An ordinal variable is an ordered categorical variable. Its categories follow a meaningful sequence, such as low, medium, and high, but the distance between one category and the next is not known to be equal. Ordinal variables commonly represent rankings, satisfaction levels, education levels, disease stages, and Likert-type responses.
Introduction
Researchers must identify a variable’s measurement level before choosing how to code, summarize, visualize, and analyze it. A method that is appropriate for heights or test times may be misleading when applied to categories such as “dissatisfied,” “neutral,” and “satisfied.”
This article explains what an ordinal variable is, how to recognize one, how it differs from other variables, and which descriptive statistics, graphs, tests, and regression models may be appropriate. It also addresses Likert scales, software coding, artificial intelligence, and common reporting mistakes.
Key takeaways
- An ordinal variable contains categories with a meaningful order.
- The gaps between adjacent categories are unequal, unknown, or not directly measurable.
- Numeric codes such as 1, 2, and 3 preserve order but do not automatically create equal intervals.
- Frequencies, percentages, the mode, cumulative percentages, and a median category are often useful summaries.
- Statistical analysis should reflect the research design, outcome type, assumptions, ties, and intended interpretation.
- Ordinal logistic regression is often appropriate when an ordered category is the outcome of a multivariable analysis.
What is an ordinal variable?
An ordinal variable is a categorical variable whose possible values can be placed in a meaningful order. The categories tell the researcher whether one observation is lower, higher, better, worse, earlier, later, more severe, or less severe than another. They do not necessarily show how much greater or smaller it is.
For example, a satisfaction variable might contain five categories:
- Very dissatisfied
- Dissatisfied
- Neither satisfied nor dissatisfied
- Satisfied
- Very satisfied
The order is meaningful. “Satisfied” represents more satisfaction than “dissatisfied.” However, the difference between categories 1 and 2 is not proven to be the same as the difference between categories 4 and 5.
Ordinal is traditionally described as the second of four levels of measurement: nominal, ordinal, interval, and ratio (Stevens, 1946). Modern statisticians sometimes criticize rigid applications of this four-level framework, but “ordered categorical variable” remains an important and widely used data type.
Main characteristics of an ordinal variable
An ordinal variable has several defining features.
The categories are distinct
Each observation belongs to a category such as mild, moderate, or severe. Well-designed categories should be clear, mutually exclusive, and sufficiently comprehensive for the intended population.
The categories have an intrinsic order
The order must reflect the meaning of the variable, not an arbitrary sorting rule.
For example:
- Beginner, intermediate, advanced: meaningful order.
- Red, blue, green: no intrinsic order.
- First, second, third: meaningful order.
- France, Brazil, Japan: no intrinsic order.
Alphabetical order does not turn nominal categories into ordinal categories.
The category intervals are not established as equal
Suppose educational attainment is coded as:
- 1 = primary education
- 2 = secondary education
- 3 = bachelor’s degree
- 4 = postgraduate degree
The difference between codes 1 and 2 cannot automatically be interpreted as the same amount of educational progression as the difference between codes 3 and 4.
Numeric labels may be codes rather than quantities
A value of 4 does not necessarily represent twice as much of the underlying characteristic as a value of 2. The numbers indicate sequence unless a separate measurement model justifies a metric interpretation.
The direction must be documented
Higher codes do not always mean better outcomes.
- In a satisfaction scale, 5 might mean very satisfied.
- In a disease-stage variable, 5 might mean more severe disease.
- In a race ranking, 1 represents the highest finishing position.
Researchers should store category labels and direction alongside the numeric codes.
Examples of ordinal variables
| Field | Ordinal variable | Ordered categories |
|---|---|---|
| Education | Highest qualification | Primary, secondary, bachelor’s, master’s, doctorate |
| Survey research | Satisfaction | Very dissatisfied to very satisfied |
| Psychology | Agreement with a statement | Strongly disagree to strongly agree |
| Healthcare | Disease severity | Mild, moderate, severe, critical |
| Medicine | Cancer stage | Stage I, II, III, IV |
| Employment | Seniority | Junior, intermediate, senior, lead |
| Economics | Income category | Low, middle, high |
| Sociology | Perceived social status | Lower, lower-middle, middle, upper-middle, upper |
| Hospitality | Hotel rating | One star to five stars |
| Education assessment | Performance category | Below basic, basic, proficient, advanced |
| Sport | Finishing position | First, second, third, and so on |
| Risk analysis | Risk level | Negligible, low, moderate, high, extreme |
| Public policy | Priority rating | Not a priority, low, medium, high priority |
| Customer research | Purchase likelihood | Very unlikely to very likely |
| Human resources | Performance rating | Unsatisfactory, developing, effective, exceptional |
Verbal ordinal categories
Some ordinal variables use words:
- Poor, fair, good, excellent
- Never, rarely, sometimes, often, always
- Low, medium, high
- Mild, moderate, severe
The words carry the order.
Numeric rating categories
Other ordinal variables use numbers whose meanings are defined by labels:
- One to five stars
- A 1–7 agreement item
- A zero-to-ten self-reported severity rating
- Priority levels 1 through 4
Numeric appearance alone does not prove that the intervals are equal.
Formal ranks
Competition positions and preference rankings are ordinal:
- First choice, second choice, third choice
- First place, second place, third place
- University rank 1, rank 2, rank 3
Ranks show order but not the size of the performance gap. The difference between first and second could be much smaller than the difference between second and third.
Categorized continuous measurements
A continuous variable may become ordinal when grouped:
- Age converted to 18–24, 25–34, 35–44, and 45+
- Income converted to low, middle, and high bands
- Body mass index converted to ordered clinical categories
- A continuous test score converted to fail, pass, merit, and distinction
After grouping, the new variable is ordered categorical. The original continuous precision has been reduced and generally cannot be recovered from the categories alone.
How can you identify an ordinal variable?
Use the following five-step test.
Step 1: Determine whether the values are categories
Ask whether each value identifies a group or level. If the observations are exact measurements in units such as seconds, kilograms, centimetres, or dollars, the variable is more likely quantitative.
Step 2: Ask whether the categories have a natural order
Can the categories be placed from lower to higher based on the construct being measured?
“Low, medium, high” can be ordered. “Biology, chemistry, physics” cannot be placed in an intrinsic order.
Step 3: Ask whether the distances are known and comparable
Can you defend the claim that moving from category 1 to category 2 represents the same amount of change as moving from category 3 to category 4?
If not, the variable is likely ordinal rather than interval.
Step 4: Check whether zero represents complete absence
A true zero is not required for an ordinal variable. Even when a category is labelled zero or “none,” the remaining category intervals may still be unequal.
Step 5: Read the variable definition, not only its stored values
A spreadsheet containing 1, 2, 3, 4, and 5 does not reveal the measurement level by itself. The values might represent:
- Ordered satisfaction categories.
- Unordered department codes.
- Exact counts.
- Identification numbers.
The codebook, questionnaire, and operational definition determine the meaning.
Quick identification rule
A variable is ordinal when you can make a defensible statement that one category comes before or after another, but you cannot make a defensible statement about the exact size of the difference between categories.
Ordinal variable versus other measurement levels
| Property | Nominal | Ordinal | Interval | Ratio |
|---|---|---|---|---|
| Values identify categories | Yes | Yes | Sometimes categorized for presentation | Sometimes categorized for presentation |
| Meaningful order | No | Yes | Yes | Yes |
| Equal, interpretable intervals | No | Not established | Yes | Yes |
| True zero required | No | No | No | Yes |
| Meaningful multiplication or ratios | No | No | No | Yes |
| Typical examples | Country, discipline, blood group | Satisfaction, severity, education level | Celsius temperature, some standardized scores | Age, duration, income, distance |
Ordinal versus nominal variable
Both nominal and ordinal variables are categorical. The difference is order.
A nominal variable divides observations into categories without ranking them. Examples include nationality, academic discipline, and blood group.
An ordinal variable adds an intrinsic sequence. Examples include educational attainment, service quality, and disease severity.
Coding nominal categories as 1, 2, and 3 does not create an order. Conversely, an ordinal variable remains ordered even when its categories are stored as words.
Ordinal versus interval variable
Both types have an order, but interval variables also have equal, interpretable differences.
For Celsius temperature, the difference between 10°C and 20°C has the same physical meaning as the difference between 20°C and 30°C.
For satisfaction categories, the psychological change between “dissatisfied” and “neutral” is not automatically equal to the change between “neutral” and “satisfied.”
Ordinal versus ratio variable
Ratio variables have order, equal intervals, and a meaningful zero that allows ratio statements.
A duration of 20 minutes is twice a duration of 10 minutes. It is not generally meaningful to claim that “very satisfied” contains twice as much satisfaction as “dissatisfied.”
Ordinal versus discrete variable
“Ordinal” and “discrete” describe different properties.
- Ordinal describes the information contained in the ordering.
- Discrete describes values that occur as separate countable possibilities.
An ordinal variable is usually discrete because it contains separate categories. A count such as the number of publications is also discrete, but it is a ratio variable rather than an ordinal variable.
Ordinal versus continuous variable
A continuous variable can theoretically take any value within a range, such as height or reaction time. An ordinal variable contains a limited set of ordered categories.
A researcher may convert a continuous variable into ordered categories, but doing so loses information and may reduce statistical precision.
Types and sources of ordinal variables
Directly observed ordered categories
These categories are recorded without first measuring a continuous quantity. Examples include preference rankings and self-reported agreement categories.
Categorized measurements
A metric measurement is divided into ordered ranges, such as age groups or income bands. The resulting variable is ordinal, although its source variable may be ratio-level.
Expert-rated categories
A trained evaluator assigns a stage, grade, risk level, or performance category. Inter-rater reliability and clear category criteria become important.
Self-reported rating items
Respondents select an ordered response such as never, rarely, sometimes, often, or always. Response wording, translation, cultural interpretation, and central-response tendencies can affect the observations.
Derived ordinal outcomes
Researchers sometimes combine several indicators to produce ordered states, such as low-, medium-, and high-risk groups. The derivation rule should be stated and validated.
Designing an ordinal variable
A sound analysis begins with sound measurement.
Define the construct
Specify what the variable is intended to represent: satisfaction, severity, frequency, confidence, preference, attainment, or another ordered construct.
Choose categories that respondents can distinguish
Too few categories can hide variation. Too many categories can create distinctions that respondents or raters cannot reliably make.
There is no universally correct number of categories. The choice should reflect the construct, population, mode of administration, and intended analysis.
Use an unambiguous order
Labels should make the direction obvious. “Occasionally” and “sometimes,” for example, may overlap unless they are operationally defined.
Keep missing responses separate
“Not applicable,” “do not know,” and “prefer not to answer” are not automatically points on the underlying ordinal scale. They should normally use distinct missing-value codes rather than being placed above or below substantive categories.
Consider balanced response options
For an agreement item, balanced options might include two disagreement categories, a neutral category, and two agreement categories. Whether a neutral option is appropriate depends on whether neutrality is conceptually possible and relevant.
Pilot test the variable
Cognitive interviewing, pilot surveys, inter-rater checks, and expert review can reveal ambiguous labels, reversed ordering, or unused categories.
How should ordinal variables be coded?
Numeric codes make storage and analysis easier, but the labels and order must remain available.
A satisfaction variable might be coded as:
| Code | Label |
|---|---|
| 1 | Very dissatisfied |
| 2 | Dissatisfied |
| 3 | Neither satisfied nor dissatisfied |
| 4 | Satisfied |
| 5 | Very satisfied |
| 97 | Not applicable |
| 98 | Do not know |
| 99 | No response |
The substantive codes 1–5 form the ordinal scale. Codes 97–99 must be declared missing; otherwise, software may incorrectly treat “no response” as greater satisfaction than “very satisfied.”
Coding rules
- Store category labels.
- Define the intended order explicitly.
- Use a consistent direction across related variables.
- Declare non-substantive responses as missing.
- Document any reverse-scored items.
- Retain the original variable when collapsing or recoding categories.
- Check whether every expected category appears in the data.
Reverse-coded items
Some questionnaires contain negatively worded items. If high scores normally represent a positive attitude, a negatively worded item may need reverse coding before a composite score is calculated.
For a five-category item, the transformation is often:
Revised score = 6 − original score
This transformation is appropriate only when the response categories are symmetrically ordered and the scoring plan requires a common direction. The researcher should verify the item wording rather than reversing variables automatically.
How to summarize ordinal data
Frequencies and percentages
A frequency table is usually the starting point. It shows how many observations fall in each ordered category.
| Satisfaction level | Frequency | Percentage | Cumulative percentage |
|---|---|---|---|
| Very dissatisfied | 8 | 8% | 8% |
| Dissatisfied | 17 | 17% | 25% |
| Neutral | 20 | 20% | 45% |
| Satisfied | 38 | 38% | 83% |
| Very satisfied | 17 | 17% | 100% |
Cumulative percentages use the order and help locate percentiles.
Mode
The mode is the most frequent category. In the example, the modal response is “satisfied.”
The mode is easy to interpret but ignores most of the distribution. Two datasets can have the same mode while differing substantially in the other categories.
Median category
The median category is the category containing the 50th percentile after observations are ordered. In the example, the cumulative percentage first reaches 50% in the “satisfied” category, so “satisfied” is the median category.
When an even sample has two different middle categories, calculating their arithmetic average may not have a meaningful ordinal interpretation. Researchers may report the category containing the 50th percentile, the two central categories, or another clearly defined ordinal median convention.
Quartiles and percentiles
Researchers may identify the categories containing the 25th, 50th, and 75th percentiles. These are useful for describing the distribution without assuming equal distances.
Subtracting the code of the first quartile from the code of the third quartile should not automatically be interpreted as a metric interquartile range.
Minimum and maximum categories
The lowest and highest observed categories can be reported. It is safer to report their labels rather than subtracting their numeric codes to create a conventional numerical range.
Can you calculate the mean?
A computer can calculate the arithmetic mean of numeric category codes, but that does not make the result intrinsically meaningful.
A mean assumes that numeric differences contribute in a comparable way. Strictly ordinal categories provide order but do not establish equal intervals. For a single item, frequencies, percentages, the mode, and the median category usually preserve the measurement structure more transparently.
Researchers sometimes report means for rating items or multi-item scores. That choice should be justified by the measurement model, scale construction, distribution, field conventions, robustness evidence, and intended interpretation. A sensitivity analysis comparing ordinal and approximately continuous methods can be informative.
Can you calculate a standard deviation?
A standard deviation measures numeric distances from a mean. It therefore depends on treating the category codes as having metric meaning.
For strictly ordinal categories, category frequencies, cumulative percentages, percentile categories, or an ordinal dispersion measure are more directly aligned with the data. A standard deviation may be reported for a justified composite or approximately interval score, but the assumption should be disclosed.
Suitable graphs for ordinal variables
Ordered bar chart
An ordered bar chart is usually the clearest choice. Place the categories in their substantive order rather than alphabetically.
Stacked bar chart
A stacked bar chart can compare the category distributions of several groups.
Diverging stacked bar chart
For balanced agreement or satisfaction scales, a diverging chart can place negative responses on one side, positive responses on the other, and neutral responses near the centre.
Cumulative distribution plot
A cumulative plot shows the proportion at or below each category. It can be particularly helpful when comparing ordinal outcomes between groups.
Heatmap or ordered contingency table
When two categorical variables are cross-tabulated, a heatmap can display concentrations while retaining their category order.
Graphs to use cautiously
Histograms imply numeric bins along a measured axis and are often inappropriate for a small set of verbal ordinal categories. Pie charts make ordered comparisons difficult. Line charts may imply smooth change between adjacent categories unless the horizontal axis represents time or another genuinely sequential variable.
Choosing a statistical test for ordinal data
There is no single “ordinal-data test.” The choice depends on:
- The research question.
- Whether observations are independent or paired.
- The number of groups or time points.
- Whether the ordinal variable is an outcome or predictor.
- The number and frequency of categories.
- The presence of ties.
- The estimand the researcher wants to report.
- Model assumptions and sample size.
| Research question | Possible method | Important interpretation |
|---|---|---|
| Describe one ordinal variable | Frequencies, percentages, mode, median category, cumulative percentages | Describes the ordered distribution |
| Test association in a contingency table | Chi-square test | Detects association but does not use category order |
| Test an ordered trend across groups | Trend test or ordinal regression | Uses order and requires a stated trend model |
| Compare two independent groups | Mann–Whitney U/Wilcoxon rank-sum or ordinal regression | Compares rank distributions; not automatically a pure median test |
| Compare paired observations using direction only | Sign test | Uses whether each observation increased or decreased |
| Compare paired observations when ranked difference magnitudes are defensible | Wilcoxon signed-rank | Uses direction and rank magnitude of paired differences |
| Compare three or more independent groups | Kruskal–Wallis or ordinal regression | Tests distributional differences among groups |
| Compare three or more repeated conditions | Friedman test or an ordinal mixed model | Accounts for repeated observations |
| Examine monotonic association | Spearman’s rho or Kendall’s tau-b | Measures rank-based monotonic association; tau-b handles ties explicitly |
| Model an ordinal outcome with predictors | Ordinal logistic/probit regression | Estimates effects across ordered outcome thresholds |
| Model repeated or clustered ordinal outcomes | Cumulative-link mixed model or ordinal GEE | Accounts for dependence within people, centres, or clusters |
Mann–Whitney U test
The Mann–Whitney U test compares the relative ranks of two independent groups. It can detect differences in distributions and stochastic ordering.
It should not automatically be described as a test of medians. A median-shift interpretation is most defensible when the group distributions have similar shapes and differ mainly in location.
Wilcoxon signed-rank and sign tests
The Wilcoxon signed-rank test is commonly used for paired observations. It ranks the magnitudes of paired differences as well as their directions.
When numeric differences between ordinal codes do not have a defensible magnitude, the sign test is more conservative because it uses only whether an observation moved upward or downward. An ordinal repeated-measures model may provide a fuller analysis when covariates or several time points are involved.
Kruskal–Wallis test
The Kruskal–Wallis test compares rank distributions across three or more independent groups. A statistically significant result indicates that at least one group differs, but post-hoc comparisons are required to locate the differences.
As with Mann–Whitney, the test should not be called a test of medians without considering distribution shapes and other assumptions.
Friedman test
The Friedman test compares repeated or matched observations across three or more conditions. It is useful for simple blocked designs but does not accommodate complex covariate structures as flexibly as an ordinal mixed model.
Spearman’s rho and Kendall’s tau
Spearman’s rho measures the strength and direction of a monotonic relationship using ranks. Kendall’s tau-b is often attractive when an ordinal variable has many tied observations.
Neither coefficient proves causation, and both summarize only a particular form of association.
Chi-square test
A chi-square test can examine whether two categorical variables are associated. However, the ordinary test treats the categories as labels and does not use their order.
This may be appropriate when any association is of interest. If the hypothesis concerns an increasing or decreasing trend, an ordinal model or order-sensitive test can be more informative.
Ordinal logistic regression
Ordinal logistic regression models the relationship between one or more predictors and an ordered categorical outcome.
Suppose an outcome (Y) contains (K) ordered categories. A cumulative-logit proportional-odds model can be written as:
[
\log \left[\frac{P(Y \leq j)}{P(Y > j)}\right]
= \alpha_j – \beta_1X_1 – \cdots – \beta_pX_p,
\quad j = 1,\ldots,K-1
]
Each (\alpha_j) is a threshold separating lower categories from higher categories. The predictor coefficients describe how the cumulative odds change as the predictors change.
Example
A researcher models satisfaction as:
- Very dissatisfied
- Dissatisfied
- Neutral
- Satisfied
- Very satisfied
Predictors might include waiting time, service channel, and previous experience. The model can estimate the probability that a person falls in each satisfaction category while preserving the category order.
Proportional-odds assumption
The standard model assumes that a predictor has the same coefficient at each cumulative threshold. This is known as the proportional-odds or parallel-lines assumption.
If the assumption is poor, alternatives include:
- Partial proportional-odds models.
- Generalized ordered-logit models.
- Multinomial models, although these discard the ordinal structure.
- Category-specific or continuation-ratio models.
- Adjacent-category models.
- Flexible Bayesian ordinal models.
Reporting an ordinal model
Report:
- How the outcome was ordered.
- The link function.
- Reference groups.
- Coefficients or odds ratios with confidence intervals.
- How the proportional-odds assumption was assessed.
- Predicted probabilities for meaningful predictor values.
- Missing-data handling and sensitivity analyses.
Predicted probabilities are often easier for nontechnical readers to interpret than coefficients alone.
Using an ordinal variable as a predictor
An ordinal predictor does not have to be entered as a single numeric score.
Treat it as an ordered categorical factor
This preserves the categories and allows their effects to differ. Indicator or contrast coding can compare categories without assuming equal spacing.
Fit a monotonic trend
The predictor may be represented by increasing scores when theory supports a monotonic relationship. The scores should reflect an explicit assumption, not merely the fact that the categories were stored as 1, 2, 3, and 4.
Compare coding approaches
Researchers can compare:
- A categorical model.
- A linear-score model.
- A flexible monotonic-effect model.
If results depend strongly on the coding choice, that uncertainty should be reported.
Are Likert scales ordinal or interval?
A single Likert-type item is generally treated as ordinal because its response categories have a clear order but do not establish equal distances.
A multi-item Likert scale is a score created by combining several items intended to measure the same construct. Such a composite can have many possible values and may behave more like a continuous measure. Researchers sometimes use means, standard deviations, t-tests, analysis of variance, or linear regression for well-constructed composite scales.
The literature does not provide one universal rule. One position emphasizes that ordinal categories should be analyzed with ordinal or nonparametric methods (Jamieson, 2004). Another emphasizes evidence that many parametric procedures are robust under common conditions (Norman, 2010). Sullivan and Artino (2013) recommend distinguishing individual items from composite scales and considering the scale, sample, distribution, and research purpose.
A defensible approach is to:
- Distinguish clearly between an item and a composite scale.
- Check whether the items measure a coherent construct.
- Examine category frequencies and score distributions.
- State the assumptions behind treating a composite as continuous.
- Consider an ordinal analysis or sensitivity analysis when conclusions may depend on the choice.
- Avoid treating field convention as proof of equal intervals.
How ordinal variables are used in modern research
Survey and social research
Ordinal variables measure attitudes, trust, agreement, perceived status, frequency, political preference, and service evaluation.
Education
Researchers use proficiency levels, letter-grade bands, degree levels, rubric ratings, class ranks, and teacher assessments.
Healthcare and clinical research
Ordinal outcomes include symptom severity, disease stages, functional states, pain ratings, treatment-response categories, and clinical status scales.
Ordered outcomes can preserve more information than reducing a response to a simple yes/no variable. Their analysis must nevertheless match the clinical question and model assumptions.
Business and customer analytics
Common examples include satisfaction categories, purchase likelihood, service ratings, customer-priority bands, and lead-quality levels.
Human–computer interaction and artificial intelligence
Usability, perceived quality, confidence, relevance, helpfulness, and human preference are frequently collected through ordered ratings. Modern research increasingly uses cumulative-link and mixed-effects models when ratings are repeated across users, systems, or tasks.
Machine learning
Ordinal classification differs from ordinary multiclass classification because the errors have an order. Predicting “high” when the truth is “medium” is usually less serious than predicting “very low.”
Suitable approaches include ordinal regression, threshold models, ordinal loss functions, and models that estimate cumulative category probabilities. Accuracy alone may hide the seriousness of different classification errors, so ordered error measures or category-specific results can be helpful.
Working with ordinal variables in software
SPSS
In SPSS Variable View:
- Create numeric values for the substantive categories.
- Add value labels.
- Define non-substantive codes as missing.
- Set the measurement level to Ordinal.
- Check that the stored order matches the conceptual order.
SPSS distinguishes nominal, ordinal, and scale metadata. However, selecting “Ordinal” does not automatically make every subsequent procedure appropriate. Researchers must still choose an analysis that treats the variable correctly.
SPSS provides ordinal regression through its ordinal-regression procedure, often referred to as PLUM. Review model fit, category frequencies, reference directions, thresholds, and the test of parallel lines rather than reporting only a p-value.
R
In R, an ordinal variable can be stored as an ordered factor:
satisfaction <- factor(
satisfaction,
levels = c(
"Very dissatisfied",
"Dissatisfied",
"Neutral",
"Satisfied",
"Very satisfied"
),
ordered = TRUE
)
The specified level order is essential. Alphabetical ordering may produce an incorrect analysis.
Ordinal regression can be fitted with functions designed for ordered responses, including proportional-odds and cumulative-link models. The response should be stored with its intended category order.
Python
In pandas, use an ordered categorical data type:
import pandas as pd
order = [
"Very dissatisfied",
"Dissatisfied",
"Neutral",
"Satisfied",
"Very satisfied",
]
df["satisfaction"] = pd.Categorical(
df["satisfaction"],
categories=order,
ordered=True,
)
For ordinal regression, statsmodels provides OrderedModel. An ordered pandas category is preferable because it carries the relationship between the levels instead of relying on an accidental alphabetical or numeric order.
Excel and spreadsheets
Spreadsheets can store numeric codes and labels but do not enforce a statistical measurement model. Maintain a separate codebook, validate allowed values, and sort charts using the intended category order.
Do not allow special missing codes such as 99 to enter arithmetic calculations as if they were substantive observations.
Artificial intelligence and ordinal-data analysis
Generative AI tools can assist with code drafting, explanations, data dictionaries, and sensitivity-analysis ideas. They cannot determine a variable’s measurement level reliably from numeric values alone.
For example, a column containing 1–5 might represent a rating, a count, an identifier, or a grouped measurement. An AI system needs the questionnaire wording, labels, operational definition, and research design.
When using AI-assisted analysis:
- Provide the complete codebook and category order.
- State whether the variable is an outcome, predictor, or grouping variable.
- Specify independent, paired, repeated, or clustered observations.
- Ask the tool to state the assumptions of every proposed method.
- Verify generated code against official software documentation.
- Inspect category frequencies and outputs manually.
- Do not upload confidential participant data to an unauthorized service.
- Have consequential or publication-level analyses reviewed by a qualified statistician.
AI can accelerate technical work, but responsibility for measurement validity and statistical interpretation remains with the researcher.
Advantages of ordinal variables
They represent ordered judgments directly
Ordinal categories are useful when respondents can distinguish direction more reliably than exact numeric distance.
They are understandable
Labels such as mild, moderate, and severe are often easier to communicate than an abstract numerical score.
They support more analysis than nominal categories
Order-sensitive summaries, rank associations, trend tests, and ordinal regression can use the sequence that nominal analysis ignores.
They can reduce response burden
A participant may be able to select a meaningful category even when exact measurement is impossible, intrusive, or unnecessarily difficult.
They can preserve clinically or practically meaningful states
Disease stages, proficiency bands, and risk levels may map directly to decisions or interventions.
Limitations of ordinal variables
Exact differences are unavailable
Researchers can identify direction but cannot automatically quantify the amount of change between adjacent categories.
Coding can create false precision
Means, differences, and ratios calculated from category codes may look exact even when the measurement process does not support that precision.
Categories lose information
Converting age, income, or test scores into bands discards within-category differences and can weaken analysis.
Interpretation can vary
Respondents may understand verbal labels differently. The perceived distance between categories may also vary between people or cultures.
Sparse categories can destabilize models
Very small frequencies in extreme categories can produce imprecise estimates or convergence problems.
Results can depend on assumptions
Conclusions may differ depending on whether the variable is treated as categorical, ordinal, or approximately continuous.
Common mistakes
Assuming numeric codes have equal intervals
Coding “low,” “medium,” and “high” as 1, 2, and 3 preserves order. It does not prove equal distances.
Letting software sort categories alphabetically
Alphabetical order may place “high” before “low” or “satisfied” before “very dissatisfied.” Always define the intended order.
Treating missing responses as the highest category
Codes such as 97, 98, or 99 must be declared missing.
Calling every Mann–Whitney or Kruskal–Wallis result a median difference
These tests work with ranks and distributions. A median interpretation requires additional conditions.
Using the mean without explaining the assumption
A mean of ordinal codes may be convenient, but the analysis should state why a metric interpretation is reasonable.
Confusing a Likert item with a Likert scale
One ordered response item and a composite of several items do not have identical measurement or analysis considerations.
Collapsing categories after seeing the results
Post-hoc category merging can introduce bias. Collapsing should be substantively justified, documented, and preferably planned before outcome analysis.
Dichotomizing unnecessarily
Turning five ordered categories into “low” and “high” discards ordering information and may reduce statistical power and interpretability.
Reversing the order during recoding
A model’s coefficients can appear to change direction if the outcome order is accidentally reversed.
Selecting a test only from the measurement level
Measurement level matters, but so do the study design, sample dependence, estimand, distribution, ties, clustering, missing data, and model assumptions.
How to report an ordinal variable
A concise methods statement should identify the categories, order, coding, and analysis.
Reporting template
Satisfaction was measured using a five-category ordered item ranging from 1 (“very dissatisfied”) to 5 (“very satisfied”). “Not applicable” responses were treated as missing. Category frequencies and cumulative percentages were reported. Group differences were estimated with a proportional-odds ordinal logistic regression. Higher outcome categories represented greater satisfaction. Odds ratios, 95% confidence intervals, predicted category probabilities, and an assessment of the proportional-odds assumption were reported.
For a descriptive study, the statement might instead say:
Responses were summarized using category frequencies, percentages, the modal category, and the category containing the median response. No equal-interval assumption was made.
The results section should present the full distribution whenever important differences might be hidden by a single summary statistic.
Conclusion
An ordinal variable records categories that can be meaningfully ranked but are not known to be equally spaced. Correct analysis begins by preserving the category labels and order, keeping missing values separate, and avoiding unsupported arithmetic interpretations.
Frequencies, percentages, cumulative distributions, the mode, and a median category are useful descriptive tools. Rank-based tests and ordinal regression can support inference, but their assumptions and estimands must be understood. Numeric coding, software defaults, and common practice should never replace a clear explanation of what the categories actually measure.
