Nominal and ordinal data are two of the most common types of categorical information collected in scientific research, and choosing the wrong analysis for either one can distort results or waste statistical power. Nominal data are categories with no inherent ranking, like blood type or country of origin. Ordinal data carry a rank order, like pain severity rated as mild, moderate, or severe, but the gaps between ranks are not guaranteed to be equal. That distinction sounds straightforward, yet it generates ongoing disagreement among researchers about which statistical methods are appropriate, how to design surveys, and even how many response options a questionnaire should offer.
What Makes the Two Types Fundamentally Different
The classification traces back to a framework proposed by S. S. Stevens in 1946, which sorted measurement into four levels. At the most basic level, a nominal scale uses numbers or labels purely as identifiers, where the only meaningful relationship is whether two observations fall in the same category or different ones. An ordinal scale goes a step further: it establishes a “greater than” or “less than” relationship between values, producing a rank order, but does not define equal spacing between those values.1Science. On the Theory of Scales of Measurement Stevens also described interval and ratio scales, but the nominal-ordinal boundary is the one that generates the most confusion in applied research because it determines whether you can meaningfully say one value is “higher” than another.
In practice, nominal data show up whenever participants are sorted into groups. Think diagnosis categories in a medical study, types of land cover in an ecological survey, or political party affiliation in a poll. You can count how many observations fall in each group, but averaging the groups or computing distances between them makes no sense. Ordinal data appear wherever researchers ask people to rate or rank something: satisfaction scales, disease severity grades, educational attainment levels, socioeconomic status tiers. The categories have a clear sequence, but you cannot assume the jump from “satisfied” to “very satisfied” is the same size as the jump from “neutral” to “satisfied.”
Choosing the Right Statistical Test
The division between nominal and ordinal data dictates which statistical tools are valid. For nominal variables, the workhorses are the chi-square test and Fisher’s exact test. The chi-square test evaluates whether two categorical variables are associated; Fisher’s exact test is preferred when the sample is small.2Translational Sports Medicine. Categorical variable analyses: chi-square, Fisher exact, Mantel–Haenszel When the same subjects are measured at two time points on a binary nominal outcome (say, “cured” versus “not cured” before and after treatment), the McNemar test handles that paired comparison. A mid-p version of the McNemar test performs well even with small to moderate samples, striking a balance between overly conservative exact methods and methods that rely on large-sample assumptions.3PubMed Central. The McNemar test for binary matched-pairs data: mid-p and asymptotic are better than exact conditional When three or more related proportions need comparing, Cochran’s Q test extends the McNemar approach to handle the additional groups.4PubMed Central. Biostatistics Series Module 4: Comparing Groups – Categorical Variables
For ordinal data, the standard toolkit shifts to rank-based (nonparametric) methods: the Mann-Whitney U test for comparing two independent groups, the Wilcoxon signed-rank test for paired observations, and the Kruskal-Wallis test for comparing three or more independent groups. When you need to measure correlation between two ordinal variables, Spearman’s rho and Kendall’s tau-b are the go-to choices because they assess monotonic relationships without assuming equal intervals.5Journal of Classification. A Comparison of Reliability Coefficients for Ordinal Rating Scales Applying a standard Pearson correlation to ordinal data is technically a violation, because Pearson assumes the data are measured on at least an interval scale, though the practical consequences of that violation depend on how many categories exist and how symmetric the distribution is.
The Likert Scale Controversy
No topic in applied statistics generates more reviewer comments than whether you can run parametric tests on Likert scale data. A Likert item, the familiar “strongly disagree” to “strongly agree” format, produces ordinal responses. Strictly speaking, you cannot assume equal intervals between the labeled points, which means parametric methods like t-tests and analysis of variance should not apply. Many published papers get flagged for this during peer review.
The counterargument, supported by simulation research going back to the 1930s, is that parametric statistics are robust enough to handle violations of the interval-level assumption. One widely cited review of this evidence argues that challenges to using parametric methods on Likert data are largely unfounded, because studies have consistently shown that methods like analysis of variance, regression, and correlation give reliable answers even when the underlying data are ordinal.6PubMed. Likert scales, levels of measurement and the “laws” of statistics On the other side, researchers have cautioned that ordinal responses violate statistical assumptions needed for parametric tests and that care must be taken to avoid misinterpretation.7PubMed Central. Use and Misuse of the Likert Item Responses and Other Ordinal Measures – Section: Abstract
A systematic review of 134 visualization studies that used Likert scales found that many papers were inconsistent in how they treated the data, sometimes analyzing the same responses as both discrete and continuous within a single study. The review also found that researchers sometimes sacrificed statistical power by applying nonparametric tests unnecessarily, and that subjective measures gathered through Likert scales received less rigorous reporting and analysis than the objective measures in the same study.8Computer Graphics Forum. Effective Use of Likert Scales in Visualization Evaluations: A Systematic Review The practical upshot is that neither blanket permission nor blanket prohibition is warranted. The right choice depends on how many response categories you have, how skewed the responses are, and what you are trying to conclude.
When Ordinal Methods Outperform Shortcuts
Researchers often simplify ordinal outcomes by collapsing them into binary ones, splitting a five-level severity score into “good” versus “bad,” for instance. This feels cleaner but throws away information. Simulation studies have demonstrated that purpose-built ordinal methods, such as proportional odds models and win probability estimates, generally have higher statistical power and lower bias than approaches that either collapse ordinal data into two groups or treat the data as though they were continuous, especially when the distribution of responses is skewed.9PubMed Central. Statistical analysis of Likert-based ordinal scales: a guide for clinical trialists – Section: RESULTS
A striking example comes from clinical research in liver disease. Researchers emulated a clinical trial in patients with compensated cirrhosis and found that an ordinal outcome could detect a statistically significant difference between patient groups that binary, chi-square, and time-to-event analyses all missed. Beyond just detecting real differences, the ordinal approach also allowed dramatically smaller trials: sample-size estimates were three to four times lower than those required for a binary composite endpoint.10PubMed Central. Ordinal Outcomes Are Superior to Binary Outcomes for Designing and Evaluating Clinical Trials in Compensated Cirrhosis For fields where enrolling patients is expensive and slow, that kind of efficiency gain is enormous.
Regression Models for Categorical Outcomes
When the outcome variable you care about is categorical, standard linear regression does not apply. For nominal outcomes with three or more unordered categories, multinomial logistic regression is the standard approach, modeling the probability of each category relative to a reference category. A mixed-effects version of this model also handles data that are clustered or measured repeatedly over time, which is common in longitudinal studies where the same patients are evaluated at multiple visits.11PubMed. A mixed-effects multinomial logistic regression model
For ordinal outcomes, the proportional odds model is widely used. It works by fitting a series of logistic regressions across the ordered categories and assumes that the effect of each predictor is the same regardless of where you draw the dividing line in the ranking. That proportionality assumption is the key requirement: if the effect of, say, age on severity differs depending on which severity threshold you examine, the model can give misleading results.12PubMed. Assessing proportionality in the proportional odds model for ordinal logistic regression Checking this assumption is something many applied papers skip, but it is straightforward to do and can reveal when a simpler or more complex modeling strategy is needed.
Measuring Agreement Between Raters
In many studies, two or more people independently classify or rate the same observations, and the research question is how well they agree. The appropriate agreement statistic depends entirely on whether the rating scale is nominal or ordinal. For unordered categories with two or more raters, Cohen’s kappa (for two raters) and Fleiss’ kappa (for more than two) are the standard tools.13PubMed Central. Kappa statistic considerations in evaluating inter-rater reliability between two raters: which, when and context matters – Section: RESULTS These treat every disagreement the same: if one pathologist calls a biopsy “benign” and another calls it “malignant,” that counts the same as if they disagreed over two benign subcategories.
When the categories have a natural order, weighted kappa becomes appropriate. A linear or quadratic weighting scheme penalizes large disagreements more than small ones, which makes intuitive sense. If two radiologists grade a tumor as “stage 2” versus “stage 3,” that is a less worrying disagreement than “stage 1” versus “stage 4.” However, both Cohen’s original kappa and Fleiss’ extension can behave paradoxically in certain situations, producing misleadingly low values even when agreement is high, particularly when the distribution of observations across categories is very uneven.14PubMed. Assessing the inter-rater agreement for ordinal data through weighted indexes Researchers who report kappa values without examining the distribution of their data risk misinterpreting what looks like poor agreement.
Planning Sample Sizes Around Categorical Outcomes
Power and sample-size calculations for studies with categorical outcomes require different formulas than those used for continuous measurements. For ordinal outcomes, there are dedicated methods for computing how many subjects you need. One approach for paired ordinal data provides an approximate calculation based on the expected number of discordant pairs. That approximation underestimates the required sample when there are only two categories but improves as the number of categories increases.15PubMed. Sample size calculations for paired or matched ordinal data The same work found that collapsing an ordinal variable into just two categories does not sacrifice as much power as you might expect, which partly explains why binary endpoints remain popular despite the efficiency advantages of ordinal analysis.
For ordinal data heading toward logistic regression analysis, sample-size formulas have been developed that account for the number and breadth of response categories, the possibility of misclassification, and whether stratification will be used.16PubMed. Sample size calculations for ordered categorical data For more complex generalized linear models with nominal, count, or ordinal responses, a practical method involves fitting the model to an expanded dataset with calculated weights representing response probabilities, then using the resulting variance estimates in standard power formulas.17PubMed. A practical approach to computing power for generalized linear models with nominal, count, or ordinal responses These methods matter because using continuous-outcome sample-size formulas for a categorical endpoint almost always gives you the wrong number of subjects.
Psychometrics and Social Science Applications
The social sciences rely heavily on ordinal data, particularly through questionnaires built from Likert-type items. Factor analysis and item response theory models are widely used to analyze these ordinal responses, whether the items are dichotomous (yes/no) or multi-category scales.18Psychological Methods. Estimating Ordinal Factor Analysis and Item Response Theory Models: A Comparison of Full- and Limited-Information Techniques Item response theory is particularly useful because it can evaluate how well individual questionnaire items measure the underlying trait at different levels of that trait, rather than assuming every item works equally well across the full range.
An analysis of two commonly used identity development questionnaires illustrates the kind of insight this approach provides. Graded response models applied to data from nearly 5,000 students across the Netherlands and the United States showed that the instruments precisely measured only a limited range of the underlying traits, and that item quality varied considerably across the scales.19PubMed. An item response theory analysis of two frequently used measures of identity development Findings like that are invisible if you simply average the item scores and treat the result as a continuous variable, which is exactly what many studies do.
Designing Better Surveys
How many response options should a Likert-type item offer? The answer affects both the reliability and the validity of the data. A review of the literature on this question concluded that an odd number of response options above five, particularly a seven-point scale, tends to produce the best reliability and validity. If the goal is to push respondents toward taking a side rather than sitting on the fence, an even number of options, especially six, works better because it eliminates the neutral midpoint.20International Journal of Educational Methodology. Number of Response Options, Reliability, Validity, and Potential Bias in the Use of the Likert Scale Education and Social Science Research: A Literature Review The same review flagged response bias and central tendency bias as threats that can undermine Likert data regardless of how many options you offer. Central tendency bias is the tendency for respondents to cluster around the middle of the scale and avoid the endpoints.
These design choices feed directly into the statistical issues discussed earlier. A scale with more ordered categories gives ordinal methods more information to work with, which generally increases power. But more categories also increase the cognitive burden on respondents, who may start treating a ten-point scale as effectively a three-point one by only using the bottom, middle, and top. The seven-point sweet spot reflects a compromise between statistical information and human attention.
Handling Missing Ordinal Data
Missing values are a fact of life in survey-based and clinical research, and how you fill in those gaps matters more when the data are ordinal. One common shortcut is to impute missing Likert responses using the same methods designed for continuous data, then round the imputed values to the nearest whole number. Simulations have shown this approach can introduce substantial bias into some or all of the estimated parameters. Logistic regression imputation models, while better matched to the data type in principle, can also lead to biased results under certain conditions.21PubMed. A Comparison of Imputation Strategies for Ordinal Missing Data on Likert Scale Variables The safer approach is to use imputation methods that are specifically designed to respect the ordinal structure of the data, preserving the rank order without forcing assumptions about equal intervals.
Machine Learning with Mixed Data Types
Real-world datasets rarely contain only one type of variable. A hospital database might include nominal fields like diagnosis codes, ordinal fields like functional status ratings, and continuous fields like lab values. Clustering algorithms and predictive models need to handle this mix, and the nominal-ordinal distinction matters for how you calculate the “distance” between two observations.
Gower’s distance is a popular approach for mixed-type data because it can handle numeric, categorical, and ordinal variables simultaneously, computing separate distance components for each variable type and combining them. A modification of the k-prototypes clustering algorithm to use Gower’s distance has been proposed to ensure the algorithm converges reliably when ordinal variables are present alongside other data types.22Advances in Data Analysis and Classification. Clustering large mixed-type data with ordinal variables A further refinement weights features by their importance, so that a variable contributing more predictive information gets a larger say in how observations are grouped.23PubMed Central. A modified and weighted Gower distance-based clustering analysis for mixed type data: a simulation and empirical analyses – Section: METHODS
The temptation in machine learning is to convert ordinal categories into plain numbers (1, 2, 3, 4, 5) and feed them into any algorithm that accepts numeric input. This “works” in the sense that the algorithm runs, but it smuggles in an assumption of equal spacing. For tree-based models like random forests and gradient boosting, this usually does not cause problems because the algorithm only needs to identify cut points along the variable, which preserves order without requiring interval-level data. For distance-based methods like k-nearest neighbors or standard clustering, the assumption matters more, which is why approaches like Gower’s distance exist.
When Nominal Data Get Mistaken for Ordinal
One of the subtler mistakes in applied research is treating a variable as ordinal when it is actually nominal, or vice versa. Education level, for instance, looks straightforwardly ordinal: high school, bachelor’s, master’s, doctorate. But what about a variable coding religion as 1 = Christian, 2 = Muslim, 3 = Hindu, 4 = Other? If a researcher inadvertently runs a rank-based test on those codes, the test will interpret “Hindu” as greater than “Muslim,” which is meaningless. The numeric labels are arbitrary.
The reverse also happens. Some variables that seem nominal actually carry an implicit order. Tumor stage classifications (stage I through IV) are clearly ordinal, but a general-purpose coding scheme might store them as unordered categories if the database was not set up carefully. When that happens, an analysis might apply a chi-square test where a proportional odds model would be more appropriate, losing the ability to detect a trend across stages. The safest practice is to document every variable’s measurement level before any analysis begins, not after a reviewer flags it during peer review.