Modelling multilevel ordinal response under endogeneity: application to DTC patients’ outcome.

Ordinal data are widely used in medicine as they help to summarise health states or responses to treatment. Cumulative link models are useful because they explicitly consider the order in response’s categories, and they allow to model the probability of observing a value that does not exceed a specific category. In this contribution, two situations that often arise in real data analysis will be investigated: i) data with hierarchical structure; ii) presence of endogenous covariates. We describe this framework by discussing theapplication to the response to treatment at 3 years in patients with differentiated thyroid cancer (DTC) diagnosis. The application aims to evaluate potential predictors for the response at 3 years considering both repeated measurements from the same patient, and data from several clinical centres. The analysis of the response over time is of interest as it evaluates whether the improvement can be attributed to the time thatallows to obtain clearer eco images. Consequently, it could lead to a reduction in the use of radioactive iodine therapy, which is essential both to limit the side effects it produces on patients and to reduce the production of waste that are difficult to dispose of.