The Parallel-Regression Restriction in Cumulative-Link Models

Why the Parallel-Regression Restriction Matters

An ordinal response has categories with a meaningful order, such as mild, moderate, and severe. A cumulative-link model describes the response by considering several cutpoints: below or at the first cutpoint, below or at the second, and so on. The parallel-regression restriction says that a predictor has the same slope at every one of these cutpoints. Consequently, a given predictor contrast has the same cumulative odds ratio at each cutpoint.

This restriction is also called the proportional-odds assumption when the link is the logit. “Proportional” refers to the constancy of the odds ratio across cutpoints; it does not mean that the category probabilities themselves change by the same amount. The equations below make precise which odds are being compared and what is held constant.

Cumulative Odds and the Restricted Model

Let \(Y\) be an ordinal response with ordered categories \(1,\ldots,K\), where \(K\geq 2\), and let \(x\) be a \(p\)-dimensional predictor vector. For cutpoint \(j\), where \(j=1,\ldots,K-1\), the cumulative lower probability is \(P(Y\leq j\mid x)\), and the cumulative upper probability is \(P(Y>j\mid x)\). When both probabilities are positive, their ratio is the cumulative upper-versus-lower odds.

Here \(\beta\) is one common \(p\)-vector of predictor slopes, and \(\alpha_j\) is a cutpoint-specific threshold. Equivalently, the cumulative lower logit is \(\log\{P(Y\leq j\mid x)/P(Y>j\mid x)\}=\alpha_j-x^\mathsf{T}\beta\). This sign convention makes a positive coefficient increase the odds of being above, rather than at or below, any given cutpoint. The model is understood to describe a valid ordinal response distribution, with cumulative probabilities strictly between zero and one wherever the odds are used.

The “parallel” feature is seen by holding \(j\) fixed and viewing the log cumulative upper-versus-lower odds as a function of \(x\). Each cutpoint has its own vertical offset, \(-\alpha_j\), but all these linear predictors have the same slope vector \(\beta\). With one predictor, they are parallel lines; with several predictors, they share the same gradient.

Unrestricted Cutpoint-Specific Slopes

To state the restriction clearly, compare it with a cumulative-logit specification that allows a separate slope vector at each cutpoint:

The proportional-odds model is the special restriction \(\beta_1=\cdots=\beta_{K-1}=\beta\). In both expressions, the threshold \(\alpha_j\) can vary with the cutpoint. The restriction concerns the slopes, not the thresholds. The cutpoint-specific equations are used here to define the slope comparison; a set of such equations must still correspond to valid cumulative probabilities to describe an ordinal response model.

For a change from predictor vector \(x\) to \(x+d\), the odds ratio compares the cumulative upper-versus-lower odds at those two predictor values, using the same cutpoint \(j\). It is not the odds of category \(j+1\) versus category \(j\).

Deriving the Cumulative Odds Ratio

To derive the result, exponentiate the cumulative-logit equation. At predictor value \(x\), the odds at cutpoint \(j\) are \(\exp(x^\mathsf{T}\beta_j-\alpha_j)\). At \(x+d\), they are \(\exp((x+d)^\mathsf{T}\beta_j-\alpha_j)\). Thus their ratio is

The cutpoint threshold and the starting predictor value cancel. If the slope vector is common, replace every \(\beta_j\) by \(\beta\), so the resulting odds ratio does not depend on \(j\). This establishes the odds-ratio interpretation directly from the model equation.

Worked Examples

When Invariant Odds Ratios Imply Shared Slopes

The shared-slope restriction implies invariant odds ratios by the derivation above. There is also a converse, provided predictor contrasts are rich enough to identify all slope directions. This condition matters: if the observed predictor values vary only along a lower-dimensional subspace, equality of odds ratios for those observed contrasts cannot identify slope components outside that subspace.

Proof. For any contrast \(d=x'-x\), the odds ratio at cutpoint \(j\) is \(\exp(d^\mathsf{T}\beta_j)\). If the ratios are equal at cutpoints \(j\) and \(k\), then

The exponential function is one-to-one, so \(d^\mathsf{T}(\beta_j-\beta_k)=0\). By assumption, the available contrasts span \(\mathbb{R}^p\). In particular, every vector in a spanning set of \(\mathbb{R}^p\) is orthogonal to \(\beta_j-\beta_k\). A vector orthogonal to a spanning set is orthogonal to every vector in \(\mathbb{R}^p\), including itself; hence \(\beta_j-\beta_k=0\). This holds for each pair \(j,k\), so all the slope vectors are equal. Conversely, if every \(\beta_j=\beta\), then \(\exp(d^\mathsf{T}\beta_j)=\exp(d^\mathsf{T}\beta)\), independent of \(j\). \(\square\)

What Is Invariant—and What Is Not

Under proportional odds, invariance applies to the cumulative upper-versus-lower odds ratio for a fixed predictor contrast. The thresholds affect the odds at each cutpoint, as the first worked example showed, but they cancel when comparing two predictor values at that same cutpoint. The invariant ratio need not be one: a nonzero contrast can increase or decrease the odds, with the same multiplicative factor at every cutpoint.

The model does not assert that the response-category probabilities change by a common multiplicative factor, or that adjacent-category odds are constant across cutpoints. Those are different comparisons. Nor does a common slope mean that the cumulative odds themselves are equal at different cutpoints: the thresholds generally make those odds different.

A common mistake is to read \(\exp(\beta_r)\) as the odds ratio for category \(r+1\) versus category \(r\). That is not the model’s interpretation. For a one-unit increase in predictor \(x_r\), \(\exp(\beta_r)\) is the multiplier for the odds \(P(Y>j\mid x)/P(Y\leq j\mid x)\), compared at the same cutpoint \(j\). If the lower-versus-upper odds are used instead, the corresponding ratio is its reciprocal, \(\exp(-\beta_r)\). Always state the cutpoint comparison and the direction of the odds before interpreting a coefficient.