Friday, August 30 2019
11:00am - 1:00pm
Mengjie Yao - PhD Defense

Title: Dual block canonical correlation analysis with an application in genetics
Abstract: Many genetic studies have focused on complex traits, which result from multiple genes, their interaction with each other and the environment. The human face is a complex trait with strong genetic component. In order to identify genetic variation that is associated with normal facial variation measured with 3D facial imaging in African children, the application of canonical correlation analysis (CCA) and its extended methods in the area of genetic association studies was explored. Almost all CCA and its extended methods are theoretically flawed when applied in a genetics setting and fail to consider and utilize the structure inherit in genetic data. In order to address the limitations of CCA related methods and make full use of the structure of genetic data, a novel method, dual block canonical correlation analysis (dbCCA), is proposed. dbCCA derives a linear combination of two sets of variables that maximally correlate with each other when the variables in one set have highly correlated group structure. dbCCA incorporates four major components in the model: a component that models internal correlation, a component that models cross correlation, like traditional CCA, a term that controls the interaction between the internal and cross correlation components, and a sparsity term. The performance of dbCCA is explored under a variety of simulation studies treating CCA with a lasso penalty as a benchmark. The simulation results show that dbCCA outperforms CCA lasso in detecting correlated variables and excluding uncorrelated variables when there is group correlation structure present in the data. Additionally, dbCCA was applied to the facial data, which motivated this work, and detected genetic associations are consistent with previous findings.
Speaker:Mengjie Yao

Download as iCalendar