Interpreting and Testing the Proportional-Odds Assumption

Why the Shared-Slope Assumption Matters

An ordinal pharmacodynamic endpoint might record response as none, mild, moderate, or severe. A cumulative logit model describes the probability of being at or below each response threshold. Its proportional-odds assumption says that concentration has the same slope at every threshold. This is a substantive restriction: the concentration effect is not allowed to differ between, for example, the transition from none to at least mild and the transition from moderate or less to severe.

The restriction can make an ordinal model compact and interpretable, but it should not be mistaken for a consequence of the categories being ordered. It is an assumption about how the modeled concentration effect behaves across all cumulative splits. If it is badly wrong, a single reported slope can conceal important differences in response across thresholds.

Definition: Let $Y$ have ordered categories $1,\ldots,K$, and let $x$ be concentration. The cumulative logit proportional-odds model is $$ \operatorname{logit}\{P(Y\leq j\mid x)\}=\alpha_j-\beta x,\qquad j=1,\ldots,K-1. $$ The $\alpha_j$ are threshold-specific intercepts, while the slope $\beta$ is shared across all $K-1$ cumulative logits. Proportional odds refers to this common slope, not to equal threshold intercepts or equal category probabilities.

This tutorial uses the minus-sign convention in the cumulative logit. Thus, when $\beta>0$, increasing concentration lowers the cumulative probability of being at or below any given threshold, shifting probability toward higher response categories. The interpretation depends on the order in which categories are coded; the response ordering must therefore be stated and checked before interpreting the sign.

What Proportional Odds Implies

For a given threshold $j$, the cumulative odds of being at or below that threshold are $P(Y\leq j\mid x)/P(Y>j\mid x)$. Increasing concentration by $\Delta x$ changes the log of these odds by $-\beta\Delta x$. The same change applies at every threshold. Equivalently, the odds of being above the threshold rather than at or below it are multiplied by $\exp(\beta\Delta x)$ at every threshold.

\frac{\operatorname{odds}(Y\leq j\mid x+\Delta x)} {\operatorname{odds}(Y\leq j\mid x)} =\exp(-\beta\Delta x),\qquad \frac{\operatorname{odds}(Y>j\mid x+\Delta x)} {\operatorname{odds}(Y>j\mid x)} =\exp(\beta\Delta x).

The threshold intercepts can differ, so baseline cumulative odds are allowed to vary across splits. What is held constant is the concentration-related odds ratio for a specified concentration change. This is not a claim that every individual category probability changes by the same amount: category probabilities are differences between cumulative probabilities, so their changes can differ in size and even in direction.

Worked Example: Interpreting a Shared Concentration Slope

Hypothetical model. Suppose response has four ordered categories, from none (1) to severe (4), and concentration $x$ is measured in mg/L. A fitted cumulative logit model has $\beta=0.040$ per mg/L. Interpret an increase of 10 mg/L.

For each threshold $j=1,2,3$, the odds of being above that threshold are multiplied by $$ \exp(\beta\Delta x)=\exp(0.040\times 10)=\exp(0.4)=1.492. $$ Thus, for a 10 mg/L increase, the model estimates 1.492 times the odds of a response above the threshold, compared with at or below it. This is a 49.2% increase in those odds, and the same odds ratio applies at all three thresholds.

In the opposite direction, the odds of being at or below a threshold are multiplied by $\exp(-0.4)=0.6703$. The estimates are about 0.6703 times as large, or 32.97% lower. These are odds comparisons, not probability increases of 49.2 percentage points. The probabilities depend on the relevant threshold intercept and starting concentration.

Why the Odds Ratio Is Shared

The shared odds ratio follows directly from subtracting the cumulative logits at two concentrations. For a fixed threshold $j$, the model gives a logit of $\alpha_j-\beta x$ at $x$ and $\alpha_j-\beta(x+\Delta x)$ at $x+\Delta x$. Their difference is $-\beta\Delta x$; the threshold intercept cancels. Exponentiating the difference gives the cumulative odds ratio. Because neither $\beta$ nor $\Delta x$ depends on $j$, this ratio is the same at every threshold.

\begin{aligned} &[\alpha_j-\beta(x+\Delta x)]-[\alpha_j-\beta x]\\ &\qquad=-\beta\Delta x,\\ &\frac{\operatorname{odds}(Y\leq j\mid x+\Delta x)} {\operatorname{odds}(Y\leq j\mid x)} =\exp(-\beta\Delta x). \end{aligned}

This short derivation also shows why the assumption is about parallel cumulative logits: as concentration changes, each cumulative logit has the same slope, while its threshold-specific intercept sets its level. A plot of the fitted cumulative logits against concentration therefore gives straight, parallel lines when the concentration effect is linear and proportional odds holds.

Threshold-Specific Estimates and Graphical Checks

A useful diagnostic is to fit or estimate a concentration slope separately for each cumulative split, temporarily treating $Y\leq j$ versus $Y>j$ as a binary outcome for $j=1,\ldots,K-1$. Under the proportional-odds model, these slopes should be reasonably compatible with one common value, allowing for sampling uncertainty. These separate fits are diagnostic rather than a replacement for the ordinal likelihood: their outcomes overlap, so the estimates are correlated. Comparing their confidence intervals by eye is not a formal test of equality.

Graphical checks can expose patterns that a single summary statistic hides. Plot observed cumulative proportions, or their logits when the proportions are not zero or one, against concentration. For a continuous predictor, binning concentration can help display the pattern, but the resulting points are descriptive and depend on how the bins are chosen. A model-based plot can instead overlay fitted cumulative logits or cumulative probabilities for each threshold. Look for systematic changes in slope across thresholds, not just noisy separation among points. Sparse categories, especially at the extremes, can make empirical logits unstable or undefined.

Worked Example: Reading Threshold-Specific Estimates

Hypothetical diagnostic. For a four-category endpoint and concentration measured per 10 mg/L, separate cumulative-logit fits give slope estimates $0.40$, $0.15$, and $0.05$ at the three thresholds, with standard errors $0.12$, $0.12$, and $0.15$, respectively. For comparison, the proportional-odds model estimates a shared slope of $0.21$.

Approximate 95% Wald intervals for the separate slopes are: $$ \begin{aligned} 0.40\pm1.96(0.12)&=[0.165,\ 0.635],\\ 0.15\pm1.96(0.12)&=[-0.085,\ 0.385],\\ 0.05\pm1.96(0.15)&=[-0.244,\ 0.344]. \end{aligned} $$

The estimates decrease across thresholds, which is a possible sign that a single slope does not describe all splits well. But the individual intervals are wide, and their overlap does not demonstrate that the slopes are equal. Nor would non-overlap alone be a suitable formal test, because the estimates are correlated. The appropriate next steps are to inspect the cumulative-response pattern and use a formal comparison that accounts for the model structure.

Formal Tests of Proportional Odds

A common formal approach compares the proportional-odds model with a nested cumulative-logit alternative that has a separate concentration slope at each threshold. With $K$ response categories and one concentration predictor, the proportional-odds model has one slope and the alternative has $K-1$ slopes. The likelihood-ratio statistic is twice the difference in maximized log likelihoods. Under the null hypothesis of proportional odds, and under the usual large-sample regularity conditions, it is approximately chi-square with $K-2$ degrees of freedom.

The comparison must use the same observations, response categories, concentration representation, and other model terms. If the concentration effect is represented by a nonlinear function, proportional odds means that the coefficients for that function are shared across thresholds; a test should compare that model with an alternative that relaxes those shared effects, not silently change the shape of the concentration response. A likelihood-ratio test also relies on a valid, comparable likelihood and adequate information for the fitted parameters.

Worked Example: Likelihood-Ratio Test With Four Categories

Hypothetical test. An independent-observation data set has a four-category ordinal response and one continuous concentration predictor. The proportional-odds fit has log likelihood $-198.4$; a nested fit with three threshold-specific concentration slopes has log likelihood $-194.1$. The models use the same observations and all other terms are identical.

The likelihood-ratio statistic is $$ G^2=2\{(-194.1)-(-198.4)\}=2(4.3)=8.60. $$

There are $K-1=3$ slopes in the less restricted model and one in the proportional-odds model, so the difference is $3-1=2$ degrees of freedom. Under $H_0$, $G^2$ is compared with a chi-square distribution with $df=2$. The p-value is $$ P(\chi^2_2\geq8.60)=\exp(-8.60/2)=\exp(-4.30)=0.0136. $$

At significance level $\alpha=0.05$, this p-value is below 0.05, so the test provides evidence against the shared-slope restriction. It does not identify which threshold causes the discrepancy or quantify its clinical importance. Those questions require examination of the estimates, uncertainty, and predicted response patterns.

A score test or a Wald test can also be used to assess restrictions on threshold-specific effects. The Brant test is a familiar test of the proportional-odds restriction in cumulative-logit models. The particular implementation and its covariance assumptions matter, especially when data are sparse or observations are not independent. For repeated observations, a test based on an independent-observation likelihood is not justified merely because each row has a concentration and response; the dependence must be handled by an appropriate model and inferential procedure.

Assumptions, Diagnostics, and Alternatives

A formal rejection is meaningful only if the compared models are otherwise sensible. Check that categories are correctly ordered and coded, that concentration is measured on the stated scale, and that its functional form is adequate. A curved concentration effect forced into a straight line can produce apparent threshold differences. Also inspect category counts and fitted probabilities: very sparse categories or separation can make threshold-specific slopes unstable and standard likelihood approximations unreliable. For each concentration value in the relevant range, cumulative probabilities should remain ordered, so that category probabilities derived from them are nonnegative.

The null model’s key assumption is that the concentration effect on every cumulative logit is the same. Its alternative in the likelihood-ratio comparison relaxes that restriction. For more than one predictor, proportional odds usually means the slope for each predictor is shared across thresholds; an omnibus test of all such restrictions adds $(K-2)p$ degrees of freedom when there are $p$ predictors and the corresponding unrestricted slopes are identifiable. Tests of individual predictor restrictions can be more targeted, but multiple testing and limited precision should be considered.

Approach	What it contributes	Main caution
Threshold-specific fits	Show the direction and magnitude of slope estimates at individual cumulative splits.	The estimates are correlated; separate fits do not, by themselves, provide a valid joint test.
Graphical checks	Reveal visible changes in cumulative-logit slopes or systematic mismatch in cumulative response patterns.	Plots are descriptive, can be noisy, and may be sensitive to binning.
Formal score, Wald, or likelihood-ratio test	Tests a specified set of shared-slope restrictions while accounting for model structure.	A p-value does not measure clinical importance or prove the assumption when nonsignificant.
Multinomial logit	Models category probabilities without the proportional-odds restriction.	It does not use the ordering in the same way and may require more parameters.

The main limitation of a formal test is that it answers a narrow question: whether the data provide evidence against specified equality restrictions, under the fitted models and their assumptions. With a small sample, low power can leave substantial departures undetected. With a very large sample, a minor departure can be statistically significant even if the resulting predictions are practically similar. Evaluate how the restrictions affect estimated cumulative probabilities and clinically meaningful response contrasts, not only the test result.

Common pitfall: “The proportional-odds test was not significant, so the assumption has been proven.” This reasoning is incorrect. A nonsignificant result means the data did not provide sufficient evidence against the tested restriction at the chosen $\alpha$; it does not establish that all threshold slopes are equal. Consider the precision of threshold-specific estimates, the diagnostic plots, sample size, model form, and the clinical consequences of plausible departures.

If the restriction appears inadequate, the next modeling decision should be guided by the source and pattern of the departure. The following material develops partial proportional-odds models, which relax selected shared effects while retaining the ordinal structure. For the present assessment, the essential distinction is between diagnosing a shared-slope assumption and choosing a more flexible model: threshold-specific fits and tests provide evidence about the restriction, but they do not by themselves settle which alternative is appropriate.

Check Your Understanding

Answer each question, then compare your reasoning with the solutions.

In the model $\operatorname{logit}\{P(Y\leq j\mid x)\}=\alpha_j-\beta x$, what does a positive $\beta$ imply for the odds of a response above threshold $j$ when concentration increases?
A four-category endpoint has one concentration predictor. How many degrees of freedom does the likelihood-ratio comparison use when testing one shared slope against separate slopes at all thresholds?
Why is overlap among separate threshold-specific slope confidence intervals not a formal demonstration of proportional odds?
A formal test gives $p=0.18$ at $\alpha=0.05$. State the conclusion without claiming that the assumption is true.
Name one way a misspecified concentration functional form could affect an assessment of proportional odds.

Solutions. (1) The odds of being above the threshold are multiplied by $\exp(\beta\Delta x)$ for an increase of $\Delta x$; if $\beta>0$, those odds increase. (2) There are three threshold-specific slopes versus one shared slope, so $df=3-1=2$, equivalently $K-2=2$. (3) The estimates are correlated, and overlap of separate intervals is not a joint test of equality. (4) At the 0.05 level, there is insufficient evidence to reject the shared-slope restriction; this does not prove proportional odds. (5) For example, forcing a curved concentration effect to be linear can create apparent differences among threshold slopes, even if a more suitable shared nonlinear effect would fit adequately.