The following investigator is involved in two Broadly Applicable Statistical Methods projects: David Dunson.
Density regression: Traditional statistical regression models focus on describing how the center of a response distribution changes with predictors. Such methods can produce misleading results when the shape of the response distribution changes with environmental and genetic factors. Such changes are a natural consequence of gene by environment interaction and biological constraints on the range of a health response. The Branch developed semiparametric Bayesian methods for density regression, using generalizations of Dirichlet process mixture models.
Bayesian meta-analysis: In important applications researchers need to combine evidence from multiple sources. For example, studies may have been carried out under similar designs, in different species, with different technologies, or at multiple sites. Traditional meta analysis methods may be insufficiently flexible by requiring normality. As a more flexible approach, the Branch developed a semiparametric Bayes method based on a matrix stick-breaking process (MSBP). The MSBP has clear advantages over previously-proposed Dirichlet process methods, as the Branch illustrated through application to an international validation study of the uterotrophic bioassay.