories.
• It may be nominal, with no natural order, or ordinal, with a meaningful order.
• Numeric category codes are labels unless genuine numerical distances exist.
• A nominal polytomous outcome is often analyzed with multinomial logistic regression.
• An ordinal polytomous outcome may be analyzed with ordinal logistic regression.
• Converting several categories into two groups can discard useful information and should require a substantive justification.

What is a polytomous variable?

A polytomous variable is a categorical variable that can take one of at least three mutually distinguishable values or categories. “Polytomous” comes from terminology meaning that the variable is divided into many possible outcomes, in contrast to a dichotomous variable, which has only two.

Categorical variables with exactly two possible values are commonly described as binary or dichotomous. Those with more than two possible categories are described as polytomous. Polytomous categories can be nominal or ordinal.

Examples include:

• Blood group: A, B, AB, or O
• Employment status: employed, unemployed, student, retired, or other
• Preferred transport: car, bus, train, bicycle, or walking
• Satisfaction: very dissatisfied, dissatisfied, neutral, satisfied, or very satisfied
• Examination score category: fail, pass, merit, or distinction

The categories should normally be defined so that each observation belongs to one category for that variable.

How can you identify a polytomous variable?

A variable is polytomous when all four of the following conditions are met:

1. It represents categories rather than an unrestricted numerical measurement.
2. It has at least three possible categories.
3. Each recorded value indicates category membership.
4. The categories may be nominal or ordinal.

Consider the variable transport_mode with the values:

• Car
• Bus
• Train
• Bicycle
• Walking

This is polytomous because it contains five categorical outcomes.

By contrast, journey_time_minutes is quantitative. It is not polytomous merely because many different journey times are possible.

A quick identification test

Ask these questions:

• Do the values name groups or levels?
• Are there three or more possible groups?
• Would arithmetic on the assigned codes be meaningless or require strong assumptions?
• Does each participant usually receive one value for the variable?

When the answers are yes, the variable is probably polytomous.

Nominal and ordinal polytomous variables

The most important distinction is whether the categories have a meaningful order.

Nominal polytomous variable

A nominal polytomous variable has three or more categories that do not have an inherent ranking.

Examples include:

• Blood group: A, B, AB, O
• Field of study: biology, history, engineering, economics
• Device type: desktop, laptop, tablet, smartphone
• Primary diagnosis category
• Preferred political party

The categories are different, but one is not inherently higher or lower than another.

Ordinal polytomous variable

An ordinal polytomous variable has three or more categories arranged in a meaningful order. However, the distances between adjacent categories are not necessarily equal.

Examples include:

• Satisfaction: very dissatisfied to very satisfied
• Education level: primary, secondary, undergraduate, postgraduate
• Disease severity: mild, moderate, severe
• Agreement: strongly disagree to strongly agree
• Socioeconomic status: low, middle, high

Ordinal values can be ranked, but the difference between the first and second category cannot automatically be assumed equal to the difference between the second and third. UCLA’s statistical guidance emphasizes that order does not establish equal spacing.

Feature	Nominal polytomous	Ordinal polytomous
Number of categories	Three or more	Three or more
Natural order	No	Yes
Equal distance assumed	No	No
Example	Blood group	Satisfaction level
Typical outcome model	Multinomial logistic regression	Ordinal logistic regression
Useful summary	Frequencies, proportions, mode	Frequencies, cumulative percentages, median category

Examples of polytomous variables in research

Education research

A researcher records students’ preferred learning format as:

• Face-to-face
• Live online
• Recorded online
• Blended
• Self-directed

The variable is nominal and polytomous.

A course grade reported as fail, pass, merit, or distinction is ordinal and polytomous.

Health research

A diagnosis recorded as viral infection, bacterial infection, autoimmune condition, metabolic condition, or other is nominal.

Self-rated health recorded as poor, fair, good, very good, or excellent is ordinal.

Business and marketing research

A customer’s preferred purchasing channel—physical store, company website, mobile application, online marketplace, or telephone—is nominal.

Customer satisfaction measured from very dissatisfied to very satisfied is ordinal.

Social-science research

Marital status, employment status, housing tenure, and principal source of information are often nominal polytomous variables.

Political-interest levels such as not at all interested, slightly interested, moderately interested, and very interested are ordinal.

Educational and psychological testing

An item scored 0, 1, 2, or 3 according to the quality of a response is a polytomously scored item. The scores usually represent increasing performance and are therefore ordered.

Polytomous versus dichotomous variables

A dichotomous variable has two categories; a polytomous variable has three or more.

Characteristic	Dichotomous variable	Polytomous variable
Number of categories	Two	Three or more
Examples	Yes/no, present/absent	Blood group, transport mode, satisfaction
Possible measurement level	Usually nominal; an ordered binary distinction is also possible	Nominal or ordinal
Typical outcome model	Binary logistic regression	Multinomial or ordinal logistic regression
Number of reference-coded indicators	1	$K-1$K−1 for $K$K categories

The distinction helps determine the probability model and the number of parameters required. Binary logistic regression handles two-category outcomes, whereas extensions are needed when the outcome has more than two categories.

Polytomous versus multinomial

The terms are related but not always interchangeable.

Polytomous describes a variable with more than two categorical outcomes.

Multinomial may describe:

• A probability distribution for counts across several mutually exclusive categories.
• A regression model for an unordered categorical outcome.
• A multicategory response more generally.

A nominal polytomous outcome is commonly modeled through a multinomial distribution or multinomial logistic regression. The variable itself remains a categorical variable; “multinomial” describes the associated probability structure or model.

Polytomous versus discrete variables

A polytomous variable is categorical, while a discrete quantitative variable represents counts or separated numerical quantities.

For example:

• Number of children: discrete quantitative
• Household type: polytomous categorical
• Number of hospital visits: discrete quantitative
• Type of health service used: polytomous categorical

Having more than two possible values does not by itself make a variable polytomous. The values must represent categories rather than numerical quantities on which arithmetic has substantive meaning.

Polytomous versus multiple-response variables

A standard polytomous variable usually records one category per observation. A multiple-response question allows a participant to select several options.

For example:

Which devices do you use? Select all that apply.

A respondent might select a laptop, tablet, and smartphone. This is not a single ordinary polytomous variable because the categories are not mutually exclusive. It is usually stored as several binary variables:

• Uses laptop: yes/no
• Uses tablet: yes/no
• Uses smartphone: yes/no

Treating every possible combination as a separate category can create an unmanageable number of rare categories.

Can a polytomous variable be independent or dependent?

Yes. “Polytomous” describes the number and nature of the categories, not the variable’s causal or analytical role.

Polytomous independent variable

A polytomous predictor divides observations into three or more groups.

Example:

• Teaching method: lecture, seminar, blended, or online
• Outcome: continuous examination score

The teaching-method variable can be represented through indicator variables and included in an ANOVA or linear regression model.

Polytomous dependent variable

A polytomous outcome has three or more possible response categories.

Example:

• Outcome: preferred teaching format
• Predictors: age, employment status, prior online-learning experience

A nominal outcome may require multinomial logistic regression. An ordered outcome may require ordinal logistic regression.

This distinction is essential because the correct analysis depends on the outcome’s measurement level, not simply on whether any variable in the dataset is polytomous.

How should polytomous categories be designed?

Well-designed categories should be:

Mutually exclusive

Each observation should fit one category unless the study intentionally uses a multiple-response structure.

Poor categories:

• Undergraduate
• Science student
• Postgraduate

A person could be both an undergraduate and a science student.

Better categories:

• Undergraduate
• Postgraduate taught
• Postgraduate research
• Other

Collectively exhaustive

The available categories should cover plausible responses. Depending on the construct, this may require an “other” option, an open-text follow-up, or a documented missing-value code.

Clearly labelled

Avoid vague categories such as “regularly” unless the questionnaire defines what regularly means.

Substantively ordered when classified as ordinal

Researchers should not impose an order simply because categories have numeric codes. For example, coding regions as 1, 2, 3, and 4 does not make region ordinal.

Appropriate in number

Too few categories can hide meaningful distinctions. Too many can produce sparse cells, unstable estimates, respondent confusion, and difficult interpretation.

How do you code a polytomous variable?

Category codes should be treated as labels rather than numerical measurements unless a defensible quantitative scale exists.

Example data dictionary:

Code	Category	Meaning
1	Car	Private car is the main transport mode
2	Bus	Bus is the main transport mode
3	Train	Train is the main transport mode
4	Bicycle	Bicycle is the main transport mode
5	Walking	Walking is the main transport mode
99	Missing	No usable response

The values 1–5 identify categories. Code 5 is not “five times” code 1, and the difference between codes 2 and 3 has no quantitative interpretation.

Dummy or indicator coding

When a nominal polytomous predictor has $K$K categories, reference coding usually represents it with $K-1$K−1 indicator variables.

Suppose transport mode has four categories:

• Car
• Bus
• Train
• Bicycle

With car as the reference:

Category	Bus indicator	Train indicator	Bicycle indicator
Car	0	0	0
Bus	1	0	0
Train	0	1	0
Bicycle	0	0	1

The coefficients compare each represented category with the omitted reference category. UCLA’s regression guidance confirms that a $K$K-category predictor is usually represented using $K-1$K−1 dummy variables.

Most modern statistical packages create these indicators automatically when the variable is declared categorical.

Choosing a reference category

A useful reference may be:

• The control group
• The most common category
• The conventional or standard condition
• The lowest category in a meaningful sequence
• A category of particular policy interest

The choice changes coefficient interpretation, but it does not normally change the fitted probabilities or overall model fit.

How should missing responses be coded?

Missingness should not automatically be treated as a substantive category.

For example, “no response” may reflect:

• Refusal
• Lack of knowledge
• Survey routing
• Data-entry failure
• Inapplicability

These reasons can have different meanings. Researchers should retain documented missing-value codes during data cleaning and decide how to handle them during analysis.

“Prefer not to answer” may sometimes be a meaningful response category, but it should not be combined automatically with unavailable or accidentally missing data.

How do you summarize a polytomous variable?

The most useful descriptive summaries are:

• Frequency for each category
• Percentage or proportion
• Valid percentage after excluding defined missing values
• Mode
• Bar chart
• Cross-tabulation with another categorical variable

For an ordinal variable, researchers may also report:

• Cumulative frequencies
• Cumulative percentages
• Median category
• Percentiles or quantiles where meaningful

Frequency tables and bar charts are appropriate for nominal and ordinal categorical variables.

Example frequency table

Satisfaction level	Frequency	Percentage
Very dissatisfied	18	9%
Dissatisfied	30	15%
Neutral	44	22%
Satisfied	72	36%
Very satisfied	36	18%
Total	200	100%

A bar chart should normally preserve the logical order for ordinal categories.

Which statistical method should you use?

The method depends on:

1. Whether the variable is a predictor or outcome.
2. Whether it is nominal or ordinal.
3. The measurement level of the other variables.
4. The study design.
5. Whether observations are independent, repeated, or clustered.

Research purpose	Role of polytomous variable	Common method
Describe one variable	Descriptive	Frequencies, percentages, bar chart
Test association between two categorical variables	Either variable	Chi-square test of independence
Compare a continuous outcome across categories	Predictor	ANOVA or linear regression with categorical coding
Predict an unordered multicategory outcome	Outcome	Multinomial logistic regression
Predict an ordered multicategory outcome	Outcome	Ordinal logistic regression
Analyze repeated or clustered categorical outcomes	Outcome	Mixed-effects or population-averaged categorical model
Examine several categorical variables symmetrically	No designated outcome	Log-linear analysis
Model ordered responses to test or questionnaire items	Item responses	Polytomous item response theory
Explore latent associations among ordinal items	Measurement variables	Polychoric correlations and suitable factor/SEM methods

No single statistical test is universally correct for every polytomous variable.

Chi-square analysis

A chi-square test of independence can assess whether two categorical variables are associated.

Example:

• Transport mode: car, bus, train, bicycle
• Residential area: urban, suburban, rural

The observations are arranged in a contingency table. The test compares observed cell counts with counts expected under independence.

Researchers should inspect sparse and empty cells. When expected counts are too small, an exact method, category redesign, or carefully justified category combination may be more appropriate.

ANOVA and linear regression

ANOVA may be appropriate when:

• The polytomous variable is a categorical predictor.
• The dependent variable is continuous.
• Model assumptions are sufficiently satisfied.

Example:

• Predictor: teaching method with four categories
• Outcome: examination score

A regression model with $K-1$K−1 indicator variables produces the same core group-comparison framework as a standard one-way ANOVA.

ANOVA is not ordinarily used when the dependent variable itself is nominal and polytomous.

Multinomial logistic regression

Multinomial logistic regression models a nominal outcome with three or more unordered categories. It estimates each non-reference outcome relative to a selected reference category.

For outcome categories $1,\ldots,J$1,…,J, with category $J$J as the reference:$$\log\left[\frac{P(Y=j\mid X)}{P(Y=J\mid X)}\right] = \beta_{j0}+X^\mathsf{T}\beta_j, \qquad j=1,\ldots,J-1.$$log[P(Y=J∣X)P(Y=j∣X)​]=βj0​+XTβj​,j=1,…,J−1.

Each comparison receives its own coefficient set.

Suppose the outcome is transport mode:

• Car, reference
• Bus
• Train
• Bicycle

The model estimates:

• Bus versus car
• Train versus car
• Bicycle versus car

Penn State describes this as an extension of binary logistic regression in which a response has more than two categories. UCLA and XLSTAT similarly distinguish multinomial models for nominal outcomes from ordinal models for ordered outcomes.

Interpreting results

Exponentiating a coefficient gives an outcome-specific relative-odds or relative-risk-ratio interpretation, depending on the software’s terminology.

Because interpretation becomes difficult when several categories and predictors are present, researchers should also report:

• Predicted probabilities
• Marginal effects
• Confidence intervals
• Clearly stated reference categories
• Graphs of adjusted predictions

Important considerations

Researchers should assess:

• Adequacy of observations in each outcome category
• Influential observations
• Multicollinearity
• Linearity assumptions for continuous predictors on the logit scale
• Model specification
• Whether the independence-of-irrelevant-alternatives assumption is defensible for the chosen multinomial-logit formulation

Alternative models may be needed when choices are structurally related, repeated, nested, or clustered.

Ordinal logistic regression

Ordinal logistic regression models an ordered categorical outcome while preserving its ranking. A common form is the proportional-odds cumulative logit model.

For cut point $j$j:$$\log\left[ \frac{P(Y\leq j\mid X)} {P(Y>j\mid X)} \right] = \alpha_j-X^\mathsf{T}\beta.$$log[P(Y>j∣X)P(Y≤j∣X)​]=αj​−XTβ.

The cut points $\alpha_j$αj​ vary across category boundaries, while the predictor coefficient vector $\beta$β is shared under the proportional-odds assumption. Some software uses the opposite sign convention, so the documentation must be checked when interpreting coefficients.

Ordinal modeling can be more parsimonious than ignoring the order and fitting a nominal multinomial model. Penn State notes that a proportional-odds model uses ordinal information to reduce the number of parameters, while MathWorks describes cumulative probabilities and threshold-based ordinal models.

The proportional-odds assumption

The standard model assumes that a predictor’s relationship with the outcome is consistent across cumulative category divisions.

For a five-category satisfaction outcome, those divisions include:

• Very dissatisfied versus all higher categories
• Very dissatisfied or dissatisfied versus all higher categories
• Up to neutral versus satisfied or very satisfied
• Up to satisfied versus very satisfied

If this assumption is not reasonable, researchers may consider:

• Partial proportional-odds models
• Generalized ordered-logit models
• Adjacent-category models
• Continuation-ratio models
• Multinomial logistic regression
• Models tailored to the data-generating process

Model choice should follow the research question and substantive meaning, rather than being based only on a preliminary significance test.

Are Likert responses polytomous variables?

A single Likert-type item with three or more response options is an ordinal polytomous variable.

For example:

• Strongly disagree
• Disagree
• Neither agree nor disagree
• Agree
• Strongly agree

The categories are ordered, but equal distance between adjacent responses is not guaranteed.

A multi-item Likert scale is more complicated. Researchers may sum or average several items to create a composite score, but whether that score can be treated as approximately continuous depends on the scale’s construction, distribution, reliability, research question, and modeling assumptions.

Recent methodological guidance continues to recommend methods designed for ordinal outcomes when the analysis concerns an individual Likert-type response, particularly when category probabilities and ties are substantively important.

Polytomous variables in item response theory

In item response theory, a polytomous item permits more than two possible scores or response categories.

Examples include:

• A constructed-response question scored 0–4
• A rating item scored from strongly disagree to strongly agree
• A task awarding no, partial, or full credit
• A nominal multiple-choice item where options provide diagnostic information

Common polytomous IRT models include:

• Graded response model
• Partial credit model
• Generalized partial credit model
• Rating scale model
• Nominal response model

The appropriate model depends on whether categories are ordered, how item discrimination is modeled, and whether threshold or step parameters are allowed to differ across items. ETS and methodological reviews identify graded-response, partial-credit, generalized partial-credit, rating-scale, and nominal-response approaches as major polytomous IRT families.

These models can help researchers examine:

• Category functioning
• Item difficulty or threshold locations
• Item discrimination
• Information across a latent trait
• Whether response categories are used as intended
• Whether adjacent categories should remain separate

Advantages of polytomous variables

They preserve meaningful distinctions

Several categories can represent real-world heterogeneity better than a binary division.

They support richer interpretation

Researchers can distinguish several response patterns rather than reporting only presence versus absence.

Ordered categories preserve severity or progression

Ordinal variables can represent levels such as mild, moderate, and severe without claiming exact equal intervals.

They can improve measurement

Partially scored and graded items can provide more information than correct-versus-incorrect scoring when the response process supports partial credit.

They support probability-based predictions

Regression models can estimate the probability of every category rather than reducing the outcome to a single binary event.

Limitations and practical challenges

More categories require more data

A larger number of categories produces more cell counts, parameters, comparisons, and possible sparse-data problems.

Interpretation becomes more complex

A multinomial model with many outcomes and predictors may generate a large set of coefficients.

Category definitions can be ambiguous

Respondents may interpret vague or overlapping categories differently.

Rare categories can cause instability

Very small categories may produce large standard errors, convergence problems, or quasi-separation.

Ordinal models require additional assumptions

The proportional-odds model, for example, imposes a common-effect structure across category thresholds.

Numeric coding can be misleading

Software may mistakenly treat codes as a continuous predictor unless the variable is explicitly declared categorical.

Should categories be combined?

Categories should be combined only when there is a defensible substantive or measurement-based reason.

Potentially defensible reasons include:

• Categories are conceptually indistinguishable for the research question.
• A validated scoring manual requires combination.
• Respondents cannot reliably distinguish adjacent response options.
• Category-response analysis shows that options do not function as intended.
• A preregistered policy threshold requires a particular grouping.

Weak reasons include:

• A desired result was not statistically significant.
• One category produces an inconvenient coefficient.
• Categories are combined repeatedly until a preferred model is obtained.

When categories are combined, researchers should report:

• The original categories
• The revised categories
• The timing of the decision
• The substantive justification
• Any sensitivity analysis using the original coding

Should a polytomous variable be converted to binary?

Usually, it should remain polytomous unless a scientifically meaningful binary decision is central to the research question.

For example, a clinical decision may legitimately distinguish whether a patient meets a treatment threshold. Even then, the original multicategory or continuous measure may remain valuable for secondary analysis.

Dichotomization can discard distinctions, reduce statistical power, obscure nonlinear relationships, and make results depend on an arbitrary cut point. Methodological literature has repeatedly warned about these costs.

Common mistakes

Mistake 1: Assuming any variable with several values is polytomous

Age, income, and counts can have many values but are not categorical unless they have been grouped.

Mistake 2: Treating category codes as measurements

Codes such as 1, 2, and 3 may be convenient labels. They do not automatically indicate equal differences.

Mistake 3: Ignoring category order

Using a nominal model for a clearly ordered outcome may discard useful structure.

Mistake 4: Imposing an artificial order

Blood group or transport mode should not be treated as ordinal because the categories happen to be coded numerically.

Mistake 5: Using binary logistic regression for an unchanged multicategory outcome

Binary logistic regression requires a two-category outcome. A multicategory outcome requires another formulation unless it is deliberately and justifiably recoded.

Mistake 6: Confusing a polytomous predictor with a polytomous outcome

The analysis is selected primarily according to the outcome variable and study design.

Mistake 7: Ignoring sparse categories

Category frequencies should be examined before fitting a model.

Mistake 8: Failing to state the reference category

Coefficients in reference-coded models cannot be interpreted clearly without knowing the comparison group.

Mistake 9: Treating missingness as an ordinary response

Missing, inapplicable, and refused responses require explicit documentation.

Mistake 10: Reporting coefficients without probabilities

Adjusted predicted probabilities are often easier for readers to understand than multiple log-odds coefficients.

How polytomous variables are used in modern research

Polytomous variables are common in:

• Electronic questionnaires
• Patient-reported outcome measures
• Educational assessments
• Administrative records
• Consumer-preference studies
• Machine-learning classification
• Public-policy research
• Ecological and epidemiological classification
• Natural-language annotation
• Human-resource and labor-market analysis

Modern research increasingly uses models that retain category detail rather than automatically converting outcomes into two groups. Current statistical platforms support nominal, ordinal, clustered, repeated, imputed, and latent-variable formulations. Stata, for example, documents multinomial, ordered, mixed-effects, and multiple-imputation methods for categorical outcomes, while MATLAB provides ordinal multinomial modeling and cumulative predicted probabilities.

Digital tools for analyzing polytomous variables

SPSS

SPSS can fit multinomial logistic regression and ordinal regression and can automatically code categorical predictors. Researchers must verify the selected reference category and category ordering.

R

Common R workflows use factor variables for nominal categories and ordered factors for ordinal categories. Packages support multinomial regression, cumulative-link models, mixed models, IRT, predicted probabilities, diagnostics, and visualization.

Stata

Stata provides commands for multinomial logit, ordered logit, ordered probit, mixed-effects categorical models, marginal predictions, and multiple imputation for categorical variables.

SAS

SAS supports generalized-logit and proportional-odds models through procedures designed for categorical responses.

MATLAB and similar tools

MATLAB supports ordinal multinomial models, several link functions, category probabilities, cumulative probabilities, and confidence bounds.

How artificial intelligence can assist

Generative AI can assist researchers by:

• Drafting preliminary R, Python, SPSS, Stata, or SAS syntax
• Creating a data-dictionary template
• Identifying inconsistent category labels
• Suggesting visualizations
• Translating statistical output into preliminary plain-language explanations
• Generating reproducible documentation from an approved analysis plan

AI should not independently decide:

• Whether categories have a meaningful order
• Whether categories should be combined
• Which outcome is scientifically primary
• Whether a model assumption is substantively reasonable
• Whether missing responses form a valid category
• Whether an automated interpretation is statistically correct

Researchers should inspect generated code, use official documentation, validate results against known examples, protect confidential data, and retain responsibility for every analytical decision.

How to report a polytomous variable

A clear methods section should state:

1. The variable name and construct.
2. Every category and its coding.
3. Whether categories are nominal or ordinal.
4. Which category served as the reference.
5. How missing and rare categories were handled.
6. Whether categories were combined and why.
7. The descriptive statistics.
8. The statistical model and link function.
9. Relevant model assumptions and diagnostics.
10. Effect estimates, uncertainty intervals, and predicted probabilities.

Example reporting statement

Primary transport mode was recorded as car, bus, train, bicycle, or walking and analyzed as a nominal categorical outcome. Car was the reference category. We fitted a multinomial logistic regression model and reported adjusted relative risk ratios, 95% confidence intervals, and predicted probabilities for each transport category.

Example for an ordinal outcome

Satisfaction was recorded on a five-category scale from very dissatisfied to very satisfied. Categories were retained in their original order. A cumulative-logit proportional-odds model was fitted, and the proportional-odds assumption and adjusted category probabilities were examined.

Conclusion

A polytomous variable is a categorical variable with three or more possible values. It can be nominal or ordinal and can function as a predictor, outcome, or scored item. Correct analysis begins by preserving the meaning and ordering of its categories, declaring it properly in statistical software, and selecting a model that matches its role, measurement level, and study design.