Statistics Triple Crown 2016


 You are invited to participate in an afternoon of statistics seminars at NCAR.

Date: Thursday, July 14, 2016   (Bastille Day!)

Time: 1:00 pm-4:00 pm

Location: National Center for Atmospheric Research in Boulder, Colorado

Mesa Lab, Main Seminar Room

PROGRAM (download a copy. pdf)

Each presentation will be followed by a short discussion and a coffee break.

Covariate-Driven Nonstationary Spatial Modeling
Catherine Calder, Department of Statistics, The Ohio State University

In Gaussian process (GP) models for spatial data, the covariance function is typically assumed to belong to a parametric class of stationary covariance models.  Various authors have proposed methods for relaxing the second-order stationarity assumption in GP models using constructive techniques for specifying valid parametric and nonparametric covariance functions. While these approaches are elegant in their flexibility, model fitting can be problematic due to the high dimensionality of the parameter space and weak identifiability of model parameters. To overcome these issues, we build on the growing literature of covariate-driven nonstationary spatial modeling. We propose a Bayesian model for continuously-indexed spatial data based on a flexible covariance regression structure for a convolution-kernel covariance matrix, which allows the components of the convolution-kernel matrix to vary smoothly over space according to spatially-varying covariate information. We explore the properties of this model, including a description of the implied spatially-varying covariance function, and demonstrate that our parsimonious model provides a compromise between stationary and overly parameterized nonstationary models that do not perform well in practice. We illustrate our approach through an analysis of precipitation data. This presentation is based on joint work with Mark Risser.


Attribution of climate records in a multi-model max-stable framework
Philippe Naveau, Laboratoire des Sciences du Climat et l’Environnement (LSCE) CNRS

Nowadays, there is not one climate model but a plethora of them (see, e.g., Jones et al., 2013; Taylor et al., 2012). Each climate model has its strength and drawbacks depending on the region of interest and the atmospheric variable at hand. In this context, climate model variability, as well as error model, has to be taken account. In addition, it is difficult to postulate that the observational vector can be considered as a perfect and simple random draw of any climate model. Still, it is worthwhile to compare, in a relative way, how climate model outputs mimic the behavior of the observational time series. To explore the question of attribution for records, we assume that we have mclimate models at our disposal, and any of those climate models has been run under two different scenarios, say natural and all forcings respectively. By leveraging Multivariate Extreme Value Theory, we infer and interpret the probabilities of how record events change from the counterfactual to the factual world. Our computation are based on the concept of max-stable process. This framework lends to manageable computation time and simple inference schemes. Simulations and examples illustrate our approach.

Some thoughts on the use of extreme value theory for temperature extremes
Michael Stein, Department of Statistics and the College, University of Chicago

Applying extreme value theory to meteorological data is made difficult by seasonality and by dependence in both space and time. One way to avoid explicit modeling of seasonality is to consider block extremes, such as annual maximum temperatures at a location. However, I will argue that existing extreme value theory does not justify the usual assumption that annual maxima should approximately follow a generalized extreme value (GEV) distribution. I will consider possible asymptotic formulations that more properly account for seasonality and that, depending on how one takes limits, does or does not lead to a limiting GEV distribution. To investigate to what extent these results and others from extreme value theory are relevant to actual temperature extremes, I will consider 140 years of daily temperatures from Central Park in New York City. In particular, I will examine to what extent the influence of temporal dependence becomes negligible as one considers increasingly extreme temperatures.


 After the seminar sessions, please join the group for an informal discussion at Under the Sun at 5:00pm.

Doug Nychka will lead a voluntary hike to Under the Sun where hikers get to hear about Doug's ideas on the creation of the 1st data trail!

Dorit Hammerling and Doug Nychka, Organizers