Experiments and Causal Inference
Experimental design / Statistical analysis / Communicating results /
Cleaning data / Mining and exploring data
The most interesting decisions we make are decisions where we believe the input will change some output: redesign a customer experience to increase retention; advertise to users using this message to increase conversions; enroll in UC Berkeley data science to learn. And yet, most data is ill equipped to actually answer these questions.
This course introduces students to experimentation and design-based inference. Increasingly, large amounts of data and the learned patterns of association in that data are driving decision-making and development in the marketplace. This data is often lacking the necessary information to make causal claims.
This course teaches how to collect data in a way that is creative, forward-looking, and contains information that is necessary to know that the relationship is causal. First, students build a basis for understanding when a relationship has a causal meaning and when it has only an associational meaning. Faculty equips students with the tools to avoid making inappropriate decisions based on association-only data. Second, the course introduces the analytic framework of potential outcomes; this formal system of thinking serves as the basis for causal claims across all decision domains. Third, students develop the theoretical and technical skills to estimate causal quantities using randomization inference and regression. Fourth, students examine the common problems in implementation with data-driven case studies and an extended student project that collects original data.