Causal modeling with many experiments

Sean Taylor, PhD, Lyft, Research Science Manager

Causal Modeling with Many Experiments

Causal inference is often developed in the context of binary treatments and one-time, discrete decisions. Real business decisions can be dramatically more complex — policy spaces can be high dimensional and effects can interact or vary over time. Directed acyclic graphs (DAGs) provide a convenient tool for understanding causal relationships, but it is not straightforward to produce counterfactual estimates from graphs without additional assumptions and interventional data. In this talk I will present some (in progress) work on fitting structural causal models in order to estimate complex counterfactuals. I will show how we can conveniently combine human domain knowledge with results from many experiments to identify a model suitable for studying “what if” scenarios and for policy optimization.

Bio: Sean J. Taylor is a research scientist manager at Lyft, where he works on causal inference, experimentation, and structural modeling. Previously, he led the Statistics team within Facebook’s Core Data Science team, where he worked on experimentation, survey modeling, forecasting, and machine learning. Prior to Facebook, he earned his PhD in Information Systems from NYU’s Stern School of Business as well as a BS in Economics from Wharton School. He specializes in using machine learning methods and randomized experiments for measurement, prediction, and policy decisions.  Sean’s research spans a wide range of topics: online social influence, social networks, statistical methodology, causal inference, and Bayesian modeling.  He is also an avid engineer who enjoys putting research into practice by building tools and services.