Mixed models for the analysis of categorical repeated measures
Biostatistical Centre, K.U.Leuven, Belgium
Whenever a response is measured repeatedly within a set of units (study-participants, animals, patients, ...)† correct statistical inference should account for the correlation of the repeated measurements within subjects. For continuous outcomes, the linear mixed model is nowadays well-established. For categorical outcomes, a variety of models has been proposed in the statistical literature. In this presentation, focuss will be on mixed-effects models, which can be viewed as a direct extension of the linear models for continuous outcomes to generalized linear or non-linear models. The models will be introduced, and estimation will be discussed. The various estimation methods will be explained and compared from an intuitive point of view. These include Laplace transformation, marginal and penalized quasi-likelihood, as well as adaptive and non-adaptive Gaussian quadrature. Advantages and disadvantages will be discussed, and all methods will be extensively illustrated using real data. Emphasis will also be on the interpretation of the parameters in the models. Illustrations will be based on data from a two-arm randomized longitudinal clinical trial as well as data from an animal toxicity study for the comparison of several doses of exposure.