ORDINAL LOGISTIC REGRESSION · PROPORTIONAL ODDS · ORDERED PROBIT · BRANT TEST · PREDICTED PROBABILITIES · RUNS IN YOUR BROWSER · NOTHING UPLOADED

What moves a rating up or down the scale, and does one model fit the whole scale?

Paste one row per person or item with an ordered outcome (a 1 to 5 rating, a Likert answer, a grade, a disease stage) and what might explain it: numbers, or groups such as a treatment arm. You get an ordinal logistic regression (the proportional-odds model) or an ordered probit: each predictor's odds ratio with its interval, the cut-points, a likelihood-ratio test and pseudo R², Brant's test of whether one odds ratio really fits every cut of the scale, and the chance of each level as a predictor changes.

Nothing you enter leaves this browser tab. Statistical output from data you supply, not advice. Not medical, legal, tax, or financial advice.

Your data

Rows with a missing outcome or predictor are skipped and counted.

Predictors, and how each is read

Paste data to choose predictors.

Method and citation

The model. Think of each answer as a hidden continuous score cut into levels at thresholds θ₁ < θ₂ < … The proportional-odds model (McCullagh 1980; Walker & Duncan 1967) says P(Y ≤ level j) = F(θⱼ − x′β) with F the logistic distribution, one slope per predictor shared by every cut. A positive coefficient pushes the outcome up the scale, and exp(β) is the factor by which one more unit multiplies the odds of being above any level rather than at or below it. The ordered probit uses the normal distribution for F, so a coefficient is a shift in standard deviations of the hidden score.

The fit. Maximum likelihood by Newton-Raphson with the exact gradient and Hessian, step halving, and thresholds kept in order, starting from the thresholds that reproduce the overall shares of each level. Predictors are centred and scaled during the fit and mapped back afterwards, so a column measured in thousands and one measured in thousandths fit equally well. Standard errors come from the observed information; intervals are Wald intervals, exponentiated for odds ratios, which are also given per standard deviation of each numeric predictor so they can be compared. Text predictors become dummy variables against a reference level (the most common one unless you choose). The likelihood-ratio test compares the model with one that only has the thresholds; McFadden's (1974) R² is 1 − log L / log L₀.

Brant's test. Proportional odds is an assumption. Brant (1990) fits a separate binary logistic regression for each cut (above level j or not) and tests whether their slopes are equal with a Wald test, using the covariance between two binary fits on the same rows, (X′WⱼX)⁻¹ X′WⱼₗX (X′WₗX)⁻¹ with weights πₗ − πⱼπₗ. It is given overall and for each coefficient, with the per-cut slopes so you can see where a predictor acts differently. Like any test it flags small, harmless differences in large samples, so read the slopes, not only the p-value.

Verified by tests in the source against an independent computation (eval/reference.py): on the wine example, a simulated Likert survey with a three-level group and a simulated data set built to violate proportional odds, the log-likelihoods match statsmodels 0.15.0's OrderedModel to 10⁻¹¹, the coefficients and cut-points to 10⁻⁷ and the standard errors to 10⁻⁵ (statsmodels differentiates the likelihood numerically, the engine exactly), for both links; Brant's statistics match a numpy implementation built on statsmodels' binary logits to 10⁻⁹. On 400 data sets simulated under proportional odds, Brant's test rejected at the 5% level in 6.0% of them.

BibTeX

@article{mccullagh1980regression,
  author  = {McCullagh, Peter},
  title   = {Regression models for ordinal data},
  journal = {Journal of the Royal Statistical Society: Series B (Methodological)},
  volume  = {42}, number = {2}, pages = {109--142}, year = {1980},
  doi     = {10.1111/j.2517-6161.1980.tb01109.x}
}
@article{walker1967estimation,
  author  = {Walker, Strother H. and Duncan, David B.},
  title   = {Estimation of the probability of an event as a function of several independent variables},
  journal = {Biometrika},
  volume  = {54}, number = {1/2}, pages = {167--179}, year = {1967},
  doi     = {10.1093/biomet/54.1-2.167}
}
@article{brant1990assessing,
  author  = {Brant, Rollin},
  title   = {Assessing proportionality in the proportional odds model for ordinal logistic regression},
  journal = {Biometrics},
  volume  = {46}, number = {4}, pages = {1171--1178}, year = {1990},
  doi     = {10.2307/2532457}
}
@incollection{mcfadden1974conditional,
  author    = {McFadden, Daniel},
  title     = {Conditional logit analysis of qualitative choice behavior},
  booktitle = {Frontiers in Econometrics},
  editor    = {Zarembka, Paul},
  publisher = {Academic Press}, pages = {105--142}, year = {1974}
}
@book{agresti2010analysis,
  author    = {Agresti, Alan},
  title     = {Analysis of Ordinal Categorical Data},
  edition   = {2nd}, publisher = {Wiley}, year = {2010},
  doi       = {10.1002/9780470594001}
}

Limits

Questions

When should I use ordinal regression instead of linear regression?

When the outcome is a set of ordered categories, such as a 1 to 5 rating, a Likert answer, a letter grade or a disease stage. Linear regression treats the gaps between levels as equal and can predict impossible values like 5.7; multinomial regression throws the order away. Ordinal regression uses the order and nothing more.

How do I read an odds ratio from an ordered logit?

An odds ratio of 2 for a predictor means each extra unit doubles the odds of being above any given level rather than at or below it: the odds of a 4 or 5 against 1 to 3, and equally the odds of a 2 or higher against a 1. Above 1 pushes the outcome up the scale, below 1 pushes it down.

What is the proportional odds assumption?

That one odds ratio per predictor holds at every cut of the scale: the predictor shifts the odds of being above level 1 by the same factor as the odds of being above level 4. When it fails, a predictor matters more at one end of the scale than the other, and its single odds ratio is an average that can hide that.

What does Brant's test do?

It fits a separate binary logistic regression for each cut of the scale (above level 1 or not, above level 2 or not, and so on) and tests whether their slopes are equal, overall and for each predictor. A small p-value says proportional odds fails for that predictor; the table of per-cut slopes shows where.

My proportional odds test is significant. What now?

Look at the per-cut slopes before doing anything. With large samples the test flags differences too small to matter. If the slopes really differ, report the per-cut slopes or the predicted probabilities alongside the odds ratio, merge sparse levels, or move to a partial proportional odds or multinomial model in a statistics package.

Logit or probit?

They nearly always agree on direction and significance; probit coefficients are about 0.55 to 0.6 times the logit ones. Logit gives odds ratios and Brant's test; probit reads as shifts on a normal latent scale, which some fields prefer. Compare their AIC if you want the data to choose.

Is my data uploaded anywhere?

No. The file is read and analysed by JavaScript in this browser tab. There are no accounts, cookies or analytics vendors.

Example data: the Wine Quality data set of Cortez, Cerdeira, Almeida, Matos and Reis (2009), UCI Machine Learning Repository, doi:10.24432/C56S3T, licensed CC BY 4.0, retrieved on 2026-11-19: 1,599 red Vinho Verde wines, each with the median of at least three experts' quality grades and its laboratory measurements. Kept: the quality grade and six of the eleven measurements, unchanged (data/build_example.py).