Jyotiska Datta

Biography

I am a second-year postdoctoral associate in the Department of Statistical Science at the Duke University and SAMSI (Statistical and Applied Mathematical Sciences Institute), currently working under the supervision of Prof. David B. Dunson and Dr. Sandeep S. Dave. At SAMSI, I was a part of the year-long "Beyond Bioinformatics" (2014-15) program. I graduated from Purdue University with a Ph.D. in Statistics in August, 2014, under the guidance of Prof. Jayanta K. Ghosh. My dissertation was on "Some Theoretical and Methodological Aspects of Multiple Testing, Model Selection and Related Areas"

Research Interest

My research focuses on developing novel statistical methods motivated by large and complex structured data, arising from high-throughput biological experiments. An over-arching focus is to develop scalable, multi-resolution Bayesian methods for whole-genome sequencing studies, that gain statistical power by pooling information across sites and sources and incorporating biological information (e.g. functional annotation) about rare mutations. This methodological advance applies to a broad range of disease models, and can be adapted to other sources of sparse and structured count data, such as imaging genetics, ecological survey or high-energy particle physics. A cornerstone of my research is building methods guided by theoretical investigation of optimality in handling high-dimensional sparse data, pervasive in many modern fields. Such theoretical insights have led to improvement of existing estimator. In the course of my Ph.D. research, I have proved the first asymptotic optimality properties of the popular 'Horseshoe' prior in a decision theoretic framework. Extending this effort, I have developed an ultra-sparse signal detection method called the 'Horseshoe+' prior in collaboration with Prof. Nick Polson and Dr. Anindya Bhadra. My subsequent work has also shed light on the excellent performance of heavy-tailed priors for low-dimensional functions of multi-dimensional parameters, and has settled a long-standing problem about a default prior for such problems.