Statistics Seminar - Spring 2022
Seminars are held on Thursdays from 4:00 - 5:00pm on Webex unless otherwise noted. For access information, please contact the Math Department via email at math@njit.edu.
For questions about the seminar schedule, please contact Zuofeng Shang
March 24
Alex Dytso, Department of Electrical and Computer Engineering, NJIT
Location: CKB 120 and WebEx
Some Aspects of Totally Positive Kernels Useful in Estimation and Information Theory
This talk will present the role of totally positive kernels and Po'lya type distributions to information theory. In particular, it will be discussed how the variational diminishing property of positive definite kernels, which is captured by the oscillation theorem, can be used to characterize the structure of least favorable prior distributions in a variety of different problems in estimation theory and information theory.
April 19
Yongzhao Shao, Division of Biostatistics, Departments of Population Health & Environmental Medicine, NYU Grossman School of Medicine
*Please note this seminar will be held on a Tuesday*
Location: WebEx
Multicellular Network-based Survival Models for Drug Resistance and Predicting Patient Survival
Cancer immunotherapies have saved numerous lives of cancer patients in the last decade. However, currently, only a portion of patients have durable responses to these treatments as drug resistance still exists widely when treating many late-stage cancers. An important task in precision medicine is to build effective statistical models to predict which patient will likely respond to the given therapy and with long survival or who will experience treatment resistance with short survival. It is known that interactions of tumor cells and immune cells in the tumor-associated microenvironment play important roles in tumor progression and drug resistance to immunotherapies. We have developed a multicellular gene network approach aimed at investigating the prognostic role of interactions between tumor cells and macrophage cells in tumor progression and drug resistance in advanced brain tumors (gliomas). Multicellular gene networks connecting macrophages and tumor cells were constructed from samples of RNA-seq data in mice gliomas treated with BLZ945. Subsequently translated into human gliomas and a gene network-based penalized Cox regression model was built to identify a risk signature using a cohort of 310 glioma samples. A large independent validation set of 690 glioma samples from The Cancer Genome Atlas (TCGA) database was used to test the prognostic significance and accuracy of the gene signatures in predicting survival of glioma patients. Limitations of our approach and the data sets used as well as remaining challenges and possible further research directions will be discussed. The talk is based on joint work Drs. Xiaoqiang Sun and Xinwei He at Zhangshan University School of Medicine and Yuting Lu of New York University Grossman School of Medicine.
April 28
Hu Sun, Department of Statistics, University of Michigan
Location: WebEx
Auto-regressive Model for Matrix-Valued Time-Series with Auxiliary Vector-Valued Time-Series Data
Matrix-valued time-series, different from scalar & vector time-series, contains individual time-series identified by multiple additional non-temporal dimensions placed on a regular 2-D grid. Examples of matrix-valued time-series can be found in various scientific domains such as astronomy and biology in the form of spatio-temporal process. To provide forecasts for future matrices in the sequence, given the history of matrices observed in the past, the standard vector auto-regressive model (VAR) fails to take advantage of the matrix nature of the data. The modeling problem is made even more complicated when auxiliary vector-valued predictors associated with the matrix time-series are present. To undertake the matrix time-series forecasting problem, we propose an extension to the traditional vector auto-regressive model to formulate a multi-lag matrix auto-regressive model. We accommodate the auxiliary predictors in the same model by augmenting the vector predictors with matrix basis functions to allow for both matrix and vector predictors to contribute to forecasting future matrices. Model estimation is done via alternating convex optimization and asymptotic properties for the estimators are provided. Both simulation and application to an astronomical dataset demonstrate the superior performance of the model. Finally, we also propose a stochastic updating scheme to make the method scalable to large spatio-temporal datasets.
Updated: April 4, 2022