Seminar Announcements for FALL 2012

________________________________________

Title: Small Area Confidence Bounds on Small Cell Proportions in Survey Populations
Speaker:  Dr. Eric V Slud, Professor, Statistics Program, Department of Mathematics, University of Maryland
Abstract:   Abstract: Motivated by the problem of `quality filtering' of estimated counts in American Community Survey (ACS) tables, and of reporting small-domain coverage results from the 2010 decennial-census Post-Enumeration Survey (PES), this talk describes methods for placing confidence bounds on estimates of small proportions counts within cells of tables estimated from complex surveys. While Coefficients of Variation are generally used in measuring the quality of estimated counts, they do not make sense for assessing validity of very small or zero counts. The problem is formulated here in terms of (upper) confidence bounds for unknown proportions. We discuss methods of creating confidence bounds from small-area models including synthetic, logistic, beta-binomial, and variance-stabilized (arcsine square root transformed) linear models. The model-based confidence bounds are compared with single-cell bounds derived from arcsine-square-root transformed binomial intervals with survey weights embodied in the "effective sample size". The comparison is illustrated on county-level data about Housing-Unit Erroneous Enumeration status from the 2010 PES. The primary methods of the talk are "small area estimation", a kind of empirical Bayes model-based prediction relevant to survey problems, with some discussion of parametric-bootstrap methods for interval estimation. This talk is based on joint work with Aaron Gilary and Jerry Maples of the Census Bureau.

Date: Friday, September, 14
Time: 11:00-12:00
Location: Phillips Hall, Room 416 (801 22nd Street, NW, Washington, DC 20052)
________________________________________

Title: Interdisciplinary Methods for Prediction and Confidence Sets
Speaker:  Dr. Sherri Rose, NSF Postdoctoral Fellow, Johns Hopkins Bloomberg School of Public Health
Abstract:   Abstract: The incorporation of methodology from disparate fields to answer scientific questions has become increasingly common in the age of massive data sets. This talk will discuss two areas of statistics, prediction and confidence sets, and interdisciplinary approaches that can be used to answer specific problems within these subspecialties. First, we generate a prediction function in an epidemiology study using a flexible machine learning ensembling approach that combines multiple algorithms into a single algorithm, returning a function with the best cross-validated mean squared error. Second, we discuss an algorithm for the construction of valid confidence regions for the optimal regime in sequential decision problems using linear programming.

Date: Friday, September, 28
Time: 11:00-12:00
Location: Phillips Hall, Room 108 (801 22nd Street, NW, Washington, DC 20052)
________________________________________

Title: Longitudinal High-Dimensional Data Analysis
Speaker:  Dr. Vadim Zipunnikov, Assistant Professor in the Department of Biostatistics at Johns Hopkins School of Public Health
Abstract:   Abstract: We introduce a flexible inferential framework for the longitudinal analysis of ultra high dimensional data. Typical examples of such data structures include, but are not limited to, observational studies that collect imaging data longitudinally on large cohorts of subjects. The approach decomposes the observed variability into three high dimensional components: a subject-speci c random intercept that quanti es the cross-sectional variability, a subject-speci c slope that quanti es the dynamic irreversible deformation over multiple visits, and a subject-visit speci c imaging deviation that quanti es exchangeable or reversible visit-to-visit changes. The model could be viewed as the ultra high dimensional counterpart of random intercept/random slope mixed e ects model. The proposed inferential method is very fast, scalable to studies including ultra-high dimensional data, and can easily be adapted to and executed on modest computing infrastructures. The method is applied to the longitudinal analysis of diff usion tensor imaging (DTI) data of the corpus callosum of multiple sclerosis (MS) subjects. The study includes 176 subjects observed at a total of 466 visits. For each subject and visit the study contains a registered DTI scan of the corpus callosum at roughly 30,000 voxels

Date: Friday, October, 12
Time: 11:00-12:00
Location: Phillips Hall, Room 108 (801 22nd Street, NW, Washington, DC 20052)
________________________________________