Advanced Causal Mediation Analysis

Causal mechanism is a fundamental component of social science theories. Social scientists are not only interested in whether one variable causally affects another, but also how such a causal relationship arises. However, most of the existing empirical strategies aim at merely establishing causal effects between variables, resulting in a “black box” approach to causality. This course introduces students to the concepts and techniques that help empirical researchers to unpack the black box of causality and formally evaluate the process in which a causal variable of interest influences an outcome.

On the first day, the class will begin with an introduction to causal inference under the Rubin causal framework, or potential outcome framework. Students are introduced to concepts such as potential outcome and average treatment effect (ATE) that are necessary to understand causal mediation. With these foundations, we will then conceptualize causal mechanism as a decomposition of the causal effect into a direct effect and an indirect effect, or causal mediation effect (ACME). We then move on with the assumptions required for the identification of ACME. This is the most crucial part of causal mediation analysis as the validity of estimation by and large depends on the extent to which the assumptions hold. For that, we introduce the idea of sensitivity analysis, which allows us to evaluate the robustness of the results to potential violations of the identification assumptions. As you can see, Day 1 is filled with theoretical and conceptual discussions on causal mediation.

What then are the methods used to identify causal mediation effect? Day 2 introduces the experimental approach to mediation analysis. After Day 1’s lesson, you will understand that both the treatment and the mediator variables need to be randomized for the mediation effect to be identified. Thus, Day 2’s lesson introduces several experimental designs that help researchers identify mediation effect. Compared with the conventional experimental design, the key difference here is that the experimentalist needs to manipulate both the treatment and the mediator. Some of the designs require perfect mediator and others can be used even with imperfect manipulation. Examples of social science experiments are used to illustrate the key ideas that underlie each of the designs.

Day 3’s lesson focuses on mediation analysis using observational data. As you will see, observational data differs from experimental data in that the data generating process is not randomized for the former. Therefore, the researcher can manipulate neither the treatment nor the mediator in observational studies. The most important message of Day 3’s lesson is that observational studies should use experimental designs introduced in Day 2 as templates. We will see, for example, how the cross-over (experimental) design can be extended and applied to observational studies. Other strategies that exploit instrumental variables and interaction terms will also be introduced.

Day 4’s lesson has two parts. The first part looks a little bit into the technical aspect of estimation. We will study the structural equation framework and the Baron-Kenny procedure in establishing causal mediation. We will see that a major limitation of the linear structural equation model is that it cannot be directly applied to mediators and outcomes that are measured with discrete variables. Against this background, students are introduced to a newly developed general algorithm that can be applied to any type of data. The second part is a practical session whereby students will learn to use the “mediation” package in R. Students may carry out causal mediation analysis with their own data.

On the last day, students will present their research or research proposals that employ causal mediation analysis. Students will receive feedbacks from the instructor and their peers on their research design and data analysis, if any.

This course builds on and makes use of several fields and techniques in social science methodology. Students should possess basic knowledge and skills in some – if not all – of the following areas: 1) causal inference and Rubin causal framework; 2) experimental methods; 3) regression analysis (and preferably structural equation modeling), 4) instrumental variable method, 5) Bayesian statistics and 6) basic skills in R.

Students are not discouraged from signing up for this course if they do not fully meet the above criteria. However, they are strongly encouraged to take the relevant courses offered by the Summer School before or concurrently with this course. Given that this is an advanced level course, the instructor may adjust the content depending on the general background of the participants.