Categories → Causal Inference , Method
Mediation is a statistical technique used in causal inference to understand the mechanisms by which a predictor variable (X) affects an outcome variable (Y).
Mediation is the act of modifying or relaying a causal effect.
If the effect of a Treatment variable X on an Outcome variable Y occurs indirectly, through an intermediate variable (M), then M is known as a mediator variable.
M "mediates" the effect of X on Y.
Mediation can be full (the entire effect is produced only via M) or partial (there are other paths from X to Y not via M, possibly including a direct path from X to Y).
Mediation analysis aims to disentangle total treatment effects into direct and various indirect effects so that these relationships can be understood. It can also be used in observational studies, and to estimate causal effects.
If a mediating variable fully mediates the effect of Treatment on Outcome, and there is no backdoor path from treatment to outcome, the frontdoor adjustment can be used to identify (and consequently estimate) Causal Effects. This is obviously incredibly useful where no backdoor adjustment exists.
The process of exploring mediation usually involves the following steps:
Identify the causal relationship between X and Y: Before conducting mediation analysis, it is important to establish a causal relationship between X and Y. This can be done using various causal inference techniques such as randomized controlled trials, natural experiments, or observational studies with appropriate statistical adjustments for confounding variables.
Determine if there is a potential mediator: A mediator is a variable that explains how X influences Y. To determine if there is a potential mediator, one needs to assess whether there is a significant association between X and M, and between M and Y.
Test for mediation: To test for mediation, a statistical model is fitted with X as the predictor variable, Y as the outcome variable, and M as the mediator variable. The total effect of X on Y is decomposed into the direct effect of X on Y and the indirect effect of X on Y through M. The indirect effect is estimated using the product of coefficients method or the difference in coefficients method.
Assess the strength of mediation: The strength of mediation can be assessed using various measures such as the proportion mediated, which is the ratio of the indirect effect to the total effect of X on Y. The strength of mediation can also be assessed using the bootstrap method, which estimates the confidence intervals for the indirect effect.
Interpret the results: The results of mediation analysis can help to understand the mechanism by which X affects Y. If the indirect effect is significant, it suggests that the mediator variable M partially or fully explains the relationship between X and Y. If the indirect effect is not significant, it suggests that there is no evidence of mediation, and X affects Y directly.
In Causal Wizard, Mediator variables may provide a mechanism to identify and estimate a Causal effect via the Frontdoor adjustment. However, mediation analysis (another feature - analysis of the effect of an intermediate variable and its degree of responsibility for the causal effect) is not currently supported. Let us know if you need this feature!
In summary, mediation analysis is a useful tool in causal inference to understand the mechanisms by which a predictor variable influences an outcome variable through an intermediate (mediator) variable.