Download as pdf or txt
Download as pdf or txt
You are on page 1of 4

SIGNIFICANCE

TESTING THESTATISTICAL
and
t ehe Sobel test (Sobel, 1982, 1986) and its variants, the Arojan test (Aroian, 1947)
the Goodman test (Goodman, 1960). Partly bccause the sampling distribution of the
indirect effect tends to be somewhat skewed rather than meeting the assumption that
received
it is normal (causing the tests to lose some statistical power), these tests have
some criticismin the professional community(e.g., Hayes, 2009; MacKinnon, Lockwood,
Hoffman, West, & Sheets, 2002; Preacher & Hayes, 2004). Nonetheless, these tests are
renorted extensively in the research literature and we will apply one of the members of
this family, the Aroian test, here, in that it was the variation of the Sobel family of tests
popularized by Baron and Kenny (1986).
The cquation for the Aroian test is presented in Figure 40.8 together with a summary
of the calculations. All three tests comprising the Sobel evaluations are computed in a
similar manner and differ only in how the expression (SE,* SE,') at the end of the
square root expression in the denominator of the equation is treated:

. The Arojan test adds the expression to the other terms under the square root sign
in the denominator.
. The Goodman test subtracts the expression from the other terms under the square
root sign in the denominator.
. The Sobel test does not include the expression under the square root sign in the
denominator.

In the equation shown in Figure 40.8, the letters represent the unstandardized regres
sion coefficients associated with the paths shown in Figure 40.7, and the SE values of the
unstandardized regression weights are shown, with the subscripts indicating the reference
coefficient. The outcome of the equation is a z value with an alpha level of .05 indicated
by a value of 1.96 or better. Here, we obtained a value of approximately 5.79 with a hand
calculator. It is also possible to use the calculator available on Kristopher J. Preacher's
web site http://quantpsy.org/sobel/sobel.htm, which yields the same approximate value.
The value of 5.79 is sufficient to reject the null hypothesis that the indirect path is not
different from zero; instead, we conclude that the indirect effect of beckdep through
regard on negafect is statisticaly significant.

40.9 TESTING THE STATISTICAL SIGNIFICANCE


OF THE DIFFERENCE BETWEEN THE DIRECT PATHS
IN THE UNMEDIATED AND THE MEDIATED MODELS

The Freedman-Schatzkin test (Freedman & Schatzkin, 1992) compares the relative
strengths of the paths from the independent variable to the outcome variable in the
unmediated model versus the mediated model. The equation is shown in Figure 40.9
together with a summary of the calculations; it produces a tvalue tested against a Student

(d"e)
The Aroian test:z =

Ve'sE' +d'*sE,'+(SE SE,))


(-.814) *(-.110)
Aroian z=

I-10' (.062))+ [(-.814)' *(.01z'] +[(062)' *(.o17)1


FIGURE 40.8
Aroian z= 5.7918704 The Aroian (est.
SIMPLE MEDIATION

The Freedman-Schatzkin test: t =


V(SE,?+SE,) -(2*SE," SE,"V(1-r))

(.276) -(.186)
Freedman-Schatzkin t=
(L022)' +(.025)) - (2) *(.022)" (.025)V0-S395)
FIGURE 40,9
The Freedman-Schatzkin test. Freedman-Schatzkin t==6.6621265

tdistribution with N-2 degrees of freedom. Here, we obtained a value of approximately


6.66 with a hand calculator. With 419 degrees of freedom (our Nwas 421), it is sufficient
to reject the nullhypothesis that the coefficients are not significantly different (p <.001).
We therefore conclude that the mediated path coefficient is significantly lower than the
unmediated path coefficient, indicating that we have obtained a partial mediation effect.

40.10 DETERMINING THE RELATIVE STRENGTH


OF THE MEDIATED EFFECT

We now calculate the relative strength of the mediated effect using the
beta coeficients
that are associated with the paths in our mediation model. This is
of the strength of the indirect effect to the strength of the direct computed the ratio
as
as follows:
effect and is calculated

The strength of the indirect effect is the product of the beta


with paths beckdep to regard and regard to negafect coeffhcients associated
mediated model. Here it is equal to (-.539) * (-.312), or .168.(paths d and e) in the
The strength of the isolated direct effect is the beta
model (with avalue of .515) where beckdep is the coefficient in the unmediated
It can also be calculated as the sum of the single predictor of negafect.
for beckdep predicting negafect in the mediatedindirect eftect and the beta coeficient
.515. model (path : .168 +.347, 0r
. The relative strength of the
mediated effect is equal to the idirect effect divided
by the direct effect. Here it is equal to
.168/.515, or .326.
We may then conclude thal about a
third (32.6%) of the effect of
negafect is mediated through regard. beckdep
ASSESSING THE STATISTICAL SIGNIFICANCE OFTHE INDIRECT EFFECTS 395

age

-,31 -.11
.35 .33

-29
optimism
.53
quality of life

FIGURE 41.9
43
The path model with the path
SES coefficients and the R' values in
round-edged rectangles to the upper
right part of each endogenous variable.

41.8 PATH ANALYSIS USING


REGRESSION: SYNTHESIS MULTIPLE
The completed path model is shown in Figure 41.9. The
SES as well as the standardized path coefficients are correlation between age and
and the adjusted R values for the outcome variable placed in their respective paths,
quality _of_life and the endogenous
variable optimism are shown in round-edged rectangles near their upper
Based on the multiple regression results, researchers might be inclined right corners.
to trim this model
by rernoving the nonsignificant path from age to quality _of_life should it be
theoretically
reasonable to do so. Such an action stretches the use of path analysis from confirmatory
to that borders on exploratory, but some researchers would take that action based on
the
principle of parsimony.

41.9 ASSESSING THE STATISTICAL SIGNIFICANCE


OFTHE INDIRECT EFFECTS
Our visual inspection of the path coefficients suggests that both age and SES exerted
an indirect effect on quality_of_life through optimism as a mediating variable. The
statistical significance of the indirect effects can be evaluated by the Aroian (1947) test
as described in Chapter 40:

" Using the raw partial regression coefficients and SEs associated with the paths
age to optimism (b=-.634, SE=.112) and optimism to quality_of lite
(b=.190, SE=.021), we obtain (based on Kristopher J. Preacher's web site
htp:/quantpsy.org/sobel/sobel.htm) an Aroian z value of approximately -4.78,
p<.001.
" Using the raw partial regression coefficients and SEs associated with the paths
SES to optimism (b=3.187, SE=.399) and optimism to quality_of life
(b=190, SE=.021), we obtain (based on Kristopher J. Preacher's web site
http://quantpsy.org/sobel/sobel.htm) an Aroian z value of approximately 5.97,
p<.001.

We may therefore conclude that the indirecteffects of age and SES through optimism
in explaining quality_of life are statistically significant.
396
PATH ANALYSIS USINGMULTIPLE
REGRESSION
41.10 ASSESSING THESTRENGTH OF EACH INDIRECT EFFECT
We can assess the strength of each indirect effect by multiplying the
coefficients associated with each segment of each path. Thus, the absolutestandardized (heta)
indirect cffect of age through optimism on quality_of_life is -.31 *.53 = value of the
16. and the
value of the indirect effect of age through optimism on
quality_of_life
Both of these values are quite substantial for indirect effects and is 43 * 53 =23
being of practical wornh. Our general conclusion from this should be treated as
analysis
indirectly influence quality of life through the mediated influence of is that age and SES
optimism.
41.11 EVALUATINGTHE POSSIBILITY OF MEDIATION
Although the path from age to
in the model, it does raise the quality_of_life did not achieve statistical significance
interesting
ty_of_life in isolation. If that is the case, question
then
of whether age can predict quali
we may have observed full mediation
with the inclusion of optimism in the model. To
an additional regression analysis using age in test that possibility, we have performed
results, shown in Figure 41.10, indicate that ageisolation to predict quality_of_life. The
on its own does predict
(p<.001); the unstandardized regression coefficient is --248 (with SE quality_of_life
standardized regression =45) and the
coefficient is -.332.
41.12 TESTING THE STATISTICAL
OF THE SIGNIFICANCE
IN THE DIFFERENCE BETWEEN THE DIRECT PATHS
UNMEDIATED AND THE MEDIATED MODELS
Although it may be obvious what the result will be, it is
Freedman-Schatzkin test, as described in Chapter 40, to instructive to perform the
of the paths from age to quality compare
of life in the unmediated and the the relative strengths
When we solve the mediated modeis.
against a Student t Freedman-Schatzkin
distribution with
equation, we obtain a value of 8.699. Evaluated
or 242, degrees of freedom), our N-2 degrees of freedom (here we have 244-2,
result indicates that the coefficients are
different (p <.001). We may therefore conclude that we have obtained a fullsignificantly
effect (given that the path in the mediated
model is not statistically significant).mediation

Model Summary
Adjusted R Std. Error of Change Statistics
Model R Square R Square
Square the Estimate
1 .332 .110 .106 4.225
Change FChange dfl df2 Sig. F Change
.110 29.943
a. PrediCIors (Constant, age 1 242 .000

Coefficients
Standardized
Unstandardized Coefficients Coefficie nts
Model Sd. Error Beta
Sig,
Correlations
(Constarnt) 20.098 1.372
Zero-order Partial Part
14.649 .000
age .248 .045 .332 -5.472 .000 -.332 -.332 -,332
a. Depe ndent Variable: quality of ife
FIGURE 41.10 The Model Sunmary and Coefieients tables using age to predict
quality of life.

You might also like