Statistical Methods for Causal Inference and Prediction

Clinical trials in biomedical research aim to determine whether a treatment is effective. Do patients receiving a COVID vaccine fare better than those in the control group who do not receive it? Similarly, A/B tests in online business research seek to evaluate whether a new website design boosts sales: Do visitors viewing page version A spend more than those viewing version B? Social scientists ask causal questions, too: When households face real-time electricity prices that vary by the hour, do they really use less electricity when the price is high? By how much does a feed-in tariff promote investment in sustainable energy sources? Does raising the minimum wage increase unemployment? Does banning digital devices in classrooms improve student performance?

In all these cases, the goal is to estimate a causal effect, and the preferred method is the randomized experiment. However, not all experiments researchers wish to conduct can be done or should be done – there are practical and ethical constraints. As a result, researchers make do with data collected through passive observation. But this raises a challenge: How can you distinguish causation from mere correlation?

The adage “correlation does not imply causation” is true but not always actionable. Scientists and practitioners strive to go beyond correlations to make robust causal claims. Over the past three decades, the “causal revolution” has reshaped empirical research in the health, social, and behavioral sciences. By leveraging counterfactual thinking and tools like causal diagrams, researchers can draw credible causal inferences from observational data.

The modern causal inference toolkit is widely used not only in science but also in business and public policy. Uber and many other companies routinely apply causal methods to enhance customer satisfaction and achieve their strategic goals. During the COVID crisis, causal epidemiological studies based on observational data guided critical policymaking. Governments and international organizations, as part of the evidence-based policy making agenda, conduct ex-post impact evaluations to assess whether policies exert causal effects on outcomes of interest.

This master-level course equips students with the theoretical foundations and practical tools needed to quantify cause-and-effect relationships in complex socio-technical systems. It covers the linear regression model and ordinary least squares estimation, the potential outcomes framework, causal diagrams, and instrumental variables estimation.

The course teaches the statistical theory upon which the statistical methods are based. To build your understanding, the ideas are illustrated with example applications drawn from recent research in different domains (e.g., transport, energy, health) and scientific disciplines (e.g., economics, political science, and epidemiology).

You will acquire coding skills along the way as a by-product of working through the examples and doing a replication. A replication, loosely speaking, is a follow-up study with the goal of evaluating the validity of a previous scientific study. You will replicate the results of a published scientific article. This means downloading the raw data, processing it, and analyzing it to check whether the published results hold up to scrutiny.

By default, the final exam determines your final grade.

The replication group assignment is non-graded but participation is mandatory, i.e., in order to pass the course, you have to enroll in a group and submit your work.

The optional mid-term assessment can only improve your final grade but not worsen it. If your mid-term grade is higher than your final exam grade, it will count 25 percent toward your final course grade. If you do not participate in the mid-term assessment or if the mid-term grade hurts you, the mid-term will not count. The mid-term assessment is a bonus opportunity and offered only once – you cannot re-take it.

To sum up: depending on your participation and performance in the final exam and the mid-term, the percentage composition of your final grade will be 100-0 (final exam, mid-term), or 75-25.

The mid-term and the final exam are closed-book assessments in the sense that no textbooks or other published learning materials are allowed. However, you may bring your own notes. You are allowed to bring up to 10 double-sided pages (A4) of printed or handwritten notes.