I have developed a 3-day workshop on causal mediation analysis with longitudinal data. This workshop was taught for the first time in Helsinki, Finland, May 7 to 9. The master’s students, PhD students, and postdocs taking the course (mainly sociologists and demographers, but also a few epidemiologists and psychologists) evaluated the course very well: it received an average overall score of 4.5 out of 5. Students especially praised the clear course material (e.g. annotated R-code) and teaching style. I am available and quite happy to teach this workshop at other universities or institutes, and can tailor the workshop to the needs of the specific group of participants. Please contact me if you are interested.
Mediation analysis refers to the techniques used to investigate the causal mechanisms or pathways by which a determinant affects an outcome. For example, to what extent does education affect a woman’s risk of childbirth directly, and to what extent indirectly by first affecting her employment career? To answer such a question, we take concepts and methods from the counterfactual causal inference framework (also known as the potential outcomes framework), which is a common framework in biostatistics and epidemiology and now becoming more popular in the social sciences.
The course starts with an introduction to Directed Acyclic Graphs (DAGs), distinguishing mediation, confounding (including confounders that are also mediators), and colliding. Then covers various concepts of effect decomposition that exist for mediation, such as natural (in)direct effects and controlled (in)direct effects. Afterwards, methods for estimating these concepts will be discussed, including a brief comparison with more traditional approaches (e.g. KHB, Baron & Kenny, Oaxaca-Blinder). We will look at mediation both in cross-sectional and in longitudinal settings, and at settings with multiple mediators. The course will not focus on analytical methods or solutions for mediation. Instead, we will focus on estimation methods that use empirical data in conjunction with simulation; these provide both a more intuitive understanding of the process under investigation, and a more flexible approach to mediation. This approach provides a general solution to many mediation settings, including settings with mixed variable types, i.e. where variables can be continuous, categorical, count, and/or binary. To this end, it is important that participants have basic knowledge of the R-programming language. However, references to various STATA packages, offering similar (but less flexible) solutions, will be made. The course ends with the topic of mediation analysis in dynamic longitudinal settings using the parametric g-formula.