Estimating Causal Effects from High-Dimensional Observational Data

11 Oct
Thursday, 10/11/2018 12:00pm to 1:00pm
Computer Science Building, Room 150/151
Machine Learning and Friends Lunch

Abstract:  Everyone wants to make better decisions. The impact of a decision on an outcome of interest is called a causal effect, and is traditionally estimated by performing randomized experiments. However, large data sources such as electronic medical records or insurance claims present opportunities to study causal effects of interventions that are difficult to evaluate through experiments. One example is the management of septic patients in the ICU. This typically involves performing several interventions in sequence, the choice of one depending on the outcome of others. Successfully evaluating the effect of these choices depends on strong assumptions, such as having adjusted for all confounding variables. While many argue that having high-dimensional data increases the likelihood of this assumption being true, it also introduces new challenges: the more variables we use for estimating effects, the less likely that patients who received different treatments are similar in all of them. In this talk, we will discuss causal effect estimation and treatment group overlap through the lens of domain adaptation and off-policy reinforcement learning. We will introduce the potential outcomes framework, classical methods for estimating causal effects, as well as new ones, tailored for working with large datasets.

Bio:  Fredrik Johansson is a postdoctoral associate in David Sontag's Clinical Machine Learning Group at MIT. He completed his Ph.D. at Chalmers University of Technology, Sweden in 2017, working on machine learning methods for network data. His current research is focused on theory and methodology for estimating causal effects and learning policies from observational data, often inspired by problems in the medical domain.