AI Glossary / AI and machine learning / Intermediate
Ordinal regression
Definition
Ordinal regression models an outcome whose categories have a meaningful order but whose gaps need not be equal. Examples include low, medium and high, or ordered satisfaction ratings. It uses the ordering without treating category labels as ordinary continuous measurements.
What it means
A rating of four is higher than a rating of three, but the difference between them need not equal the difference between one and two in a person's experience. Ordinal models preserve the ranking while avoiding that simple equal-interval assumption.
The term covers a family of models rather than one formula. A common example is the cumulative-logit, or proportional-odds, model. It models probabilities of being at or below successive category thresholds.
Ordinal regression differs from multinomial regression, which treats categories as unordered. It also differs from simply predicting a number and rounding it. The model produces category probabilities under a structure designed for ordered outcomes.
How it works
Training uses ordered labels and input features. In a cumulative model, thresholds divide an underlying score into outcome regions. A link function connects the score and thresholds with cumulative probabilities.
The proportional-odds version uses the same predictor effects across cumulative splits, while allowing different threshold intercepts. This shared-effect assumption can be useful but must be assessed for the application.
Category probabilities are obtained from differences between successive cumulative probabilities. Evaluation considers both probability quality and the ordering of errors: confusing adjacent categories may have different consequences from confusing opposite extremes. Sparse categories and inconsistent label ordering require particular attention.
Three practical examples
- A survey analysis models ordered satisfaction responses while preserving their rank.
- A quality-research project predicts inspection grades such as poor, acceptable and excellent.
- A service-planning study estimates ordered workload severity categories and evaluates whether mistakes tend to be small or extreme.
Think of it this way
Think of a staircase whose steps have a clear order but whose heights are not assumed identical. The model respects which steps are above others without assigning equal meaning to every gap.
The analogy is limited because an ordinal model estimates probabilities rather than locating a person at a perfectly known height. Its thresholds and predictor effects are statistical constructs, not necessarily direct psychological or physical measurements.
Worked example
Recovering category probabilities
Suppose a fitted model for low, medium and high produces P(Y at or below low) = 0.20 and P(Y at or below medium) = 0.75 for one observation.
The probability of low is 0.20. The probability of medium is 0.75 minus 0.20, or 0.55. The probability of high is 1 minus 0.75, or 0.25. The three probabilities sum to 1.
Medium is the most likely category, but uncertainty remains. If high outcomes have especially important consequences, the decision rule may consider the 25% high probability rather than relying only on the most likely label.
The example assumes coherent ordered cumulative probabilities. It illustrates prediction interpretation, not a fitted dataset or a recommendation for a particular threshold policy. The cost of decisions and the calibration of the probabilities still need separate evaluation.
Illustrative case study
Fictional case: analyzing service ratings
Meadowline Services, an invented company, studies voluntary ratings from one to five. Analysts use an ordinal model to examine associations with response time and service type.
They discover that the proportional-odds assumption may not describe every predictor well. They compare a more flexible model and report predicted category probabilities rather than treating a coefficient as a direct causal effect on satisfaction.
The case illustrates modeling choices and interpretation. It does not claim a real customer-service improvement or show that changing response time alone would cause the estimated change in ratings.
Common misconceptions and mistakes
Numbers used as category labels do not automatically create an interval scale. Ordinal regression is also not limited to logistic links; different models use different assumptions.
The proportional-odds assumption should not be confused with equal category probabilities or equal spacing between thresholds. A statistically significant coefficient does not establish causation, and a high overall accuracy can conceal frequent errors at rare but important extremes.
Business use
Ordinal regression is useful when a business records ranked categories and wants interpretable probability estimates. It can preserve useful ordering while using fewer parameters than a fully unordered model in suitable settings.
Confirm that the ordering is meaningful and consistently recorded. Inspect category frequencies, missingness and whether respondents use scales comparably. Evaluate assumptions and compare with reasonable alternatives. Report probabilities or clearly explained effects rather than presenting label codes as precise measurements. For operational use, choose decision thresholds according to consequences and workload, and validate the model on data representative of future cases.
People also ask
Must the gaps between categories be equal?
No. That is a key reason to use an ordinal model.
Is every ordinal model proportional odds?
No. Proportional odds is one common formulation.
Can a model return probabilities for every category?
Yes. Those probabilities can support decisions beyond simply selecting the most likely label.
Check your understanding
A five-level rating is coded 1 through 5. Does that coding prove the difference between 1 and 2 equals the difference between 4 and 5?
Answer: No. The numbers establish order, but equal spacing requires an additional justification that ordinal modeling does not assume automatically.
Related terms
- Threshold (for decision trees)
- Calibration
- Ordered categories
- proportional odds
- cumulative logit
- threshold
- multinomial regression
- ordinal data
Learn next
Learn next how cumulative probabilities differ from individual category probabilities. Then examine whether a shared predictor effect across all thresholds is plausible for a particular ordered outcome before choosing a proportional-odds model.
More in AI and machine learning
- Optimal discriminant analysis and classification tree analysis - Optimal discriminant analysis is a statistical framework that selects classification rules to optimize a specified accuracy criterion.
- Optimality criterion - An optimality criterion is the rule used to decide which candidate model or solution counts as best.
- Optimizer - An optimizer is the procedure that adjusts a model's parameters during training to reduce its errors, step by step, until the model fits the data as well as...
- Ordination (statistics) - Ordination is a family of statistical methods that arrange observations in a small number of dimensions to summarize relationships in multivariate data.
- Organoid intelligence - Organoid intelligence is a research field exploring biological computation using three-dimensional cultures of brain cells, called brain organoids...
- Orthogonal Matching Pursuit - Orthogonal Matching Pursuit is a method for building a sparse model: it selects the few input variables that matter most, one at a time, each time picking...
Sources
Primary documentation: statsmodels OrderedModel, https://www.statsmodels.org/dev/generated/statsmodels.miscmodels.ordinal_model.OrderedModel.html ; ordinal-regression example, https://www.statsmodels.org/stable/examples/notebooks/generated/ordinal_regression.html . The OrderedModel documentation and example were consulted for ordered labels, shared predictor effects, thresholds and cumulative-probability differences. Probability calculation and fictional case are original.
Previous: Ordinal probit regression Next: Ordinal scale