Вы находитесь на странице: 1из 27

Mediation

Paul Jose
PSYC 325
Thought for the day:
You know the world is going crazy when
the best rapper is a white guy, the best golfer
is a black guy, the tallest guy in the NBA is
Chinese, the Swiss hold the Americas Cup,
France is accusing the U.S. of arrogance, and
Germany doesnt want to go to war.
Chris Rock
Are mediation and moderation cousins?

People are thrown off a bit by the fact that


moderation and mediation sound alike. It
makes it seem that they are very similar.
In actuality, they are somewhat similar.
Well, its confusing, I admit.
Statistically they are cousins, but they are
designed to answer different questions.
Similarities and differences

Similarities:
They both involve three variables;
You can use regression to compute both;
There is no statistics programme (other than ModGraph)
designed to tackle either one.
Differences:
You create a product term in moderation; not in mediation;
You dont have to centre anything in mediation;
Moderation can be used on concurrent data, but mediation
is best used on longitudinal data.
So lets focus on mediation now

The question you wish to answer is whether


the effect of the IV is at least partially
mediated by the mediating variable on the
DV.
You can answer this question with two
regressions (and a correlation matrix).
Lets consider a specific example.
Does rumination mediate the effect of
stress on depression?

Stress Depression

Rumination
The theories

Susan Nolen-Hoeksema believes that an individual


who does more rumination ends up with more
depression. Simple direct effect.
I dont disagree with her, but I think that this
simple effect should be embedded within the stress
and coping context.
In other words, stress affects both rumination and
depression. Can the effect of rumination on
depression be contextualized by examining the
larger model (previous page)?
Baron and Kenny (1984)

Baron and Kenny are the people responsible for


systematizing the research approach to
mediation.
They claim that three preconditions must exist
before one tests for mediation:
1. Significant correlation between IV and DV;
2. Significant correlation between IV and mediator; and
3. Significant correlation between mediator and DV.
Do we have data that meet these
preconditions?

.473***
Stress Depression

.481*** .475***
Rumination

Seems so. So now what?


The test

Baron and Kenny say that if the relationship


between IV and DV is reduced to non-
significance when the potential mediator is
included in the regression equation, then
mediation has been demonstrated.
This is somewhat misleading, but I will
show you what they mean.
Hierarchical regression again

First step: regress depression on stress (emuch).


Second step: add the mediating variable to the
equation (rumination).
Check to see whether the beta weight of stress
goes down at the second step. If it does, then you
have mediation. Easy to tell when the beta
becomes non-significant, but what if it stays
significant (see next page)?
Results of regression

Coefficientsa

Unstandardized Standardized
Coefficients Coefficients
Model B Std. Error Beta t Sig.
1 (Constant) -.560 .031 -17.863 .000
EMUCH 3.117E-02 .001 .476 23.439 .000
2 (Constant) -1.374 .064 -21.467 .000
EMUCH 2.116E-02 .001 .323 14.669 .000
RUMINATE 4.253E-02 .003 .316 14.363 .000
a. Dependent Variable: CDIZ
Different levels of mediation

There are actually three types of mediation:


No mediation (no change in beta);
Partial mediation (a significant drop in beta, but it stays
significant); and
Full mediation (a significant drop in beta, and it
becomes non-significant).
The first type is quite frequent; the second type
occurs occasionally; and the last type is as rare as
hens teeth.
What we have here is the second type.
But how do we know that its a significant
drop?
Well, one uses a test called the Sobels t-test.
You can find it in stat texts, but you can also find
it on-line at Preacher and Leonardellis web-site:
http://www.unc.edu/~preacher/sobel/sobel.htm
The programme computes whether the drop in
beta (or B) is significant or not.
You have to give it very specific information to
obtain a result, however. You have to compute two
regressions (see next page):
Two regressions

1. First regression: regress the mediator on the IV.


Take the unstandardized regression coefficient
(the B, not the ) and the standard error (se).
2. Second regression: regress the DV on both the
IV and the mediator simultaneously, and take the
B and se for the mediator on the DV.
3. Then you can input these four numbers into the
window, and it will compute Sobels t for you.
4. So lets do it.
First regression

a
Coefficients

Unstandardized Standardized
Coefficients Coefficients
Model B Std. Error Beta t Sig.
1 (Constant) 19.018 .224 84.744 .000
EMUCH .235 .010 .481 24.659 .000
a. Dependent Variable: RUMINATE
Second regression

a
Coefficients

Unstandardized Standardized
Coefficients Coefficients
Model B Std. Error Beta t Sig.
1 (Constant) .222 .496 .449 .654
EMUCH .164 .011 .323 14.669 .000
RUMINATE .329 .023 .316 14.363 .000
a. Dependent Variable: CDI
Input

B (IV to rum.): .235


B (rum. to DV): .329
SE (IV to rum.): .010
SE (rum. to DV): .023

Sobels t = 12.22, p < .00001


So what does it mean . . . ?
Well, I have several problems with this web-site,
despite the fact that it computes the t quickly and
accurately:
You dont know what kind of mediation you have: full
or partial.
You dont know what the new beta weight is.
So Ive written something called MedGraph that
will answer all of those questions, plus more.
Youll probably eventually see advertisements on
late-nite TV for ModGraph/MedGraph (like the
vegetable slicers, knives that cut through steel,
CDs for classic rock music, etc.).
MedGraph will do this: ta-da!

.473***
Stress Depression
(.323**)

.481*** .475***
Rumination (.316**)

Sobels t = 12.22, p < .00001; partial mediation; indirect effect = .120


Now, isnt that much better?
You can see in a glance whether and how much the
beta decreased.
You can see what the Sobels t value is, and
whether it is significant.
You will learn whether you have no, partial, or full
mediation.
You will see an estimate of the indirect effect (how
much variance in the DV does the IV explain
through the mediator).
It doesnt slice and dice vegetables yet, but Im
working on that.
So is everybody clear on mediation
and moderation now?
Let me summarize both, and give you a chance to ask
questions.
Both approaches are used to explore the
interrelationships among 3 variables.
Two variables? Do a simple correlation.
Having three variables means that one can examine
their various relationships in more complicated ways.
Mediation and moderation are tests of association, but
structured so that particular questions can be
answered.
Why do moderation and mediation?
Researchers are increasingly using these
approaches because they wish to
understand mechanisms of how
variables affect each other.
However, most researchers:
1. Are confused about what mediation and
moderation do;
2. Use one or the other incorrectly;
3. Overuse one or the other (they have a
favourite approach); and/or
4. Misinterpret the results.
Structural equation modeling

Some researchers think that mediation is the


pinnacle of statistical achievement in an
investigation of mechanisms. Wrong.
Mediation is a special case of path analysis,
and more researchers should be examining
mechanisms by including more than 3
variables in their analyses.
Here is an example.
The case for path analysis
I predicted that four coping strategies might
mediate between Individualism and/or
Collectivism and Negative adjustment.
I and C are higher order constructs descriptive of
how people relate to others in their particular
culture.
Western societies are high in I, and Asian societies
are high in C, according to Harry Triandis.
So I could conduct 8 separate mediation analyses
(I to Rumination to NegAdj; etc.) to see whether
mediation occurred.
The problem is that this inflates Type 1 error.
The answer: Path analysis
-.07*

Rumin.
.16* .65**

Individ- .14*
Extern. .29*
ualism -.23*
Neg.
adjustment
.13*
Collect- .23* Problem-
ivism solving How many
-.20* mediations are
.44** shown here?
Social
support Answer: 4
Why is this better?
Its better insofar as one is considering all of these
variables together.
Whenever one does an associational analysis
(correlation, regression, etc.), one must consider
the third variable problem.
When one does a mediation analysis with three
variables, one should think about whether one has
included all relevant variables.
Inclusion of other variables can wash out (or
intensify) relationships identified with a simple
mediation.
The real world has more than 3 variables in it.
The world of mediation and moderation
The distinctions between these two techniques
should be clearer for you now. For more
information check out my Help Centre. It
presents some clear examples.
Confidence in this area is obtained by doing both
techniques on the same dataset (permissible), and
examining the results.
Keep experimenting and pushing the limits.
And watch for advertisements for ModGraph and
MedGraph!

Вам также может понравиться