Epidemiological Methods for Causal Mediation Analyses

Course Number
5400 (Epidemiology)
Course Instructor(s)
Peter M Smith, Selahadin Ibrahim, Olli Saarela

Examining the chain of relationships between measures is central to many observational investigations. For example, researchers may be interested in the relative importance of different pathways through which an independent variable might impact on an outcome, in order to determine which variables have the greatest potential for future interventions. A variable located on a causal pathway between two variables is referred to as a mediating variable. While methods to examine mediation have been around for decades in the social and psychological sciences they have not been extensively used in the field of epidemiology until more recently. In addition, common methods to assess mediation (e.g. using a series of regression models with and without the mediating variable of interest) are not always appropriate when the outcome of interest is not normally distributed (e.g. for dichotomous and time-to-event outcomes), and these situations are more common in epidemiology.

The purpose of this course is to introduce students to conceptual issues and analytical approaches to assess mediation in observational study designs. At the completion of the course students should have developed an understanding of the merits of different statistical approaches to assess mediation, and the theoretical and statistical assumptions behind these approaches. In turn, they will be able to determine the most appropriate approach to assessing mediation, given their data source, available measures and research question.

Students taking the course will require a basic understanding of confounding, mediation and the principles of measurement such as reliability and validity, as this knowledge will be assumed in the course materials. It is also assumed that students will have experience in running linear and logistic multivariable analyses.  When examples are given in class, they will be provided with either R, SAS or Mplus. Students are free to use other statistical programs throughout the course, but the instructors may not be able to provide assistance with programming errors encountered.

The course will begin with an overview/review of key concepts in causation and mediation. We will then introduce methods to derive directed acyclic graphs (DAGs) for research questions, and briefly outline how DAGs can be used to better understand questions of bias (including confounding) and mediation. The course with then focus on two approaches to assess mediation: path and structural equation modelling; and counterfactual approaches. Each of these approaches allow researchers to estimate the total effect, and the indirect and direct effects, between an independent variable and an outcome of interest. However, they differ in key ways, which will be discussed as the course progresses.