Why Relax the Proportional-Odds Assumption?
An ordinal pharmacodynamic endpoint might distinguish no response, partial response, and strong response. A cumulative-logit model represents the probabilities of being at or below successive category thresholds. Under the proportional-odds assumption, exposure has the same log-odds slope at every threshold. That compact model can be useful, but the earlier material on interpreting and testing the proportional-odds assumption establishes why a common slope may not describe every threshold adequately.
A partial proportional-odds model relaxes that restriction for selected predictors while retaining shared slopes for others. In exposure-response work, for example, exposure may have threshold-specific slopes while baseline covariates retain common slopes. This flexibility is not automatic permission to fit unrelated cumulative models: their cumulative probabilities must still be ordered so that the implied category probabilities are valid.
The sign convention matters. With this definition, a positive \(\beta_j\) means that increasing exposure decreases the probability of being at or below category \(j\), holding \(z\) fixed. Thus, it shifts probability toward higher categories at that threshold. Whether higher categories represent benefit or harm depends on how the endpoint is coded.
Ordering Thresholds and Preserving Valid Probabilities
The cumulative probabilities must satisfy \(F_1\leq F_2\leq\cdots\leq F_{K-1}\) for every modeled covariate value. Because the logistic function is strictly increasing, this is equivalent to requiring the cumulative-logit predictors \(\eta_j=\alpha_j-\beta_jx-\gamma^Tz\) to be ordered in the same way. Shared covariate terms cancel when adjacent predictors are compared.
If exposure is modeled only over a bounded interval \(x\in[L,U]\), each adjacent difference above is linear in \(x\). It is nonnegative throughout that interval if and only if it is nonnegative at both endpoints. This endpoint check is specific to a linear exposure effect. A nonlinear difference may attain its minimum inside the interval, so its full range must be checked.
Once the cumulative probabilities are ordered, category probabilities follow by subtraction:
These probabilities are nonnegative and sum to one. Nonnegativity follows from the ordering, and the sum telescopes to \(F_1+(F_2-F_1)+\cdots+(1-F_{K-1})=1\). If two adjacent cumulative probabilities are equal, their difference is zero; that is a boundary case of a valid distribution, not a crossing.
Worked Example: Checking a Model and Calculating Category Probabilities
Suppose a hypothetical four-category response is modeled over exposure \(x\in[0,2]\) mg/L, with \(z\) held fixed at its reference value. Let \((\alpha_1,\alpha_2,\alpha_3)=(-1,0.4,1.5)\) and \((\beta_1,\beta_2,\beta_3)=(0.3,0.8,0.5)\), where slopes are in L/mg. Check whether the model is valid throughout this exposure range, then calculate its category probabilities at \(x=1\) mg/L.
For the first adjacent pair, the difference in cumulative logits is \(1.4-0.5x\). Its endpoint values are \(1.4\) at \(x=0\) and \(0.4\) at \(x=2\), both nonnegative. For the second pair, the difference is \(1.1+0.3x\), with endpoint values \(1.1\) and \(1.7\). Both adjacent differences are positive throughout the range, so the cumulative probabilities are ordered.
At \(x=1\) mg/L, the three cumulative-logit predictors are \(-1-0.3=-1.3\), \(0.4-0.8=-0.4\), and \(1.5-0.5=1.0\). Applying the logistic function gives \(F_1=0.2142\), \(F_2=0.4013\), and \(F_3=0.7311\).
Therefore the four category probabilities are \(0.2142\), \(0.4013-0.2142=0.1871\), \(0.7311-0.4013=0.3298\), and \(1-0.7311=0.2689\). They are all nonnegative and sum to \(1.0000\), up to rounding.
Fitting and Interpreting the Partial Model
The parameters are typically estimated by maximum likelihood using the probability of each observed category. For independent observations conditional on their predictors, the likelihood is the product of those category probabilities. The exposure range and the restrictions on the \(\beta_j\) values should be specified as part of the model: a fit that crosses outside the observed range may still be useful within the intended range, but it should not be treated as valid for arbitrary extrapolation.
A practical strategy is to retain a common exposure slope at most thresholds and estimate separate slopes only where the proportional-odds assumption is not adequate and there is enough information to support the extra parameters. Enforce the ordering conditions during fitting over the intended exposure range. For linear exposure effects, these are linear inequality constraints in the parameters, evaluated at the two exposure endpoints. Do not fit unconstrained threshold-specific models and then assume their category probabilities will be coherent.
Specify the thresholds with separate exposure slopes and the thresholds sharing a slope. Keep the choice tied to the scientific question and available data.
State the interval over which predictions are intended. Centering exposure can make threshold intercepts easier to interpret, but the ordering constraints still have to hold over the full interval.
Constrain every adjacent cumulative-logit difference to be nonnegative throughout the domain. For linear exposure terms, check the two endpoints.
Calculate category probabilities over the exposure range, check for crossings or near-zero probabilities, and assess whether the extra slopes improve fit enough to justify their complexity.
Worked Example: Detecting and Repairing a Crossing
Consider a hypothetical three-threshold model for a four-category endpoint over \(x\in[0,2]\) mg/L. Let \((\alpha_1,\alpha_2,\alpha_3)=(-0.5,0.4,1.3)\) and initially \((\beta_1,\beta_2,\beta_3)=(0.2,0.8,0.5)\). The first adjacent predictor difference is \(0.9-0.6x\). At \(x=2\), it equals \(0.9-1.2=-0.3\), so the cumulative logits cross within the range. At that exposure, \(\eta_1=-0.9\) and \(\eta_2=-1.2\), which implies \(F_1>F_2\) and a negative probability for category 2.
To see the problem numerically, \(F_1=\operatorname{logistic}(-0.9)=0.2891\) and \(F_2=\operatorname{logistic}(-1.2)=0.2315\). Thus \(P(Y=2)=F_2-F_1=0.2315-0.2891=-0.0576\), which cannot be a probability.
Keeping these intercepts and \(\beta_1=0.2\), the first-pair condition at the upper endpoint requires \(0.9-(\beta_2-0.2)2\geq0\), or \(\beta_2\leq0.65\). Choose \(\beta_2=0.6\) and retain \(\beta_3=0.5\). The first gap is now \(0.9-0.4x\), which is \(0.1\) at \(x=2\). The second gap is \(0.9+0.1x\), which is positive at both endpoints. The model is therefore ordered over the full range.
At \(x=2\), the repaired predictors are \(-0.9\), \(-0.8\), and \(0.3\). Their cumulative probabilities are \(0.2891\), \(0.3100\), and \(0.5744\). The category probabilities are \(0.2891\), \(0.3100-0.2891=0.0209\), \(0.5744-0.3100=0.2644\), and \(1-0.5744=0.4256\). All are nonnegative and sum to \(1.0000\), up to rounding.
The constraint could instead be met at equality by setting \(\beta_2=0.65\): the first gap would then be zero at \(x=2\), giving category 2 probability zero there. That boundary distribution is valid, although it may be undesirable if the category should remain possible or if observed category-2 responses occur at that exposure.
Worked Example: A Likelihood Contribution
Use a hypothetical three-category model with \(F_j(x)=\operatorname{logistic}(\alpha_j-\beta_jx)\), \(\alpha_1=-1\), \(\alpha_2=1\), \(\beta_1=0.4\), and \(\beta_2=0.7\). Suppose \(x\) ranges from 0 to 2 mg/L. The adjacent logit gap is \(2-0.3x\), which is at least \(1.4\) on this range, so the model has ordered cumulative probabilities.
For three hypothetical independent records with exposures 0, 1, and 2 mg/L, suppose the observed categories are 1, 2, and 3, respectively. At \(x=0\), \(F_1=0.2689\) and \(F_2=0.7311\), so the probability of the observed category 1 is \(0.2689\). At \(x=1\), the cumulative logits are \(-1.4\) and \(0.3\), giving \(F_1=0.1978\) and \(F_2=0.5744\); the probability of category 2 is \(0.5744-0.1978=0.3766\). At \(x=2\), the cumulative logits are \(-1.8\) and \(-0.4\), giving \(F_2=0.4013\); the probability of category 3 is \(1-0.4013=0.5987\).
The likelihood contribution for these three observations is the product of their observed-category probabilities: \(L=0.2689(0.3766)(0.5987)\approx0.0606\). The log-likelihood contribution is approximately \(\log(L)=-2.803\). In an actual fit, the parameter values would be chosen to maximize the sum of log-likelihood contributions across the data, subject to the ordering constraints.
Model Comparisons, Diagnostics, and Limitations
| Model | Exposure slopes | Main trade-off |
|---|---|---|
| Proportional odds | One shared slope at every threshold | Parsimonious, but can impose an unsuitable common effect |
| Partial proportional odds | Shared at some thresholds and separate at selected thresholds | Targeted flexibility, with ordering constraints and added estimation uncertainty |
| Fully threshold-specific | A separate slope at every threshold | Most flexible, but uses more parameters and can produce crossing cumulative probabilities |
Useful checks include plotting each fitted \(F_j(x,z)\) and each category probability across the modeled exposure range, evaluating adjacent differences at the endpoints, and examining whether some predicted probabilities are extremely small. A negative category probability or a crossing is a model-validity failure, not a minor numerical irregularity. Very small probabilities can also signal weak support for a category, sparse data, or a model strained by its constraints. A fitted slope with a large standard error, or estimates that change markedly when a few observations are removed, can indicate that the data do not support the chosen degree of flexibility.
The proportional-odds model remains a reasonable alternative when a shared exposure effect is adequate; it is simpler and avoids threshold-specific slope parameters. A fully threshold-specific model can be considered when the data and scientific question justify that flexibility, but it still needs coherent ordering constraints. Model comparisons should account for added complexity, and a better fit alone does not establish that every threshold-specific slope is scientifically meaningful.
A partial proportional-odds model has important limits. Its conclusions depend on the chosen categories, exposure scale, covariates, and exposure domain. Linear slopes on the logit scale may not capture a nonlinear exposure-response relationship. Results should not be extrapolated beyond the domain over which ordering was checked. Finally, this model specification addresses the ordinal response probabilities; it does not by itself account for dependence between repeated observations from the same individual.
Common Mistake
- Wrong reasoning: “Every adjacent category probability must be strictly positive, so every cumulative-logit gap must be strictly positive at every exposure.”
- Correction: A valid distribution requires nonnegative category probabilities. Therefore cumulative probabilities need only be nondecreasing. Equality between adjacent cumulative probabilities gives a zero probability for the intervening category, which is valid, though strict ordering is needed if that category must have positive probability throughout the range.
- Another warning: Checking that the threshold intercepts are ordered at one reference exposure is not enough when slopes differ. Check the adjacent predictor differences across the full modeled exposure range.
Check Your Understanding
Answer each question, then compare your reasoning with the solutions.
- In the model \(\operatorname{logit}\{F_j\}=\alpha_j-\beta_jx-\gamma^Tz\), what does a positive \(\beta_j\) imply about \(F_j\) as exposure increases?
- For adjacent thresholds, let \(\alpha_{j+1}-\alpha_j=1.2\), \(\beta_{j+1}-\beta_j=0.4\), and \(x\in[0,2]\). Is the model ordered for this pair throughout the interval?
- If \(F_j=F_{j+1}\) at one exposure, what is the probability of category \(j+1\) there, and is the distribution invalid?
- Why is checking threshold intercept order alone insufficient when exposure slopes vary by threshold?
Solutions. (1) Since the cumulative-logit predictor contains \(-\beta_jx\), a positive \(\beta_j\) makes that predictor, and hence \(F_j\), decrease as \(x\) increases. (2) The gap is \(1.2-0.4x\); it equals \(1.2\) at \(x=0\) and \(0.4\) at \(x=2\), so it is positive throughout. (3) The category probability is \(F_{j+1}-F_j=0\). The distribution remains valid because nonnegative, not strictly positive, probabilities are required. (4) With different slopes, adjacent predictor differences change with exposure. They may become negative away from the reference exposure, so the differences must be checked over the full modeled range.