Applied Causal Inference

This course begins with an introduction to causal inference adapted for applied science (see description below, several weeks, including practice slots). Students then work in groups to select a small project and develop a project plan over several weeks. They are supervised during the lecture and practice slots. The groups then present their project plan to the instructor and receive feedback. Following this, the projects are implemented and a poster is created. The last slots of the semester are reserved for poster presentations, with each group member presenting a portion of their project.

Causal inference deals with the detection and quantification of causal relationships from observational data and model assumptions. A causal effect is defined as a change in a variable Y when an intervention is made in a variable X and the goal of causal inference is to learn such effects from purely observational data without manipulating the system. When quantifying causal effects, the model assumption usually consists of qualitative knowledge about causal dependencies in the form of graphs over nodes (X,Y,Z,...) with directed arrows indicating causal relationships, and the goal is to use this qualitative graph to quantify the precise influence of a node X on a node Y. When detecting causal relationships, i.e., reconstructing causal networks (graphs), more abstract assumptions about the underlying processes come into play, for example, that direct, indirect, or common-cause connections are also reflected in statistical dependencies, and vice versa. Causal inference then addresses not only algorithms that determine when a causal effect can be calculated (identified), but also the practical problem of statistical estimation. The initial lectures will focus on time series as the underlying data and will mainly explain the concepts and illustrate them with lots of Python tutorials.

Prerequisites: Some experience with Python (loading data, plotting with matplotlib, numerical packages such as numpy and optionally sklearn and tigramite).