Adjacent-Category Logits for Ordered Labels

Why Adjacent Categories Need an Explicit Direction

An ordered response has categories with a meaningful order, such as mild, moderate, and severe. A comparison of neighboring categories asks a focused question: how much more probable is one category than the next one? The answer depends on two choices that should never be left implicit: the order of the labels and which category is placed in the numerator.

An adjacent-category logit is a log ratio of probabilities for two neighboring categories. It is a contrast between those two categories, not a probability and not, by itself, a regression model. When covariates are introduced, a model may specify how these logits vary with covariates; the definition of each logit comes first.

Definition: Let $Y$ have $K$ ordered categories, written $1<2<\cdots<K$, and let $\pi_j(x)=P(Y=j\mid x)$ be the probability of category $j$ at a given covariate value $x$. For $j=1,\ldots,K-1$, the forward adjacent-category logit is $$ \lambda_j(x)=\log\left\{\frac{\pi_{j+1}(x)}{\pi_j(x)}\right\}. $$ The numerator is the higher of the two neighboring categories, and the denominator is the lower. The corresponding adjacent odds are $\pi_{j+1}(x)/\pi_j(x)=\exp\{\lambda_j(x)\}$.

The definition requires $\pi_j(x)>0$ and $\pi_{j+1}(x)>0$ for a finite logit. If one probability is zero and the other is positive, the log ratio is interpreted as an infinite limit; if both are zero, the ratio is undefined. In applications, a finite adjacent logit therefore refers to a pair of categories with positive probabilities.

The sign is determined by the stated numerator and denominator. A positive $\lambda_j(x)$ means category $j+1$ is more probable than category $j$; a negative value means category $j$ is more probable. A value of zero means they are equally probable. These statements compare the two probabilities, not either probability with the sum of all category probabilities.

What the Log Ratio Says

Exponentiating an adjacent-category logit returns the odds ratio of the two category probabilities. For example, $\lambda_j(x)=\log(2)$ means that category $j+1$ is twice as probable as category $j$. This does not say that the probability of category $j+1$ is $2$, or that it is twice the probability of all other categories combined.

There is also a conditional-odds interpretation. Restrict attention to observations in either category $j$ or category $j+1$. The probability of category $j+1$ conditional on being in this pair is $\pi_{j+1}/(\pi_j+\pi_{j+1})$; the corresponding conditional probability for category $j$ is $\pi_j/(\pi_j+\pi_{j+1})$. Their odds equal the original adjacent odds.

Theorem: If $\pi_j(x)>0$ and $\pi_{j+1}(x)>0$, then the odds of category $j+1$ versus category $j$, conditional on $Y\in\{j,j+1\}$, equal $\exp\{\lambda_j(x)\}$.

Proof. By the definition of conditional probability, $$ \frac{P(Y=j+1\mid Y\in\{j,j+1\},x)} {P(Y=j\mid Y\in\{j,j+1\},x)} = \frac{\pi_{j+1}(x)/\{\pi_j(x)+\pi_{j+1}(x)\}} {\pi_j(x)/\{\pi_j(x)+\pi_{j+1}(x)\}} = \frac{\pi_{j+1}(x)}{\pi_j(x)} = \exp\{\lambda_j(x)\}. $$ The common conditioning denominator cancels, proving the result. $\square$

Worked Example: Reading the Signs Across Four Ordered Categories

Suppose an ordered response has categories 1, 2, 3, and 4, with probabilities $0.10, 0.25, 0.40,$ and $0.25$, respectively. Calculate the forward adjacent-category logits and interpret their signs.

For categories 1 and 2, $$ \lambda_1=\log(0.25/0.10)=\log(2.5)\approx 0.9163. $$ The odds of category 2 rather than category 1 are $2.5$, so category 2 is more probable.

For categories 2 and 3, $$ \lambda_2=\log(0.40/0.25)=\log(1.6)\approx 0.4700. $$ The odds of category 3 rather than category 2 are $1.6$, again favoring the higher category.

For categories 3 and 4, $$ \lambda_3=\log(0.25/0.40)=\log(0.625)\approx -0.4700. $$ The odds of category 4 rather than category 3 are $0.625$, so category 3 is more probable. The three signs describe three local comparisons; they do not imply that probability increases at every step through the categories.

Ordering and Reference Direction

A numerical code is not enough to define an ordinal comparison: the analyst must say what the codes mean. If category 1 is “mild” and category 3 is “severe,” the forward logit compares the more severe neighbor with the less severe neighbor. In a different scientific setting, the chosen order might run from best to worst, or from least exposure to most exposure. The meaning of a positive logit changes with that choice.

For a fixed category order, one could instead use the reverse logit $\log\{\pi_j(x)/\pi_{j+1}(x)\}$. It compares the same pair but reverses numerator and denominator, so its value is the negative of the forward logit. This is a convention change, not a change in the underlying probabilities. Stating the direction prevents readers from mistaking a negative sign for evidence favoring the opposite outcome when the convention is simply reversed.

Theorem: For every adjacent pair with positive probabilities, the reverse-direction logit is the negative of the forward-direction logit. If the entire category order is reversed, the adjacent logits appear in reverse sequence with their signs changed.

Proof. For a single pair, the reverse logit is $$ \log\left\{\frac{\pi_j(x)}{\pi_{j+1}(x)}\right\} = -\log\left\{\frac{\pi_{j+1}(x)}{\pi_j(x)}\right\} = -\lambda_j(x). $$ Now reverse the full order, so the new category $r$ is the old category $K+1-r$. The new forward logit at position $r$, for $r=1,\ldots,K-1$, is $$ \log\left\{\frac{\pi_{K-r}(x)}{\pi_{K+1-r}(x)}\right\} = -\lambda_{K-r}(x). $$ Thus the sequence is reversed and each sign changes. $\square$

Worked Example: Reversing Mild, Moderate, and Severe

Suppose $P(Y=\text{mild})=0.50$, $P(Y=\text{moderate})=0.30$, and $P(Y=\text{severe})=0.20$. With the order mild $<$ moderate $<$ severe and the higher category in the numerator, the logits are $$ \lambda_1=\log(0.30/0.50)=\log(0.6)\approx -0.5108,\qquad \lambda_2=\log(0.20/0.30)=\log(2/3)\approx -0.4055. $$ Both are negative because each more severe category is less probable than its neighbor below.

Now write the same labels in reverse order: severe $<$ moderate $<$ mild. The first new logit compares moderate with severe, and the second compares mild with moderate: $$ \lambda'_1=\log(0.30/0.20)=\log(1.5)\approx 0.4055=-\lambda_2, \qquad \lambda'_2=\log(0.50/0.30)=\log(5/3)\approx 0.5108=-\lambda_1. $$ The probabilities have not changed. Only the order and thus the stated direction of comparison have changed.

Adjacent Comparisons and More Distant Categories

Although an adjacent-category logit directly compares only neighbors, a sequence of adjacent odds can be combined to compare categories farther apart. This fact is an algebraic consequence of the ratios; it does not turn any one adjacent logit into a direct comparison with every other category.

Result: If all probabilities $\pi_r(x)$ for $r=a,\ldots,b$ are positive, where $a<b$, then $$ \frac{\pi_b(x)}{\pi_a(x)} = \prod_{r=a}^{b-1}\exp\{\lambda_r(x)\}, \qquad \log\left\{\frac{\pi_b(x)}{\pi_a(x)}\right\} = \sum_{r=a}^{b-1}\lambda_r(x). $$

Proof. By definition, $\exp\{\lambda_r(x)\}=\pi_{r+1}(x)/\pi_r(x)$. Multiplying these equalities from $r=a$ through $r=b-1$ gives $$ \prod_{r=a}^{b-1}\exp\{\lambda_r(x)\} = \frac{\pi_{a+1}(x)}{\pi_a(x)} \frac{\pi_{a+2}(x)}{\pi_{a+1}(x)} \cdots \frac{\pi_b(x)}{\pi_{b-1}(x)} = \frac{\pi_b(x)}{\pi_a(x)}. $$ Every intermediate probability cancels. Taking logarithms gives the second equality. $\square$

Worked Example: Combining Adjacent Odds Across Four Stages

Suppose the ordered probabilities for stages 1 through 4 are $0.20, 0.50, 0.20,$ and $0.10$. The three adjacent odds, in the forward direction, are $$ \frac{\pi_2}{\pi_1}=\frac{0.50}{0.20}=2.5,\qquad \frac{\pi_3}{\pi_2}=\frac{0.20}{0.50}=0.4,\qquad \frac{\pi_4}{\pi_3}=\frac{0.10}{0.20}=0.5. $$ Their product is $2.5(0.4)(0.5)=0.5$. Directly, the odds of stage 4 versus stage 1 are $\pi_4/\pi_1=0.10/0.20=0.5$, as required. On the log scale, the adjacent logits are $\log(2.5)$, $\log(0.4)$, and $\log(0.5)$; their sum is $\log(0.5)$, the log odds of stage 4 versus stage 1.

Common Mistakes and Practical Checks

Adjacent-category logits are useful when scientific questions concern transitions or comparisons between neighboring ordered labels. They retain the local structure of the ordering: the comparison of categories 2 and 3 is not defined by a reference category at either endpoint of the whole scale. That makes the ordering itself part of the analysis, rather than a cosmetic choice of labels.

Leaving the numerator direction unstated. A logit can be written in either direction. Name the numerator and denominator, or state that the higher category is in the numerator.
Calling a logit a probability. A logit is a logarithm of a probability ratio. Exponentiating gives the ratio, not the probability of the numerator category.
Treating category codes as measurements. The labels identify an order; adjacent logits do not assume equal numerical spacing between categories.
Reading all signs as one trend. Each logit describes its own neighboring pair. The signs can differ across the ordered scale.
Using a nonadjacent comparison as if it were one local logit. A comparison of distant categories involves the product of intervening adjacent odds, or equivalently the sum of intervening logits.

Before interpreting an adjacent-category logit, verify the order, identify the two labels it compares, check which one is in the numerator, and then translate its sign or exponentiated value into that context. These steps keep the mathematical contrast aligned with the scientific question.

Check Your Understanding

Use the stated forward convention: the higher category is in the numerator.

For probabilities $0.60$ and $0.40$ in neighboring categories 1 and 2, calculate the forward adjacent-category logit and interpret its sign.
If a reverse-direction logit for a pair is $0.7$, what is the forward-direction logit for the same pair?
Explain why an adjacent-category logit of zero does not mean both category probabilities are zero.
For four ordered categories, express the log odds of category 4 versus category 1 as a sum of adjacent logits.
When the full category order is reversed, what happens to the order and signs of the forward adjacent logits?