Re: st: Would _sgmediation_ work for dichotomous dependent variable?

[Philip Ender <ender97@gmail.com>]

Bo MacInnis <bo@macinnis.org> wrote:

Another question on effect size: I am an economist and trained in
being interested in marginal effects rather than coefficients in the
nonlinear model like logit or probit but I am working on a project for
a psychologist. Would the mediational analysis be extended to marginal
effect rather than coefficients in the probit model? e.g., the
marginal effect of X on Y is mfx_c, the marginal effect of mediator M
on Y is mfx_b, and the marginal effect of X on mediator M is mfx_a,
would it make sense if I calculate the effect size based on the ratio
of these marginal effects, rather than coefficients?

-----

One way of estimating indirect effects for OLS regression to use the
product of coefficients method.  David Kenny
(http://davidakenny.net/cm/mediate.htm) discusses a method for using
the product of coefficients for models with dichotomous variables.
One of the concerns for these types of models is that OLS coefficients
are scaled very differently from logit or probit coefficients.
-binary_mediation- accounts for this by standardizing all of the
variables.  Continuous variables are standardized in the normal way
while the binary variables are standardized using the variances of the
latent variables. So yes, standardization is required for this
approach.

I don't think the product of coefficients approach generalizes the
marginal effects.

I should also point out that -binary_mediation- does not directly
compute standard errors like -sgmediation- but makes use of the the
bootstrap command.

-- 
Phil Ender
UCLA Statistical Consulting Group
*
*   For searches and help try:
*   http://www.stata.com/help.cgi?search
*   http://www.stata.com/support/statalist/faq
*   http://www.ats.ucla.edu/stat/stata/