STATISTICS WEEK AT ASU – SPRING 2009

Our annual statistics week activities for spring 2009 will be held Wednesday and Thursday, 4 and 5 March. We will have two seminars by our statistics week visitor, Dr. David M. Steinberg of Tel Aviv University. Dr. Steinberg is also the current editor of Technometrics.

We are also pleased to have the participation of JMP this year. In addition to an exhibit during the statistics fair, there will also be a presentation illustrating the new features in the recent release of JMP (v8). Dr. Bradley Jones, Director of Research & Development in the JMP division of SAS, will give a presentation on choice experiments, one of the new capabilities of JMP 8.

Seminar and Statistics Fair Schedule

Wednesday, 4 March

1:00 pm GWC 487

“Sequential Bayesian Design of Experiments for Glm's” by Dr. David M. Steinberg

2:00 Refreshments

JMP Presentation and “Choice Experiments”, by Dr. Bradley Jones

Thursday, 5 March

9:30-11:30 am Statistics Fair, Hayden Lawn

1:30 pm GWC 487

“Sensitivity Analysis and Computer Experiments”, by Dr. David M. Steinberg

Abstracts and biographical information are at the end of this announcement.

Dr. Steinberg will be on campus on Monday and Tuesday 2 and 3 March and is available to meet with interested faculty on Monday and Tuesday afternoons. To schedule an appointment, please contact Theresa Chai in the Industrial, Systems and Operations Engineering Department, 5-3185.

SENSITIVITY ANALYSIS AND COMPUTER EXPERIMENTS

David M. Steinberg

Department of Statistics and Operations Research

Tel Aviv University

Scientific work is often carried out today on a computer screen rather than in the field or the laboratory. We study phenomena like the crash-worthiness of an automobile, the spread of pollutants from a waste dump or the anticipated performance of the national economy by running computer code that simulates the actual physics, biology or economics. These “computer experiments” have much common statistical ground with conventional laboratory experiments: we need to decide at what settings of the input factors to run the simulator and, having obtained output, to model its relationship to the factors. There are also differences – often there is no “random error”, factor settings can be assigned a large number of values and reducing bias is the key issue in modeling.

This talk will give a broad background to the field of computer experiments. I will discuss and illustrate ideas for design and for analysis that take advantage of the unique features of these studies. I will describe numerous applications and will present a detailed analysis of a computer simulation of a nuclear waste repository.

The ideas reflect collaboration with many colleagues in statistics and in applied science.

SEQUENTIAL BAYESIAN DESIGN OF EXPERIMENTS FOR GLM'S

David M. Steinberg

Department of Statistics and Operations Research

Tel Aviv University

Many experiments involve binary or count responses that don't follow a normal distribution. Great advances have been made for modeling the data from these experiments using Generalized Linear Models (GLM's). Knowledge of how to design these experiments efficiently has lagged far behind. For example, classical two-level factorial designs can be very inefficient in binary response experiments. An important, and problematic, aspect is that good designs depend on knowledge of the unknown model parameters. One natural way to resolve this dilemma is to adopt a Bayesian approach and finding designs that are efficient in an average sense with respect to a prior distribution on the model parameters.

I will present some effective, and practical, algorithms for designing experiments that achieve a high degree of robustness to initial assumptions about the model parameters. I will briefly present ideas that can be used in static experiments, in which all runs are determined at the outset, and will then turn to sequential experiments, in which the data collected thus far are used in planning subsequent observations. One algorithm for static experiments takes advantage of a fast scheme for generating locally optimal designs together with a clustering procedure. The second algorithm uses a sophisticated quadrature approach. The algorithm for the sequential case uses a sampling and weighting approach, rather than direct computation, to represent the posterior. I will illustrate the performance of the algorithms on several applications. For the sequential case, I compare the results from our algorithm with those from the classic ``Bruceton'' method on an actual sensitivity test conducted at an industrial plant.

This is joint work with Hovav Dror, Chris Gotwalt and Bradley Jones.

David M. Steinberg is Professor of Statistics at Tel Aviv University. He has a B. Sc. in mathematics from Tel Aviv University and a Ph. D. in statistics from the University of Wisconsin-Madison, where he was a student of George Box. Professor Steinberg's research has focused on the statistical design of experiments. Some achievements include a Bayesian criterion for scaling response surface designs, a justification of the use of noise factors for finding dispersion effects in robust design experiments, a method for adding runs to fractional factorial experiments and results on optimal deployment of stations in a seismic monitoring network. In recent years his work has emphasized design of experiments for computer simulations and for generalized linear models, the topics of the two talks at Arizona State University. Professor Steinberg has been actively involved in applications. He served as head of the Statistical Laboratory at Tel Aviv University for 10 years and has consulted in a diverse array of fields, from engineering to medicine to seismology to public opinion surveys. He represented the State of Israel at three meetings on event screening for the Comprehensive Nuclear-Test-Ban Treaty. Professor Steinberg was the section editor on design of experiments for the Encyclopedia of Statistics in Quality and Reliability. He is currently Editor of the leading journal Technometrics.

Dr. Bradley Jones is Director of Research & Development in the JMP division of SAS. At JMP he developed the Custom Designer, a general and powerful tool for generating optimal experimental designs. Before joining JMP, he was the Chief Statistician at The MathWorks, Inc. where he wrote the Statistics Toolbox for MATLAB. He was a founding partner of Catalyst Inc. where he created the first interactive graphical computer tool for the design and analysis of experiments. He is the inventor of the prediction profile plot - an interactive graph for exploring multivariate response surfaces. He also holds a patent on the use of designed experiments for minimizing registration error in multi-layer laminated circuit boards.