Specification of outcome model using mediate

[Janine Stubbs]

Dear Statalisters,

In my understanding, causal mediation analysis involves estimating coefficients (lower case letters) for;

1. the outcome model in which the mediator (M) is conditioned upon, to give the direct effect of the exposure (A) on the outcome (Y)

e.g. E[Y] = a + bA + cM, where the direct effect = b

1. the mediator model of the exposure (A) on the mediator (M)

e.g. E[M] = d + eA

Following these, the total effect of A on Y can be partitioned into direct (A -> Y) and indirect (A -> M -> Y) effects.
e.g. if the total effect is:

E[Y] = f + gX = a + bA + cM = a + bA + c(d + eA) = a + bA + cd + ceA = a + cd + (b + ce)A

Such that g = b+ce

And if the direct effect = b, then the indirect effect = ce

If the above is correct, then why does the Stata documentation for the mediate not require that the outcome model must include the mediator? Does this command automatically condition the outcome model on the mediator?

Thanks,

Janine

[Janine Stubbs]

After closely reading the Stata documentation, I see that above I have described the Baron and Kenny (1986) "product method" approach to mediation analysis, which is not the basis of mediate.

mediate uses a potential outcomes framework, whereby coding for the inclusion of covariates proceeds as follows:

1. the outcome model includes confounders of the treatment -> outcome and mediator -> outcome models (and includes baseline outcome in the case of imbalance)
2. the mediator model includes confounders of the treatment -> mediator model (and includes baseline mediator in the case of imbalance)

Janine