Speaker: Jim Berger
Title: P-values - Bad and Not-So-Bad
I will review the main problem with p-values, through demonstration of a new applet (developed by German Molina). The applet (or variants) could be quite useful in elementary courses. Then I will discuss the fact that, if p-values are used as qualitative rather than quantitative measures of evidence, they lead to (conditional) frequentist error probabilities in testing that are exactly equal to default Bayesian posterior probabilities of hypotheses. That p-values, which are neither Bayesian nor frequentist, provide the link towards unification of Bayesian and frequentist testing is quite surprising.
Speaker: Herbie Lee
Title: Introduction to Neural Networks
Neural networks can be viewed statistically as a method of nonparametric regression or classification. I will briefly discuss some of the history of neural networks, and then show how they fit into modern statistics, including results on asymptotic consistency. Many people claim the parameters are interpretable, but I beg to differ, and I will show a simple example of noninterpretability, which is a good motivation for a noninformative prior. Finally I present some examples of neural networks, including examples with model selection.
Speaker: Doug Nychka
Title: Wavelet Representations for Nonstationary Spatial Fields
Spatial analysis for large nonstationary processes poses challenges in both modeling and computation. A promising way to represent nonstationary covariance structure is by expanding the field in terms of a wavelet basis and then building a simple, sparse model for correlations and variances among the wavelet coefficients. In this talk a nonorthogonal wavelet basis (the W-transform) is presented that not only appears to fit a variety of standard covariance models but is well suited to the computation of Kriging estimates and conditional distributions. From a more conventional perspective, this wavelet-based model provides an reasonable blending between an EOF representation (principle components of the sample covariance matrix) and a stationary, parametric family. This approach is illustrated using output from a run of the Regional Oxidant Model, an EPA pollution simulation.
Speaker:Todd Graves
Title: Opportunities in the Statistical Sciences Group at Los Alamos National Laboratory
Los Alamos National Laboratory's Statistical Sciences Group is hiring this year and is always looking for graduate research assistants and faculty guests. I will discuss what it's like to work at Los Alamos and introduce several interesting classes of statistical problems that frequently arise (namely reliability, computer model evaluation, integration of diverse sources of information, statistical population bounding, and Monte Carlo methods). I will also discuss a problem from my own research: using statistical modeling to devise appropriate rule bases for agents to use in agent-based computer simulations.