BMD Modeling and Multiple Endpoints. L. Ryan, Harvard University, Biostatistics, DFCI, 44 Binney St., Boston, MA 02115
Although the broad principles of BMD calculation are clear, common aspects of non-cancer risk assessment (especially reproduction and development) raise technical statistical issues that must be addressed and solved before BDs can be widely used in practice. Perhaps the most challenging and interesting of these issues is that of multiple outcomes. To address this problem, several authors in recent years have proposed the use of multivariate modelling techniques. After briefly reviewing several available methods, we will discuss the advantages of these multivariate approaches over simpler, "naive" methods that classify each individual as normal or abnormal. Some extensive data analyses will be used to illustrate the arguments.