Statistics Seminar, 2009-2010
Fall 2009 Talks
(Fall 2009, Seminar No. 1)
SPEAKER: Prof. Nikolai Chernov
University of Alabama at Birmingham,
Birmingham, AL, 35294, U.S.A.
TITLE:
Errors-in-variables regression models: Parameter estimates often have no
moments
TIME AND PLACE:
Thursday, September 10, 2009, 3:30pm
Room 1313, Math Bldg
ABSTRACT:
In the studies of linear and nonlinear regression problems when
both variables are subject to errors, it often happens that maximum
likelihood estimates have infinite mean values and infinite variances, but
they work well in practice. I will discuss these facts and their
methodological
implications.
(Fall 2009, Seminar No. 2)
SPEAKER: Terrence Moore, PhD Candidate
US Army Research Laboratory
Adelphi, MD 20783, U.S.A.
TITLE:
Constrained Cramer-Rao Bound
TIME AND PLACE:
Thursday, September 17, 2009, 3:30pm
Room 1313, Math Bldg
ABSTRACT:
Hero and Gorman developed a simple expression for an information
inequality under differentiable equality constraints (a constrained
Cramer-Rao bound) in terms of a full rank Fisher information and the
Jacobian of the constraint function. Later,
Stoica and Ng presented a more general expression applicable to
singular
Fisher
information matrices. This bound is particularly useful in measuring
estimation
performance in communications models, in which the design or control
of constraints is entirely plausible. Here, I will present a very
simple proof of this constrained Cramer-Rao bound and detail an
example of its utility in communications research.
BIO:
Terrence Moore is a mathematician working for the Tactical
Communications Networks Branch at the Army Research Lab in Adelphi
on statistical signal processing issues
of interest to the Army. He received his B.S. and M.A. degrees
in mathematics from the American University in Washington, DC,
in 1998 and 2000, respectively, and he is
currently a Ph.D. candidate in mathematics at the University of Maryland.
(Fall 2009, Seminar No. 3)
SPEAKER: Dr. Lance Kaplan, Team Leader
US Army Research Laboratory
Adelphi, MD 20783, U.S.A.
TITLE:
Monotonic Analysis for the Evaluation of Image Quality Measures
TIME AND PLACE:
Thursday, September 24, 2009, 3:30pm
Room 1313, Math Bldg
ABSTRACT:
A number of image quality measures have been proposed to quantify the
perceptual quality of images. When humans are given a task to
interpret a set of images, the "performance" of these humans over the
imagery should correspond to the image quality values in a monotonic
fashion. To this end, we consider two tests to determine whether or not
the monotonic relationship exists. First, the monotonic correlation
test assumes that human performance measurement errors are Gaussian, and
it is computed from the R2 value when using isotonic regression. The
second test, the diffuse prior monotonic likelihood ratio test, assumes
that the performance measurements follow a binomial distribution. This
talk will discuss properties of these two tests and apply these tests to
evaluate the effectiveness of image quality measures to "score" fused
images.
BIO:
Lance M. Kaplan received the B.S. degree with distinction from Duke
University, Durham, NC, in 1989 and the M.S. and Ph.D. degrees from the
University of Southern California, Los Angeles, in 1991 and 1994,
respectively, all in electrical engineering. From 1987 to1990, he was a
Technical Assistant at the Georgia Tech. Research Institute. He held a
National Science Foundation Graduate Fellowship and a University of
Southern California (USC) Dean's Merit Fellowship from 1990 to 1993, and
was a Research Assistant in the Signal and Image Processing Institute at
USC from 1993 to 1994. Then, he worked on staff in the Reconnaissance
Systems Department of the Hughes Aircraft Company from 1994 to 1996.
>From 1996 to 2004, he was a member of the faculty in the Department of
Engineering and a Senior Investigator in the Center of Theoretical
Studies of Physical Systems (CTSPS) at Clark Atlanta University (CAU),
Atlanta, GA. Currently, he is a Team Leader in the Networked Sensing and
Fusion branch of the U.S. Army Research Laboratory. Dr. Kaplan serves as
an Associate Editor-In-Chief and EO/IR Systems Editor for the IEEE
Transactions on Aerospace and Electronic Systems (AES). In addition, he
is the tutorials editor for the IEEE AES Magazine, and he also serves on
the Board of Governors of the IEEE AES Society. He is a three-time
recipient of the Clark Atlanta University Electrical Engineering
Instructional Excellence Award from 1999 to 2001. His current research
interests include signal and image processing, automatic target
recognition, data fusion, and resource management.
(Fall 2009, Seminar No. 4)
SPEAKER: Prof. Jian-Jian Ren
University of Central Florida,
Orlando, FL 32816, U.S.A.
TITLE:
Full Likelihood Inferences in the Cox Model
TIME AND PLACE:
Thursday, October 1, 2009, 3:30pm
Room 1313, Math Bldg
ABSTRACT:
We derive the full likelihood function for regression
parameter $\beta_0$ and baseline distribution function $F_0$ in the
continuous Cox model. Using the empirical likelihood parameterization,
we explicitly profile out nuisance parameter $F_0$ to obtain the
full-profile likelihood function and the maximum likelihood estimator
(MLE) for $\beta_0$. We show that the log full-likelihood ratio has an
asymptotic chi-squared distribution, while the simulation studies indicate
that for small or moderate sample sizes, the MLE performs favorably
over Cox's partial likelihood estimator. Moreover, we show that the
estimation bias of the MLE is asymptotically smaller than that of Cox's
partial likelihood estimator. In a real dataset example, our full likelihood
ratio test leads to statistically different conclusions from Cox's partial
likelihood ratio test. Part of this work is joint with Mai Zhou.
(Fall 2009, Seminar No. 5)
SPEAKER: David Judkins, Senior Statistician
WESTAT Inc.
Rockville, MD 20850-3195, U.S.A.
TITLE:
Using Longitudinal Surveys to Evaluate Interventions
TIME AND PLACE:
Thursday, October 8, 2009, 3:30pm
Room 1313, Math Bldg
ABSTRACT:
Longitudinal surveys are often used in evaluation studies conducted to
assess the effects of a program or intervention. They are useful for
examining the temporal nature of any effects, to distinguish between
confounding variables and mediators, and to better control for
confounders in the evaluation. In particular, the
estimation of causal effects may be improved if baseline data are
collected before the intervention is put in place. This presentation
will provide an overview of types of interventions, types of effects,
some issues in the design and analysis of
evaluation studies, and the value of longitudinal data. These points
will be illustrated using three evaluation studies: the U.S. Youth
Media Campaign Longitudinal Survey (YMCLS), conducted to evaluate a
media campaign to encourage 9-to 13-year-old Americans to be
physically active; the National Survey of Parents and
Youth (NSPY), conducted to evaluate the U.S. National Youth Anti-Drug Media
Campaign; and the Gaining Early Awareness and Readiness for
Undergraduate Programs (GEAR UP) program, designed to increase the
rate of postsecondary education among
low-income and disadvantaged students in the United States.
Based on:
Piesse, A., Judkins, D., and Kalton, G. (2009). Using longitudinal
surveys to evaluate interventions. In P. Lynn (Ed.), Methodology of
Longitudinal Surveys (pp. 303-316). Chichester: Wiley.
(Fall 2009, Seminar No. 6)
SPEAKER: Dr. Ben Klemens
United States Census Bureau
Suitland, MD 20746, U.S.A.
TITLE:
Using Agent-Based Models as Statistical Models
TIME AND PLACE:
Thursday, October 15, 2009, 3:30pm
Room 1313, Math Bldg
ABSTRACT:
Agent-based models (ABMs) involve the simulation of hundreds to millions
of individual agents, each making simple decisions. The results of these
decisions are often striking and make a direct qualitative statement.
However, ABMs can also be used like traditional statistical models for
quantitative analysis. I give the example of an ABM that explains a common
anomaly in the distribution of equity prices.
Click here to see the paper.
(Fall 2009, Seminar No. 7)
SPEAKER: Prof. David Stoffer,
Program Director, Probability and Statistics Program
University of Pittsburgh and National Science Foundation
Arlington, VA 22230-0002, U.S.A.
TITLE:
Spectral Magic
TIME AND PLACE:
Thursday, October 22, 2009, 3:30pm
Room 1313, Math Bldg
ABSTRACT:
The problem of estimating the spectral matrix of a multivariate time
series that has slowly changing dynamics has become a recent interest of
mine. The problem is difficult and had to be broken into smaller pieces.
I will discuss the first two pieces; there are at least two more
pieces to the puzzle. In the first place, estimating the spectral
density matrix of vector-valued stationary time series is not easy
because different degrees of smoothness are typically needed for
different components; this problem must be balanced with the fact that
the matrix must be positive semi-definite. I will discuss our approach
and then move on to the harder task of estimating the slowly changing
spectral density of a univariate locally stationary time series.
(Fall 2009, Seminar No. 8)
SPEAKER: Prof. Mei-Ling Ting Lee
University of Maryland,
College Park, MD 20742, U.S.A. (MLTLEE@UMD.EDU)
TITLE:
Threshold Regression for Time-to-event Data: with Applications in
Proteomics, Cancer Research, and Environmental Health.
TIME AND PLACE:
Thursday, October 29, 2009, 3:30pm
Room 1313, Math Bldg
ABSTRACT:
Threshold regression (TR) methodology is based on the concept that
health degradation follows a stochastic process and the onset of
disease, or death, occurs when the latent health process first
reaches a failure state or threshold (a first hitting time). Instead
of calendar time, the analytical time is considered. The model is
intuitive and generally does not require the proportional hazards
assumption and thus provides an important alternative for analyzing
time-to-event data. Connections with proportional hazard models will
be discussed. Examples and extensions will be discussed.
(Fall 2009, Seminar No. 9)
SPEAKER: Prof. Abram Kagan
University of Maryland
College Park, MD 20742, U.S.A.
TITLE:
A Class of Multivariate Distributions Related to Distributions with a Gaussian Component
TIME AND PLACE:
Thursday, November 5, 2009, 3:30pm
Room 1313, Math Bldg
ABSTRACT:
Abstract
(Fall 2009, Seminar No. 10)
SPEAKER: Prof. Antai Wang
Georgetown University
Washington, DC 20057, U.S.A.
TITLE:
Archimedean Copula Tests
TIME AND PLACE:
Thursday, November 12, 2009, 3:30pm
Room 1313, Math Bldg
ABSTRACT:
In this talk, we propose two tests for parametric models belonging to
the Archimedean copula family, one for uncensored bivariate data and
the other one for right-censored bivariate data. Our test procedures
are based on the Fisher transform of the correlation coefficient of
a bivariate $(U, V)$, which is a one-to-one transform of the
original random pair $(T_{1}, T_{2})$ that can be modeled by an
Archimedean copula model. A multiple imputation technique is applied
to establish our test for censored data and its $p$ value is
computed by combining test statistics obtained from multiply imputed
data sets. Simulation studies suggest that both procedures perform
well when the sample size is large. The test for censored data is
carried out for a medical data example.
(Fall 2009, Seminar No. 11)
SPEAKER: Prof. Jordan Stoyanov
Newcastle University
Newcastle upon Tyne, United Kingdom, NE1 7RU
TITLE:
Moment Analysis of Distributions: Recent Developments
TIME AND PLACE:
Thursday, November 19, 2009, 3:30pm
Room 1313, Math Bldg
ABSTRACT:
We start with classical conditions under which a distribution with finite
moments is M-determinate (unique), however our goal is to focus on recent
results providing conditions under which a distribution is
M-determinate or M-indeterminate (non-unique). Thus we will be
able to analyze Box-Cox functional transformations of random data,
before and/or after transforming, and characterize the moment
determinacy of their distributions. Popular distributions such as
Normal, Skew-Normal, Log-normal, Skew-Log-Normal, Exponential,
Gamma, Poisson, IG, etc. will be used as examples. Distributions
of Random walks and of Stochastic processes such as the Geometric
BM and the solutions of SDEs will also be considered.
We will illustrate the practical importance of these properties
in areas such as Financial modelling and Statistical inference problems.
Several facts will be reported. It seems, some of them are not
so well-known, they are a little surprising and even shocking.
However they are all quite instructive.
The talk will be addressed to professionals in Statistics/Probability,
Stochastic modelling, and also to Doctoral and Master students in
these areas. If time permits, some open questions will be discussed.
(Fall 2009, Seminar No. 13)
SPEAKER: At the Invitation of the Hiring Committee:
Dr. Richard Simon
Chief, Biometric Research Branch, National Cancer Institute
Rockville, MD 20852, U.S.A.
TITLE:
TBA
TIME AND PLACE:
Thursday, December 10, 2009, 3:30pm
Colloquium Room 3206, Math Bldg
ABSTRACT:
TBA
(Spring 2010, Seminar No. 1)
SPEAKER: Prof. Galit Shmueli
University of Maryland
College Park, MD 20742, U.S.A.
TITLE:
To Explain or To Predict?
TIME AND PLACE:
Thursday, February 11, 2010, 3:30pm
Room 1313, Math Bldg
ABSTRACT:
Statistical modeling is at the core of many scientific disciplines.
Although a vast literature exists on the process of statistical modeling,
on good practices, and on abuses of statistical models, the literature
lacks the discussion of a key component: the distinction between modeling
for explanatory purposes and modeling for predictive purposes. This
omission exacts considerable cost in terms of advancing scientific
research in many fields and especially in the social sciences, where
statistical modeling is done almost entirely in the context of
explanation. In this talk, I describe how statistical modeling is used in
research for causal explanation, and the differences between explanatory
and predictive statistical modeling.
(Spring 2010, Seminar No. 3)
SPEAKER: Prof. William Rand,
Director of Research, Center for Complexity in Business
University of Maryland
College Park, MD 20742, U.S.A.
TITLE:
Inferring Network Properties from Aggregate Diffusion Data
TIME AND PLACE:
Thursday, February 25, 2009, 3:30pm
Room 1313, Math Bldg
ABSTRACT:
Where do fads come from? Why are urban myths popular? Which of our
friends tells us about the next must-have gadget? Underlying all of
these questions is a process of diffusion, that is how do ideas,
concepts, and best practices spread through a population of
individuals? We examine these questions using a combination of agent-
based modeling, social network analysis, and machine learning.
Beginning with a replication of a traditional model using an agent-
based approach, we move on to explore diffusion processes on social
networks. After that we examine an application of these techniques to
a marketing application, specifically the role of customer
interactions in product success. Anecdotal evidence from business
practitioners and numerous academic studies have shown the importance
of word-of-mouth communication for product adoption. However, rarely
are interaction networks observable. Nevertheless, most firms do have
a significant amount of dynamic, aggregate marketing data, such as
customer purchases, attitudes, and information queries. We present a
new technique for inferring general network properties from this
aggregate-level data. We propose a Bayesian model selection approach
in combination with agent-based modeling to infer properties of the
unobserved consumer network, and show that it has the ability to
distinguish between various classes of networks on the basis of
aggregate data.
|
