The perils of machine-learning for model selection: A case-study
Details
Machine learning (ML) is increasingly being used for data-driven tasks in public health and epidemiology, including for model selection. However, this approach differs radically from the typical principles of model selection in this field, and naive application of ML techniques to this task can lead to unexpected or unrealistic model results.
In this presentation, I will discuss the common principles of model selection in epidemiology, give an example of a recent application of ML-based model selection by The Economist, and discuss the consequences of naive application of data-driven model selection approaches in this complex field.
Biography
Stuart Gilmour is professor of biostatistics and bioinformatics at the Graduate School of Public Health, St. Luke’s International University, Japan. Stuart was born in New Zealand and raised in Australia by British parents, and has lived in Japan since 2006. Stuart obtained his undergraduate degree in mathematical physics at the University of Adelaide, completed a Masters in Public Health and a Masters in Statistics in Australia, and his PhD at the University of Tokyo. His research interest is in quantitative health system assessment, the use of statistics to improve our understanding of policies and interventions that can improve population health.