Dalla Lana School of Public Health

Links

• For more information

Prof. Miguel Hernan, Director of CAUSALab, and Professor from the Department of Biostatistics and Epidemiology, Harvard T.H. Chan, joins the Data Sciences Speaker Series (DSSS).

The DSSS is intended as a collaborative marquee series, whose focus is data science and its intersection with other fields. Together, we seek to advance knowledge in the field of data science by featuring world-class speakers from academic, healthcare, industry, finance, technology, and other sectors and industries. In doing so, we hope to facilitate the exchange of ideas, information, and knowledge among researchers, practitioners, and other professionals, and to enhance educational opportunities for students and trainees.

This talk is co-sponsored by the Data Sciences Institute and Dalla Lana School of Public Health at the University of Toronto.

Title: Without causal inference, AI is FI (fake intelligence)

Abstract: The word “artificial” may be interpreted as “made by humans” (as in artificial flavor) or as “fake” (as in artificial smile). In this talk, I will argue that the second meaning is more appropriate for AI in 2023. A combination of clever algorithms, large amounts of data, and increasingly powerful computers has led to impressive advances in so-called AI. However, in the health and social sciences, these advances are largely concerned with tasks that use data for prediction (e.g., pattern recognition, curve fitting) rather than for counterfactual prediction. The latter is a key feature of intelligent beings and the basis of causal inference. When dealing with the complex systems of the health and social sciences, counterfactual prediction based on the integration of data and causal models remains a future goal for AI agents.

Bio: Prof. Miguel Hernan’s research is focused on methodology for causal inference, including comparative effectiveness of policy and clinical interventions. In an ideal world, all policy and clinical decisions would be based on the findings of randomized experiments. For example, public health recommendations to avoid saturated fat or medical prescription of a particular painkiller would be based on the findings of long-term studies that compared the effectiveness of several randomly assigned interventions in large groups of people from the target population that adhered to the study interventions. Unfortunately, such randomized experiments are often unethical, impractical, or simply too lengthy for timely decisions. His collaborators and himself combine observational data, mostly untestable assumptions, and statistical methods to emulate hypothetical randomized experiments. They emphasize the need to formulate well defined causal questions, and use analytic approaches whose validity does not require assumptions that conflict with current subject-matter knowledge. For example, in settings in which experts suspect the presence of time-dependent confounders affected by prior treatment, they do not use adjustment methods (e.g., conventional regression analysis) that require the absence of such confounders.