Ordinal Logistic Regression
Also known as: proportional-odds model, cumulative link model, ordered logit, OLR
Ordinal logistic regression — most commonly the proportional-odds model — estimates the relationship between one or more predictors and an ordered categorical outcome (e.g., Likert scales, disease severity grades, educational attainment levels). It models cumulative log-odds across the ordered categories while assuming a single shared effect of each predictor at all thresholds.
Key highlights
- Handles ordered categorical outcomes naturally without discarding the rank information.
- Produces a single, interpretable cumulative odds ratio per predictor when the proportional-odds assumption holds.
- More statistically efficient than collapsing an ordinal outcome into binary categories.
- Widely implemented in standard software (R: MASS::polr, Stata: ologit, SPSS: PLUM, Python: statsmodels).
- Extends gracefully to generalised ordinal models when the proportional-odds assumption fails for some predictors.
Intuition
This section is available to Pro members. Upgrade to Pro
How it works
This section is available to Pro members. Upgrade to Pro
When to use it
Use ordinal logistic regression when the outcome has three or more naturally ordered categories and treating them as continuous (OLS) would be inappropriate, or collapsing them to binary would discard information. It is well-suited for Likert-scale survey items, pain or severity ratings, education or income brackets, and graded clinical outcomes. Require at least 10–15 events per predictor to avoid overfitting. Do not use it when the proportional-odds assumption is severely violated (use partial-proportional-odds or multinomial logistic instead), when the outcome is truly continuous (use OLS), or when only two categories are present (use binary logistic regression).
Strengths & limitations
- Handles ordered categorical outcomes naturally without discarding the rank information.
- Produces a single, interpretable cumulative odds ratio per predictor when the proportional-odds assumption holds.
- More statistically efficient than collapsing an ordinal outcome into binary categories.
- Widely implemented in standard software (R: MASS::polr, Stata: ologit, SPSS: PLUM, Python: statsmodels).
- Extends gracefully to generalised ordinal models when the proportional-odds assumption fails for some predictors.
- The proportional-odds assumption is frequently violated in practice, requiring diagnostic testing before trusting results.
- Coefficient interpretation is less intuitive than OLS; cumulative odds ratios need careful explanation to non-specialist audiences.
- Requires a moderate-to-large sample; with very few observations per category, threshold estimates become unstable.
- Cannot model non-monotonic or U-shaped predictor–outcome relationships without additional transformations.
Common pitfalls
This section is available to Pro members. Upgrade to Pro
Applications
This section is available to Pro members. Upgrade to Pro
Frequently asked
How do I test the proportional-odds assumption?
Apply the Brant test (available in R via brant() after polr()) or the score test of parallel lines (Stata: omodel after ologit). A significant result suggests the single-coefficient constraint is violated for at least one predictor; consider a partial-proportional-odds model or a multinomial logistic alternative.
How do I interpret the odds ratio from an ordinal logistic model?
The exponentiated coefficient exp(β) is a cumulative odds ratio. A value of 1.5 for predictor x means that a one-unit increase in x multiplies the odds of being in a higher category (above any threshold) by 1.5, assuming the proportional-odds constraint holds.
What is the minimum sample size required?
A common guideline is at least 10–15 observed outcomes per predictor entered in the model, and each category should have at least 20–30 observations. Very sparse categories inflate standard errors and should be merged if theoretically defensible.
When should I use multinomial logistic regression instead?
When the proportional-odds assumption fails badly for most predictors, or when the categories are nominally ordered but the underlying construct is better treated as unordered, multinomial logistic regression is more appropriate, at the cost of requiring more parameters and larger samples.
Can I include both continuous and categorical predictors?
Yes. Ordinal logistic regression handles mixed predictor types just as binary logistic regression does. Continuous predictors enter linearly by default; categorical predictors are dummy-coded. Interactions and polynomial terms can be added in the usual way.
Sources
- 1.McCullagh, P. (1980). Regression models for ordinal data. Journal of the Royal Statistical Society: Series B (Methodological), 42(2), 109–142.
- 2.Agresti, A. (2010). Analysis of Ordinal Categorical Data (2nd ed.). John Wiley & Sons.ISBN 978-0470082898
You have read it. What now?
Cite this page
ScholarGate. (2026, June 3). Ordinal Logistic Regression. ScholarGate. https://scholargate.app/statistics/ordinal-logistic-regression