Why use a proportional-odds model?
An ordinal variable has categories with a meaningful order, but the distances between categories are not necessarily equal. For example, a symptom rating of “none,” “mild,” “moderate,” or “severe” is ordered, but it is not generally sensible to treat the move from none to mild as numerically identical to the move from moderate to severe. A cumulative-logit model respects the ordering without requiring equal spacing.
In MICE, the conditional model for an ordinal variable should reflect its variable type. The proportional-odds model is one such choice: it models the probability of being at or below each category boundary, while using the same predictor slopes at all boundaries. The thresholds differ, but the effect of a predictor on the cumulative logits is shared.
What the thresholds and shared slopes mean
For a threshold \(k\), the cumulative event is \(Y\leq k\): the response is in category \(k\) or below. Its complement is \(Y>k\). The threshold \(\alpha_k\) sets the cumulative log-odds when all predictors equal zero. With predictors present, the linear predictor shifts that log-odds. Thresholds are not category-specific regression slopes: they mark the successive boundaries in the response scale.
The “proportional odds” feature follows directly from the common slope. For a single predictor \(x\) with slope \(\beta\), compare two values \(x\) and \(x+1\), holding all other predictors fixed. At every threshold \(k\), the cumulative log-odds difference for \(Y\leq k\) is \(-\beta\). Thus, the ratio of the odds of being above the threshold to the odds of being at or below it is \(\exp(\beta)\), regardless of which threshold is chosen.
The same model also implies that the ratio of the odds of being at or below the threshold is \(\exp(-\beta)\). These statements are conditional on the other predictors in the model. An odds ratio describes a change in odds, not a change in probability, and it is not a causal effect unless the study design and assumptions support a causal interpretation.
Worked Example: Reading a Shared Slope Across Thresholds
Suppose, hypothetically, that a four-level symptom response is modeled using a predictor \(x\), measured in units where a one-unit increase has a practical interpretation. Consider the fitted cumulative logits $$ \operatorname{logit}\{P(Y\leq k\mid x)\}=\alpha_k-0.55x, \qquad (\alpha_1,\alpha_2,\alpha_3)=(-1.00,0.20,1.10). $$ Assume the response categories are correctly ordered and the common-slope form is an adequate working model.
At threshold \(k=2\) and \(x=2\), the cumulative logit is \(0.20-0.55(2)=-0.90\). The odds of being at or below category 2 rather than above it are therefore \(\exp(-0.90)=0.4066\), or about 0.407 to 1.
For a one-unit increase in \(x\), the odds of being above category 2 are multiplied by \(\exp(0.55)=1.733\). This is an increase of about 73.3% in those odds, conditional on the model’s other predictors. The same multiplier applies to being above category 1 and to being above category 3. The model does not say that the probability increases by 73.3 percentage points.
For the at-or-below event, the odds ratio is \(\exp(-0.55)=0.5769\). The reciprocal relationship is consistent: \(1/0.5769\approx1.733\). Because the model uses one slope, the threshold changes the baseline odds through \(\alpha_k\), not the multiplicative effect of \(x\).
Assumptions and model specification
The equation is useful only when its response structure and conditional model are appropriate. Before fitting it as an imputation model, check the following conditions.
- Ordered response categories. The \(K\) categories must have a substantively defensible order. If categories are merely distinct labels, an ordinal model imposes an order that is not warranted.
- Correct category coding. The numeric codes must follow the intended order. Reversing the coding reverses the direction of the interpretation.
- Ordered thresholds. The cutpoints must satisfy \(\alpha_1<\cdots<\alpha_{K-1}\). This reflects the nested events \(Y\leq1\), \(Y\leq2\), and so on.
- Common slopes across thresholds. A predictor has the same coefficient at each cumulative split. This is the proportional-odds restriction; it is substantive, not merely a convenient notation.
- Appropriate predictor form. A continuous predictor entered linearly is assumed to have a linear relationship with the cumulative log-odds. Nonlinear relationships or interactions require corresponding terms in the model.
- A suitable conditional specification. The predictors should include relevant information for the ordinal variable and for analyses that will use it. As with other imputation models, the model is a conditional model and does not by itself establish that missingness is ignorable.
Ordinal regression software estimates the thresholds and shared slopes together. The threshold estimates locate the cumulative boundaries, while each slope summarizes the predictor’s common association across them. In a MICE setting, the conditional model is fitted using cases where the ordinal variable is observed; the model must then support draws in its ordered categories. The present focus is the cumulative-logit specification and its coefficients, rather than the subsequent conversion of cumulative logits into category probabilities.
Worked Example: Comparing Cumulative Odds in Two Groups
Imagine a hypothetical survey with a four-category ordered rating and a binary predictor \(x\), where \(x=1\) denotes membership in a group and \(x=0\) denotes the comparison group. Suppose the model is $$ \operatorname{logit}\{P(Y\leq k\mid x)\}=\alpha_k-0.70x, \qquad (\alpha_1,\alpha_2,\alpha_3)=(-0.80,0.30,1.00). $$ Assume each response is independent conditional on the predictors, the categories are ordered as coded, and a common slope is a reasonable model for illustration.
At \(k=1\), the comparison group’s cumulative log-odds are \(-0.80\), so its odds of \(Y\leq1\) rather than \(Y>1\) are \(\exp(-0.80)=0.4493\). In the \(x=1\) group, the log-odds are \(-0.80-0.70=-1.50\), giving odds \(\exp(-1.50)=0.2231\). The ratio of these lower-tail odds is \(0.2231/0.4493=0.4966\), which is \(\exp(-0.70)\).
At \(k=2\), the comparison group’s odds of \(Y\leq2\) are \(\exp(0.30)=1.350\). The \(x=1\) group’s odds are \(\exp(0.30-0.70)=\exp(-0.40)=0.6703\). Their ratio is \(0.6703/1.350=0.4965\), again approximately \(\exp(-0.70)=0.4966\); the small difference is rounding.
Equivalently, at each threshold the odds of being above the boundary are multiplied by \(\exp(0.70)=2.014\) for \(x=1\) relative to \(x=0\). The calculation illustrates what is shared: the odds ratio across groups. The threshold-specific baseline odds are not the same.
Checking fit and recognizing failure modes
A fitted model can return coefficients even when the proportional-odds restriction or another part of the specification is poor. Diagnostics should therefore address both computation and the structure of the model.
- Check category order and counts. Confirm that the observed labels match the intended order and that categories are represented in the data. A coding error can make an otherwise stable fit misleading.
- Check convergence and estimates. Look for convergence warnings, extreme coefficients, unusually large standard errors, or threshold estimates that are not ordered. These may signal weak information, coding problems, or a model that is too complex for the observed data.
- Examine the shared-slope restriction. Compare observed patterns across cumulative splits or use an appropriate proportional-odds diagnostic. Strongly different predictor associations at different thresholds suggest that one common slope may be inadequate. The diagnostic is evidence about model fit, not proof that the restriction holds.
- Assess predictor form. If a continuous predictor’s association is curved on the cumulative-logit scale, a single linear term can misrepresent the conditional distribution. Consider a justified transformation, nonlinear term, or interaction rather than interpreting the linear coefficient literally.
- Check conditional predictions and imputation behavior. Implausible patterns across predictor values can indicate misspecification. In repeated imputation, also review whether the fitted model is stable across iterations and whether the imputed category distribution is credible in context.
A common mistake is to interpret \(\exp(\beta)\) as the odds ratio for one specific category, such as “severe” versus “moderate.” That is not what this coefficient represents. It is the common odds ratio for being above any cumulative threshold versus at or below that threshold. Category-specific odds are different quantities.
Another mistake is to regard the shared slope as automatically true because the model has one coefficient. The common slope is an assumption imposed by the model. If the association changes materially across cutpoints, the fitted coefficient may conceal that variation rather than resolve it.
How this model compares with alternatives
The proportional-odds model is parsimonious: it uses \(K-1\) thresholds and one slope per predictor, rather than estimating a different slope at every split. This can be attractive when the ordering is meaningful and a common association is a reasonable approximation. Its compactness also makes the coefficient interpretation consistent across thresholds.
| Approach | How it differs | When it may be considered |
|---|---|---|
| Proportional-odds cumulative logit | Uses ordered thresholds and shared predictor slopes. | When category order matters and the common-slope restriction is plausible. |
| Separate binary models at each cutpoint | Fits each cumulative split independently, so slopes can vary but fitted cumulative probabilities may not be coherent with one another. | As a diagnostic comparison, with care about the lack of a single coherent ordinal model. |
| Multinomial logistic regression | Treats categories as nominal and does not use their order in the same way. | When the category ordering is not defensible or an ordinal structure is unsuitable. |
| Other ordinal link or slope structures | Can alter the link or relax the shared-slope restriction. | When diagnostics or subject-matter reasoning indicate that the proportional-odds form is inadequate. |
No alternative is automatically superior. A less restrictive model may fit more flexibly but require more parameters and data. A proportional-odds model may be stable and interpretable yet still be inadequate if the common-slope restriction is badly contradicted. The imputation choice should reflect the response scale, available information, and intended analyses.
Check Your Understanding
Use the stated sign convention, \(\operatorname{logit}\{P(Y\leq k\mid x)\}=\alpha_k-\beta x\), to answer the questions. Brief solutions follow each problem.
- If \(\beta=0.40\), what is the multiplier for the odds of being above any threshold for a one-unit increase in \(x\)? Solution: \(\exp(0.40)=1.492\). Those conditional odds are multiplied by about 1.49 at every threshold.
- For the same coefficient, what is the multiplier for the odds of being at or below a threshold? Solution: \(\exp(-0.40)=0.6703\), the reciprocal of 1.492 up to rounding.
- Why are there \(K-1\) thresholds for \(K\) ordered categories? Solution: There is one boundary after each of the first \(K-1\) categories; the final category has no further boundary within the response scale.
- A colleague says a positive slope means the odds of being in category \(K\) rather than category \(K-1\) are multiplied by \(\exp(\beta)\). Is this interpretation correct? Solution: No. The coefficient describes the odds of being above a cumulative threshold versus at or below it, with the same odds ratio at each threshold. It is not a category-to-category odds ratio.
- What does a strongly different predictor association at different cumulative splits suggest? Solution: It may indicate that the shared-slope proportional-odds restriction is inadequate. A diagnostic and a suitable alternative ordinal specification should be considered.