Chao Chang

Dept. of Mathematics, Washington University in St.Louis

Title :
Bayesian Quantile Regression
via Dirichlet Process Mixture of Logistic Distributions

Date and Time: May 1, 2014 - 11 am. to 12 pm.
Location: Cupples I, Room 6

Abstract: Abstract: In this talk I will present a new nonparametric Bayesian approach to solve quantile regression for a single quantile. One innovation is that the error distribution is modeled by using the Dirichlet process mixture of logistic distributions, which have the desired feature of being smooth and having a close-formed quantile function. Further, with logistic distributions as the kernel, our mixture model can provide great flexibility by mixing over both the location parameter and the scale parameter. As a result, the scenarios of multi-modal error distributions and clustered data can be handled by our model. Monte Carlo Markov chain algorithm is provided to do posterior inference. The performance of our approach is evaluated using simulated data and real data. I might also discuss some theories about convergence of Dirichlet process mixture if time permits.