Proportional-Odds Assumptions in Ordinal NLME

Why a Shared Effect Across Thresholds Matters

An ordinal response has ordered categories, such as none, mild, moderate, and severe. A cumulative link model describes the probability of being at or below each category boundary. The proportional-odds constraint says that a covariate shifts every one of these cumulative boundaries by the same amount on the log-odds scale. It is a compact assumption: one effect summarizes how a predictor relates to the whole ordered response, rather than assigning it a separate effect at every boundary.

In an NLME model, this constraint applies to the conditional response model, given the individual’s random effects. It does not say that a covariate changes the probability of every category by the same amount. Category probabilities are differences between cumulative probabilities, so they can respond differently even when the cumulative log-odds effect is shared.

Definition: Let $Y_{ij}$ be an ordinal response for individual $i$ at observation $j$, with ordered categories $1,\ldots,K$. A proportional-odds cumulative-logit NLME model is $$ \operatorname{logit}\{P(Y_{ij}\leq k\mid b_i)\} =\alpha_k-\left(x_{ij}^{T}\beta+z_{ij}^{T}b_i\right), \qquad k=1,\ldots,K-1. $$ Here $\alpha_1<\cdots<\alpha_{K-1}$ are ordered thresholds, $x_{ij}$ is the fixed-effect covariate vector, $\beta$ is its coefficient vector, and $b_i$ is the individual random-effects vector, commonly modeled as $b_i\sim N(0,\Omega)$. The same fixed-effect coefficient vector $\beta$ appears at every threshold; this shared-effect restriction is the proportional-odds assumption.

The minus sign is a common cumulative-logit convention. With this convention, a positive value of $x^T\beta+z^Tb_i$ lowers the probability of being at or below a given threshold and shifts probability toward higher categories. Reversing the sign convention is possible, but the interpretation of coefficient signs must then reverse as well.

What Is Proportional?

For a particular threshold $k$, the cumulative probability is $P(Y_{ij}\leq k\mid b_i)$, and its complement is $P(Y_{ij}>k\mid b_i)$. The odds of being above that threshold rather than at or below it are therefore

\frac{P(Y_{ij}>k\mid b_i)}{P(Y_{ij}\leq k\mid b_i)} =\exp\left(x_{ij}^{T}\beta+z_{ij}^{T}b_i-\alpha_k\right).

For a one-unit increase in a scalar covariate $x$, holding other covariates and the random effects fixed, the odds above threshold $k$ are multiplied by $\exp(\beta_x)$. Crucially, the multiplier is the same for every $k$. That is the proportional-odds condition. It is not a claim that the probabilities, or the odds themselves, are equal at the different thresholds: the thresholds $\alpha_k$ give different baseline odds.

The assumption is conditional on $b_i$. In repeated-measures NLME models, the usual likelihood also treats observations as conditionally independent given the individual random effects and covariates. The random effects represent individual-level variation; they do not, by themselves, relax the shared fixed-effect restriction across thresholds. After averaging over a random-effects distribution, population-averaged odds ratios need not equal the conditional odds ratio. In general, averaging probabilities and then forming odds is not the same operation as averaging conditional odds.

From Cumulative Probabilities to Category Probabilities

Write $F_k=P(Y\leq k\mid b_i)$ for the cumulative probability at threshold $k$. The category probabilities follow by subtraction: $P(Y=1\mid b_i)=F_1$, $P(Y=r\mid b_i)=F_r-F_{r-1}$ for interior categories, and $P(Y=K\mid b_i)=1-F_{K-1}$. Ordered thresholds ensure $F_1\leq\cdots\leq F_{K-1}$, so these differences are nonnegative.

Worked Example: A Common Effect Across Three Thresholds

Suppose a hypothetical four-category response has thresholds $\alpha_1=-1$, $\alpha_2=0.5$, and $\alpha_3=2$. A covariate has coefficient $\beta=0.7$, and consider an individual with random effect $b_i=0$. Compare the predicted categories at $x=0$ and $x=1$.

At $x=0$, the linear predictor is $0$. Using $\operatorname{logit}^{-1}(u)=1/(1+e^{-u})$, the cumulative probabilities are $\operatorname{logit}^{-1}(-1)=0.268941$, $\operatorname{logit}^{-1}(0.5)=0.622459$, and $\operatorname{logit}^{-1}(2)=0.880797$. Subtracting the unrounded values gives category probabilities $0.268941$, $0.353518$, $0.258338$, and $0.119203$, or approximately (0.2689, 0.3535, 0.2583, 0.1192).

At $x=1$, the linear predictor is $0.7$. The cumulative logits are $-1-0.7=-1.7$, $0.5-0.7=-0.2$, and $2-0.7=1.3$. Their inverse logits are approximately $0.154465$, $0.450166$, and $0.785835$. Thus the category probabilities are $0.154465$, $0.450166-0.154465=0.295701$, $0.785835-0.450166=0.335669$, and $1-0.785835=0.214165$, or (0.1545, 0.2957, 0.3357, 0.2142).

At each threshold, increasing $x$ by one multiplies the odds of being above that threshold by $\exp(0.7)=2.014$. Yet the four category probabilities do not all move by the same amount: the first two decline, while the last two rise. The common odds ratio describes cumulative comparisons, not a uniform category-probability shift.

A Short Derivation of the Shared Odds Ratio

The proportional-odds interpretation follows directly from the cumulative-logit equation. For a scalar predictor $x$, hold the other covariates and $b_i$ fixed, and increase $x$ by $\Delta$. At threshold $k$, the log odds of being above the threshold are

\log\left\{\frac{P(Y>k\mid b_i)}{P(Y\leq k\mid b_i)}\right\} =x\beta+\text{constant}-\alpha_k.

After the increase, subtracting the original log odds leaves $\Delta\beta$; the threshold and all held-fixed terms cancel. Exponentiating the difference gives the odds ratio:

\frac{ P(Y>k\mid x+\Delta,b_i)/P(Y\leq k\mid x+\Delta,b_i) }{ P(Y>k\mid x,b_i)/P(Y\leq k\mid x,b_i) } =\exp(\Delta\beta).

There is no $k$ in the right-hand side. Thus, for any two thresholds, the model imposes the same covariate odds ratio. This derivation also shows why the interpretation is conditional: it holds for a fixed value of $b_i$, not automatically for probabilities averaged over individuals.

What Departures Look Like

A useful way to express nonproportional effects is to allow a separate coefficient $\beta_k$ at each threshold:

\operatorname{logit}\{P(Y_{ij}\leq k\mid b_i)\} =\alpha_k-\left(x_{ij}^{T}\beta_k+z_{ij}^{T}b_i\right).

This extension lets a covariate have different associations with different cumulative splits. For example, treatment could have a modest association with moving out of the lowest category but a larger association with avoiding the highest category. Such a pattern may be scientifically plausible, but it requires more parameters and careful checking: the resulting cumulative probabilities must still be ordered for covariate values and random effects in the intended prediction range.

Worked Example: Threshold-Specific Effects and Category Probabilities

Consider hypothetical data with the same four categories and thresholds $(-1,0.5,2)$, but suppose a predictor has threshold-specific coefficients $(0.2,0.8,1.1)$. Set $x=1$ and $b_i=0$. The three cumulative logits are $-1-0.2=-1.2$, $0.5-0.8=-0.3$, and $2-1.1=0.9$.

Applying the inverse logit gives cumulative probabilities $F_1=0.231475$, $F_2=0.425557$, and $F_3=0.710950$. They are ordered, so the category probabilities are valid: $F_1=0.231475$, $F_2-F_1=0.194082$, $F_3-F_2=0.285392$, and $1-F_3=0.289050$. Rounded only after subtraction, these are (0.2315, 0.1941, 0.2854, 0.2891).

The odds ratios for being above each threshold are $\exp(0.2)=1.221$, $\exp(0.8)=2.226$, and $\exp(1.1)=3.004$. Unlike the proportional-odds model, this model does not impose one common odds ratio. The resulting category probabilities remain interpretable, provided cumulative ordering holds throughout the predictor and random-effects range where predictions are used.

For the proportional model, ordered thresholds guarantee ordered cumulative logits at every covariate value because the same predictor term is subtracted at each threshold. With threshold-specific coefficients, this guarantee disappears. For instance, if the first cumulative logit exceeds the second at a particular covariate value, then $F_1>F_2$ and the implied probability $P(Y=2)=F_2-F_1$ is negative. A fitted threshold-specific model must therefore be examined for crossing cumulative probabilities, especially near the edges of the observed covariate range and when extrapolating.

Conditional Effects in an NLME Model

Worked Example: The Same Odds Ratio at Different Random Effects

Use the proportional model with thresholds $(-1,0.5,2)$, fixed-effect coefficient $\beta=0.7$, and $x=1$. Compare two hypothetical individuals with random intercepts $b_i=0.4$ and $b_i=-0.4$. Their linear predictors are $0.7+0.4=1.1$ and $0.7-0.4=0.3$.

For $b_i=0.4$, the cumulative logits are $(-2.1,-0.6,0.9)$, giving cumulative probabilities approximately $(0.109097,0.354344,0.710950)$. The category probabilities are therefore $0.109097$, $0.245247$, $0.356606$, and $0.289050$, or (0.1091, 0.2452, 0.3566, 0.2891).

For $b_i=-0.4$, the cumulative logits are $(-1.3,0.2,1.7)$, giving cumulative probabilities approximately $(0.214165,0.549834,0.845535)$. The category probabilities are (0.2142, 0.3357, 0.2957, 0.1545), calculated from the unrounded cumulative probabilities.

The individuals’ predicted category distributions differ because their random intercepts differ. Nevertheless, conditional on either random intercept, a one-unit increase in $x$ multiplies the odds of being above each threshold by $\exp(0.7)=2.014$. These conditional predictions should not be confused with a population-averaged prediction, which integrates the category probabilities over the random-effects distribution.

Assessing the Assumption and Choosing an Approach

The proportional-odds assumption is a modeling choice, not a consequence of the response being ordinal. Its plausibility is assessed by asking whether one shared effect gives an adequate description across the cumulative splits. Useful evidence includes threshold-specific effect estimates, their uncertainty and dependence, and whether fitted probabilities track observed category frequencies across predictor values. A large difference between threshold-specific estimates may indicate that a shared effect is too restrictive; uncertainty matters, so small numerical differences alone are not conclusive.

A direct alternative is a partial proportional-odds model: retain common effects for predictors whose effects are adequately represented by one coefficient, and allow selected predictors to vary by threshold. A fully threshold-specific model is more flexible but uses more parameters and can be unstable with sparse categories or limited information at some thresholds. The proportional model is more parsimonious and guarantees ordered cumulative probabilities when its thresholds are ordered. Either model still requires sensible category coding, adequate data support, and a random-effects structure appropriate to the repeated observations.

In an NLME analysis, inspect both the fixed-effect restriction and the implied conditional predictions. Check that the estimated thresholds are ordered, calculate category probabilities from unrounded cumulative probabilities, and evaluate cumulative ordering over the covariate and random-effects values used for inference or prediction. If an estimated threshold-specific model produces negative category probabilities or crossing cumulative curves, those predictions are not a valid ordinal distribution.

Common Mistakes and Limitations

Calling it equal category effects. Proportional odds means a common covariate effect on cumulative log odds, not equal changes in the probability of each category.
Interpreting the coefficient without specifying the split. Under the shared-effect model, $\exp(\beta)$ multiplies the odds of being above each threshold versus at or below it, conditional on the random effects and other covariates.
Assuming the conditional odds ratio is population averaged. The model’s coefficient describes conditional odds. Averaging over random effects changes probabilities, and the resulting marginal odds ratio need not be the same.
Subtracting rounded cumulative probabilities. Round only after calculating the category-probability differences. Otherwise, displayed values may not correctly represent the model-implied probabilities.
Allowing threshold-specific effects without checking ordering. Flexibility does not guarantee valid category probabilities; cumulative probabilities must remain nondecreasing over the relevant prediction range.

The proportional-odds model is limited when a predictor genuinely has different associations at different severity boundaries. Conversely, a more flexible model may be difficult to estimate when some categories are rare or the sample provides little information for particular thresholds. The choice should balance fit, interpretability, data support, and valid prediction rather than treating either specification as universally preferable.

Check Your Understanding

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

In a proportional-odds model with $\beta=0.4$, what is the conditional odds ratio for a one-unit increase in the predictor at each threshold?
For a four-category outcome, suppose the first two cumulative probabilities are $0.30$ and $0.65$. What is the probability of category 2?
Why can a threshold-specific model produce an invalid category probability even when each cumulative probability is between zero and one?
Does a shared conditional odds ratio imply that the category probabilities change by the same amount at every threshold? Explain.

Solutions. (1) It is $\exp(0.4)=1.492$ at every threshold, conditional on the random effects and other covariates. (2) The category-2 probability is $0.65-0.30=0.35$. (3) Validity also requires cumulative probabilities to be nondecreasing. If, for example, $F_2<F_1$, then $P(Y=2)=F_2-F_1<0$, despite both values being individually between zero and one. (4) No. The common effect applies to cumulative log odds; category probabilities are differences between cumulative probabilities and may change by different amounts.