Mediation analyses involve looking at whether the effect of an exposure on an outcome works through one or more other variables (the mediators). For example, we might be interested in evaluating the effect of childhood socioeconomic status (the exposure) on later life income (the outcome) and whether this possibly works through the child’s adult occupation (the potential mediator); other mediators might include the child’s educational attainment or adult mental health. Mediation analyses are inherently causal because the mediation question sets out to explain a mechanism through which the exposure causally operates to affect the outcome.
Mediator variables are also known as intervening variables or process variables. This is because they provide insight into how the exposure is working, which allows us to identify potential interventions and evaluate whether associations reflect an injustice. For example, inequities are often described as health differences that are due to social injustice. Differences in health between men and women mediated by gender discrimination would be inequities (e.g., differences in salaries between men and women with equal qualifications, i.e. the Gender Pay Gap), but effects mediated by biological mechanisms (e.g., genetic differences in risk of experiencing breast cancer) would not have the same equity implications. Thus, mediation analyses can indicate which factors to intervene on to reduce or eliminate an inequity.
Putting Evidence Into Practice
In one E4A-funded study, the investigators examined the physical and mental health effects of several social policies, including the Deferred Action for Childhood Arrivals (DACA) program. The investigators found that DACA-eligible individuals experienced a reduction in psychological distress compared with DACA-ineligible individuals. Moreover, investigators found that DACA eligibility was also associated with increased enrollment in the Women, Infants, and Children (WIC) program. While not tested in this research project, the results suggest that we might consider WIC enrollment as a potential mediator of the effect of DACA eligibility (the exposure) on psychological distress (the outcome). That is, WIC enrollment might be one process through which DACA eligibility reduces psychological distress among those eligible for WIC. If WIC was found to be a mediator, it would offer one mechanism that we could intervene on to reduce psychological distress (e.g., campaign to increase WIC enrollment among undocumented immigrants). Social justice might then be advanced by preventing disproportionate levels of psychological distress among undocumented immigrants.
Example Research Questions
Mediation analyses that might be of interest when evaluating DACA eligibility would include:
- Does DACA eligibility reduce psychological distress for undocumented immigrants because it leads to increased use of the WIC program, which in turn decreases psychological distress? Or are other mechanisms, such as economic opportunities and protection from deportation, more important?
A comprehensive understanding of mechanisms would allow us to answer subtly distinct questions such as:
- In a setting where the effect of DACA eligibility on WIC enrollment was eliminated, but different people still enroll in WIC with varying frequency, how large would the remaining effect of DACA eligibility on psychological distress be?
- In a setting where the effect of DACA eligibility on WIC enrollment was eliminated because everyone enrolled in WIC, how large would the remaining effect of DACA eligibility on psychological distress be?
The answers to each of these questions may differ, and depending on the universe of feasible policy responses, one question might be more relevant than others.
The most common method to conduct mediation analysis is the Baron and Kenny approach. In this approach, the effect of the exposure on the outcome is modeled, first in a regression without the mediating variable (to quantify the total effect of the exposure on the outcome) and second in a regression including the mediating variable (to quantify the direct effect of the exposure on the outcome that does not act through the mediator). The difference between the two effect estimates is considered the effect of the exposure on the outcome that acts through the mediator (the indirect effect). However, this approach is only valid if there is no uncontrolled confounding between the mediator and the outcome, there is no interaction between the exposure and the mediator, and the measures of effect are linear contrasts (e.g., a coefficient from a linear regression).
Modern causal mediation approaches have been developed based on counterfactual or potential outcome frameworks (e.g., g-estimation and inverse probability weighting; see our blog on time-varying confounding and resources below for more details) to overcome challenges faced by the traditional Baron and Kenny approach. These methods contrast potential outcomes under hypothetical interventions on the exposure and mediator. For example, the controlled direct effect of an exposure on an outcome is a contrast between counterfactual outcomes when fixing a mediator to a specific value (for further details see resources below on causal mediation and Dr. Justina Avila-Rieger’s talk on Facebook).
Tools and Resources
Mediation analyses can offer more nuanced insights into the mechanisms through which programs, policies, or practices are affecting outcomes, which can help better target changes to achieve racial equity. For more on mediation see the following online resources and journal articles. Also, keep a watch out for the findings of the E4A-funded Physical Activity and Redesigned Community Spaces (PARCS) Youth Cohort in which the researcher team plans to explore the role of mediators of the effect of a citywide park redesign and renovation, on physical activity, park usage, psychosocial and mental health, and quality of life in underserved neighborhoods.
- Introduction to Mediation
- Causal Mediation
- David Kenny Blog
- An Interactive Tool for Mediation Tests
- Dr. Justina Avila-Rieger’s talk on Facebook (starting at minute 8)
- Lange, T., Vansteelandt, S., & Bekaert, M. (2012). A simple unified approach for estimating natural direct and indirect effects. American Journal of Epidemiology, 176(3), 190-195.
- Hafeman, D. M., & Schwartz, S. (2009). Opening the Black Box: a motivation for the assessment of mediation. International Journal of Epidemiology, 38(3), 838-845.
- VanderWeele, T. J. (2009). Marginal structural models for the estimation of direct and indirect effects. Epidemiology, 20(1),18-26.