John Carlin, University of Melbourne

Choice and Interpretation of Multilevel Models for Longitudinal Binary Outcome Data: A Case Study on Adolescent Smoking

Building statistical models to address questions about change over time in binary or ordinal outcomes is substantially more difficult than for continuous normally distributed outcomes. A large literature has grown up on the use of multilevel or hierarchical models for normal outcomes. Similar models are gaining increasing use for binary outcomes, but a number of questions remain about their effective application. For example, is it realistic to assume that heterogeneity between subjects can be adequately represented by a normally distributed random intercept (and/or slope)? What implications does this assumption have for substantive conclusions that may be drawn from estimated model parameters, and to what extent can the assumption be checked from the data? I will illustrate and discuss these questions in the context of a specific problem concerning smoking behavior in a longitudinal study of adolescent health, by comparing the results of fitting a standard logistic-normal model with those of fitting an alternative mixture model.