AmStat Section News, March 2005

Jennifer Hoeting, Publications Chair

The Statistics and the Environment Section is sponsoring two continuing education courses at the upcoming the Joint Statistical Meetings in Minneapolis, August 2005. The courses, “Semiparametric Regression” and “Computational Statistics: Methods for Optimization and Monte Carlo Integration” are described briefly below. For more information and to register for these courses, see www.amstat.org/meetings/jsm/2005/

Semiparametric Regression

Professors David Ruppert (Cornell University) and Ciprian Crainiceanu (Johns Hopkins University) will be presenting a course based on the book Semiparametric Regression by Ruppert, Wand, and Carroll, Cambridge University Press, 2003.

Parametric regression involves fitting a curve to a data set within the confines of parametric families: for example, lines, parabolas and exponentials. Nonparametric regression, often called smoothing, only imposes the condition that the curve be smooth; the shape of the curve depends primarily upon the data. Semiparametric regression combines nonparametric and parametric models. For example, in the analysis of time series data on mortality and air pollution it is common for the air pollution effects to be modeled linearly, and confounders such as time, temperature and humidity to be nonparametric. The course will focus on semiparametric models based on penalized splines viewed as linear mixed models (LMMs), or, for binary, Poisson, and other non- Gaussian outcomes, as generalized linear mixed models (GLMMs). Attendees will learn how semiparametric models can be built using spline basis expansion and within a mixed model framework. Topics include scatterplot smoothing, especially by penalized splines; LMMs and penalized splines as best linear unbiased predictors; automatic selection of the degree of smoothing, especially estimation of the penalty parameter of a spline by restricted maximum likelihood; additive models; inference; generalized semiparametric regression and GLMMs; spatial statistics; measurement error; and Bayesian semiparametric regression. The methodology will be illustrated with numerous examples. Attendees will be introduced to software (SAS and Splus or R) available for implementing these models and by the end of the course they will be able to apply these models to their own data.

Computational Statistics: Methods for Optimization and Monte Carlo Integration

Professors Geof Givens and Jennifer Hoeting (Colorado State University) will present a course given in two parts: a morning session on Monte Carlo integration strategies including MCMC and an afternoon session on optimization methods. The course will survey a variety of techniques, ranging from classic to state-of-the-art. The course will be based on the book Computational Statistics, by Givens and Hoeting, Wiley, 2005, which covers these and other topics in greater detail.

Many problems in statistics require the evaluation of integrals that cannot be solved analytically, particularly in Bayesian statistics. The morning session will cover Monte Carlo integration, importance sampling and variance reduction techniques, and Markov chain Monte Carlo methods. Optimization also plays a central role in statistics, particularly in numerical maximum likelihood estimation. The afternoon session will focus on various optimization strategies including Newton-like methods, Gauss-Seidel iteration, tabu algorithms, simulated annealing, genetic algorithms, the EM algorithm and its variants. The goal of the course is to give students a practical understanding of how and why existing methods work, enabling students to use modern statistical methods effectively. Examples are drawn from diverse fields including bioinformatics, ecology, and medicine. The course is targeted for quantitative scientists and statisticians who are unfamiliar with these methods. Upper division undergraduate mathematical literacy is recommended. No computer programming experience is necessary.