Professional Documents
Culture Documents
INDIRECT EFFECTS
IN STRUCTURAL
EQUATION MODELS
KennethA. Bollen
UNIVERSITY OF NORTH CAROLINA
Decomposing the total effects of one variable on another into direct and
indirect effects has long been of interest to researcherswho use path
analysis. In this paper, I review the decompositionof effects in general
structural equationmodels with latent and observedvariables. I present
the two approaches to defining total effects. One is based on sums of
powers of coefficientmatrices. The other defines total effects as reducedform coefficients. I show the conditionsunderwhich these two definitions
are equivalent. I also compare the different types of specific indirect
effects. These are the influencesthat are transmittedthroughparticular
I wish to thank Barbara Entwisle, John Fox, Clifford Clogg, and the
anonymous referees for valuable comments on previous drafts of this paper.
37
38
KENNETH
A. BOLLEN
Since its introduction into sociology, path analysis has been used
to decompose the influences of one variable on another into total,
direct, and indirect effects (Duncan 1966, 1975; Finney 1972; Alwin
and Hauser 1975; Greene 1977; Fox 1980). Several researchers have
generalized the decomposition procedures from observed variable models to structural equations with latent variables (Schmidt 1978; Graff
and Schmidt 1982; Jbreskog and S6rbom 1981, 1982). This latter
development opens up new applications. For instance, the total effects
of a second-order factor on indicators can reveal which measures are
most closely related to it. Or, if a measure has a factor complexity of
two or more, an effect decomposition can show which latent variable
has the greatest effect on the measure, not just the largest direct effect.
The first purpose of this paper is to review the decomposition of effects
for these general models.
The second purpose is to clarify the definitions of total effects.
Some have defined total effects as sums of powers of the coefficient
matrices; these effects are defined only if certain stability conditions are
met (Bentler and Freeman 1983; Joreskog and Sorbom 1981). Others
have used the reduced-form coefficients to define total effects (Alwin
and Hauser 1975). Most researchers have used some combination of
these (e.g., Graff and Schmidt 1982). In this paper, I will explain the
relationship between both definitions.
Decompositions typically provide the total indirect effects, which
consist of all paths from one variable to another mediated by at least
one additional variable. Alwin and Hauser (1975), Greene (1977), and
Fox (1985) provide techniques to estimate specific indirect effects that
are transmitted by selected variables rather than by all variables.
Although they use the same term, their definitions of specific effects
differ both conceptually and operationally. As these techniques become
more widely used, this is sure to be a source of confusion, since
estimates of specific indirect effects will differ depending on the definition. The third purpose of this paper is to describe the alternative
meanings of specific indirect effects and the techniques for calculating
them.
TOTAL,
DIRECT,
39
1 = BTi+ rI + (,
(1)
X=A^x,+8,
y =Ay
+ ,
(2)
(3)
40
KENNETH
A. BOLLEN
B.
(4)
k=l
0
0
0
0 .
/33
0
B=
(5)
1.
'i
variables.
TOTAL,
41
O
0
O
0 +
f2
i3
0_13)
O
0
O
0
O
0 +0+
...
(6)
Clearly, since all Bk for k > 3 equal zero, the series converges and the
total effects are defined. The first term in the series represents the direct
effects of X1 on i1. The second and higher-power terms represent the
indirect effects of varying lengths. In Figure 1, the only indirect effect
of length 2 is ,j83, the influence of ql on 73 mediated by l2. The zero
values for B3 and the higher-power terms indicate that all indirect
effects of length 3 or greater equal zero. Summing the right-hand side
of (6) gives
0
0
0
0
0 .
A3
(7)
T,,=
_2 +tjM13
In general, the indirect effects are the differences between the total
effects and the direct effects. Subtracting B from T,, yields B2, the
indirect effects (I,,) for this model.
FIGURE 2.
42
KENNETH
A. BOLLEN
B= 1
B-[?
(8)
1o2]
r'}.
B2= Pf1f2
20
B
08,
B4=
22
(9)
T2
leads to a /f/I2
shift in j1. To
determine the total effects, T,,, we let k - cooand sum all Bk:
OC
T,= L Bk=
k=l
x
E
iL(j-1
'=1
j=l
(10)
(
E=
Ei(Z()a(
L (3I12)2
j=1
Each element of T,
E(1
i=l
TOTAL,
INDIRECT
AND
DIRECT,
43
EFFECTS
element is
00
Z (9Il2)i=
132)3 + **
+ (132)2+ (2
3112
(ii)
i=1
converges,
(P,132)?
00
(PIP2)Y=
+ (P12)1
(13I2)
+ (1332)2
+ *
(12)
i=0
13
(2)(1
1332 + (132)2
+ .**
= 1+ (1l2)i) (213i2)+l.
(13)
Since 11l1,21 is less than one, (#l,#2)i+
goes to zero as i goes to infinity.
This leaves one on the right-hand side of (13). For the product on the
left-hand side to equal one, (1 + P12 + (31132)2 + ... +(332)i)
1
oo. Thus,
(1 - 1
2)- as i
00
E (f1
Y=2)=(1-
1 and i- o.
(14)
0=o
(l3,12)0)
E0
i=l
1A i
--313
PIPl2
(15)
This is the convergent value for equation (11) and the total effect of 1
on itself. To sum up, if 131f321 < 1, the (1,1) element of (10) converges
to 1i1i2/(1 - f1fi2). Since the (1,1) and (2,2) elements are identical,
the (2, 2) element also converges to this value. When we apply a similar
argument to the other elements of (10), we see that under the same
condition of f1,81i21 < 1, the (1,2) element converges to 32/(1 - 3112)
and the (2,1) element converges to Al/(1- 1312). Thus, the total
44
KENNETH
effects of
A. BOLLEN
on TIare
T, =
ZB
T
k=1
1 P12
=(11,132))
12
(16)
P1P-2J
122
[/l12
1 2]
(17)
P132J
If l/3l121 equals or exceeds unity, then the total and indirect effects
among the latent endogenous variables in Figure 2 are not defined.
In general, for the total effects T,, (= EI= Bk) to be defined, Bk
must converge to zero as k -* oo (Ben-Israel and Greville 1974). The
matrix is convergent if and only if the modulus or absolute value of the
largest eigenvalue of B is less than one (Bentler and Freeman 1983, p.
144).2 To continue the generalization of the previous results, we next
search for the value to which T, converges. First, we add I (= B?) to
- + Bk by (I- B):
T,n and then premultiply the sum I + B + B2 + .
(18)
(I- B)(I + B + B2 + * + Bk) = I- Bk+l.
Since
Bk+ - 0 as k -oo,
zero,
in (18) to equal I,
leaving I. For the product of the two left-hand terms
*
to
+
+
must
B
B2
(I- B) as k oo. If I is
converge
(I +
+BBk)
subtracted from this value, T,, results:
Tn,
= (I- B)-
-.
(19)
- I - B.
(20)
The similar canonical form of B is useful in studying Bk and in understanding the justification for placing this condition on the eigenvalues of B. See
Searle (1982, pp. 282-89) and Luenberger (1979) for further discussion.
TOTAL,
DIRECT,
45
Bi +r
= B(Bq+ rE+
) + rP+
= B2i + (I + B)(rF + {)
= B2(BI +rt + ) + (I + B)(rt + )
(21)
= B3l + (I + B + B2)(F + {)
=
B*k
+ (I + B + B2 + ... + Bk-)(r
).
The total effects of t on 1 are in the coefficient matrix for t in the last
line of (21): (I + B + B2 + *- +Bk-')r.
The (I + B + B2 +
* * +B*-1) term is an infinite sum that converges to (I - B) -1 under
the same conditions stated above. That is, the absolute value or
modulus of the largest eigenvalue of B must be less than one. The T,}
matrix under this condition is
T,
= (I- B)-'F.
(22)
r=
((I-
B) 1-)l)r.
(23)
*- +Bk-'1)+e.
(24)
B)-Ir.
(25)
46
KENNETH
A. BOLLEN
The indirect effects also equal (25), since t has no direct effects on y.3
By the same logic, the total effects of '1 on y are
T, =A
+
(I+B+B2
+Bk)
(26)
= Ay(I-
B)-,
Ay=
A((I
I).
(27)
Total and indirect effects are defined only under certain conditions,
i.e., when the absolute value or modulus of the eigenvalues of B is less
than one.4 The eigenvalues are not always readily available. However,
two shortcuts can identify convergent B. I have already discussed the
first: If B is a lower triangular matrix, then Bk is zero for k > m. The
second applies to nonrecursive models: If the elements of B are positive
and the sum of the elements in each column is less than one, the
absolute values of the eigenvalues are less than one (Goldberg 1958,
pp. 237-38). This is a sufficient condition, but it is not necessary for
stability. Given its ease of calculation, it should prove useful in many
situations.
The preceding approach defines the total effects as infinite sums
of functions of the coefficient matrices, and it requires that stability
conditions be met for these effects to exist. A second strategy uses
reduced-form equations to identify total effects (Alwin and Hauser
1975; Fox 1980, p. 25; Graff and Schmidt 1982). In classical econometric models, reduced
form refers to solutions of structural equations in
which all endogenous variables are written as functions of only exogenous (or predetermined) variables, structural coefficients, and disturbances. My use of the concept of reduced forms is similar to Graff
and Schmidt's (1982). It differs from the classical econometric concept
in that i, the vector of exogenous variables, is a vector of latent
variables rather than directly observed variables.
3J6reskog and Sorbom (1981, p. 39; 1982, p. 409) incorrectly list the direct
effect of i on y as Ay and the indirect effects as AY(I - B)- F - Ayr. Freeman
(1982), in his dissertation, also notes this error.
4 This condition is also necessary for the existence of T,,, but it is not
necessary for the existence of Ty,, Ty1, and T, (see Sobel 1986, p. 166). In
practice, we would rarely, if ever, be interested in these latter total effects if T,, did
not exist.
TOTAL,
DIRECT,
47
(I- B)-'r
(28)
+ (I- B)-1 .
(29)
(30)
48
KENNETH
A. BOLLEN
are zero, and the indirect and total effects are equal. The indirect
effects of 1 on y are all the influences that T exerts on y; therefore, Ty
is A (I - B) -. Thus, when the reduced-form and structural equations
are used, all the total and indirect effects can be identified.
Defining total effects as the sum of an infinite series and
defining them in terms of reduced-form coefficients appear to yield the
same results. However, there is one important difference. To illustrate,
consider T . For the total effects Ek=Bk to be defined, the convergence condition for B must be met. To write the structural equation for
i1 in reduced form, (I - B)-1 must exist. The existence of (I- B)-'
need not imply the convergence condition. For example, specify B as
0]
B=[
(31)
(32)
N::1-[?
12
K
[jq
0 ][Y2
1[::1?
++H
h T'l
[42]
J2Y
1,l
(3 )
X4-
A,
0X
0
~1
0O
A [
4-
A+
2
8
4~
(34)
TOTAL,
DIRECT,
49
and
y,~
Y2
X5
X6
0
0
+2 rl
Y3
Y4.
X8
*
.(35)
3(35)3
-4-
-(I-B)
T, (1-=A)
-(i
-r
A2)
11Y1
[:2
l 1
][
y]
(36)
2Y2 ]
The first term on the right-hand side of this expression, (1 - fi1 2)-1,
results from the reciprocal relationship between 71 and 21,which may
be activated by the effect of ^l (or 2) on 'll (or q2)' For this feedback
loop to converge, 1Af21 must be less than one. The direct effects of t
on aq appear as coefficients in F. The indirect effects of t on 1 equal
((I - B)-1 - I)r, which for Figure 2 are
I,}=(i-=
(-
3):)l
1[1
2Y2Y
(37)
Similarly, the total and indirect effects for the remaining variables can
be determined by substituting the B, F, and Ay in Figure 2 into the
appropriate decomposition formula derived above.
Table 1 summarizes the decomposition of effects for the general
structural equation model with latent variables. It is well known that
the classical econometric simultaneous equation model, confirmatory
factor analysis, and the MIMIC model are special cases of the LISREL
model (Joreskog and S6rbom 1982; Joreskog and Goldberger 1975).
Thus, Table 1 can be applied to these models given the appropriate
substitutions. For example, if x were substituted for i, and y for al, the
decomposition of effects in Table 1 would match those in Fox (1980),
except for a slight difference in notation.
50
KENNETH A. BOLLEN
TABLE 1
Direct, Indirect, and Total Effectsof E and n on XI,y, and x
Effects on:
1
Effects of i
Direct
Indirect
Total
Effects of ii
Direct
Indirect
Total
0
Ay(I-B)-r
Ay(I- B)-
(I-B)-r-
(I - B)(I-B)-
B
'-I-B
(I-B)-
-I
A,(I-
Ay
Bf)- -Ay
Ay(I-B)-1
A,
0
A,
0
0
SPECIFIC EFFECTS
The indirect effects comprise all the indirect paths from one
variable to another. Consequently, the contribution of particular
mediating variables can be obscured. Alwin and Hauser (1975), Greene
(1977), and Fox (1985) have proposed strategies to analyze specific
indirecteffects,i.e., those effects transmitted by a particular variable or
group of variables. Their conceptions of specific indirect effects differ.
In this section, I contrast their definitions of these effects and their
procedures for calculating them.
Alwin and Hauser (1975, p. 42) define specific indirect effects as
those influences mediated by the variable in question and thoseafterit.
They exclude from the definition those paths running through endogenous variables that precede the variable of interest. Consider Figure 3
(also used by Alwin and Hauser), which depicts an exactly identified
recursive model with three exogenous variables (x1, x2, x3) and three
endogenous variables (y, Y2, Y3). According to Alwin and Hauser's
definition, the specific indirect effect of xl on Y3 mediated by Y2 in this
model is the compound path x, - Y2 -*y3. They exclude the indirect
path operating through y, and Y2 (i.e., x -y-*Yl -Y2
3). The
specific indirect effects of xl on Y3 through y, appear in two com- Y
*->
The second
3.
pound paths: x~~y~
3 and x,- y-y
Y2
is
as
the
effect of y,
included
of
indirect
path through y2
part
specific
because in the causal sequence, Y2 comes after yl. Alwin and Hauser
TOTAL,
DIRECT,
51
X23Y3
conceptualize specific indirect effects as increments in the causal sequence of variables in a model. I call these indirect effects incremental
specificefects.
Incremental specific effects are restrictive, i.e., they do not
include compound paths operating through prior endogenous variables. Greene's (1977) idea of specific effects is even more restrictive. It
includes as specific effects only those influences mediated by the
intervening variable or variables of interest.5 For instance, according to
his definition, the specific effect of x1 on y3 operating through Yl is the
compound path x1
yI->Y3 (see Figure 3). This differs from the
incremental specific effects for Yl because it excludes the path going
-y32 3). Greene's notion of specific indirect
through Y2 (i.e., x,-, y
effects focuses on those effects that operate through Yl but no other
endogenous variable. Alwin and Hauser's concept of incremental
specific effects includes all compound paths subsequent to the endogenous variable of interest. I call Greene's specific indirect effect exclusive
specificeffects.
5
Greene's (pp. 377-78) emphasis is on presenting a computational method
for computing specific effects for recursive systems. He does not give a detailed
definition of specific effects nor does he explain how his use of the term differs from
Alwin and Hauser's. Greene's computation method suggests a particular definition
of specific effects, which is what I discuss.
52
KENNETH
A. BOLLEN
and
Y2 consists of two compound paths: x1 -> y, -Y2 y-3
y3 (see Figure 3). The incremental and the exclusive specific
TOTAL,
53
DIRECT,
affects another, its coefficient is listed in its row and in the column of
the variable it affects. To determine an exclusive specific effect, the
rows and columns corresponding to endogenous variables other than
those of interest are first removed, creating a new and smaller square
matrix. This new matrix is premultiplied by itself again and again
until the resulting matrix is null. When each matrix product,beginning
with the first, is summed, the end result is a matrix of the exclusive
specific effects.
To see how this procedure works, consider the matrix representation of Figure 3:
Yl
Y2 =
Al
Y3
Yl
Y3
Yi
Y
3
_2
442
77
72
73
'1
X1
75 +
Ys
9s
X3
(38)
J3
0
0
y1
y,
Y7
Y2
75
78
O
o0 0
O
0
3 Y,
0
1
00
O
O
[0
[0
,79
2
I']
(39)
B' i3
54
KENNETH
A. BOLLEN
M*=
y,
Y2 Y8
O O O
0 0
?0
'3
0
00
y7
59 '
(40)
0
0
0
0000
0000
0
0
0
0
O y/A2
0
0
Y1I2
Y2fi2
(41)
TOTAL,
DIRECT,
55
and B*, where the "*" indicates the matrices with the proper row(s) or
column(s) absent. Then, he forms a new total effects matrix,
(I* - B*)-'1*. He subtracts the (I* - B*)-1'* matrix from the original Tyx matrix with the appropriate row(s) removed to determine
inclusive specific effects through the variable(s) being analyzed.7 He
follows a similar series of steps to find the inclusive specific effects of y
on y.
A generalizationof specificeffects.The procedures reviewed above
share one limitation: They are oriented toward variablesrather than
paths. Suppose we want to know that part of the indirect or total effects
transmitted by a path or combination of individual paths. The preceding methods would be helpful only if these paths happened to correspond to incremental, exclusive, or inclusive specific indirect effects.
This need not be the case, of course. In this subsection, I explain a
technique that is coefficient (or path) oriented. Since the coefficient or
path is a more elementary unit than the variable, this procedure is
more general than the existing procedures. By choosing appropriate
combinations of paths, we can determine the specific indirect effects of
variables. In the next few pages, I use this idea to show that the specific
indirect effects proposed by Alwin and Hauser, Greene, and Fox are
special cases of effects estimated with this alternative method. Then, I
apply this method to a type of specific effect that falls outside the scope
of these other procedures. Unlike Alwin and Hauser's procedures for
estimating specific indirect effects, this method is not restricted to
exactly identified recursive models nor to incremental specific effects.
Like Greene's and Fox's procedures, it uses matrix operations; but
unlike their procedures, it does not require row and column removal.
The method has several steps. The last is optional.
1. Identify the changes needed in the coefficient matrices.
2. Modify the coefficient matrices.
3. If B is changed, check the modulus or absolute value of the largest
eigenvalue of the new B to insure that it is less than one.
4. Calculate the direct, indirect, or total effects with the modified
matrices.
5. Subtract the new decompositions from the old.
7Fox also showshow Alwinand Hauser'sincremental
specificeffectscan
be computedwith his procedure.
56
KENNETH
A. BOLLEN
B()=
0
0
0
0
0
0
/33
Y4
Y5
Y6
(42)
r(/2)=
Y7 Y8 Y9
where 1 = {(., 1) = 0} and 12= {(1,.) = 0}. The (11) subscript indicates a set of modifications that are given in the statement defining /1.
In (42), the first column of B is set to zero, which is represented by
(., 1) = 0. Similarly, the first row of r is set to zero, and this is
symbolized as Fr2), with 12 representing (1,.) = 0. In this notation, the
"." in the row position indicates that all rows in the stated column are
zero. The "." in the column position defines all columns in the stated
row as zero.
Before calculating indirect effects with B(, ), we need to check
whether it satisfies the convergence condition presented earlier. That is,
the modulus or absolute value of the largest eigenvalue must be less
than one. In general this is done because B(/) can fail this condition
even if B does not (Fisher 1970; Sobel 1986). In this example, B and
B(/I) are lower triangular and the condition is satisfied.
TOTAL,
DIRECT,
57
The indirect effects Iyx based on the original B and F (see [38])
are
(I-B)-'FfiYi
1Y3
i(Y2
+
(32 + 1i3)Y2?3Y5
The indirect effects based on the B(1,) and F(12) of (42), in which those
paths into or out of y, have been eliminated, are
(I-B(1)
0
0
I2)-(2=
Bi374
0
0
0
0
3Y75
(44)
f3Y6
f1Y2
l1Y3
(2 + 133)Y1 (2 +
1/i3)Y2
(45)
+
(2 ?1f3)Y3
58
KENNETH
A. BOLLEN
B(,)0=
Y1
r(2)
Y2
Y3
0,
(46)
where I/=((.,2)=0,
and 12={(2,.)=0,
The
(.,3)=0}
(3,.)=0}.
in
in
nonzero
elements
the
first
column
and
only
represent
B(l,) appear
the direct effects of y, on Y2 and Y3. The nonzero elements in r(l2)
denote effects of the exogenous variables on y. The modified versions
of the coefficient matrices replace the original matrices in the indirect
effects formula, yielding
0
(I-
B(,,) -r12)-
r(2)=
(47)
As the first row of (47) shows, x has no exclusive specific indirect effects
on y1; all its effects are direct. The second row contains the specific
effects of x on Y2exclusively through y,, and the third row contains the
exclusive specific effects of x on Y3. The procedure illustrated above
can be applied to other types of models and can be used to find the
exclusive specific effects of more than one endogenous variable at a
time.8
8Alwin and Hauser's
(1975) incremental specific effects can also be calculated with this method. However, these effects are not clearly defined when a
nonrecursive model is analyzed because with a feedback or reciprocal relationship,
it is unclear which variable precedes or follows in the causal hierarchy.
TOTAL,
DIRECT,
59
B(,
(48)
(49)
~lB
'
Y2
(I-B)-
r=(
2)
[Yi
12Y21
(50)
When equation (49) is subtracted from (50), we get that part of the
total effects that is attributable to the direct influence of '2 on / 1:
(I-
B)-lr-
(IB(l,))
r=(l-l( 2)-
[f ,f2?1
2(
]. (51)
Equation (51) shows, for instance, that //lfl2Y/(l - f1I2) of the total
effects of (l on /l results from the effect of 112 on 1,.
60
KENNETH
A. BOLLEN
TOTAL,
DIRECT,
61
df
(52)
~2 = '1'1 + ~2'312
In (53), assume that l, and
with ~,, that E('i)
'2
h8171
+ ~2
~(53)
/1,
2, and
their means. Define y,fi, the indirect effect of ~l on 12, as the single
element of f(O), with 0 containing y, and p,. The df/0O is [/P, y,]',
9 In addition to
requiring a consistent and asymptotically normal estimator
of 0, f(O) must be continuously differentiable with respect to 0 in the neighborhood
of the true parameter, and I assume the continuity of N-1V(0) with respect to 0
in the neighborhood of the true parameter. See Bishop et al. (1975) for additional
discussion of these conditions.
62
KENNETH A. BOLLEN
v( I
(54)
+ yl V(#i)].
(55)
EMPIRICAL ILLUSTRATION
To illustrate the preceding methods, I utilize a model relating
objective and subjective components of stratification for individuals.
Figure 4 depicts this model. Define xl as income, x2 as occupational
prestige, -/l as subjective income, r7, as subjective occupational
prestige, and 773 as subjective overall class ranking. Let yl, Y2, and Y3
be measures of the 771, 12, and r/3 latent variables. In the model,
subjective
dependent upon their objective counterparts and are reciprocally related to each other. In turn, the subjective income and occupation
variables (7q1 and /2) have direct effects on overall subjective class
(173). The data are taken from a white subsample of 432 individuals
from Indianapolis, Indiana (Kluegal, Singleton, and Starnes 1977).
TOTAL,
DIRECT,
63
FIGURE
4.
A model of objective
}I
and subjective
stratification.
EI
E2
.288
(.168)
0
.101
(.016)
0
(.006)
.037
(.011)
.064
(.013)
(.0014)
.0007
(.0004)
.0048
(.0014)
0
.330
r=
.007 . (56)
(.090)
(001)
.399
.535
0
?
?
(.139)
(.201)
The standard errors are in parentheses. All coefficient estimates carry
the expected positive signs, and all are statistically significant (p < .05,
one-tailed test), though /? is barely significant. The goodness of fit of
the overall model is excellent (X2 = 0.27, df= 1, GFI = 1.00, AFGI =
.996).
The estimated indirect effects of 'q on '1 and x on 1 are10
.011
0
.0022
.030
.105
B=
T1q
(.067)
.035
(.026)
.237
(.105)
(.036)
.105
(.067)
.183
(.127)
0
0
(57)
10 In this
model, x is equivalent to i, and Iy, and I.x are of little interest,
since AY is an identity matrix. Also, the estimate of the eigenvalues for B are
considerably less than one. Therefore, the convergence condition is met.
64
KENNETH A. BOLLEN
In the I1x matrix, all the indirect effects of x on q except the indirect
effect of occupational prestige (x2) on subjective income (71l) are
statistically significant (p < .05, one-tailed test). In the IT, matrix, in
contrast, none of the indirect effects of T] on i1 except the effect of
subjective income (-/1) on subjective overall class (7q3) are statistically
significant. This suggests that with one exception, the most significant
effects of the subjective variables on one another are direct, not
indirect. Statistically significant direct effects need not be accompanied
by significant indirect effects.
To explain the largely nonsignificant indirect effects of q1on q,
we can begin with ,l, the coefficient of the path from subjective
occupational prestige (r2) to subjective income (Xq1).Relative to 2,,the
influence of subjective income on prestige, /l is somewhat weak. In
substantive terms, people who perceive their income as high are likely
to perceive their occupational prestige as higher than those who
perceive their income as low. Yet, individuals with positions they
perceive as prestigious are less likely to believe that their incomes are
high. Since the /, path channels a portion of the indirect effects, this
may partially explain the relatively small magnitude of I 's compared
to its estimated asymptotic variance. I investigate this possibility with
the generalization of specific-effects methodology that I proposed
earlier. First, I estimate the indirect effects transmitted through the
path from r/2 to m/1(i.e., 1P). If the above ideas are correct, these new
specific indirect effects from 1 to i1 should be nonsignificant. Second, I
estimate the indirect effects that do not cross the /, path. If these are
significant, the idea that the largely nonsignificant IT, are due at least
partially to the weak link from T2 to r1i would receive further support.
With B(/) defined so that l= { / = 0}, the estimated indirect effects
transmitted by the P, path are
.105
.030
(.067)
.035
(.026)
.061
(.044)
(.036)
.105
(.067)
.183
(.127)
0
0
0
^
nx
.011
(.006)
.003
(.002)
.006
(.004)
.0022
(.0014)
.0007
(.0004)
.0012
(.0008)
(8)
In (58), the "*" indicates that these are indirect effects sent over
of (58) to
? we compare I*
particular paths-in this case, for I,. When
~~~~~~T/r
TOTAL,
DIRECT,
65
of (57), we see that all the indirect effects of T1on Tqexcept the
indirect effects of subjective income (q1,) on subjective overall class ('q3)
are transmitted through the P, path. The indirect effect of 71Ton 73
passed through the P, path is not statistically significant, although as
shown in (57), the total indirect effect of m on i3 is statistically
significant.
The indirect effects of x on q through the /,l path are even more
revealing. As I* shows, the only indirect effects through the ll path
that are statistically significant are the effect of income (xl) on subjective income (q71) and the effect of occupational prestige (x2) on its
subjective counterpart (q2). The remaining indirect effects of income
and occupational prestige via the P, path are not significant. When I *
is contrasted with I,X in (57), the secondary role that the influence of
subjective occupational prestige on subjective income plays in generating total indirect effects of x on T is highlighted.
Finally the indirect effects from T to T and x to 1 that do not
cross the path from q2 to m, (i.e., fl) are
I1,
7x
0
I
A^7
.7
.177
0 0
0
0
0
0
I*x
(.089)
(-.09
.033
(.010)
.058
.0036
(.012)
(.0013)
(59)
The only indirect effect of T on T is the one from m71to 73 through iq2.
This effect is statistically significant. In ,* , all the indirect effects of x
on T not mediated by the path associated with fl are statistically
significant. Thus, the indirect effects that cross the ,l path tend to be
nonsignificant, and the indirect effects that do not cross the path are
significant.
The example illustrates that not all specific indirect effects that
comprise the total indirect effects have the same statistical significance.
Examining the statistical significance of only the total indirect effects
can give a misleading portrayal of the significance of the individual
indirect paths. In other cases, if all the components of an effect are not
significantly different from zero, then the delta method is inappropriate. For instance, if Y, and P, from equation (53) are essentially
zero, then the delta method should not be applied to find the standard
error of 'Yl 1.
66
KENNETH
A. BOLLEN
do
doI)
= V(V (I-B)
((I-B)-1))
(A2)
+Vr(
?
((I- -B)
Vy(((I- A)
B)-\
+V((I-
B)-' ?(Ay(I-B)-')')
d veclr
I~
ado '
- I)')
I)
/(A3)
Ip|
(A3)
TOTAL,
67
DIRECT,
dey
= VB((I-B)-lr
(A^y(I-B)-l)')
(AY(I-B)-
+VIr(I
(A4)
)')
11?Ir )
+V%((I-B)
where
[B
VB =
vecl
Vr =
vec --,
a[
vec
aB1
vec-
vec
ar
,...,vec
dos
d02
[
VA
A
dB
, vec d
vec
vec
...
dO,
80,
90O
O
1
0
0
O
0
1
0
O
0
0
0
O
O
O
O
O
O
0
O
O
1
O
O
0
O
O
0
O
O
0
0
0
0
_0
0
0
0
0
0
0
0
0
0
0
0
00
0
0
0
0(A
0
0
1
0
VB=
,0
O
0
0
0
O
0
0
0
(A5)
and
0
0
0
(A6)
These are used to estimate the standard errors for I and Ix. The
standard errors of the specific effects were estimated with appropriate
modifications.
68
KENNETH A. BOLLEN
REFERENCES
ALWIN,
DUANE
M. W.
BROWNE,
Covariance
37:62-83.
DUNCAN,
OTIS D.
Sociology 72:1-16.
.1975.
Academic Press.
York:
FINNEY,
Equa-
FOX,
Calculation
14:81-95.
of Effect Decomposition Methods for Two General Structural Covariance Modeling Systems." Ph.D. diss.,
University of California at Los Angeles.
GOLDBERG,
GRAFF, j., AND P. SCHMIDT. 1982. "A General Model for Decomposition
of Effects."
and
Indirect
in
Under
Observation:
131-48
Prediction,
Structure,
System
Causality,
Pp.
edited by Karl G. Joreskog and Herman Wold. Amsterdam: North-Holland.
VERNON L. 1977. "An Algorithm for Total and Indirect Causal Effects."
Political Methodology44:369-81.
KARL G. 1978. "Structural Analysis of Covariance and Correlation
JORESKOG,
Matrices." Psychometrika43:443-77.
GREENE,
JORESKOG,
DIRECT,
TOTAL,
69
JORESKOG, KARL G., AND DAG SORBOM. 1981. LISREL. Analysis of Linear Structural
KLUEGAL,
1977. "Subjec-
PETER.
SHAYLE R. 1982. Matrix Algebra Useful for Statistics. New York: Wiley.
MICHAEL.
Structural
Equation