Development of an Ordinal NMA Model Allowing for Testing of the Proportional Odds Assumption

OBJECTIVES: The NICE TSDs provide standardized code for probit ordinal network meta-analysis (NMA) models when outcomes are ordinal. However, these models present challenges in interpretation: (1) The treatment effect scale is unfamiliar to clinicians; (2) Testing the proportional odds assumption is difficult with current models; and (3) NICE TSD models borrow information from the control arm cumulative probabilities across trials. We describe the development and validation of a logistic ordinal NMA model allowing for the testing of the proportional odds assumption.

METHODS: The NICE TSD ordered probit was modified as follows: (1) Use of a logit link; (2) study-specific intercepts for the control arm in each study for all cumulative thresholds. (3) Treatment effects are included threshold. Models are compared via DIC. Complete data are simulated under models assuming proportional odds and under models not assuming proportional odds; results are compared to fitting separate binomials. A case study involving data from irritable bowel syndrome is used to assess the impact of the model in an applied scenario.

RESULTS: The new model aligns with separate binomial models while leveraging the correlation in endpoints and providing increased precision in cases where some studies report one but not the other endpoint. This is consistent with expectations based on other findings from multivariate meta-analysis. The new model also successfully identifies cases where the proportional odds assumption is violated, and the comparison of estimated effects from the two models provides a summary of the potential clinical relevance of that difference.

CONCLUSIONS: We developed a novel NMA model for ordinal outcomes that leads to valid tests of the violation of the proportional odds assumption while summarizing treatment effects on a more familiar scale. The new model offers advantages over separate binomial outcomes when some studies report only some criteria or when estimates are subsequently used in economic models.