Abstract

In this talk, we introduce a parametrization of a correlation matrix based on partial correlations. This accounts for conditional dependence, allowing a more informed/structured model of multivariate distributions. This can be applied to model multiple outcomes jointly, and we will illustrate the practical use when modeling with an extra source of dependency, such as longitudinal or spatial. We will illustrate the flexibility of our implementation in both how to specify a graph-based correlation, its parametrization, and how to use it as a prior distribution in complex models.