|
Abstract:
|
Ideally, priors for Bayesian analysis should be based on the best available information about model parameters. In observational epidemiology, much prior information is structural, concerning functional dependencies of target parameters on other parameters or variables. This information is captured naturally in a hierarchical (multilevel) model for the observations. In bias modeling, variables determining the parameters include methodological features of the study, and the model is not identified. As a consequence, results are sensitive to the structure as well as to the prior distributions. Thus sensible results demand sensible structural input. An illustration is given in which unknown classification parameters are modeled as a function of the location and spread of a continuous classification criterion. The model is then applied to a study in which the key exposure indicator is latent.
|
