LOCATION:   319 Snedecor

SPEAKER:  Nidhan Choudhuri, Case Western Reserve University, Cleveland, Ohio

TITLE: Fully nonparametric Bayesian analysis in regression model with
           serially correlated errors

ABSTRACT:

Consider the regression model Y=f(X)+e, where X is possibly a multivariate predictor. When the data is collected over the time, the errors are likely to be serially correlated. Most of the methodologies for analyzing such data assumes a parametric from for the correlation structure. While the parametric modeling of the regression function is sometimes useful in explaining the functional relationship between the variables, such assumption on the correlation structure is solely technical. This talk describes a Bayesian approach to nonparametric modeling of the correlation structure. A prior is put on the spectral density of the auto-covariances rather than the auto-covariances itself. The advantage of shifting into the frequency domain is that one may now use the Whittle likelihood for updating the prior. COFFEE:  3:45 p.m., 104 Snedecor Hall