Résumé
Fitting binary and ordinal correlated responses is presently an important field of development<br />in epidemiology. The longitudinal study of disability in elderly people and the search for risk<br />factors of disability is a crucial issue for public health. In this context, we compared marginal<br />logistic models and random effects logistic models considering disability as a binary response<br />to illustrate the following aspects: choice of the covariance structure, impact of missing data,<br />importance of time-dependent covariates, and interpretation of the results. The random effects<br />model was used to compute a predictive score from a large number of risk factors available in<br />the Epidos cohort. The mixed ordinal logistic models were then described and compared. We<br />showed how they allow identifying differential effects of the factors on the stages of<br />disability.