Cumulative Logit Models for Ordered Pharmacodynamic Outcomes

Why Model Ordered Outcomes Cumulatively?

A pharmacodynamic endpoint may be recorded in ordered categories rather than as a continuous measurement or a yes/no response. For example, a symptom score might be classified as none, mild, moderate, or severe. These categories have a meaningful order, but their numerical labels do not guarantee that the distance from none to mild equals the distance from moderate to severe. A cumulative-logit model uses the ordering without treating category codes as equally spaced measurements.

The model describes probabilities of being at or below each category boundary. Its distinctive feature is that the exposure effect is shared across all those boundaries: a one-unit exposure increase changes each cumulative log-odds by the same amount. This shared-slope structure gives a compact model, but it is an assumption about how exposure relates to the whole ordered response.

Definition: Let $Y_i$ be the ordered response for individual $i$, with categories $1,\ldots,K$ from lowest to highest, and let $x_i$ be that individual’s exposure. A cumulative-logit model with ordered thresholds $\alpha_1<\cdots<\alpha_{K-1}$ and shared exposure slope $\beta$ is $$ \operatorname{logit}\{P(Y_i\leq j\mid x_i)\}=\alpha_j-\beta x_i,\qquad j=1,\ldots,K-1, $$ where $\operatorname{logit}(p)=\log\{p/(1-p)\}$. The cutpoint-specific $\alpha_j$ values are thresholds; $\beta$ is the common exposure slope across all cumulative logits.

The sign convention above makes a positive $\beta$ correspond to a shift toward higher categories as exposure increases. In fact,

\operatorname{logit}\{P(Y_i>j\mid x_i)\}=\beta x_i-\alpha_j.

Thus, for every cutpoint $j$, the odds of being above rather than at or below that cutpoint are multiplied by $\exp(\beta)$ for a one-unit increase in exposure. This interpretation concerns cumulative odds, not the odds of occupying one particular category.

Thresholds, Exposure, and Model Assumptions

Each threshold locates a boundary on the cumulative-logit scale. At $x_i=0$, $\alpha_j$ is the log-odds of being in category $j$ or lower rather than above $j$. If exposure is centered, zero refers to the chosen centering value; if it is not centered, zero refers to zero exposure. Thresholds are not category-specific treatment effects.

The thresholds must increase with $j$. For an individual exposure value, the cumulative probabilities must satisfy $P(Y_i\leq 1\mid x_i)\leq P(Y_i\leq 2\mid x_i)\leq\cdots\leq P(Y_i\leq K-1\mid x_i)$. The inverse-logit function is increasing, so the model ensures this ordering when $\alpha_1<\cdots<\alpha_{K-1}$, because the same $\beta x_i$ is subtracted at every cutpoint. If thresholds cross, the fitted cumulative probabilities can violate their required ordering and no longer define a coherent ordinal distribution.

The shared-slope condition is often called the proportional-odds assumption. It requires the exposure slope to be the same at every division of the categories into “at or below $j$” and “above $j$.” It does not require the thresholds to be equal. The model also relies on response categories being mutually exclusive and exhaustive, exposure being defined consistently, and the observations contributing to the likelihood being conditionally independent given their modeled predictors. A repeated-response data set may violate that independence assumption; simply including multiple records for one person does not make them independent.

The response scale must genuinely be ordered. If categories cannot be ranked meaningfully, a cumulative model imposes an order that the endpoint does not support. Conversely, replacing a genuine ordinal scale with a binary endpoint discards distinctions among categories. These are modeling choices about the scientific outcome, not fixes for a model that fits poorly.

Worked Example: Specifying Thresholds and a Shared Exposure Slope

Suppose a hypothetical study records nausea after dosing as category 1 (none), category 2 (mild), category 3 (moderate), or category 4 (severe). Let $x_i$ be plasma concentration in mg/L. Write a cumulative-logit model and interpret its exposure slope if $\beta=0.70$ per mg/L.

There are $K=4$ ordered categories and therefore $K-1=3$ thresholds. The model is

\begin{aligned} \operatorname{logit}\{P(Y_i\leq 1\mid x_i)\}&=\alpha_1-0.70x_i,\\ \operatorname{logit}\{P(Y_i\leq 2\mid x_i)\}&=\alpha_2-0.70x_i,\\ \operatorname{logit}\{P(Y_i\leq 3\mid x_i)\}&=\alpha_3-0.70x_i, \end{aligned} \qquad \alpha_1<\alpha_2<\alpha_3.

The concentration slope is shared across the three cumulative logits. For a 1 mg/L increase, the cumulative odds of being above any one of the three cutpoints, rather than at or below it, are multiplied by $\exp(0.70)=2.014$. This does not say that the probability of severe nausea doubles: probabilities depend on the thresholds and the starting concentration as well as on the slope.

The units of $\beta$ are inverse mg/L, so that $\beta x_i$ is dimensionless. The thresholds are also dimensionless log-odds parameters. Before interpreting the slope this way, the analyst must regard the category order as meaningful and the common-slope restriction as scientifically and empirically tenable.

From Cumulative Probabilities to an Individual Likelihood

The likelihood needs the probability of the category actually observed, not just a cumulative probability. Write $F(z)=1/(1+\exp(-z))$ for the inverse-logit function. The model gives $P(Y_i\leq j\mid x_i)=F(\alpha_j-\beta x_i)$. The probability of an interior category is the probability of being at or below its upper boundary minus the probability of being at or below the preceding boundary. For the first and last categories, the corresponding cumulative boundary is zero or one.

Result: For a response $y_i=k$, define $\alpha_0=-\infty$ and $\alpha_K=+\infty$. Its category probability and individual likelihood contribution are $$ P(Y_i=k\mid x_i) =F(\alpha_k-\beta x_i)-F(\alpha_{k-1}-\beta x_i), \qquad k=1,\ldots,K, $$ using $F(-\infty)=0$ and $F(+\infty)=1$. For conditionally independent observations, the sample likelihood is the product of the individual observed-category probabilities.

This result follows directly from nesting: for an interior category $k$, the event $Y_i=k$ is the event $Y_i\leq k$ with the event $Y_i\leq k-1$ removed. The events are nested, so their probability difference is the probability of category $k$. At the bottom category, only $P(Y_i\leq 1\mid x_i)$ is needed; at the top category, the probability is one minus $P(Y_i\leq K-1\mid x_i)$. Conditional independence then gives the product rule for the joint probability of the observed responses.

Worked Example: One Individual’s Category Probability

Consider a hypothetical three-category endpoint, ordered from low to high, with thresholds $\alpha_1=-0.50$ and $\alpha_2=1.00$, and shared slope $\beta=0.80$ per mg/L. At exposure $x_i=0.50$ mg/L, calculate the probability that the observed response is category 2.

The threshold ordering holds because $-0.50<1.00$. The two cumulative logits at this exposure have linear predictors $-0.50-(0.80)(0.50)=-0.90$ and $1.00-(0.80)(0.50)=0.60$. Therefore,

\begin{aligned} P(Y_i=2\mid x_i=0.50) &=F(0.60)-F(-0.90)\\ &=\frac{1}{1+\exp(-0.60)}-\frac{1}{1+\exp(0.90)}\\ &=0.6457-0.2891\\ &=0.3566. \end{aligned}

The likelihood contribution is $0.3566$ if this individual’s observed category is 2. It is not the likelihood of a threshold or slope by itself; it is the probability of the observed response evaluated at the specified parameter values and exposure.

Worked Example: Likelihood for Several Observed Categories

Using the same hypothetical model, suppose three independent individuals have exposure-response pairs $(0\text{ mg/L},1)$, $(0.50\text{ mg/L},2)$, and $(1.00\text{ mg/L},3)$. Calculate the likelihood and log-likelihood at the specified thresholds and slope.

For the first observation, category 1 has probability $F(-0.50)=0.3775$. For the second, the category 2 probability was calculated as $0.3566$. For the third observation, category 3 is above category 2, so

\begin{aligned} P(Y_3=3\mid x_3=1.00) &=1-F\{1.00-(0.80)(1.00)\}\\ &=1-F(0.20)\\ &=1-0.5498\\ &=0.4502. \end{aligned}

Conditional independence makes the likelihood the product of the three observed-category probabilities:

\begin{aligned} L &=(0.3775)(0.3566)(0.4502)\\ &\approx 0.0606,\\ \ell &=\log(L)\\ &=\log(0.3775)+\log(0.3566)+\log(0.4502)\\ &\approx -0.974-1.031-0.798\\ &\approx -2.803. \end{aligned}

This is the likelihood and log-likelihood at one specified parameter set, not an estimate or a goodness-of-fit conclusion. Parameter estimation searches for threshold and slope values that make the observed categories comparatively likely. A single likelihood value has no standalone interpretation as the probability that the model is true.

Model Checks, Alternatives, and Limitations

A fitted model should be examined for both numerical and structural problems. Threshold estimates should remain ordered, and fitted cumulative probabilities should increase across cutpoints. Sparse or absent response categories can make some thresholds weakly informed or lead to numerical instability. Very large slope estimates or standard errors, failed convergence, and fitted probabilities that are nearly zero for observed categories are warnings to investigate the data and specification rather than automatic evidence for a particular remedy.

The shared-slope assumption can be considered by examining whether the exposure-response pattern appears broadly compatible across cumulative splits. Strongly different trends across splits suggest that one slope may be too restrictive. Formal assessment of that assumption is a distinct inferential task; a good overall fit at a few exposure values does not by itself establish that the common-slope restriction is appropriate. Likewise, dependence among repeated observations requires a model or analysis that accounts for the within-individual structure.

The main alternatives answer different modeling needs. A binary logistic model is appropriate when the scientific endpoint is a prespecified binary distinction, but collapsing categories loses information. A nominal multinomial model does not use category order and allows category-specific relationships, at the cost of not exploiting that order. Separate binary models for several cutpoints do not automatically form a coherent set of cumulative probabilities. A cumulative-logit model is attractive when order is meaningful and a shared exposure slope is plausible, but it may be too restrictive when exposure relates differently to different parts of the scale.

Common pitfall: Treating $\exp(\beta)$ as the odds ratio for each individual response category is incorrect. It is the multiplier in the cumulative odds of being above a cutpoint versus at or below it, and the same multiplier is imposed at every cutpoint. The probability of a particular category is a difference of two cumulative probabilities; its change need not have the same odds-ratio interpretation.

The model also does not make category labels equally spaced, establish a causal exposure effect, or account automatically for measurement error in exposure, omitted predictors, or subject-level heterogeneity. Its likelihood is only as appropriate as its response definition, dependence assumptions, and predictor structure. These limitations matter especially when translating an estimated association into clinical claims.

Check Your Understanding

Work each question using the stated cumulative-logit model. Solutions follow each question.

For five ordered categories, how many thresholds are needed, and what ordering must they satisfy?
Solution: There are $K-1=4$ thresholds, and they must satisfy $\alpha_1<\alpha_2<\alpha_3<\alpha_4$.
In the model $\operatorname{logit}\{P(Y\leq j\mid x)\}=\alpha_j-\beta x$, what does a positive $\beta$ imply for the cumulative odds of being above cutpoint $j$ when exposure increases by one unit?
Solution: Those odds are multiplied by $\exp(\beta)$, for every cutpoint $j$. This describes cumulative odds, not the odds or probability of one particular category.
For three categories, write the probability of category 2 in terms of the cumulative probabilities.
Solution: $P(Y=2\mid x)=P(Y\leq2\mid x)-P(Y\leq1\mid x)$, because category 2 is included in the first event but not the second.
Two independent individuals have observed-category probabilities $0.40$ and $0.25$ under a specified model. What is their joint likelihood contribution?
Solution: By conditional independence, it is $(0.40)(0.25)=0.100$.
Why is a shared slope not equivalent to saying that exposure has the same effect on the probability of every category?
Solution: The shared slope specifies the same exposure effect on each cumulative log-odds. Individual-category probabilities are differences between cumulative probabilities, so their exposure changes can differ in size or direction.

Cumulative Logit Models for Ordered Pharmacodynamic Outcomes

What You'll Learn