|
Abstract:
|
For a single observation, one may observe multiple sparse functions, i.e., sampled on an irregular grid of points, where each function is correlated. Most Bayesian functional approaches are designed to handle observations on a single function, and though some approaches exist to handle multiple functions very few can handle hundreds or even tens of curves per observation. This talk presents a Bayesian nonparametric approach that assumes a latent model on the functional loadings when there are many sparse functions per chemical. It is applied to bioassays taken from the US EPA’s ToxCast database that studies the relationship between chemical exposure and various biological targets. By sharing information across assays, the method allows for better predictions of individual concentration-response curves given a single assay. Additionally, it allows the development of assay correlation matrix prediction across exposed concentrations for different covariates. That is one can look at correlations between in vitro assays across the concentration levels, which allows for better understanding of the in
|
