Wednesdays 1:15 pm - 2:05 pm
922 Hunter East
695 Park Avenue
New York, NY 10065
Main entrance (Visitors Center): Southwest corner of Lexington Ave. and 68th St.
Warren Tai (wtai AT gradcenter.cuny.edu)
Olympia Hadjiliadis (olympia.hadjiliadis AT gmail.com) Jiangtao Gou (jiangtao.gou AT hunter.cuny.edu)
October 4, 2017
Speaker: Lei Xie
Professor, Department of Computer Science, CUNY Hunter College and CUNY Graduate Center
Title: Multi-scale high-dimensional latent variable models for biological system
Abstract: Technological advances in high-throughput recording of biological systems have generated vast, dynamic, heterogeneous, and noisy data sets. These data provide new opportunities for addressing unmet medical needs, but impose new challenges in data integration and predictive modeling. Specifically, the high dimensionality of the biological system often prevent sufficient sampling of state space. New statistical machine learning methods are needed to gain information and discover knowledge from these high-dimensional sparse data sets. In this talk, I will introduce two methods for the predictive modeling of biological systems. The first method integrates multiple biological data sets across different organismal hierarchies, different temporal scales, and different species into a multi-layered latent variable network model. Novel relations are inferred by jointly optimizing within-layer and cross-layer latent space. The second method extends generalized additive model (GAM) to enable the simultaneous sampling and selection of both individual node and pairwise node variables in a network, and use of context-specific network-based variables for predictive modeling. Both of these methods have been successfully applied to solve real-world problems including drug repurposing and disease diagnosis.
November 22, 2017
Speaker: Jiangtao Gou
Assistant Professor, Department of Mathematics and Statistics, CUNY Hunter College
Title: Hierarchical testing in a group sequential trial when the primary endpoint data become available earlier than the secondary endpoint data
Abstract: We consider the problem of hierarchical testing a primary and a secondary endpoint using a group sequential design with multiple interim analyses. The secondary boundary can be refined if the hypothesis test follows the stagewise hierarchical rule or the partially hierarchical rule. When the primary endpoint data can be assessed earlier than the secondary endpoint data, the refinement can boost the secondary power significantly. Under the stagewise hierarchical rule, we provide a feasible region of information fractions when alpha-level boundary can be directly used in testing the secondary hypothesis at each stage. For a trial using the partially hierarchical rule, we recommend using the O'Brien-Fleming boundary for both the primary and the secondary endpoint.
November 29, 2017
Speaker: Wenge Guo
Associate Professor of Statistics, Department of Mathematical Sciences, the New Jersey Institute of Technology
Title: Analysis of Error Control in Large Scale Two-Stage Multiple Hypothesis Testing
Abstract: When dealing with the problem of simultaneously testing a large number of null hypotheses, a natural testing strategy is to first reduce the number of tested hypotheses by doing screening or selection, and then to simultaneously test selected hypotheses. The main advantage of this strategy is to greatly reduce the severe effect of high dimensions. However, the first screening or selection stage must be properly accounted for in order to maintain some type of error control. In this talk, we will introduce a selection rule based on the selection statistic which is independent of the test statistic when the tested hypothesis is true. Combining this selection rule and the conventional Bonferroni procedure, we can develop a powerful and valid two-stage procedure. The suggested procedure has several nice properties: (i) completely remove the selection effect; (ii) reduce the multiplicity effect; (iii) do not waste any samples while carrying out both selection and testing. Asymptotic power analysis and simulation studies illustrate that this proposed method provides higher power compared to usual multiple testing methods while controlling the type 1 error rate. Optimal selection thresholds are also derived based on our asymptotic analysis. This is a joint work with Joseph Romano from Stanford University.