|
Starting from a brief introduction on the linear mixed model for continuous longitudinal data, extensions will be formulated to model outcomes of a categorical nature, including counts and binary data. Based on Verbeke and Molenberghs (2004), several families of models will be discussed and compared, from an interpretational as well as computational point of view. First, models will be discussed for the full marginal distribution of the outcome vector. This allows model fitting to be based on maximum likelihood principles, immediately implying inferential tools for all parameters in the models. The main disadvantage of such models is that they require complete specification of all higher-order interactions, which is often based on unrealistic assumptions, and often lead to computational problems, especially in examples with many repeated measurements per subject. Therefore, alternatives have been formulated in the statistical literature. First, following the reasoning in the linear mixed models, a full marginal model can be obtained from a random effects approach, where association between repeated measurements within the same subject is believed to be generated by underlying unobserved random effects. Alternatively, semiparametric methods can be used which do no longer require full specification of the likelihood, only of the first moments or of the first and second moments. This leads to the so-called generalized estimating equations. For both approaches, estimation and inference will be discussed and illustrated in full detail, and it will be extensively argued that both approaches yield parameters with completely different interpretations. Advantages and disadvantages of both will be discussed in full detail. Finally, when analysing longitudinal data, one is often confronted with missing observations, i.e., scheduled measurements have not been made, due to a variety of (known or unknown) reasons. It will be shown that, if no appropriate measures are taken, missing data can cause seriously biased results, and interpretational difficulties. Methods to properly analyze incomplete data, under flexible assumptions, are presented. Key concepts of sensitivity analysis are introduced. Without putting too much emphasis on software, some examples will be given on how the different approaches can be implemented within the SAS software package. Throughout the course, it will be assumed that the participants are familiar with basic statistical modelling, including linear models (regression and analysis of variance), as well as generalized linear models (logistic and Poisson regression). Moreover, pre-requisite knowledge should also include general estimation and testing theory (maximum likelihood, likelihood ratio). OPTIONAL TEXTBOOK AVAILABLE
|
= Applied Session,
= Theme Session,
= Presenter