Chairholder | Adrian E. Raftery

Adrian is Blumstein-Jordan Professor Emeritus of Statistics and Sociology at the University of Washington. He is also a faculty affiliate of the Center for Statistics and the Social Sciences and the Center for Studies in Demography and Ecology. He works on the development of new statistical methods for the social and environmental sciences. An elected member of the U.S. National Academy of Sciences, he was identified as the world's most cited researcher in mathematics for the decade 1995-2005 by Thomson-ISI. He has supervised 36 Ph.D. graduates, of whom 21 hold or have held tenure-track university faculty positions, and has 170 academic descendants.

Professor Raftery works on two main topics. The first is the development of new methods for Bayesian model selection and uncertainty quantification for inference in artificial intelligence (AI) models. The second is the development of new Bayesian statistical methods for integrated climate assessment.

1 - Bayesian model selection and uncertainty assessment for deep learning
Many current AI procedures are based on complex statistical models, notably neural net- work models, especially the versions that underly deep learning. The specification of neural network models involves many choices, such as the choice of network architecture, and of activation function. Each combination of choices can be viewed as corresponding to a different statistical model. Bayesian model selection (BMS) provides a general principled approach to this task, that is optimal in an inferential sense (Madigan & Raftery, 1994) and also optimizes cross-validatory predictive performance (Gneiting & Raftery, 2007). Bayesian model averaging (BMA) provides a principled way of accounting for uncertainty in such settings, including uncertainty about model choice.

BMS and BMA both depend on the marginal likelihood of a model. This is hard to compute, which has been an obstacle to the general adoption of BMS in deep learning. Bayesian inference is often carried out using Markov chain Monte Carlo (MCMC), but computing the marginal likelihood from MCMC output is often difficult. Recently, a promising and easy-to-compute estimator of the marginal likelihood from MCMC output has been proposed, called the truncated harmonic mean estimator, or THAMES (Metodiev et al., 2023). 

2 - Bayesian methods for integrated climate assessment
Integrated climate assessment is typically based on the IPAT equation, or Kaya identity, in which critical
climate quantities, such as carbon emissions for each country, are represented as a product of population, GDP per capita, and carbon intensity (i.e. carbon per unit of GDP). These three quantities are projected into the future to project carbon emissions, which are then combined with climate models to obtain projections of climate outcomes such as global average temperature (Intergovernmental Panel on Climate Change, 2021). This has often been done using subjective scenarios, which have been criticized as lacking scientific objectivity. More recently, progress has been made on developing Bayesian statistical models for this task (Raftery et al., 2017; Liu & Raftery, 2021).

Professor Raftery works on improving these models in various ways. One focus will be to improve methods for forecasting international migration, which are critical to the population part of the issue. A second focus will be to improve the use of climate model output, using the updated data which has become available. A third focus will be to investigate the possible use of AI models in the various forecasting sub-tasks.