Causal inference and machine learning for prediction

Course goals

• Students have developed knowledge and understanding of causal inference (Knowledge and understanding) 
• Students have developed knowledge and understanding of diverse causal inference frameworks, notably the Rubin causal model and structural causal graphs (Knowledge and understanding)
• Students have learned to apply techniques from causal frameworks to determine whether we can identify a causal effect (Learning skills)
• empirical data (Learning skills)
• Students have learned to apply techniques from causal frameworks to estimate causal effects from empirical data (Learning skills)
• Students can interpret the results obtained with their analyses (Applying knowledge and understanding)
• Students can critically reflect on the choices they made in their analyses (Judgment)

Content

Two topics that are often covered in this course are causal inference  and machine learning for prediction. Both prediction and causal inference are at the core of many research questions. In this course we teach you about both related topics in two parts.

Causal Inference part
Causal inference is about making inferences about whether and to what extent variables cause each other. Making such inferences is quite challenging, especially when due to ethical or practical limitations, we cannot do a randomized controlled trial. In the Causal Inference part of the course we introduce you to a set of tools that researchers can use to answer causal questions using non-experimental data. In other words, tools that they can use to learn about causation from correlation! In three lectures we take you through the basics of two causal inference frameworks, and some of the possible techniques from those frameworks. In the accompanying lab meetings we use R to practice the theory and apply the techniques.

Machine Learning for Prediction part
How can we go beyond classical statistical methods such as linear or logistic regression in order to achieve the best predictive performance using modern machine learning methods? Why and how do we use cross-validation? How does hyper-parameter tuning affect out-of-sample predictions? In the Machine Learning for prediction part of the course, you will learn how to justify, evaluate, and implement various modern machine learning techniques for supervised learning problems. Topics include random forest, tree-based boosting methods, (deep) neural networks. In addition, we discuss forecasting methods as a case of prediction-into-the future with their associated particularities and pitfalls. The course is framed in the statistical learning context, with an eye for data-generating processes, sampling, and the trade-off between bias and variance (or fit and complexity).