Senior Data Science Fellow, Center for Statistics and the Social Sciences
Health Sciences, Social Sciences, Statistics
Applications in Social Sciences
Working Spaces and Culture
Thomas Richardson’s main area of research, graphical models, has developed at the interface between statistics and computer science, but has origins and applications in econometrics, epidemiology, genomics, psychology, and sociology. In simple terms, graphs (composed of vertices and edges) provide a language for formulating complex hypotheses. These hypotheses are subtle as they often encode both causal and statistical information.
Much of Richardson’s work has been concerned with the problem of inferring causal structure from observational and experimental data. This has motivated the construction of new classes of graphical model that allow for the possibility that there are unmeasured ‘confounding variables’ without explicitly modeling them – as this often occurs in practice.
Richardson has also developed novel algorithms for learning the parameters and structure of these new models. One algorithm, developed in conjunction with Diego Colombo, Marloes Maathuis and Markus Kalisch (ETH Zurich) is capable of learning the structure of a graph with hundreds of variables.
In another recent project, with James Robins (Harvard School of Public Health), and Ilya Shpitser (Southampton UK), Richardson formulated a fast generalized Mobius transform algorithm that facilitates the efficient computation of intervention distributions from causal Bayes nets with high-dimensional hidden variables; the first computationally efficient algorithm for this task.