Estimating causal effects from observational and limited experimental data using graphical models

Abstract:
Much of intelligent behavior involves predicting the causal effects of actions. The large and increasing stores of electronic data raise the prospect of learning many cause-effect relationships from such data. Observational data are plentiful in many domains, such as electronic health record (EHR) data in the healthcare domain. However, the use of observational data to estimate causal effects may produce biased estimates. Experimental data, such as randomized controlled trials (RCTs) in healthcare, provide relatively unbiased estimates of causal effects; however, they are much scarcer. This talk will describe a method that uses experimental data to determine whether and how to use observational data to estimate causal effects. One application of the method is personalized patient-outcome prediction from EHR and RCT data. Another potential use is the combination of observational user data with the results of more limited quantity A/B experiments to predict the effects of system actions on user behavior.

Bio:
Gregory Cooper, M.D., Ph.D. is a faculty member in the Department of Biomedical Informatics and in the Intelligent Systems Program at the University of Pittsburgh. His research involves the use of causal modeling, decision theory, Bayesian statistics, machine learning, and artificial intelligence to develop new methods for addressing biomedical informatics research problems. He is currently involved in projects on causal discovery from observational and experimental data, infectious disease outbreak detection and characterization, personalized cancer diagnosis and outcome prediction, clinical alerting based on machine learning from an electronic-medical-record (EMR) archive, and an EMR system that learns to highlight the most useful patient information.

Sofia Triantafyllou, Ph.D., is am assistant professor at the Department of Mathematics and Applied Mathematics in the University of Crete. Her research involves causal discovery and inference from multiple sources of data, and applications of causal inference. Her recent work includes tuning causal discovery algorithms, and integrating observational and experimental data, collected under different conditions, to estimate causal effects.