Modeling migraine severity with autoregressive ordered probit models

Longitudinal data with ordinal outcomes routinely appear in medical applications. An example is the analysis of clinical diaries where patients are asked to score the severity of their symptoms. In this framework, a class of dynamic models for ordinal repeated responses with subjectspecific random effects and distinguished correlation structures for different groups of patients is presented. Model specifications having marginal or conditional interpretations are investigated and compared. Since likelihood inference for the proposed class of models is computationally infeasible because of high dimensional intractable integrals, a pseudolikelihood approach to both inference and model selection is preferred. The methodology is motivated and confirmed by the analysis of a large longitudinal study on the determinants of migraine severity.