In this QCast episode, co-hosts Jullia and Tom unpack the proportional odds assumption in ordinal logistic regression. They explain what proportional odds means in practice, how it underpins the analysis of ordered endpoints, and why it matters for the credibility of trial conclusions. They walk through how the model works, practical ways to check whether the assumption is reasonable, what to do when it fails, and how to communicate results clearly to clinical and regulatory stakeholders.
The proportional odds assumption says that a predictor, such as treatment, has the same effect across all cut points of an ordered outcome. Instead of fitting separate models for each category, the proportional odds model uses cumulative logits and a single odds ratio that describes how treatment shifts patients towards better categories across the whole scale. When the assumption is reasonable, this offers an efficient, information rich summary of benefit that respects the ordering of the endpoint and supports clear interpretation in clinical trials.
In an ordinal setting, the model focuses on cumulative probabilities, such as being at or below each category boundary of a rating scale. These cumulative probabilities are transformed into log odds and expressed as an intercept for each threshold plus a shared slope for each predictor. Software then estimates the intercepts and common slopes, which can be transformed back into probabilities for each category. The single reported odds ratio reflects a constant treatment effect across all thresholds and is often used to summarise results for patient reported outcomes and severity scales.
Assessing whether proportional odds is plausible should be part of routine model diagnostics. Formal score tests compare the proportional odds model to a more flexible alternative and flag when the constrained structure may not fit well. Alongside tests, teams can fit a series of binary logistic regressions at different cut points of the scale and compare treatment odds ratios across these models. If those effects are broadly similar, the assumption looks reasonable; if they vary in ways that matter clinically, proportional odds may not be an appropriate simplification.
When diagnostics suggest that proportional odds does not hold, there are options beyond discarding the ordinal framework. Generalised ordinal models allow treatment effects to vary across thresholds, and partial proportional odds approaches relax the assumption only for selected variables while keeping others common. In some cases, it may be clinically sensible to combine adjacent categories on very granular scales to improve stability. If needed, teams can move to models such as multinomial logistic regression, accepting additional complexity in exchange for a more faithful description of how treatment affects different outcome levels.
For studies using ordered endpoints, it is helpful to plan the analysis strategy early and specify how proportional odds will be used and assessed. During analysis, do more than report a single odds ratio: review outcome distributions, run diagnostics, and examine how treatment effects behave across thresholds. When presenting results, explain what the odds ratio means in terms of shifting the overall distribution and support it with predicted probabilities for each category. Treating the proportional odds assumption as something to test, interpret, and, where necessary, adapt around leads to more transparent, reliable, and decision ready evidence.
Jullia
Welcome to QCast, the show where biometric expertise meets data-driven dialogue. I’m Jullia.
Tom
I’m Tom, and in each episode, we dive into the methodologies, case studies, regulatory shifts, and industry trends shaping modern drug development.
Jullia
Whether you’re in biotech, pharma or life sciences, we’re here to bring you practical insights straight from a leading biometrics CRO. Let’s get started.
Tom
So Jullia, today we are looking at the proportional odds assumption. It pops up in ordinal logistic regression outputs all the time, but many people are not fully comfortable with what it means. To begin, how would you define the proportional odds assumption in practical terms for someone working with ordered outcomes in a trial?
Jullia
So in practical terms, the proportional odds assumption says that a predictor has the same effect across all cut points of an ordered outcome.
Think of an endpoint with ordered categories such as mild, moderate and severe. In a proportional odds model we do not estimate separate treatment effects for each pair of categories. Instead, we model cumulative logits, such as the log odds of being mild or better versus worse, then mild or moderate versus severe, and so on.
The model includes several intercepts, one for each threshold between categories, but only one slope for each explanatory variable. That slope corresponds to a single odds ratio for treatment, which applies consistently across all those cumulative splits. On the log odds scale, the fitted lines for different thresholds are parallel, which is why people talk about “parallel slopes”.
So if treatment halves the odds of being in a worse category at one threshold, the assumption says it halves the odds at every threshold of that ordered scale.
Tom
That makes sense. Now. staying with that clinical lens, many trials now use ordinal outcomes like symptom scores or global impression scales. Why does this particular assumption matter so much when we are analysing those kinds of endpoints, from both a scientific and regulatory perspective?
Jullia
It matters because the assumption underpins how we summarise treatment benefit on an ordered scale.
When proportional odds holds, we can describe how treatment shifts patients towards better categories with a single odds ratio. That makes full use of the ordering and avoids throwing away information by collapsing the scale to a responder or non-responder analysis. Clinicians then have one coherent measure that summarises improvements across the entire endpoint.
From a regulatory perspective, current guidance expects that methods are appropriate to the endpoint and that key modelling assumptions are checked. If proportional odds are clearly violated but we still present a single pooled odds ratio as if it were valid everywhere on the scale, we risk misrepresenting the treatment effect.
So this isn’t just a technical detail. It affects whether the analysis reflects what’s happening across the full range of patient outcomes, which is what regulators and decision makers care about.
Tom
Let’s unpack the mechanics a bit more. People see several intercepts, one treatment coefficient and then predicted probabilities for each category, but the link between them can feel opaque. How is the model actually working behind the scenes, without us needing to write down equations?
Jullia
A helpful way to see it is to start from cumulative probabilities.
Suppose our endpoint has five ordered categories. Instead of modelling the probability of each category directly, we first consider the probability of being at or below each category boundary. So we look at the probability of being in category one, two, four, or better. The top category is implied, because the cumulative probability at the top must be one.
For each of those cumulative probabilities we compute log odds, which transforms them onto an unbounded scale. The model then expresses each cumulative log as an intercept for that boundary plus a common slope multiplied by the predictors, such as treatment group or baseline covariates.
When we fit the model, the software estimates that set of intercepts and the shared slopes. To obtain something more intuitive, we convert the fitted cumulative log odds back to cumulative probabilities, then subtract adjacent cumulative probabilities to get the probability for each specific category.
The proportional odds assumption is exactly the idea that the odds ratio for treatment, comparing groups at any chosen cut point of the scale, is the same. That single odds ratio is what people typically report, and it’s all driven by this cumulative structure.
Tom
Once you have the model in place, the next question is whether that proportionality really holds in your data. When you’re working on an analysis, what are the main tools or checks you use to assess whether the proportional odds assumption is reasonable?
Jullia
I like to combine a formal test with some simple, visual diagnostics.
Many software packages provide a score test for proportional odds. Under the null hypothesis, the model with common slopes fits as well as a more flexible model that allows each threshold to have its own coefficients. A small p value suggests that the constrained model may not be adequate.
However, I wouldn’t rely on that test alone. A practical check is to fit a series of standard logistic regressions that each represent a different cut point of the scale. For each cut, you define a binary outcome such as “at or below this category” and then estimate the treatment odds ratio. If those odds ratios are broadly similar across the different cut points, that supports the proportional odds assumption.
On the other hand, if treatment has a strong effect at one threshold and little or no effect at another, especially in a way that matters clinically, then the assumption may not hold. Looking at model based predicted probabilities for each category by treatment group can also help you see where any differences arise.
In the end, it’s a combination of statistical evidence and clinical judgement. Small deviations are inevitable in real data. What we’re really asking is whether the assumption is a reasonable simplification.
Tom
Suppose you’ve done those checks and there is clear evidence that proportional odds is not a good simplification. At that point, what options do you have, and how do you decide which way to go without losing the advantages of working with an ordered scale?
Jullia
So, the first option is to relax the assumption rather than abandon the framework altogether.
Generalised ordinal models allow treatment effects to vary across thresholds. In many implementations you can specify that some variables, such as treatment, have non proportional effects while others, such as baseline covariates, retain common slopes. That’s often called a partial proportional odds approach, and it lets you focus flexibility where it’s needed.
Another way is to review the outcome coding. If the scale is very granular, it might be sensible, based on clinical input, to combine adjacent categories that are rarely used or are difficult to distinguish in practice. Doing this in a pre-planned way can improve both estimation stability and the plausibility of proportional odds.
If these strategies don’t resolve the issue, you may need to consider alternative models, such as multinomial logistic regression that treats categories as nominal. This removes the proportional odds assumption but can require larger sample sizes and gives multiple effect estimates instead of a single summary.
Regardless, whichever route you choose, it’s important to describe the approach in the analysis plan where possible, document the diagnostics, and explain the rationale clearly when reporting results. Regulators are generally comfortable with flexible models when the reasoning is transparent.
Tom
So, for listeners working with ordinal endpoints in their trials, what would you highlight as the key practical takeaways around proportional odds, both in terms of what to do and what to avoid?
Jullia
I would focus on four points.
First, think about the ordinal endpoint at the design stage. Make sure the categories are clinically meaningful, ordered in a sensible way, and likely to be applied consistently by sites. If a proportional odds model will drive your main analysis, say so upfront and outline how you plan to assess the assumption.
Second, when you analyse the data, don’t just fit the model and report the odds ratio. Look at the distribution of the outcome by treatment group, use at least one formal test, and check how estimated treatment effects behave across different cut points of the scale.
Third, be precise in how you interpret the odds ratio. It describes how treatment shifts the overall distribution towards better categories, under the proportional odds assumption. It’s not the same as a risk ratio for a single responder definition, so mixing those concepts can cause confusion.
Fourth, if diagnostics suggest the assumption is doubtful, address it rather than ignoring it. That might mean using a partial proportional odds model, modestly adjusting the categorisation in a justified way, or moving to a different modelling family. The important thing is that the chosen method reflects the data and that the reasoning is clear to clinical and regulatory audiences.
Tom
To close off today’s episode, let’s give people a short recap. When someone sees an ordinal logistic regression table in a report, and they know proportional odds is involved, what are the two or three things you’d like them to keep in mind?
Jullia
I’d suggest they remember that proportional odds is a helpful simplification, not a guarantee.
When it holds, it provides an efficient way to use the full ordering of an endpoint and to summarise treatment benefit with a single, interpretable odds ratio. That’s a real strength in modern trials that rely on rating scales and patient reported outcomes.
But that strength depends on the assumption being roughly true. So they should look for evidence that the assumption has been checked, and for supporting information such as predicted category probabilities that make the treatment effect easier to understand. If those elements are missing, it’s reasonable to ask how robust the conclusions really are.
Treating the proportional odds assumption as something to examine and explain, rather than something to take for granted, will make analyses more transparent and more useful for decision making.
Jullia
With that, we’ve come to the end of today’s episode on the proportional odds assumption in clinical trials. If you found this discussion useful, don’t forget to subscribe to QCast so you never miss an episode and share it with a colleague. And if you’d like to learn more about how Quanticate supports data-driven solutions in clinical trials, head to our website or get in touch.
Tom
Thanks for tuning in, and we’ll see you in the next episode.
QCast by Quanticate is the podcast for biotech, pharma, and life science leaders looking to deepen their understanding of biometrics and modern drug development. Join co-hosts Tom and Jullia as they explore methodologies, case studies, regulatory shifts, and industry trends shaping the future of clinical research. Where biometric expertise meets data-driven dialogue, QCast delivers practical insights and thought leadership to inform your next breakthrough.
Subscribe to QCast on Apple Podcasts or Spotify to never miss an episode.