JOINT UCONN/UMASS STATISTICS COLLOQUIUM
Aaron Sarvet
University of Massachusetts, Amherst
Title: The outperformance of machine learning by human intuition: resolving a paradox with unmeasured confounding
Abstract: In a precision-medicine system, decision rules might be algorithmically individualized based on an optimal regime previously learned from data. However, there is some resistance to the notion that such a system will result in better outcomes, compared to existing human-decision rules: existing care providers often will have access to relevant information for decision-making that is not recorded in the observed data. This is the essence of unmeasured confounding. I will present methodology for leveraging human intuition, given by the intended treatment values, by identifying a super-optimal regime using data generated by either nonexperimental or experimental studies, and clarify when a fusion of such data is beneficial. The super-optimal regime would indicate to a care provider precisely in which cases expected outcomes would be maximized if the care provider would override the optimal regime recommendation and, importantly, those cases when the optimal regime recommendation should be followed regardless of the care-provider’s assessment. Furthermore, I will discuss how super-optimal regime methodology may be alternatively applied for the systematic surveillance of public health paradoxes. As an example, I will consider the historic case of Ignaz Semmelweis, whose observations of paradoxes with puerperal fever precipitated a hygiene revolution in clinical medicine.
Bio: Dr. Sarvet is an Assistant Professor of Biostatstics, at the University of Massachusetts, Amherst, MA, USA. Aaron received a PhD in population health sciences from the Harvard School of Public Health, and conducted postdoctoral research at the École polytechnique fédérale de Lausanne. Aaron's research focuses on the development of innovative methods for statistical causal inference in medicine and epidemiology, which can be applied to large-scale and passively collected longitudinal data, e.g., from electronic health records. A primary objective is to expand the set of policy questions that can be explicitly posed within the language of formal causal and statistical frameworks. Aaron's research also applies a critical lens to the dominant paradigms in statistics and causal inference that structure epidemiologic thought and practice.
Date: Wednesday, April 15, 2026, 3:30 PM, AUST 434
WebEx link: https://uconn-cmr.webex.com/uconn-cmr/j.php?MTID=m403ab81d70d234a6fd0b4bb38a09c983
Coffee will be available at 3:00 PM in the Noether Lounge (AUST 326)
For more information, contact: Yuwen Gu at yuwen.gu@uconn.edu