A categorical variable places observations into groups or categories, such as blood type, employment status or satisfaction level. Its values describe group membership rather than measurable quantities. The main types are nominal and ordinal variables, while binary variables are a special case containing exactly two categories.
Introduction
Categorical variables appear throughout academic research. A medical researcher may classify patients by treatment group, an education researcher may record degree level, and a business researcher may measure customer satisfaction using ordered response categories.
Correctly identifying a categorical variable matters because variable type influences how data should be collected, coded, summarized, visualized and analyzed. Treating category codes as ordinary numbers can produce meaningless averages, incorrect graphs and unsuitable statistical models.
This article explains:
- What categorical variables represent.
- How nominal, ordinal, binary and polytomous variables differ.
- How to distinguish categorical variables from quantitative and discrete variables.
- How to design, code and clean categories.
- Which tables, graphs, tests and regression models are appropriate.
- How categorical variables are handled in statistical software, machine learning and AI-assisted research.
Key takeaways
- Categorical variables represent group membership, labels or ordered classifications.
- Nominal categories have no inherent order; ordinal categories have a meaningful order but not necessarily equal intervals.
- Binary variables contain two possible categories, while polytomous variables contain more than two.
- Numbers such as 0, 1 and 2 may be category codes rather than numerical measurements.
- Categorical data are usually summarized with counts, percentages, modes, bar charts and contingency tables.
- The appropriate statistical method depends on whether the categorical variable is an outcome, predictor, grouping variable or control variable.
What Is a Categorical Variable?
A categorical variable is a variable whose values place observations into distinct groups or categories. The categories describe qualities, types, states, identities or classifications rather than numerical amounts.
For example, a researcher recording participants’ employment status might use the categories:
- Employed
- Unemployed
- Student
- Retired
- Not currently seeking work
These values identify groups. It would not make sense to add “student” to “retired” or calculate their arithmetic average.
Categorical variables are also called qualitative variables, although “categorical” is often the clearer term because categories may be represented by words, letters or numerical codes.
A note about terminology
Most introductory statistics texts use categorical variable as an umbrella term covering both nominal and ordinal variables. Some sources use “categorical” more narrowly as a synonym for “nominal” and discuss ordinal variables separately.
Researchers should therefore define their terminology when writing a dissertation, thesis or research paper. In this article, categorical variables include both nominal and ordinal variables.
Categorical Variable Versus Categorical Data
A variable is the characteristic being recorded. Data are the observed values of that variable.
For example:
- Variable: blood type
- Possible categories: A, B, AB and O
- Observed data: A, O, B, A, AB, O and so forth
Similarly:
- Variable: satisfaction level
- Possible categories: very dissatisfied, dissatisfied, neutral, satisfied and very satisfied
- Observed data: each participant’s selected response
The terms are closely related, but they are not identical. A categorical variable defines what is measured, while categorical data are the recorded category values.
Characteristics of Categorical Variables
A well-designed categorical variable generally has the following characteristics.
Categories represent membership
Each value identifies the category or categories to which an observation belongs.
Categories should be clearly defined
Researchers should provide category labels that respondents, data collectors and analysts can interpret consistently.
Categories should be mutually exclusive when only one response is allowed
For a single-response variable, one observation should not fit into two categories at the same time.
For example, the age groups “18–25” and “25–35” overlap at age 25. Better boundaries would be:
- 18–24
- 25–34
- 35–44
Categories should be collectively exhaustive where practical
The available options should cover all plausible responses. This may require options such as:
- Other, please specify
- Not applicable
- Prefer not to answer
- Unknown
However, these labels should not automatically be treated as equivalent. “Prefer not to answer” is different from “not applicable,” and both differ from a value that is missing because of a data-entry problem.
Arithmetic operations are usually not meaningful
Category codes may be counted, compared for equality and used in statistical models. They should not ordinarily be added or averaged as though they represented measured quantities.
Types of Categorical Variables
The two fundamental types are nominal and ordinal variables. Binary and polytomous variables describe the number of available categories and can overlap with the nominal–ordinal classification.
Nominal Variable
A nominal variable contains categories that have no inherent rank or order.
Examples include:
- Blood type: A, B, AB or O
- Mode of transport: car, bus, bicycle, train or walking
- Academic discipline: biology, economics, history or engineering
- Operating system: Windows, macOS, Linux or another system
- Marital status: single, married, divorced or widowed
The order in which nominal categories appear in a questionnaire or dataset is arbitrary. Alphabetical order may improve presentation, but it does not create a substantive ranking.
For nominal variables, useful summaries include:
- Frequency
- Percentage or proportion
- Mode
- Bar chart
- Contingency table
A mean or median is not meaningful for an unordered nominal variable.
Ordinal Variable
An ordinal variable contains categories with a meaningful sequence or ranking. However, the differences between adjacent categories are not necessarily equal.
Examples include:
- Satisfaction: very dissatisfied to very satisfied
- Disease severity: mild, moderate or severe
- Education level: secondary school, bachelor’s, master’s or doctorate
- Agreement: strongly disagree to strongly agree
- Class position: first, second, third and so forth
- Socioeconomic classification: low, middle or high
The ordering provides more information than a nominal classification, but it does not establish equal distances.
For example, the difference between “dissatisfied” and “neutral” may not be psychologically equal to the difference between “satisfied” and “very satisfied.”
Ordinal data may be summarized with:
- Frequencies and percentages
- Mode
- Median and percentiles when the ordering is defensible
- Ordered bar charts
- Ordinal association measures
- Ordinal regression models
Researchers should be especially careful with individual Likert-type items. A single response item is ordinarily treated as ordinal. A multi-item composite scale may sometimes be analyzed approximately as a quantitative score, but that decision requires justification based on measurement properties, study conventions and analytical assumptions (Sullivan & Artino, 2013).
Binary or Dichotomous Variable
A binary variable, also called a dichotomous variable, has exactly two possible categories.
Examples include:
- Yes or no
- Present or absent
- Pass or fail
- Treatment or control
- Positive or negative test result
- Purchased or did not purchase
A binary variable may be nominal or ordinal.
For example:
- Blood-test result—positive or negative—is generally nominal.
- Disease status—absent or present—may have an implicit progression but is commonly modelled as a binary outcome.
Binary variables are often coded as 0 and 1. The values are still category indicators unless the research context gives them a separate quantitative interpretation.
Polytomous Variable
A polytomous variable has more than two categories.
Examples include:
- Blood group with four categories
- Voting preference with several parties
- Employment status with five categories
- Satisfaction with five ordered levels
A polytomous variable may be nominal or ordinal:
- Unordered categories produce a nominal polytomous variable.
- Ordered categories produce an ordinal polytomous variable.
Multiple-response or Multi-label Variable
A multiple-response question allows an observation to belong to more than one category.
For example:
Which research software do you use? Select all that apply.
A participant might select R, SPSS and Python. This differs from an ordinary single-response categorical variable because the categories are not mutually exclusive.
A common coding method is to create a separate binary indicator for each option:
| Participant | Uses R | Uses SPSS | Uses Python |
|---|---|---|---|
| 1 | 1 | 1 | 1 |
| 2 | 0 | 1 | 0 |
| 3 | 1 | 0 | 1 |
Researchers must clarify whether reported percentages are based on participants or total selections. In a multi-response question, percentages based on participants may sum to more than 100%.
Summary of Categorical Variable Types
| Type | Defining feature | Example | Is order meaningful? |
|---|---|---|---|
| Nominal | Unordered categories | Blood type | No |
| Ordinal | Ranked categories with uncertain intervals | Satisfaction level | Yes |
| Binary | Exactly two categories | Pass/fail | Depends on meaning |
| Polytomous | More than two categories | Employment status | May or may not be |
| Multiple-response | More than one category can apply | Software used | Depends on individual options |
Binary and polytomous describe the number of categories. Nominal and ordinal describe whether the categories have a meaningful order.
How to Identify a Categorical Variable
Use the following process.
Step 1: Ask what each value represents
Does the value represent a group, label, type or status? If so, the variable may be categorical.
Does it represent an amount that can be meaningfully added or subtracted? If so, it is probably quantitative.
Step 2: Ignore the storage format temporarily
Do not assume a variable is quantitative simply because it contains numbers.
For example:
- Postal code 10001 is a label.
- Participant ID 247 is an identifier.
- Department code 3 is a category code.
- Satisfaction code 5 may mean “very satisfied.”
None of these numbers necessarily represents a measurable amount.
Step 3: Test whether arithmetic is meaningful
Ask whether adding, subtracting or averaging the values answers a sensible research question.
The average of income values can be meaningful. The average of postal codes is not.
Step 4: Determine whether the categories have an inherent order
If there is no natural ranking, the variable is nominal.
If there is a meaningful ranking but the intervals are uncertain, the variable is ordinal.
Step 5: Count the possible categories
Exactly two categories indicate a binary variable. More than two indicate a polytomous variable.
Step 6: Check whether several categories can apply
When respondents can select multiple answers, the data may require several binary variables rather than one conventional categorical column.
Categorical Versus Quantitative Variables
A categorical variable identifies groups. A quantitative variable records numerical amounts or measurements.
| Feature | Categorical variable | Quantitative variable |
|---|---|---|
| Main purpose | Classifies observations | Measures or counts an amount |
| Typical values | Labels, names or coded categories | Meaningful numerical values |
| Examples | Blood type, faculty, satisfaction | Height, income, age, number of publications |
| Arithmetic | Usually not meaningful | Usually meaningful |
| Common summaries | Counts, percentages, mode | Mean, median, standard deviation, range |
| Common graphs | Bar chart, mosaic plot | Histogram, box plot, scatterplot |
| Common models | Logistic, multinomial or ordinal models | Linear and related numerical models |
Can the same concept be categorical or quantitative?
Yes. The classification depends on how the concept is measured.
Age can be recorded as:
- Exact age in years: quantitative
- Age group such as 18–24, 25–34 and 35–44: ordinal categorical
Income can be recorded as:
- Exact annual income: quantitative
- Income band such as low, middle and high: ordinal categorical
The variable type is determined by the recorded values, not only by the underlying concept.
Categorical Versus Discrete Variables
Categorical and discrete variables are not synonyms.
A categorical variable identifies groups. A discrete quantitative variable records a count and usually takes separate whole-number values.
Examples of discrete quantitative variables include:
- Number of children
- Number of publications
- Number of hospital visits
- Number of errors
- Number of employees
The value “3 children” represents a quantity. It is sensible to say that one household has two more children than another.
By contrast, department codes 1, 2 and 3 may merely represent marketing, finance and operations. Subtracting department 1 from department 3 has no meaningful interpretation.
Can Categorical Variables Be Represented by Numbers?
Yes. Numbers may be used as category codes, but the coding does not automatically make the variable quantitative.
For example:
| Code | Employment status |
|---|---|
| 1 | Employed |
| 2 | Unemployed |
| 3 | Student |
| 4 | Retired |
The code 4 does not mean that “retired” is four times “employed.” It is simply a stored label.
Researchers should preserve a codebook explaining every code. Statistical software should also be told whether a coded variable is nominal, ordinal or quantitative.
Examples of Categorical Variables in Research
| Field | Variable | Categories | Type |
|---|---|---|---|
| Medicine | Treatment group | Drug, placebo, usual care | Nominal |
| Public health | Smoking status | Current, former, never | Nominal |
| Psychology | Anxiety severity | Mild, moderate, severe | Ordinal |
| Education | Degree level | School, bachelor’s, master’s, doctorate | Ordinal |
| Sociology | Employment status | Employed, unemployed, student, retired | Nominal |
| Business | Customer loyalty tier | Bronze, silver, gold | Ordinal |
| Political science | Party preference | Party A, Party B, independent, other | Nominal |
| Biology | Species | Species names | Nominal |
| Computer science | Operating system | Windows, macOS, Linux, other | Nominal |
| Clinical research | Adverse event | Present, absent | Binary |
| Marketing | Purchase decision | Purchased, did not purchase | Binary |
| Library science | Resource type | Book, article, dataset, video | Nominal |
Designing a Categorical Variable
Poor category design cannot always be repaired during analysis. Researchers should plan categories before collecting data.
1. Define the construct
State exactly what the variable is intended to represent.
“Employment” is ambiguous. “Current primary employment status” is more precise.
2. Define the unit of analysis
Clarify whether categories describe:
- People
- Households
- Schools
- Organizations
- Countries
- Transactions
- Documents
- Events
3. Decide whether one or several responses are allowed
A single-response question requires mutually exclusive categories. A select-all-that-apply question requires multi-response coding.
4. Make categories clear and non-overlapping
Avoid vague distinctions such as:
- Sometimes
- Often
- Regularly
unless each term has an operational definition.
5. Include necessary response options
Depending on the research question, useful options may include:
- Other, please specify
- Not applicable
- Do not know
- Prefer not to answer
These should be analytically distinguishable.
6. Preserve appropriate detail
Avoid collapsing categories too early. Detailed categories can be combined later when scientifically justified, but information lost during collection cannot be recovered.
7. Pilot-test the categories
Ask pilot participants whether:
- The labels are understandable.
- An appropriate response is available.
- Any categories overlap.
- The order is logical.
- Sensitive wording is respectful.
8. Document all decisions
Maintain a codebook describing labels, codes, order, missing-value rules, transformations and category-combination decisions.
Copyable Categorical-Variable Codebook Template
| Field | Entry |
|---|---|
| Variable name | |
| Variable label | |
| Research construct | |
| Unit of analysis | |
| Variable role | Outcome / predictor / control / grouping |
| Categorical type | Nominal / ordinal / binary / polytomous |
| Single or multiple response | |
| Permitted categories | |
| Numerical codes | |
| Reference category | |
| Ordered levels | |
| Missing-value codes | |
| “Other” handling | |
| Validation rules | |
| Source question or instrument | |
| Transformations or recoding | |
| Notes and limitations |
How to Summarize One Categorical Variable
The primary summaries are frequencies and proportions.
Frequency
The frequency (n_k) is the number of observations in category (k).
Proportion
The proportion in category (k) is:
[
p_k = \frac{n_k}{N}
]
where:
- (n_k) is the frequency of category (k)
- (N) is the total number of valid observations
The percentage is:
[
100 \times p_k
]
Example frequency table
Suppose 120 students report their preferred study format.
| Format | Frequency | Percentage |
|---|---|---|
| In person | 48 | 40.0% |
| Online | 42 | 35.0% |
| Hybrid | 30 | 25.0% |
| Total | 120 | 100.0% |
The mode is “in person” because it has the greatest frequency.
Appropriate graphs
A bar chart is normally the clearest graph for a categorical variable.
Useful alternatives include:
- Ordered bar chart for ordinal categories
- Stacked or clustered bar chart for two categorical variables
- Mosaic plot for contingency tables
- Dot plot for many category percentages
Pie charts can be used for simple part-to-whole displays, but bar charts usually make differences easier to compare. Histograms are designed for quantitative distributions and should not be used merely because categories have numerical codes.
How to Analyze Two Categorical Variables
Two categorical variables are commonly summarized with a contingency table, also called a cross-tabulation or two-way table.
Suppose a researcher compares study format with examination result.
| Study format | Passed | Failed | Total |
|---|---|---|---|
| Online | 42 | 18 | 60 |
| In person | 54 | 6 | 60 |
| Total | 96 | 24 | 120 |
Useful calculations include:
- Row percentages
- Column percentages
- Overall percentages
- Expected counts
- Chi-square statistic
- Odds ratio or risk ratio for suitable (2 \times 2) tables
- Cramér’s V as a measure of association
Researchers should choose row or column percentages according to the research question. Raw counts alone may hide important differences when group sizes are unequal.
Chi-Square Test of Independence
The Pearson chi-square test of independence evaluates whether two categorical variables are associated in a population.
The test statistic is:
[
\chi^2 = \sum_i \sum_j \frac{(O_{ij}-E_{ij})^2}{E_{ij}}
]
where:
- (O_{ij}) is the observed count in cell (i,j)
- (E_{ij}) is the expected count under independence
An expected count is calculated as:
[
E_{ij} =
\frac{(\text{row total})(\text{column total})}{N}
]
A small p-value indicates that the observed table would be unusual if the variables were independent. It does not establish causation or reveal the practical importance of the association.
Check the assumptions
Researchers should check:
- Whether observations are independent.
- Whether categories are mutually exclusive.
- Whether the sampling or assignment procedure supports the intended inference.
- Whether expected cell counts are adequate.
- Whether sparse or structural-zero cells are present.
When expected counts are too small, Fisher’s exact test, an exact procedure, a permutation method, category restructuring or an appropriate model may be preferable. Categories should not be combined solely to force significance or satisfy a mechanical rule.
Worked Example
Using the study-format table:
| Study format | Passed | Failed |
|---|---|---|
| Online | 42 | 18 |
| In person | 54 | 6 |
Under independence, the expected counts are:
| Study format | Expected passed | Expected failed |
|---|---|---|
| Online | 48 | 12 |
| In person | 48 | 12 |
The chi-square statistic is:
[
\chi^2 =
\frac{(42-48)^2}{48}
+
\frac{(18-12)^2}{12}
+
\frac{(54-48)^2}{48}
+
\frac{(6-12)^2}{12}
= 7.50
]
For a (2 \times 2) table:
[
df=(2-1)(2-1)=1
]
The resulting p-value is approximately .006. This provides evidence of an association between study format and examination result in the synthetic dataset.
However, the test does not prove that in-person teaching caused the difference. Prior attainment, course selection, attendance and other variables could explain part of the association.
Cramér’s V
For a contingency table, Cramér’s V can be calculated as:
[
V =
\sqrt{
\frac{\chi^2}
{N \times \min(r-1,c-1)}
}
]
For the example:
[
V = \sqrt{\frac{7.50}{120}} = 0.25
]
Cramér’s V describes the strength of association on a scale from 0 to 1. Its interpretation should consider the discipline, table dimensions, sample size and substantive consequences rather than relying only on universal labels.
Choosing a Statistical Method
The appropriate method depends on the variable’s role and the study design.
| Research situation | Common approach |
|---|---|
| Describe one categorical variable | Frequencies, proportions, mode, confidence interval for a proportion |
| Compare observed category counts with expected counts | Chi-square goodness-of-fit test or exact alternative |
| Examine two categorical variables | Contingency table and chi-square test of independence |
| Small or sparse (2 \times 2) table | Fisher’s exact test or another exact method |
| Measure association in a contingency table | Cramér’s V, odds ratio, risk ratio or other context-specific measure |
| Binary categorical outcome | Binary logistic regression |
| Nominal outcome with more than two categories | Multinomial logistic regression |
| Ordered categorical outcome | Ordinal logistic regression or another ordinal model |
| Quantitative outcome across two groups | Independent or paired t test, or a suitable alternative |
| Quantitative outcome across several groups | ANOVA or regression with categorical predictors |
| Paired binary responses | McNemar’s test |
| Repeated or clustered categorical observations | Generalized estimating equations, mixed-effects models or survey-adjusted methods |
| Several categorical variables | Log-linear models, multiple correspondence analysis or generalized linear models |
This table provides general guidance, not an automatic decision rule. The correct method also depends on independence, pairing, clustering, sample size, distributional assumptions, survey weights and the research question.
Categorical Variables in Regression
Categorical variables may appear as outcomes or predictors.
Categorical outcome
The outcome determines the model family:
- Binary outcome: binary logistic regression
- Unordered outcome with several categories: multinomial logistic regression
- Ordered outcome: ordinal logistic regression
- Count outcome: Poisson or negative-binomial models, which concern discrete counts rather than ordinary categorical outcomes
Categorical predictor
A categorical predictor must be represented in a form the model can interpret.
For a nominal predictor with (K) categories, a regression model containing an intercept commonly uses (K-1) indicator variables. One category is treated as the reference.
Suppose employment status has four levels:
- Employed
- Unemployed
- Student
- Retired
If “employed” is the reference, the model may contain indicators for:
- Unemployed versus employed
- Student versus employed
- Retired versus employed
Each coefficient is interpreted relative to the reference category, subject to the scale and link function of the model.
Reference-category choice
Changing the reference category changes the interpretation of individual coefficients but does not change the fitted values of a correctly specified model.
Choose a reference that is:
- Scientifically meaningful
- Common in the data
- A standard or control condition
- Easy for readers to interpret
Contrast coding
Dummy or treatment coding is not the only option. Alternatives include:
- Effect or sum coding
- Helmert contrasts
- Polynomial contrasts for ordered categories
- Custom hypothesis-based contrasts
The coding system should match the research question.
Should Continuous Variables Be Categorized?
Researchers sometimes convert a continuous variable into categories, such as changing exact age into “young,” “middle-aged” and “older.”
Categorization may be useful when:
- Established clinical thresholds exist.
- Policy decisions rely on defined cut-offs.
- Categories improve communication with a specific audience.
- The original measurement is unavailable.
- The research question concerns membership in recognized groups.
However, arbitrary categorization can:
- Discard information.
- Reduce statistical power.
- Hide nonlinear patterns.
- Create artificial boundaries.
- Leave residual differences within categories.
- Produce results that depend on chosen cut-points.
Dichotomizing a continuous variable at its median is especially difficult to justify without a scientific or decision-based reason (Altman & Royston, 2006). Researchers should retain the original numerical variable whenever possible, even if a categorized version is also reported.
Data Cleaning for Categorical Variables
Categorical data often require more cleaning than their simple appearance suggests.
Standardize labels
These may represent the same category:
- United States
- USA
- U.S.
- United States of America
Researchers should define one standardized label while preserving the original data where appropriate.
Remove accidental whitespace and case differences
“Online,” “online” and “Online ” may be treated as separate values by software.
Validate permitted categories
Flag values that do not appear in the codebook.
Distinguish missingness reasons
Do not automatically combine:
- Missing
- Unknown
- Not applicable
- Refused
- Prefer not to answer
- Data not collected
These values may have different methodological meanings.
Examine rare categories
Rare levels can produce unstable estimates, wide confidence intervals and sparse contingency tables. Combining them may be appropriate only when the combined category has a defensible substantive meaning.
Preserve an audit trail
Record every recoding decision, including:
- Original value
- Standardized value
- Reason for change
- Date
- Person or process making the change
Categorical Variables in Statistical Software
SPSS
SPSS allows researchers to define value labels and measurement levels such as nominal, ordinal and scale. Researchers should not rely entirely on automatic measurement-level detection, especially when numbers are category codes.
R
R commonly represents categorical variables as factors. An ordered factor can be used when levels have a meaningful sequence.
Researchers should explicitly set:
- Factor levels
- Level order
- Reference category
- Missing-value handling
Python
Python workflows commonly use:
categorydata types in pandas- One-hot encoders
- Ordinal encoders
- Formula interfaces that generate categorical contrasts
- Pipelines that apply identical preprocessing during training and prediction
Stata and SAS
Stata factor-variable notation and SAS classification statements can create indicators and contrasts from categorical predictors. Reference levels and coding conventions should be documented in the analysis plan.
Categorical Variables in Machine Learning
Many machine-learning algorithms require categories to be converted into a numerical representation.
One-hot encoding
One-hot encoding creates a binary column for each category or for each non-reference category, depending on the model and software.
It is commonly suitable for unordered nominal variables with a manageable number of levels.
Ordinal encoding
Ordinal encoding assigns ordered numbers to categories. It is suitable only when the order is meaningful.
Even then, assigning 1, 2, 3 and 4 may cause some algorithms to assume equal spacing. Researchers should understand how the selected model interprets the codes.
High-cardinality variables
A variable with hundreds or thousands of categories is called high-cardinality.
Examples include:
- Product ID
- Hospital
- Occupation
- Postal area
- Website domain
Possible strategies include:
- Scientifically justified grouping
- Rare-level handling
- Regularization
- Hashing
- Carefully validated target encoding
- Embeddings
- Models with native categorical-feature support
The choice should be tested with out-of-sample validation. Categories should not be combined merely because they are inconvenient.
Unseen categories
A model may encounter a category during deployment that was absent from the training data. The preprocessing pipeline must define how such values are handled, such as mapping them to an unknown level or using an encoder configured to accept unseen categories.
Artificial Intelligence and Categorical Data
AI tools can assist researchers with:
- Suggesting preliminary category systems.
- Classifying open-text responses.
- Detecting spelling variants.
- Identifying potentially overlapping labels.
- Producing code for frequency tables and regression models.
- Documenting a draft codebook.
However, AI-generated classifications require human review.
Researchers should check:
- Whether the category definitions match the research construct.
- Whether culturally specific responses have been misclassified.
- Whether sensitive demographic categories are represented appropriately.
- Whether the same rules are applied consistently.
- Whether confidential data were entered into an approved system.
- Whether automated classification accuracy has been evaluated.
For reproducibility, document the model or tool, version, prompt or instruction set, human-review process and disagreement-resolution procedure.
Ethical and Inclusive Category Design
Categories are not always neutral. They may reflect legal systems, institutional conventions, cultural assumptions or historical classifications.
When collecting sensitive demographic information:
- Explain why the variable is needed.
- Use terminology appropriate to the study population.
- Avoid unnecessarily restrictive categories.
- Consider self-description where practical.
- Separate analytically distinct concepts.
- Provide “prefer not to answer” when appropriate.
- Do not infer sensitive characteristics without a valid ethical and methodological basis.
- Report small cells carefully to protect confidentiality.
The correct design depends on the research question, jurisdiction, community and ethical approval.
Advantages of Categorical Variables
Categorical variables:
- Make complex characteristics easier to classify.
- Work naturally with questionnaire responses and administrative records.
- Support comparisons between groups.
- Can represent attributes that are not meaningfully measurable on a numerical scale.
- Are interpretable for decision-making and communication.
- Can be used in contingency tables, generalized linear models and classification systems.
Limitations of Categorical Variables
Categorical variables may:
- Provide less detail than continuous measurements.
- Depend heavily on category definitions.
- Create sparse data when many levels exist.
- Conceal differences within broad groups.
- Introduce arbitrary boundaries.
- Be difficult to compare across studies using different classifications.
- Require special coding in statistical models.
- Produce unstable estimates when categories contain few observations.
These limitations do not make categorical variables inferior. They show why careful operationalization and transparent reporting are necessary.
Common Mistakes
Treating numerical codes as measurements
Codes such as 1, 2 and 3 may be labels rather than amounts.
Calculating a mean for an unordered variable
An average department, blood group or nationality has no meaningful interpretation.
Ignoring ordinal information
Treating ordered responses as nominal discards their ranking.
Assuming ordinal categories have equal intervals
The difference between adjacent response categories may not be constant.
Confusing discrete counts with categories
The number of hospital visits is a discrete quantitative variable, not automatically a categorical variable.
Using overlapping categories
Overlapping age or income bands create ambiguous classifications.
Omitting important response options
Participants may be forced into inaccurate categories when no suitable option is provided.
Treating missing values as an ordinary group
“Missing,” “unknown” and “prefer not to answer” may require different treatment.
Combining rare categories without justification
Frequency alone is not a sufficient conceptual reason to merge groups.
Reporting only a p-value
Report relevant frequencies, percentages, effect sizes and confidence intervals as well.
Claiming causation from association
An association between two categorical variables does not by itself establish a causal effect.
How to Report a Categorical Variable
A clear report should state:
- The variable’s operational definition.
- All category labels.
- Whether the categories are nominal or ordinal.
- Whether one or several responses were permitted.
- The valid denominator and amount of missing data.
- Frequencies and percentages.
- Any category combination or recoding.
- The reference category used in a model.
- The statistical test and assumptions.
- An effect size and confidence interval where appropriate.
Example reporting template
Employment status was recorded as employed, unemployed, student, retired or other. The variable was treated as nominal. Frequencies and valid percentages were reported after excluding values missing because of nonresponse. For regression analysis, employed was used as the reference category. Two rare categories were combined before analysis based on their shared substantive definition, as specified in the preregistered analysis plan.
Conclusion
A categorical variable classifies observations into groups rather than measuring numerical amounts. Nominal variables contain unordered categories, whereas ordinal variables contain ranked categories with uncertain intervals. Binary variables have two possible categories, and polytomous variables have more than two.
Accurate analysis requires more than recognizing category labels. Researchers must design clear categories, distinguish codes from quantities, document missing values, choose suitable tables and statistical models, and report frequencies, uncertainty and effect sizes transparently.
References
- Agresti, A. (2018). An introduction to categorical data analysis (3rd ed.). John Wiley & Sons. https://doi.org/10.1002/9781119405283
- Altman, D. G., & Royston, P. (2006). The cost of dichotomising continuous variables. BMJ, 332(7549), 1080. https://doi.org/10.1136/bmj.332.7549.1080
- Australian Bureau of Statistics. (2023, February 2). Variables. https://www.abs.gov.au/statistics/understanding-statistics/statistical-terms-and-concepts/variables
- National Institute of Standards and Technology. (n.d.). Chi-square independence test. Retrieved June 27, 2026, from https://www.itl.nist.gov/div898/software/dataplot/refman1/auxillar/chistest.htm
- Pennsylvania State University, Department of Statistics. (n.d.-a). Relationships between categorical variables. Retrieved June 27, 2026, from https://online.stat.psu.edu/stat100/lesson/6
- Pennsylvania State University, Department of Statistics. (n.d.-b). STAT 504: Analysis of discrete data. Retrieved June 27, 2026, from https://online.stat.psu.edu/stat504/
- R Core Team. (n.d.). Factor: Factors. R Documentation. Retrieved June 27, 2026, from https://stat.ethz.ch/R-manual/R-devel/library/base/html/factor.html
- Scikit-learn developers. (n.d.). OneHotEncoder. Scikit-learn. Retrieved June 27, 2026, from https://scikit-learn.org/stable/modules/generated/sklearn.preprocessing.OneHotEncoder.html
- Statsmodels developers. (n.d.). Contrast coding systems for categorical variables. Statsmodels. Retrieved June 27, 2026, from https://www.statsmodels.org/stable/contrasts.html
- Stevens, S. S. (1946). On the theory of scales of measurement. Science, 103(2684), 677–680. https://doi.org/10.1126/science.103.2684.677
- Sullivan, G. M., & Artino, A. R., Jr. (2013). Analyzing and interpreting data from Likert-type scales. Journal of Graduate Medical Education, 5(4), 541–542. https://doi.org/10.4300/JGME-5-4-18
