Proportional Odds Versus Baseline-Category Multinomial Logit

Two Ways to Model an Ordinal Response

An ordinal response has categories with a meaningful order, such as mild, moderate, and severe. A proportional-odds model uses that order directly by modeling cumulative probabilities. A baseline-category multinomial logit instead models each nonreference category against one chosen reference category. It treats the outcome as a set of categories and does not impose a shared slope across cumulative cutpoints.

The distinction affects more than coefficient interpretation. With the same predictors, the proportional-odds model generally uses fewer parameters, while the baseline-category model can accommodate category-specific changes. Both produce a probability for every response category, but they construct those probabilities differently.

Definition: Let $Y$ take ordered values $1,\ldots,K$, where $K\geq 2$, and let $x$ be a vector of $p$ predictors, not including a separate intercept. In the proportional-odds cumulative-logit model, $$ \operatorname{logit}\{P(Y\leq j\mid x)\}=\alpha_j-x^\mathsf{T}\beta,\qquad j=1,\ldots,K-1, $$ where $\alpha_1<\cdots<\alpha_{K-1}$ are thresholds and $\beta$ is shared across cutpoints. In a baseline-category multinomial-logit model, choose category $K$ as the reference and specify $$ \log\frac{P(Y=k\mid x)}{P(Y=K\mid x)} =\gamma_k+x^\mathsf{T}\delta_k,\qquad k=1,\ldots,K-1. $$ Each category $k$ has its own intercept $\gamma_k$ and slope vector $\delta_k$.

The sign convention in the cumulative model is deliberate: a positive component of $\beta$ lowers the modeled cumulative probabilities $P(Y\leq j\mid x)$ as the corresponding predictor increases. Equivalently, it moves probability toward higher response categories. In the baseline-category model, the sign of a slope describes a change in the odds of one particular category relative to category $K$; it does not by itself describe a common movement across the ordinal scale.

Parameter Counts and the Structure They Impose

Assuming the predictor design has rank $p$, and counting freely estimated regression parameters but not a dispersion parameter, the proportional-odds model has $K-1$ thresholds and $p$ shared slopes. Its parameter count is therefore $K-1+p$. The baseline-category model has $K-1$ intercepts and $p$ slopes for each of the $K-1$ nonreference categories, for a total of $(K-1)(p+1)$.

These counts describe the specified model, not a guarantee of estimability. The data must contain adequate information about the categories and predictors; for example, empty or very sparse categories can cause instability or separation. The thresholds in the proportional-odds model must also be ordered so its cumulative probabilities define admissible category probabilities.

Worked Example: Counting Parameters

Suppose the response has $K=4$ ordered categories and the model includes $p=3$ predictor columns. Compare the number of regression parameters.

The proportional-odds model has $K-1=3$ thresholds and 3 shared slopes, giving $3+3=6$ parameters.

The baseline-category model has 3 nonreference categories. Each has one intercept and 3 slopes, giving $3(1+3)=12$ parameters. Thus it has 6 more parameters in this example. The difference in general is $$ (K-1)(p+1)-(K-1+p)=p(K-2). $$ For $K=4,p=3$, this is $3(2)=6$, agreeing with the direct count. If $K=2$, the two formulations reduce to binary logistic regression and the counts coincide.

The smaller proportional-odds parameter count comes from a substantive restriction: the same slope vector is used at every cumulative cutpoint. In the baseline-category formulation, each nonreference category can have a different slope vector. More parameters can describe richer patterns, but they also require more information from the data and can lead to less precise estimates.

How Each Model Produces Category Probabilities

Write $F_j(x)=P(Y\leq j\mid x)$, and set $F_0(x)=0$, $F_K(x)=1$. The probability of category $j$ is $F_j(x)-F_{j-1}(x)$. In the proportional-odds model, the cumulative probabilities are obtained by applying the inverse logit to the threshold equations. Ordered thresholds ensure these cumulative probabilities increase with $j$, so their successive differences are nonnegative.

Result: If $\alpha_1<\cdots<\alpha_{K-1}$, then the proportional-odds equations define nonnegative category probabilities summing to one for every finite $x^\mathsf{T}\beta$. Indeed, the inverse-logit function is strictly increasing, so $F_1(x)<\cdots<F_{K-1}(x)$; taking successive differences, together with $F_0(x)=0$ and $F_K(x)=1$, gives nonnegative probabilities. Their sum telescopes: $$ \sum_{j=1}^K P(Y=j\mid x) =\sum_{j=1}^K\{F_j(x)-F_{j-1}(x)\} =F_K(x)-F_0(x)=1. $$

Worked Example: Probabilities From a Proportional-Odds Model

Let $Y$ have three ordered categories, with thresholds $\alpha_1=-1$, $\alpha_2=1$, and one predictor $x$ having slope $\beta=0.8$. Calculate the fitted probabilities at $x=0$ and $x=1$.

At $x=0$, the cumulative logits are $-1$ and $1$. Therefore $F_1(0)=\operatorname{logit}^{-1}(-1)=0.268941$ and $F_2(0)=\operatorname{logit}^{-1}(1)=0.731059$. Subtracting successive cumulative probabilities gives $P(Y=1\mid 0)=0.268941$, $P(Y=2\mid 0)=\operatorname{logit}^{-1}(1)-\operatorname{logit}^{-1}(-1)=0.462117$, and $P(Y=3\mid 0)=1-0.731059=0.268941$. These sum to $1.000000$, up to rounding.

At $x=1$, the cumulative logits are $-1-0.8=-1.8$ and $1-0.8=0.2$. Hence $F_1(1)=\operatorname{logit}^{-1}(-1.8)=0.141851$ and $F_2(1)=\operatorname{logit}^{-1}(0.2)=0.549834$. Thus the category probabilities are $0.141851$, $0.549834-0.141851=0.407983$, and $1-0.549834=0.450166$. They sum to $1.000000$, up to rounding. As $x$ increases, both cumulative probabilities decrease, reflecting a shift toward higher categories.

The baseline-category model first specifies the log odds against the reference category and then converts those odds to probabilities. This conversion is also called the softmax transformation.

Result: For any finite linear predictors $\eta_k(x)=\gamma_k+x^\mathsf{T}\delta_k$, $k=1,\ldots,K-1$, the baseline-category equations uniquely imply $$ P(Y=K\mid x)=\frac{1}{1+\sum_{r=1}^{K-1}\exp\{\eta_r(x)\}},\qquad P(Y=k\mid x)=\frac{\exp\{\eta_k(x)\}}{1+\sum_{r=1}^{K-1}\exp\{\eta_r(x)\}}. $$ To derive this, let $\pi_k=P(Y=k\mid x)$. The logit equations give $\pi_k=\pi_K\exp\{\eta_k(x)\}$. Substituting these expressions into $\sum_{k=1}^K\pi_k=1$ yields $\pi_K[1+\sum_{r=1}^{K-1}\exp\{\eta_r(x)\}]=1$, which gives the stated reference probability and then each other probability. All are positive and sum to one.

Worked Example: Probabilities From Baseline-Category Logits

For three categories, take category 3 as the reference. Let the fitted logits be $\log\{\pi_1(x)/\pi_3(x)\}=-0.4+0.9x$ and $\log\{\pi_2(x)/\pi_3(x)\}=0.2-0.3x$, where $\pi_k(x)=P(Y=k\mid x)$. Calculate the fitted probabilities at $x=0$ and $x=1$.

At $x=0$, the two nonreference logits are $-0.4$ and $0.2$. Their exponentials are $e^{-0.4}=0.670320046$ and $e^{0.2}=1.221402758$. The normalizing denominator, including the reference category, is $1+0.670320046+1.221402758=2.891722804$. Therefore $\pi_1(0)=0.670320046/2.891722804=0.231806467$, $\pi_2(0)=1.221402758/2.891722804=0.422378921$, and $\pi_3(0)=1/2.891722804=0.345814612$. The rounded probabilities sum to $1.000000$.

At $x=1$, the logits are $-0.4+0.9=0.5$ and $0.2-0.3=-0.1$. The exponentials are $e^{0.5}=1.648721271$ and $e^{-0.1}=0.904837418$, so the denominator is $1+1.648721271+0.904837418=3.553558689$. Consequently, $\pi_1(1)=1.648721271/3.553558689=0.463963427$, $\pi_2(1)=0.904837418/3.553558689=0.254628528$, and $\pi_3(1)=1/3.553558689=0.281408045$. Again, the probabilities sum to one, up to rounding.

What the Slopes Mean

The proportional-odds slope has a common cumulative interpretation. At cutpoint $j$, the odds of being in categories $1,\ldots,j$ rather than $j+1,\ldots,K$ are $\exp(\alpha_j-x^\mathsf{T}\beta)$. Increasing predictor $x_r$ by one unit, holding the other predictors fixed, multiplies these lower-versus-upper cumulative odds by $\exp(-\beta_r)$ at every cutpoint. Equivalently, the upper-versus-lower odds are multiplied by $\exp(\beta_r)$ at every cutpoint. That common multiplier is the proportional-odds assumption.

In the baseline-category model, increasing $x_r$ by one unit multiplies the odds of category $k$ versus the reference category $K$ by $\exp(\delta_{kr})$. This interpretation is specific to that category-reference comparison. It does not imply that the same multiplier applies to the odds of being above each ordinal cutpoint.

Worked Example: Comparing Slope Interpretations

In the proportional-odds example, $\beta=0.8$. A one-unit increase in $x$ multiplies the odds of being above any fixed cutpoint, rather than at or below it, by $e^{0.8}=2.2255$ (approximately). The multiplier is the same for the split $1$ versus $2,3$ and for the split $1,2$ versus $3$.

In the baseline-category example, the odds of category 1 versus category 3 are multiplied by $e^{0.9}=2.4596$ for a one-unit increase, while the odds of category 2 versus category 3 are multiplied by $e^{-0.3}=0.7408$. These effects differ because they concern separate category-reference comparisons. The two sets of coefficient values are illustrative parameterizations, not estimates from one common data set; their interpretations show the different restrictions each model expresses.

When the Multinomial Model Is a Useful Benchmark

The baseline-category multinomial logit is useful as a flexible benchmark when the ordinal structure is scientifically meaningful but it is not clear that a shared cumulative effect is adequate. Its category-specific slopes allow the association between a predictor and one response category versus the reference to differ from the association for another category. Comparing its fitted category probabilities with those from a proportional-odds fit can reveal where the more constrained model produces notably different predictions.

The multinomial model does not automatically respect the ordering in the outcome. It guarantees valid category probabilities, but its coefficient structure does not require a predictor to move probability consistently toward higher (or lower) categories. Conversely, proportional odds is not simply a baseline-category model with the same coefficients: the former constrains cumulative logits, while the latter constrains logits against one reference category. Their probability formulas differ, and the model families are not generally nested under these linear-predictor specifications.

A multinomial fit can also be a poor practical benchmark when the sample is small relative to the number of categories and predictors. Its larger parameter count may yield unstable estimates or wide uncertainty, particularly when some categories are rare. A useful comparison therefore considers the scientific meaning of the restrictions, the adequacy of the data, and the fitted probabilities—not only which model has more flexibility.

Common Mistake

Wrong reasoning: “The multinomial model has separate coefficients, so each coefficient tells me how the predictor shifts the ordinal response upward.” A category-specific coefficient only describes odds for that category against the chosen reference. It does not directly describe a cumulative shift.
Correction: For a common upward or downward effect across ordinal cutpoints, interpret the cumulative odds in the proportional-odds model. For the baseline-category model, state the reference category and interpret each slope only as a change in the corresponding category-versus-reference odds.

Key takeaway: Proportional odds uses $K-1+p$ parameters and a shared cumulative-logit slope vector; baseline-category multinomial logit uses $(K-1)(p+1)$ parameters and category-specific reference logits. The multinomial model is a flexible comparison, but its flexibility does not encode the response ordering.

Check Your Understanding

Use the definitions and calculations above to answer these questions.

For $K=5$ categories and $p=2$ predictor columns, how many regression parameters does each model have?
In a proportional-odds model with $\beta_r>0$, what happens to the upper-versus-lower cumulative odds when predictor $x_r$ increases by one unit?
In a baseline-category model, what comparison does the coefficient $\delta_{kr}$ describe?
Why does the baseline-category model produce probabilities that sum to one even though it specifies only $K-1$ logits?
Give one reason a multinomial-logit fit may be an informative benchmark and one reason its estimates may be unstable.