Categorical Predictors and Reference Coding

Why Reference Coding Matters

A categorical predictor may distinguish treatment arms, hospitals, or exposure groups, but a cumulative-link model requires numerical predictors. Reference coding supplies those numbers without treating category labels as measurements: one level is designated as the reference, and each other level receives an indicator. The resulting coefficients compare each represented group with that reference in the model’s cumulative logits.

The choice of reference affects how parameters are described, not the comparisons the model can make. To interpret a coefficient correctly, keep track of the cumulative-logit sign convention and distinguish an odds ratio from a probability difference.

Reference Indicators in a Cumulative-Logit Model

Let the ordered response $Y$ have categories $1,\ldots,J$, and let a categorical predictor $G$ have levels $1,\ldots,K$. Choose level $r$ as the reference. For every nonreference level $g$, define an indicator $I_g$ that equals 1 when $G=g$ and 0 otherwise. The reference group has all of these indicators equal to 0.

Definition: Reference-indicator coding represents a $K$-level categorical predictor using $K-1$ indicators, one for each nonreference level. In the cumulative-logit parameterization used here, the model is $$ \log\left(\frac{P(Y\leq j\mid G)}{P(Y>j\mid G)}\right) =\theta_j-\eta(G),\qquad j=1,\ldots,J-1, $$ where $\theta_1<\cdots<\theta_{J-1}$ are thresholds and $$ \eta(G)=\sum_{g\ne r}\beta_g I_g. $$

This specification uses the cumulative-logit sign convention $\operatorname{logit}P(Y\leq j\mid G)=\theta_j-\eta(G)$. The same predictor contribution $\eta(G)$ appears at every cutpoint $j$. For the reference group, $\eta(G)=0$, so its cumulative logits are the thresholds $\theta_j$. For group $g\ne r$, the predictor contribution is $\beta_g$.

The coefficient $\beta_g$ is a contrast with the reference group on the cumulative-logit scale. Specifically, at any cutpoint $j$, the cumulative logit for group $g$ minus the cumulative logit for the reference is $-\beta_g$. Equivalently, the odds of being above cutpoint $j$, rather than at or below it, have ratio $e^{\beta_g}$ for group $g$ versus the reference. This odds ratio is the same at every cutpoint under this model.

Worked Example: Coding Four Treatment Groups

A study records an ordered clinical outcome and compares four treatment groups: Standard, A, B, and C. Use Standard as the reference group. Suppose the fitted predictor contribution is $$ \eta(G)=0.4I_A-0.7I_B+1.1I_C. $$ Write the contribution for each group and interpret the coefficients.

For Standard, all three indicators equal 0, so $\eta=0$. For A, $I_A=1$ and the other indicators are 0, so $\eta=0.4$. For B, $\eta=-0.7$, and for C, $\eta=1.1$.

At every cutpoint, the cumulative-logit contrast for A versus Standard is $-0.4$. Its cumulative odds $P(Y\leq j)/P(Y>j)$ are multiplied by $e^{-0.4}\approx0.6703$; equivalently, its odds of being above the cutpoint are multiplied by $e^{0.4}\approx1.4918$. For B, the cumulative-logit contrast is $0.7$, and the odds of being above the cutpoint are multiplied by $e^{-0.7}\approx0.4966$. For C, that odds ratio is $e^{1.1}\approx3.0042$. These are comparisons at each cutpoint, not claims that the category probabilities change by those same factors.

Two Consequences of the Coding

The coefficient interpretation follows directly by comparing two groups’ cumulative logits. The next result states the comparison precisely, including the ratio of cumulative odds and the equivalent odds for being above a cutpoint.

Theorem: Suppose the cumulative-logit model above holds, with finite thresholds and predictor contributions $\eta_a$ and $\eta_b$ for groups $a$ and $b$. At every cutpoint $j$, the cumulative-logit difference for group $a$ minus group $b$ is $-(\eta_a-\eta_b)$. The ratio of their cumulative odds $P(Y\leq j)/P(Y>j)$ is $e^{-(\eta_a-\eta_b)}$, and the ratio of their odds $P(Y>j)/P(Y\leq j)$ is $e^{\eta_a-\eta_b}$.

Proof. Write $C_{a,j}=P(Y\leq j\mid a)$ and $C_{b,j}=P(Y\leq j\mid b)$. By the model, $$ \operatorname{logit}(C_{a,j})=\theta_j-\eta_a,\qquad \operatorname{logit}(C_{b,j})=\theta_j-\eta_b. $$ Subtracting gives $\operatorname{logit}(C_{a,j})-\operatorname{logit}(C_{b,j})=-(\eta_a-\eta_b)$, since the same threshold $\theta_j$ occurs for both groups. Exponentiating this difference gives the ratio of cumulative odds. Taking the reciprocal gives the ratio of odds above the cutpoint. The calculation holds for every $j$, because the predictor contributions do not depend on $j$. $\square$

With reference coding, $\eta_r=0$ and $\eta_g=\beta_g$, so the theorem gives the coefficient interpretation stated above. It also handles comparisons between two nonreference groups: their predictor contrast is $\beta_a-\beta_b$, not either coefficient by itself.

Worked Example: From Coefficient to Fitted Probabilities

Suppose the response has three ordered categories: low, middle, and high. A binary predictor marks whether a patient received a new treatment, with no treatment as the reference. Let $\theta_1=-1$, $\theta_2=1$, and $\beta=\log 2$. Calculate the fitted category probabilities for each group.

For the reference group, $\eta=0$. The cumulative probabilities are $$ P(Y\leq1)=\frac{1}{1+e^{-(-1)}}\approx0.2689,\qquad P(Y\leq2)=\frac{1}{1+e^{-1}}\approx0.7311. $$ Thus $P(Y=1)=0.2689$, $P(Y=2)=0.7311-0.2689=0.4622$, and $P(Y=3)=1-0.7311=0.2689$, up to rounding.

For the treated group, $\eta=\log 2\approx0.6931$. Its cumulative logits are $-1-\log2\approx-1.6931$ and $1-\log2\approx0.3069$, giving $$ P(Y\leq1)\approx\frac{1}{1+e^{1.6931}}=0.1554,\qquad P(Y\leq2)\approx\frac{1}{1+e^{-0.3069}}=0.5761. $$ The category probabilities are therefore $0.1554$, $0.5761-0.1554=0.4207$, and $1-0.5761=0.4239$. They are nonnegative and sum to $1.0000$, allowing for rounding.

The coefficient also gives a direct odds comparison: at either cutpoint, the treated group’s odds of being above the cutpoint are multiplied by $e^\beta=2$. This does not mean that each category probability doubles; the calculated probabilities show why the odds interpretation must be kept distinct from a probability interpretation.

Changing the Reference Group

A different reference group produces different coefficient values because the contrasts are now taken against a different level. To preserve the same fitted cumulative logits, the thresholds must also be expressed on the new baseline. The following result gives the transformation.

Theorem: Let the original reference level be $r$, with group-specific predictor contributions $\beta_g$, where $\beta_r=0$. If level $s$ becomes the reference, define $\beta'_g=\beta_g-\beta_s$ for every group $g$, so $\beta'_s=0$, and define $\theta'_j=\theta_j-\beta_s$. Then $\theta'_j-\beta'_g=\theta_j-\beta_g$ for every group $g$ and cutpoint $j$. Hence the fitted cumulative probabilities are unchanged.

Proof. Under the new reference, the linear predictor for group $g$ is $\beta'_g=\beta_g-\beta_s$. Substituting the transformed parameters into the cumulative-logit equation gives $$ \theta'_j-\beta'_g =(\theta_j-\beta_s)-(\beta_g-\beta_s) =\theta_j-\beta_g. $$ This is exactly the original cumulative logit for group $g$ at cutpoint $j$. Equal logits give equal cumulative probabilities because the logistic function is one-to-one. Differences of successive cumulative probabilities therefore give equal category probabilities as well. $\square$

Worked Example: Re-Referencing a Three-Level Predictor

Suppose a predictor has levels A, B, and C. With A as the reference, let their contributions be $\beta_A=0$, $\beta_B=0.8$, and $\beta_C=-0.3$, and let the thresholds be $\theta_1=-1$ and $\theta_2=1$. Express the model using B as the reference.

Subtract B’s contribution, $0.8$, from every group contribution. The new contributions are $\beta'_A=0-0.8=-0.8$, $\beta'_B=0.8-0.8=0$, and $\beta'_C=-0.3-0.8=-1.1$. The new thresholds are $\theta'_1=-1-0.8=-1.8$ and $\theta'_2=1-0.8=0.2$.

Check the fitted logits for group C. With A as reference, they are $-1-(-0.3)=-0.7$ and $1-(-0.3)=1.3$. With B as reference, they are $-1.8-(-1.1)=-0.7$ and $0.2-(-1.1)=1.3$. The logits agree. For group B, both parameterizations give logits $-1.8$ and $0.2$. Thus the baseline and coefficient descriptions have changed, but the fitted predictions for every group have not.

Common Pitfalls and Interpretation

Reference coding is a parameterization of group contrasts, not a ranking of the predictor levels. The selected reference need not be the most common, untreated, or “best” group, although choosing a scientifically meaningful comparison can make results easier to communicate.

Reversing the sign interpretation. In this convention, a positive $\beta_g$ lowers the cumulative logit for $Y\leq j$ by $\beta_g$, and multiplies the odds of $Y>j$ by $e^{\beta_g}$. Always read the sign from the displayed model equation.
Calling an odds ratio a probability ratio. $e^{\beta_g}$ compares odds above a cutpoint; it does not multiply a category probability by $e^{\beta_g}$.
Comparing two nonreference coefficients as if either were the contrast. Their contrast is $\beta_a-\beta_b$, so the odds ratio for being above a cutpoint is $e^{\beta_a-\beta_b}$.
Assuming coefficients are unchanged when the reference changes. The new coefficients are contrasts with the new reference, and the threshold values shift as well. The fitted probabilities remain the same.

Reference coding matters because it makes model output interpretable in terms of explicit group comparisons. The estimates answer questions about relative cumulative odds, while the fitted probabilities show how those contrasts translate into the response categories in a particular model. Reporting the reference level and the sign convention is therefore essential.

Check Your Understanding

Use the convention $\operatorname{logit}P(Y\leq j)=\theta_j-\eta$ for these questions.

A predictor has four levels. If level C is the reference, how many reference indicators are used, and what values do they take for a subject in level C?
A nonreference group has coefficient $-0.5$. What is its odds ratio for $Y>j$ versus the reference group, at any cutpoint $j$?
Two nonreference groups have coefficients $0.6$ and $-0.2$. What is the odds ratio for being above a cutpoint for the first group versus the second?
Why does a coefficient’s odds ratio not give the ratio of the probabilities of one particular response category?
If the reference level has contribution 0.4 under the current parameterization, what must be subtracted from the thresholds when changing that level to the reference?