Abstract

Longitudinal data, where there are multiple repeated observations of an outcome variable for each individual, requires appropriate analysis to account for the non-independence of observations within individuals. A variety of methods are available, including random-effects (aka multi-level or mixed) models and marginal (GEE) models. Random-effects models, the focus of this presentation, extend the standard regression model to allow for subject-specific estimates of the baseline level of the outcome variable (random intercepts) and its rate of change over time (random slopes). Under certain circumstances both random-effects and marginal model approaches give equivalent estimates. However, in many situations they will not and it can be difficult to decide whether one approach should be favoured over the other.

This presentation will provide an introduction to the random-effects model for longitudinal data using data from the Early Rheumatoid Arthritis Study. Particular strengths of random-effects models such as how missing data on outcome variables at some assessments are handled and the clustering of individuals within hospitals will be considered. Furthermore, differences in the interpretation of coefficients between random-effects and marginal models where the outcome variable is dichotomous will be discussed, with guidance on situations where each approach is more useful.