Ordinal Logistic Regression and the Proportional-Odds Assumption

An ordinal outcome carries more information than a nominal one, and less than a continuous one — and the model has to match that middle ground exactly, or it either throws information away or invents precision that isn’t there. A five-point satisfaction rating, a disease-severity stage, a Likert-style agreement scale: these have a real order (strongly disagree is closer to disagree than to strongly agree), but the distance between adjacent categories is not assumed equal. Run a multinomial logistic regression on an ordinal outcome and the ordering is discarded entirely — the model treats “strongly agree” and “disagree” as no more or less related than any other pair of categories. Run ordinary linear regression on the category codes and the model assumes 1-to-2 means the same thing as 4-to-5, which is rarely defensible for a rating scale. Ordinal logistic regression is built for exactly this case, and it comes with one specific, checkable assumption — proportional odds — that determines whether its output means what it claims to mean.

When ordinal outcomes call for this model over multinomial logistic regression

The decision rule is simple to state and easy to get wrong under time pressure: if the outcome categories have a genuine, substantively meaningful order, use an ordinal model; if they don’t, use multinomial. A manuscript disposition (published, revised, rejected-with-resubmission-option, rejected-outright) is arguably ordered but the “distance” between adjacent outcomes isn’t consistent, so cases like this get judged on domain grounds, not just on whether a ranking exists on paper. A pain scale, a Likert agreement item, a cancer stage, a letter grade — these are ordinal in the sense the model needs: the order is the whole point of the measurement, and collapsing it into an unordered multinomial model discards information the survey or instrument was specifically designed to capture (see levels of measurement for the underlying distinction, and multinomial logistic regression for the model this one is not — that guide’s own worked examples are all genuinely unordered outcomes for exactly this reason).

Ordinal logistic regression also isn’t automatically the right ordinal-outcome model. It assumes a specific structure — proportional odds — that a different ordinal outcome might not have; that’s the entire subject of the rest of this guide.

The proportional-odds model, concretely

The standard ordinal logistic regression — the proportional-odds model, sometimes called the ordered logit or cumulative logit model — doesn’t model the probability of each category directly. It models a set of cumulative logits: for an outcome with k ordered categories, the model estimates k−1 logits, one for each way of splitting the scale into “at or below this point” versus “above this point.” For a five-point scale, that’s four cumulative splits: 1 vs. 2–5, 1–2 vs. 3–5, 1–3 vs. 4–5, and 1–4 vs. 5.

Each of those four splits gets its own intercept (its own baseline log-odds of falling at or below that cut point) — but critically, the model estimates only one coefficient per predictor, shared across all k−1 splits. That single shared coefficient is what “proportional odds” means in practice: the effect of a one-unit increase in a predictor on the odds of being in a higher category is assumed to be the same size, regardless of which cut point on the scale you’re looking at. Geometrically, this is also why the model is sometimes called the parallel-lines or parallel-regression model: plot the four cumulative logit lines against a continuous predictor and, under the proportional-odds assumption, they are parallel — same slope, different intercepts.

This is a real economy, not just a convenience. A binary logistic regression on a five-point outcome collapsed into “top two boxes vs. rest” would need a separate model, with its own separately-estimated coefficient, for every other way you might have collapsed the scale. The proportional-odds model estimates one coefficient that is claimed to hold across all of them simultaneously — which is exactly why it needs to be checked rather than assumed.

The proportional-odds assumption, explained concretely

Take a concrete case: a five-point post-training satisfaction survey (1 = very dissatisfied to 5 = very satisfied), predicted by whether a respondent completed an optional refresher module (yes/no). The proportional-odds assumption says the refresher module’s effect on the odds of scoring higher is the same size whether you’re looking at the boundary between “dissatisfied” and “neutral,” between “neutral” and “satisfied,” or between “satisfied” and “very satisfied.” In practice this is often false in a specific, interpretable way: the refresher might move people decisively off the bottom of the scale (a strong effect at the low cut points) while doing almost nothing to distinguish “satisfied” from “very satisfied” at the top (a near-zero effect at the high cut point). A single pooled coefficient averages those two very different effects into one number that accurately describes neither end of the scale — and every odds ratio the model reports for that predictor inherits the same blur.

This matters beyond statistical tidiness. A violated proportional-odds assumption doesn’t just produce a slightly-off model — it can produce a coefficient with the wrong practical implication, understating a real effect concentrated at one end of the scale while implying a uniform effect that isn’t there. That’s the reason this assumption gets tested rather than taken on faith, and why a partial-proportional-odds fallback exists for exactly the case where it fails for only some predictors.

Testing it: the Brant test

The most widely used formal check is the Brant test (Brant, 1990, Biometrics), which compares the fitted proportional-odds model against the set of separate binary logistic regressions the cumulative-logit approach is implicitly pooling. If the pooled, single-coefficient model and the separate-coefficients-per-cut-point models fit about equally well, proportional odds holds; if the separate models fit meaningfully better, it doesn’t.

The Brant test reports an omnibus chi-square across all predictors jointly, plus a separate chi-square for each predictor individually — the per-predictor breakdown is usually the more useful output in practice, because it’s common for the assumption to hold for most predictors and fail for just one or two, which is exactly the situation the partial-proportional-odds model below is built to handle. A significant result (conventionally p < 0.05) on the omnibus test, or on any individual predictor, is evidence against proportional odds for that predictor.

Running it: in R, fit the model with MASS::polr(), then pass the fitted object to brant::brant() from the dedicated brant package, which implements the test directly against a polr object and prints the omnibus and per-variable results in one call. In Stata, fit with ologit, then run the user-written brant command (community-contributed, install via ssc install brant) as a post-estimation check, or use the older built-in omodel logit command, which fits the ordinal model and reports a likelihood-ratio-based proportional-odds test in one step. SPSS’s PLUM procedure (Polytomous Universal Model, SPSS’s ordinal regression command) reports its own “Test of Parallel Lines” table by default alongside the model output — same underlying question, different implementation, no separate package needed.

Two practical cautions worth carrying into the interpretation: the Brant test, like most likelihood-ratio-family tests, gains power with sample size, so a large dataset can flag a violation that’s statistically real but small enough not to change any substantive conclusion — inspect the size of the per-predictor discrepancy, not just the p-value. And the test assumes the rest of the model (predictor set, functional form) is otherwise correctly specified; a Brant failure driven by an omitted nonlinearity or interaction can look identical to a genuine proportional-odds violation.

When it fails: the partial-proportional-odds fallback

A Brant test failure does not mean abandoning the ordinal model for multinomial logistic regression — that discards the ordering information again, the exact loss ordinal logistic regression exists to avoid, and is usually the wrong response when only one or two predictors are the problem. The more targeted fix is the partial proportional odds model (also called the generalized ordered logit model), introduced by Peterson and Harrell (1990): it relaxes the shared-coefficient constraint only for the specific predictors that fail the Brant test, letting those get a separate coefficient at each cut point, while every predictor that passed keeps a single pooled coefficient as before. The result is a model that is exactly as parsimonious as full proportional odds where that assumption actually holds, and only as flexible as it needs to be where it doesn’t — rather than discarding the ordinal structure entirely or over-fitting every predictor with unnecessary separate coefficients.

Running it: in Stata, the user-written gologit2 command (Williams, 2006) is the standard tool — its autofit option runs the Brant-style checks itself and automatically constrains only the predictors that pass, which is usually the fastest path from “the Brant test failed” to a usable model. In R, the VGAM package’s vglm() function fits the same class of model via family = cumulative(parallel = FALSE ~ x1 + x2), where the right-hand side of parallel names exactly the predictors allowed to vary by cut point; everything not named there stays constrained to a single coefficient. The ordinal package’s clm() function offers a comparable nominal = argument for the same purpose.

One caution worth stating plainly: a fully unconstrained partial-proportional-odds model, with every predictor allowed to vary freely by cut point, can produce fitted probabilities that aren’t monotonic across categories for some combinations of predictor values — the model can, in principle, imply a smaller cumulative probability at a higher cut point than at a lower one, which isn’t logically coherent for an ordinal outcome. This is exactly why the partial version — relaxing only the predictors the Brant test actually flagged, not every predictor by default — is the better-behaved choice over the fully generalized ordered logit model in most applications.

Reporting checklist

• State the outcome’s categories and confirm, on substantive grounds, that they’re genuinely ordered — not just orderable on paper.
• Report the proportional-odds (Brant test, or equivalent) result before reporting model coefficients — a reader can’t evaluate a pooled odds ratio without knowing whether pooling was justified.
• If the assumption held for all predictors, say so explicitly and report standard odds ratios with their shared interpretation across all cut points.
• If it failed for specific predictors, report the partial-proportional-odds model, naming which predictors were allowed to vary by cut point and why, rather than silently switching to a fully unconstrained or a multinomial model without explanation.
• Report the number of cut points (k−1 for a k-category outcome) and confirm cell sizes at each cut point are adequate — a proportional-odds test run on a sparsely populated tail category is unstable for the same reason any categorical test is with small cell counts.

Frequently asked questions

What’s the difference between ordinal logistic regression and running linear regression on the category codes?

Linear regression on category codes (treating a 1–5 scale as if it were measured on an interval scale) assumes the distance between every adjacent pair of categories is equal — that moving from 1 to 2 represents the same underlying change as moving from 4 to 5. Ordinal logistic regression makes no such assumption: it only uses the ordering, not any implied distance between categories, which is usually the more defensible position for a rating scale, a severity stage, or any other ordinal instrument that wasn’t explicitly validated as interval-level.

Can I just use multinomial logistic regression and ignore the ordering to avoid dealing with proportional odds?

You can, and it’s a legitimate fallback if the proportional-odds assumption fails badly across most predictors and a partial-proportional-odds model doesn’t resolve it cleanly — but it comes at a real cost: multinomial logistic regression estimates a full separate set of coefficients per category with no ordering constraint at all, which uses more degrees of freedom, is harder to summarize, and discards exactly the ordering information the outcome was designed to carry. Treat it as a fallback for a genuinely difficult case, not a way to skip testing the assumption in the first place.

Does a non-significant Brant test prove the proportional-odds assumption holds?

No — a non-significant result means the test didn’t detect a violation with the data and predictors at hand, which is a weaker claim than “the assumption is true.” In a small sample, the Brant test may lack the power to detect a real but modest violation; that’s a reason to look at the per-predictor effect sizes, not just the omnibus p-value, and to be cautious about over-interpreting a clean result from a small dataset as a strong confirmation.

Is the proportional-odds assumption the same thing as the proportional-hazards assumption in a Cox model?

They’re conceptually related — both assume a predictor’s effect is constant across a set of ordered thresholds (categories in one case, time points in the other) — but they apply to different outcome types and are tested with different tools. See the Cox proportional-hazards model for the survival-analysis version of the same underlying idea and how it’s checked.

For the unordered-outcome model this one is not, see multinomial logistic regression. For the binary case both generalize from, see logistic regression and the logit model, with software-specific implementations at logistic regression in R and logistic regression in Stata. For the measurement-level distinction that determines whether an outcome is ordinal in the first place, see levels of measurement, and for a common source of ordinal outcomes in practice, see Likert scale survey design. For the broader assumptions and reporting conventions regression models share, see regression analysis: assumptions, interpretation, and how to report it.