Abstract — Although algorithms for causal discovery from observational and experimental data have been developed and tested and improved for almost 25 years, they have rarely if ever been tested on data with measurement error, which in real science is the rule rather than the exception. In this talk, I show how even moderate amounts of measurement error can drastically reduce the performance of discovery algorithms and why the problem is non-local and complicated. I compare and contrast several strategies for handling the problem, none of which are ideal.
Richard Scheines is the Dean of the Dietrich College of Humanities and Social Sciences at Carnegie Mellon University. He has appointments in the Department of Philosophy, the Machine Learning Department, and the Human-Computer Interaction Institute. After receiving a PhD in History and Philosophy of Science from the University of Pittsburgh, Dr. Scheines joined the faculty of Carnegie Mellon in 1990. Working at the interface of computer science, statistics, and philosophy, he has published several books and dozens of articles on statistical causal models and how to efficiently search for them from statistical evidence.