In simulated two-arm randomised controlled trials evaluating ordinal outcomes, unconstrained and constrained partial proportional odds models, alongside separate logistic regression models, provide unbiased and efficient parameter estimates under non-proportional odds conditions, according to a Monte Carlo simulation study published in BMC Medical Research Methodology.
The proportional odds model represents the most common analytic approach for ordinal outcomes within randomised controlled trials. Although parameter estimates generated under departures from proportional odds can be interpreted as an average odds ratio, they can simultaneously obscure differing treatment effects that occur across the distribution of individual ordinal categories.
To evaluate potential extensions to standard modeling when deviations from the proportional odds assumption arise, researchers Chris J. Selman, Katherine J. Lee, Steven Y. C. Tong, Mark Jones, and Robert K. Mahar assessed four distinct analytical approaches via Monte Carlo simulation. Specifically, the authors fitted independent logistic regression models, the standard proportional odds model, constrained partial proportional odds models, and unconstrained partial proportional odds models to simulated ordinal outcome datasets.
The investigation evaluated model performance by measuring statistical bias, coverage, and mean square error across simulated ordinal outcome data representing a hypothetical two-arm randomised trial across multiple scenarios. As a demonstrative case study grounding the methodological evaluation, the research team drew upon the Australasian COVID-19 Trial, which had adopted multiple secondary ordinal endpoints.
The simulation results demonstrated that the standard proportional odds model performed best when outcome data were generated under proportional odds conditions. However, when non-proportional odds were present, the proportional odds model resulted in bias and poor statistical coverage, with performance deteriorating further as the effect size and the number of categories increased.
Conversely, odds ratios estimated using the unconstrained partial proportional odds model and separate logistic regression models exhibited negligible bias and good coverage across most non-proportional odds scenarios. The authors noted a key performance limitation: the unconstrained partial proportional odds model under-performed in scenarios with sparse data within certain categories, confirming that while proportional odds models remain effective when assumptions hold, constrained and unconstrained partial proportional odds and logistic regression models serve as reliable alternatives under departures from proportionality.
Subject of Research: Performance of Bayesian cumulative logistic models for ordinal trial outcomes
Article Title: Evaluating the performance of Bayesian cumulative logistic models in randomised controlled trials: a simulation study
Article References: Selman, C. J., Lee, K. J., Tong, S. Y. C., Jones, M., & Mahar, R. K. (2026). Evaluating the performance of Bayesian cumulative logistic models in randomised controlled trials: a simulation study. BMC Medical Research Methodology. https://doi.org/10.1186/s12874-026-03004-x
Image Credits: AI Generated
DOI: 10.1186/s12874-026-03004-x
Keywords: proportional odds model, partial proportional odds, ordinal outcomes, randomised controlled trials, Monte Carlo simulation, Bayesian cumulative logistic models, logistic regression, treatment effects, unconstrained PPO model, constrained PPO model, bias, mean square error, coverage, odds ratios, Australasian COVID-19 Trial, sparse data, statistical methodology, clinical trial endpoints
News Source: Ophelia Keating. (November 11, 2026). Alternative Ordinal Models Provide Unbiased Estimates When Proportional Odds Assumptions Fail. Scienmag.



