Abstracts

Clutter-Free Causal Inference

Donald B. Rubin - Temple, Tsinghua, and Harvard Universities
Friday, September 3 at 10:30 a.m. EDT

Many, if not most, data analyses aim at understanding relationships between quantities that we observe, and trying to understand what would happen if we intervened in our world in various ways.  This intuitive statement describes what I regard as the essential idea underlying causal inference, a field of research  that has become extremely popular in recent years.  The cleanest formalization of this idea, that is, the formalization that is the least cluttered with extraneous ideas and notation, is the 20th century one of potential outcomes.  Here the causal effect of an intervention relative to no intervention is defined by the comparison of two potential outcomes, the first, which would be observed if the intervention is implemented, and the second, which would be observed if the intervention is not implemented; either of these potential outcomes is observable depending on whether the intervention is implemented, but both potential outcomes cannot be simultaneously observed because we cannot go back in time to “undo” what was done.  This understanding is intellectually related to analogous concepts in quantum mechanics, where quantities are defined at the same instant of time yet are not simultaneously observable, e.g., a particle’s position and its momentum.  Adding clutter to this understanding may be helpful for communication in some situations, but it is wise to avoid clutter when focusing on basic definitions.

Single World Intervention Graphs (SWIGs): A Unification of the Graphical and Counterfactual Approaches to Causality with Applications

James M. Robins, Harvard School of Public Health
Friday, September 10 at 10:30 a.m. EDT

Counterfactuals (aka Potential Outcomes) are extensively used in Statistics, Political Science, and Epidemiology for reasoning about causation. Causal directed acyclic graphs (DAGs) are another formalism used to reason about causation extensively used in Computer Science, Bioinformatics, Sociology, and Epidemiology. In this talk I will show how these two approaches can be unified through a new type of causal graph: the SWIG. SWIGs enable researchers from “graphical” and “counterfactual” disciplines to seamlessly learn from and communicate with one another, thereby speeding up the “causal revolution” of all disciplines.  I will describe the utility of SWIGS in substantive applications in the biomedical sciences. 

What is Causal Inference? – A Logical Perspective

Judea Pearl - UCLA
Friday, September 17 at 1:30 p.m. EDT

The purpose of this talk is to explain the role of causal inference in the context of growing interests in machine learning and data science.

I will treat causal inference as a new branch of logic, thriving upon its own semantics, grammar, and computational tools, and capable of quantifying its own capabilities and limitations. I will then demonstrate how the new logic has changed the thinking in many of the sciences and how practical problems relying on causal information, including challenges in machine learning, can now be solved using elementary mathematics.

Background material:
https://ucla.in/2ZLRnyw
https://ucla.in/3iEDRVo
https://ucla.in/2HI2yyx

Statistical Learning: Causal-oriented and Robust

Peter Bühlmann, ETH Zürich
Friday, September 24 at 10:30 a.m. EDT

Reliable, robust and interpretable machine learning is a big emerging theme in data science and artificial intelligence, complementing the development of pure black box prediction algorithms. Looking through the lens of statistical causality and exploiting a probabilistic invariance property opens up new paths and opportunities for enhanced robustness and external validity, with wide-ranging prospects for various applications.

Panel Session: Causal Inference

Chair: Professor Jun Xie (Purdue University)
Panelists:  Professor Peng Ding (University of California, Berkeley), and Professors Chuanhai Liu, Nianqiao Phyllis Ju, Fei Xue (Purdue University)
Moderators: Professors Vinayak Rao, Arman Sabbaghi (Purdue University)
Friday, September 24 at 11:30 a.m. EDT

The panelists in this session will offer their summaries and perspectives on the causal inference frameworks and models presented by Professors Peter Bühlmann, Judea Pearl, James Robins, and Donald Rubin.  They will also compare and contrast the differing causal inference frameworks based on theoretical and applied considerations motivated by their own research in statistics, machine learning, and data science.