Department of Statistics Seminar
North Carolina State University
presents
Katja Remlinger, Kristi O'Grady, Ross Gosky and Yuliya Lokhnygina
North Carolina State University
A Sampler of Graduate students
Statistical Design of Pools Using Optimal Coverage and Minimal Collision*
Abstract: Discovery of a new drug involves screening large chemical
libraries to identify active compounds. Screening efficiency can
be improved by testing compounds in pools. We consider two
criteria to design pools: optimal coverage of the chemical space
and minimal collision between compounds. Five pooling designs are
applied to a public data set. We evaluate each method by
determining how well the design criteria are met. One design
uniformly dominates all other designs, but all criteria-designed
pools outperform randomly created pools. Furthermore, we discuss
blocking and synergism between compounds as other effects that
must be investigated when performing pooling experiments.
*This a joint work with J. M. Hughes-Oliver (NCSU), S. S. Young (NISS) and R. L. H. Lam
(GSK)
Combining Bayesian Networks and System Reliability to Meet Missile Defense Agency Needs
Abstract: System reliability has often been viewed as a post-production
statistical analysis. There exist systems, however, on such a large and
costly scale that warrant the performing of a reliability analysis before
initial system tests can be performed. In an effort to address this need
the research in this talk focuses on the development and implementation of
a Bayesian Network for "pre-test" reliability assessment. A motivating
example from the Missile Defense Agency will be presented. Specific
technical details include Bayesian network basics, and the
Metropolis-Hasting's algorithm.
Model Selection for Bayesian Closed Population Capture Recapture Models*
Abstract: We present Bayesian versions of 8 closed population Capture Recapture
Models. These models incorporate variation in animal capture
probabilities due to time effects, heterogeneity among animals, and
behavioral effects after the first capture. For a given dataset, choosing
which of the 8 models best fits the data is very important. Through
simulations, we show that both Akaike's Information Criterion (AIC), and
the Deviance Information Criterion (DIC) are useful tools in model
selection. In particular, we assess whether these criteria are useful in
selecting the correct model when average capture probabilities are between
20-30%.
*This is a joint work with S. K. Ghosh (NCSU)
Two-stage Optimal Group Sequential Designs*
Abstract: Clinical trials are usually designed to have a particular power to detect
some "clinically meaningful" treatment effect. It often happens, however,
that the estimated treatment effect at the interim stage is smaller than
this "clinically meaningful" alternative, but may still be important. Then
the study might not have enough power to detect this smaller treatment
difference. We propose a two-stage design which achieves the minimum
expected sample size at the particular alternative and also has the
desired power to detect a pre-specified smaller treatment difference. Our
method is a modification of the "backward induction" algorithm applied to
a Bayesian decision problem.
*This is a joint work with A. A. Tsiatis (NCSU)
Friday, September, 19, 2003
3:35 - 4:35 pm
206 Cox Hall
Refreshments will be served on the second floor of Dabney Hall (left of Room 222) at 3:00 pm.