Interpreting Adjacent-Category Coefficients

From Adjacent Logits to Local Odds

An adjacent-category model describes comparisons between neighboring response categories. Its regression coefficients therefore have a direct local interpretation: they describe how a predictor changes the odds of one category relative to the category immediately below or above it. That interpretation is useful, but it is not the same as saying how the predictor changes the probability of either category. Probabilities must sum to one, so a change in one local odds can affect several category probabilities together.

Let an outcome have $K$ ordered categories, numbered $1,\ldots,K$, and let $\pi_j(x)=P(Y=j\mid x)$. We use the adjacent-logit convention in which the numerator is the probability of the higher-numbered category.

Definition: The adjacent-category logit for categories $j$ and $j+1$ is $$ \eta_j(x)=\log\left\{\frac{\pi_{j+1}(x)}{\pi_j(x)}\right\},\qquad j=1,\ldots,K-1. $$ In a scalar-covariate regression, the model is $\eta_j(x)=\alpha_j+\beta_jx$. The coefficient $\beta_j$ is the change in the log odds of category $j+1$ versus category $j$ for a one-unit increase in $x$, holding other predictors fixed.

This definition assumes the category probabilities being compared are positive. If the predictor is a vector $x$, the usual linear form is $\eta_j(x)=\alpha_j+\boldsymbol{\beta}_j^{\mathsf T}x$. The same interpretation applies to a contrast: for a change from $x$ to $x+\delta$, the log odds change by $\boldsymbol{\beta}_j^{\mathsf T}\delta$.

Interpreting a Coefficient as an Odds Ratio

Exponentiating the logit change expresses it as a ratio of odds. For a scalar predictor, compare $x$ with $x+\Delta$. Subtracting the two linear predictors gives $\beta_j\Delta$, so the neighboring-category odds at the new value are the old odds multiplied by $\exp(\beta_j\Delta)$. This is a conditional comparison: the other predictors in the model are held fixed.

Result: Under $\eta_j(x)=\alpha_j+\beta_jx$, $$ \frac{\pi_{j+1}(x+\Delta)/\pi_j(x+\Delta)} {\pi_{j+1}(x)/\pi_j(x)} =\exp(\beta_j\Delta). $$ Thus, for a one-unit increase, $\exp(\beta_j)$ is the multiplicative change in the odds of category $j+1$ versus category $j$. If $\beta_j>0$, those odds increase; if $\beta_j<0$, they decrease.

The odds ratio is not a probability ratio. For example, an odds ratio of 2 says that the odds for one neighboring comparison double; it does not say that the probability of the higher category doubles. Nor does it say that its probability necessarily increases. The probabilities depend on all the adjacent logits, not just the one being interpreted.

Worked Example: Interpreting a One-Unit Increase

Suppose there are three ordered categories and the adjacent logits are $\eta_1(x)=\log(1/2)+0.6x$ and $\eta_2(x)=\log(2)+0.6x$. Interpret each slope for a one-unit increase in $x$.

For categories 1 and 2, the odds $\pi_2/\pi_1$ are multiplied by $\exp(0.6)$, approximately $1.822$. For categories 2 and 3, the odds $\pi_3/\pi_2$ are also multiplied by approximately $1.822$. Thus each neighboring comparison shifts in favor of its higher-numbered category.

The probability calculation illustrates why this does not mean that every higher category’s probability increases. At $x=0$, set the weight for category 1 to 1. The adjacent odds then give weights $1,\frac12,1$, which normalize to probabilities $(0.4,0.2,0.4)$. At $x=1$, the weights are $1,\frac12e^{0.6},e^{1.2}$, or approximately $1,0.9111,3.3201$. Their sum is approximately $5.2312$, so the probabilities are approximately $(0.1912,0.1741,0.6347)$. The probability of category 3 increases, but that of category 2 decreases even though both adjacent-logit slopes are positive.

Why Coefficient Signs Do Not Determine Every Probability

The probabilities are normalized quantities: increasing the relative weight of one category changes the denominator shared by all categories. A useful way to make the dependence explicit is to express each probability using the adjacent logits. Take category 1’s unnormalized weight to be 1. Then category $k$’s weight is the product of the adjacent odds from category 1 up to category $k$.

w_1(x)=1,\qquad w_k(x)=\exp\left\{\sum_{j=1}^{k-1}\eta_j(x)\right\}\quad (k=2,\ldots,K), \qquad \pi_k(x)=\frac{w_k(x)}{\sum_{r=1}^K w_r(x)}.

This representation also yields a precise derivative result. Suppose $x$ is scalar, $K$ is finite, all category probabilities are positive, and $\eta_j(x)=\alpha_j+\beta_jx$. Define $s_1=0$ and $s_k=\sum_{j=1}^{k-1}\beta_j$ for $k=2,\ldots,K$. Each $s_k$ is the slope of the log weight for category $k$. Differentiating the normalized weights gives the following result.

Result: Under the stated linear adjacent-logit model, $$ \frac{d\pi_k(x)}{dx} =\pi_k(x)\left\{s_k-\sum_{r=1}^K\pi_r(x)s_r\right\}. $$ Therefore, the probability of category $k$ increases locally when its cumulative slope $s_k$ exceeds the probability-weighted mean of all the cumulative slopes. Its direction is not determined by any single adjacent coefficient alone.

To verify the result, write $S(x)=\sum_{r=1}^K w_r(x)$. Since $\log w_k(x)=\sum_{j<k}(\alpha_j+\beta_jx)$, differentiation gives $w_k'(x)=s_kw_k(x)$, including $w_1'(x)=0$. Hence $S'(x)=\sum_r s_rw_r(x)$. By the quotient rule,

\begin{aligned} \pi_k'(x) &=\frac{w_k'(x)S(x)-w_k(x)S'(x)}{S(x)^2}\\ &=\frac{w_k(x)}{S(x)} \left\{s_k-\frac{\sum_{r=1}^K s_rw_r(x)}{S(x)}\right\}\\ &=\pi_k(x)\left\{s_k-\sum_{r=1}^K\pi_r(x)s_r\right\}, \end{aligned}

where $w_r(x)/S(x)=\pi_r(x)$. This proves the result. In particular, when the $\beta_j$ are all positive, the cumulative slopes $s_k$ increase with $k$, but a middle category can still have a cumulative slope below their probability-weighted mean. Its probability can then decrease as $x$ increases. $\square$

Worked Example: A Positive Common Slope and a Falling Middle Probability

Let three categories have probabilities $(0.1,0.2,0.7)$ at $x=0$, and suppose both adjacent logits have slope $1.1$. At $x=0$, the adjacent odds are $\pi_2/\pi_1=2$ and $\pi_3/\pi_2=3.5$. Thus the model can be written as $\eta_1(x)=\log(2)+1.1x$ and $\eta_2(x)=\log(3.5)+1.1x$.

For a one-unit increase, each neighboring odds is multiplied by $e^{1.1}$, approximately $3.0042$. The category weights at $x=1$, relative to category 1, are $1$, $2e^{1.1}$, and $7e^{2.2}$. Numerically these are approximately $1$, $6.0083$, and $63.1751$; their sum is approximately $70.1834$. The resulting probabilities are approximately $(0.0142,0.0856,0.9002)$.

The probability of category 2 falls from $0.2$ to about $0.0856$, despite both coefficients being positive. The coefficients say that both higher-versus-lower neighboring odds increase. They do not say that each category’s probability increases.

Category-Specific Slopes and Local Comparisons

Adjacent-category models need not have a common slope. If the $\beta_j$ differ, each coefficient applies to its own neighboring pair. The sign of $\beta_j$ determines the direction of change in that pair’s odds as $x$ increases, but the probability response for a category depends on the cumulative slopes on both sides of it and on all the category probabilities.

Worked Example: Different Signs Across Adjacent Comparisons

Consider four categories with $\eta_1(x)=x$, $\eta_2(x)=-x$, and $\eta_3(x)=x$. At $x=0$, all adjacent odds equal 1, so the four category probabilities are each $1/4$. Increasing $x$ from 0 to 1 multiplies the odds for category 2 versus 1 by $e$, those for category 3 versus 2 by $e^{-1}$, and those for category 4 versus 3 by $e$.

At $x=1$, the weights are $1,e,e^0,e^1$, or $(1,e,1,e)$, because the log weight for category 3 is $\eta_1+\eta_2=0$ and for category 4 is $\eta_1+\eta_2+\eta_3=1$. The sum is $2+2e$, approximately $7.4366$. The probabilities are therefore approximately $(0.1345,0.3655,0.1345,0.3655)$.

The odds of category 3 versus category 2 decrease, since $\beta_2=-1$, while the probability of category 3 also decreases from $0.25$ to approximately $0.1345$. The odds of category 2 versus category 1 increase, but that does not isolate the probability change for category 2 from the other comparisons. Each coefficient is local; the normalized probability vector reflects all three changes together.

Common Mistakes and Practical Interpretation

A clear interpretation names both categories, the predictor contrast, and the odds ratio. For instance: “Holding the other predictors fixed, a one-unit increase in $x$ multiplies the odds of category 3 versus category 2 by $\exp(\beta_2)$.” This states exactly which comparison the coefficient concerns.

Calling an odds ratio a probability ratio. $\exp(\beta_j)$ multiplies neighboring-category odds, not the probability of category $j+1$.
Inferring a category-probability direction from one coefficient’s sign. The coefficient gives the direction of change for one adjacent odds. Calculate or derive the full probability response before making a claim about a category probability.
Reversing the category order in the comparison. With the convention used here, the logit is $\log\{\pi_{j+1}/\pi_j\}$. Reversing the numerator and denominator reverses the coefficient’s sign and its verbal interpretation.
Ignoring the size of a predictor contrast. For a change of $\Delta$ units, the odds multiplier is $\exp(\beta_j\Delta)$, not necessarily $\exp(\beta_j)$.

When the scientific question concerns category probabilities, report predicted probabilities at relevant covariate values, or their contrast, rather than translating an adjacent-logit coefficient directly into a probability claim. The odds interpretation remains valuable because it identifies the model’s local comparison; it simply answers a different question from the probability interpretation.

Check Your Understanding

Use the convention $\eta_j(x)=\log\{\pi_{j+1}(x)/\pi_j(x)\}$ in your answers.

If $\beta_2=-0.4$, what is the odds multiplier for category 3 versus category 2 for a two-unit increase in $x$?
Explain why a positive slope for every adjacent logit does not ensure that every category probability increases.
In a three-category model, write the cumulative slopes $s_1,s_2,s_3$ in terms of $\beta_1,\beta_2$.
What changes in the verbal interpretation if the adjacent logit is defined with $\pi_j/\pi_{j+1}$ instead?
What should be calculated if the scientific question asks whether category 2’s probability increases for a specified predictor change?