Statistics Glossary → Proportional odds (parallel lines) assumption
Proportional odds (parallel lines) assumptionconcept
In ordinal logistic regression, the assumption that each predictor's effect (slope) is the same at every threshold of the outcome variable.
For an ordinal outcome with J categories (e.g. 'poor', 'fair', 'good', 'very good'), the cumulative logit model defines J - 1 thresholds: better than 'poor', better than 'fair or worse', and so on. The proportional odds model uses a separate intercept for each threshold but a single slope shared by all thresholds. This means a predictor's odds ratio is the same whichever threshold you look at; on the logit scale, the threshold curves are parallel.
The assumption is checked with the test of parallel lines: the proportional odds model is compared, by a likelihood-ratio (or Brant) test, with a general model in which the slopes may vary freely across thresholds. A significant result indicates a violation, but this test easily turns significant with large samples and many predictors and can be unreliable with sparse categories. When the assumption is violated, a partial proportional odds model that frees the slopes of only the problematic variables, a generalized ordinal model or multinomial logistic regression are possible alternatives.
An everyday example
A hotel satisfaction survey uses scores from 1 to 4. The proportional odds assumption holds that giving a guest a sea-view room raises both the odds of scoring '2 or higher rather than 1' and the odds of scoring '4 rather than 3 or lower' by the same factor (say, twofold). If the view rescues guests from poor scores but does not lift them to top marks, the effect differs across thresholds and the assumption is violated.
In MerQur
It is tested with the 'Test of parallel lines' extra output in Ordinal Regression; the 'Partial PO variables' option makes the slopes of the selected predictors threshold-specific.