Statistics Seminar Series

Department of Mathematical Sciences
and
Center for Applied Mathematics and Statistics

New Jersey Institute of Technology

Fall 2011

 

All seminars are 4:00 - 5:00 p.m., in Cullimore Hall Room 611 (Math Conference Room) unless noted otherwise. Refreshments are usually served at 3:30 p.m., and talks start at 4:00 p.m. If you have any questions about a particular seminar, please contact the person hosting the speaker.

 

Date

Speaker and Title

Host

Friday
September 9, 2011
4:00PM

Zhigen (Gene) Zhao, Ph.D., Department of Statistics,  Temple University

On the Credible Interval under the Zero-Inflated Mixture Prior in High Dimensional Inference

 (abstract)

Wenge Guo


Thursday

September 15, 2011
4:00PM

Errol C. Caby, Ph.D., AT&T Labs Research, Florham Park, NJ
Mining Port-level IP Data
 (abstract)

Aridaman Jain


Thursday

October 6, 2011
Cullimore Lecture Hall, Room 111
4:00PM

Yichao Wu, Ph.D., North Carolina State University
Continuously Additive Models for Functional Regression
(abstract)

Wenge Guo


Thursday

October 20, 2011
4:00PM

Matt Hayat, PhD., College of Nursing, Rutgers University
Model-based Prediction of Solar Particle Events 
 (abstract)

Sunil Dhar


Thursday

October 27, 2011
Cullimore Lecture Hall, Room 111
4:00PM

 M. Bhaskara Rao, Ph.D., University of Cincinnati Medical Centre
Algebraic Statistics and Applications to Statistical Genetics
 (abstract)

Wenge Guo


Thursday

November 3, 2011
4:00PM

Ganesh K. (Mani) Subramaniam, PhD., AT&T Labs - Research, Florham Park, NJ
Exploratory Analysis of Large-Scale Spatial-Temporal Data
 (abstract)

Sunil Dhar


Thursday

November 10, 2011
Cullimore Lecture Hall, Room 111
4:00PM

Kai Zhang, Department of Statistics, The Wharton School, University of Pennsylvania

Valid Post-Selection Inference
 (abstract)

Sunil Dhar


Wednesday

November 16, 2011
Tiernan Hall 104
2:30PM

 

Li He, Department of Statistics, Temple University
Optimal Multiple Testing Procedure Incorporating Signal Strength

(abstract)

Wenge Guo


Thursday

December 1, 2011
Cullimore Lecture Hall, Room 111
4:00PM

Fei Liu, PhD.,  Business Analytics & Mathematical Sciences, IBM Thomas J. Watson Research Center
High-Dimensional Variable Selection in Meta Analysis for Censored Data
 (abstract)

Sunil Dhar


Thursday

December 8, 2011
4:00PM

Ji Meng Loh, PhD., AT&T Labs - Research, Florham Park, NJ
K-scan for anomaly detection in spatial point patterns
 (abstract)

Manish Bhattacharjee

 

ABSTRACTS

On the Credible Interval under the Zero-Inflated Mixture Prior in High Dimensional Inference:


In this paper, we consider the construction of the credible set under the canonical Bayes model when assuming the parameter of interest θ follows a prior distribution which is a mixture of  zero with probabilityand another non-trivial distribution with probability = 1−. In modern application with high dimension,  is usually very large, implying that the posterior probability P(θ = 0| X) is also large, saying greater than 5%. The traditional approaches constructing 95% posterior credible intervals, such as HPD region or the equal-tail credible interval, will enclose zero very often and appear to be powerless.

 

In this paper, we use the decision Bayes approach to guide us constructing a mixture credible interval. When there is overwhelming evidence that  we scrutinize the distribution  and consider the HPD region. Otherwise, the interval is the union of such a HPD region and zero. The paper provides a systematic way in deciding when to include zero, guaranteeing the average coverage probability.

 

We apply this general approach to a normal mean problem with unknown and unequal variances and apply it to a real data set. It is demonstrated that the new approach is way more powerful than the traditional and other alternatives.

Assistant Professor Zhigen (Gene) Zhao, Department of Statistics, Temple University, Philadelphia, PA, 19122 ~ September 9, 2011

Mining Port-level IP Data:

In this paper, we look at and resolve some issues involved in analyzing IP traffic at the port level. For example, we develop a mapping between alternative ways of measuring circuit utilization and describe a method of normalizing utilization to accommodate changes in bandwidth. We also prescribe a class of patterns that is suited to analyzing utilization for a large number of circuits in a dynamic environment.

Errol C Caby, PhD, AT&T Labs Research, Florham Park, NJ ~ September 15, 2011
 

Continuously Additive Models for Functional Regression:

We propose Continuously Additive Models (CAM), an extension of additive regression models to the case of infinite-dimensional predictors, corresponding to smooth random trajectories, coupled with scalar responses. As the number of predictor times and thus the dimension of predictor vectors grow larger, properly scaled additive models for these high-dimensional vectors are shown to converge to a limit model, in which the additivity is conveyed through an integral. This defines a new type of functional regression model. In these Continuously Additive Models, the path integrals over paths defined by the graphs of the functional predictors with respect to a smooth additive surface relate the predictor functions to the responses. This is an extension of the situation for traditional additive models, where the values of the additive functions, evaluated at the predictor levels, determine the predicted response. We study prediction in this model, using tensor product basis expansions to estimate the smooth additive surface that characterizes the model. In a theoretical investigation, we show that the predictions obtained from fitting continuously additive estimators are asymptotically consistent. We also consider extensions to generalized responses. The proposed estimators are found to outperform existing functional regression approaches in simulations and in applications to human growth and yeast cell cycle data.

Assistant Professor, Yichao Wu, Department of Statistics, North Carolina State University, Raleigh, NC 27695-8203 ~ October 6, 2011

Model-based Prediction of Solar Particle Events:

Astronaut health may be at risk with exposure to high energy solar particle events (SPEs). This becomes a major concern during extra-vehicular activities (EVA) on the lunar and Mars surface. In this talk, I will present model based predictions used to optimize planning for future space missions. SPEs are modeled as a function of time within a solar cycle. A non-homogeneous Poisson model is applied on historical data on solar events occurring between 1954 and 2006. Statistical modeling results will be shown for prediction of frequency and severity of solar events.

Assistant Professor Matt Hayat , Biostatistician, College of Nursing, Rutgers University, Newark, NJ ~ October 20, 2011

Algebraic Statistics and Applications to Statistical Genetics:

When testing independence in the context of contingency tables, the chi-squared test is a boon. However, this is an asymptotic test the use of which requires certain conditions to be fulfilled. If the chi-squared test is not justified, one could use Fisher's exact test. Even the use of Fisher's exact test is fraught with enormous computational difficulties. To set the stage for the talk, I set about explaining how algebraic statistics softens computational issues. Similar computational difficulties arise when one sets about testing a la Fisher Hardy-Weinberg equilibrium on a multi-allelic marker. I will demonstrate how algebraic statistics in conjunction with Markov bases and MCMC solves the problem. If time permits, I will ruminate on triad designs in genetics and gene-environment interactions.

Professor M. Bhaskara Rao, University of Cincinnati Medical Centre, Cincinnati, OH 45267 ~ October 27, 2011

Exploratory Analysis of Large-Scale Spatial-Temporal Data:

AT&T's Mobility data related to voice and data usage has a spatial (cell sites in different geographies) as well as a temporal component (usage collected at regular intervals). The focus of this work is to develop tools for exploratory data analysis as it relates to spatial data. The exploratory analysis techniques will look at both spatial and temporal aspects independently and also the attributes of their joint (spatial-temporal) behavior. Visualization techniques in particular to identify outliers (local and global), pockets of non-stationarity and other additional insights that are not otherwise apparent will be demonstrated. Some examples from the AT&T data will be provided.

Ganesh K. (Mani) Subramaniam, PhD., AT&T Labs - Research, Florham Park, NJ ~ November 3, 2011

Valid Post-Selection Inference:

It is common practice in statistical data analysis to perform data-driven model selection and derive statistical inference from the selected model. Such inference is generally invalid. We propose to produce valid “post-selection inference” by reducing the problem to one of simultaneous inference. Simultaneity is required for all linear functions that arise as coefficient estimates in all submodels. By purchasing “simultaneity insurance” for all possible submodels, the resulting post-selection inference is rendered universally valid under all possible model selection procedures. This inference is therefore generally conservative for particular selection procedures, but it is always less conservative than full Scheffe protection. We describe the structure of the simultaneous inference problem and give some asymptotic results.

Kai Zhang, Department of Statistics, The Wharton School, University of Pennsylvania, Philadelphia, PA 19104 ~ November 10, 2011

Optimal Multiple Testing Procedure Incorporating Signal Strength:

This research focuses on incorporating signal strength into a multiple testing procedure using a decision theoretic approach. Specifically, we propose a general loss function which incorporates the severity of Type II errors. We also propose error rates that incorporate signal strength: the weighted marginal false discovery rate and the weighted marginal false nondiscovery rate. We then derive optimal procedure which minimizes the weighted marginal false nondiscovery rate, while controlling the weighted marginal false discovery rate. The numerical studies show that, by incorporating signal strength, much power can be gained.

Dr. Li He, Department of Statistics, Temple University, Philadelphia, PA, 19122 ~ November 16, 2011

High-Dimensional Variable Selection in Meta Analysis for Censored Data:

This talk considers the problem of selecting predictors of time to an event from a high-dimensional set of candidate predictors using data from multiple studies. As an alternative to the current multi-stage testing approaches, we propose to model the study-to-study heterogeneity explicitly using a hierarchical model to borrow strength. Our method incorporates censored data through an accelerated failure time (AFT) model. Using a carefully-formulated prior specification, we develop a fast approach to predictor selection and shrinkage estimation for high-dimensional predictors. For model fitting, we develop a Monte Carlo Expectation Maximization (MC-EM) algorithm to accommodate censored data. The proposed approach, which is related to the relevance vector machine (RVM), relies on maximum a posterior (MAP) estimation to rapidly obtain a sparse estimate. As for the typical RVM, there is an intrinsic threshold property in which unimportant predictors tend to have their coefficients shrunk to zero. We compare our method with some commonly used procedures through simulation studies. We also illustrate the method using the gene expression barcode data from three breast cancer studies.

Fei Liu, PhD, Research Staff Member, Statistical Analysis & Forecasting, Business Analytics & Mathematical Sciences, IBM Thomas J. Watson Research Center Yorktown Heights, NY 10598 ~ December 1, 2011

 

K-scan for anomaly detection in spatial point patterns:

We consider the problem of detecting hotspots in spatial point patterns observed over time while accounting for inhomogeneous background intensity. For example, in disease surveillance, the interest is often in identifying regions of unusually high incidence rate given a background incidence rate that may be spatially varying due to underlying variation in population density, say. I will present a K-scan method that uses components of the inhomogeneous K function to identify such anomalies or hotspots. The significance of detected hotspots is assessed using either bootstrap or a p-value approximation based on a Gumbel distribution. I will show some results from a simulation study, as well as applications of this method to dead bird sighting data from Contra Costa county in California and to fast food restaurant location data in New York City.

Ji Meng Loh, PhD., AT&T Labs - Research, Florham Park, NJ ~ December 08, 2011