Professional Documents
Culture Documents
.
JSTOR is a not-for-profit service that helps scholars, researchers, and students discover, use, and build upon a wide range of
content in a trusted digital archive. We use information technology and tools to increase productivity and facilitate new forms
of scholarship. For more information about JSTOR, please contact support@jstor.org.
Wiley and John Wiley & Sons are collaborating with JSTOR to digitize, preserve and extend access to
Sociological Methodology.
http://www.jstor.org
ASYMPTOTIC CONFIDENCE
INTERVALS FOR INDIRECT
EFFECTS IN STRUCTURAL
EQUATION MODELS
MichaelE. Sobel
INDIANA
UNIVERSITY
Since itsgeneral introductionto the sociologicalcommunity(Duncan, 1966), path analysishas become one of the most
widelyused tools in sociological research,and through its use
sociologistshave become more sophisticatedabout the general
For commentson an earlier draftof thischapter and for detailed advice I am indebted to Robert M. Hauser, Halliman H. Winsborough,and
Toni Richards,severalanonymousreviewers,and the editorof thisvolume. I
also wish to thankJohn Raisian, Nancy Rytina,and Barbara Mann for their
commentsand Mark Wilson for able research assistance. The opinions expressed here are the sole responsibilityof the author.
29()
ASYMPTOTIC
CONFIDENCE
INTERVALS
291
292
MICHAEL
E. SOBEL
ASYMPTOTIC
CONFIDENCE
293
INTERVALS
Yi = Z,S,
+ E,
(1)
moreover, Zj = (Xj, Yj) is the full-rankmatrix of predetermined and endogenous variables,respectively,and 6j is the corresponding parameter vector (7*,
xr+ E
(2)
B =
Y = (y1,
. . ., XK)
where
, YM),
X=(X1,
(f81h. . ., /3M)is the nonsingular matrixof coefficientsof the
endogenous variables,r = (Vi, . . .y, m)is the K x M matrix
of coefficients for the exogenous variables, and E =
(El, . * *,E(M). We assume that the restrictionson r include
only zero restrictions,whereas for B the restrictionsinclude
zero restrictionsas well as the restrictionsthatthe diagonal elementsare set to 1. Finally,we assume thateach row of E is independentlyand identicallydistributedN(O', I).
294
MICHAEL
E. SOBEL
AN(O, [1(6)]1)
8)
(3)
convergesin law (distribution)to the quanis the maximumlikelihood estimatorof 6 based on a sample of size n. In the
special case whereB is an upper-triangularmatrixand t is diagonal, the system(2) is recursive(Theil, 1971, p. 461). It is well
knownthatforthiscase (Rothenbergand Leenders, 1964; Malinvaud, 1970; Theil, 1971) maximum-likelihoodestimationof
the systemreduces to equation-by-equationleast-squares estimationbecause the log-likelihoodfunctionis merelythe sum of
the M likelihood functionsfor the structuralrelationshipsspecifiedby (1). That is, the log-likelihoodfunctionfor the system
(2) is given by
That is, n1/2(6n -8)
(n/2) log
I (B')
1lB-1
2)
E
m'=1
E
M=1
'M(ym
(n/2)
- Zm'6m')'
(Y M
Zm6m)
which reduces to
L(81m
M,
) =
c+
j=1
log
AP-'
(4)
(Yj - ZjSj)4rj-
ASYMPTOTIC
CONFIDENCE
295
INTERVALS
0
Q21
Q'1
Q-1
. . .
0
0
. . . Q-1
withcomponentsQf1 = plimTjj(Z;zj/n)-1.3That is, fora recursive model the vector an obtained by combining the M
least-squaresestimatorssatisfies
n1/2(&
Hence
-6)
N(O, [Q(8)]-1)
(5)
is asymptoticallynormallydistributed.
Now that the asymptoticdistributionof the coefficient
vector has been established, it is necessary to consider the
asymptoticdistributionof the indirecteffects.As previouslyindicated, the multivariate-delta
method is used forthispurpose.
Because this technique may be unfamiliarto some readers, a
descriptionof the procedure follows.Readers who are familiar
with this technique may wish to skip the exposition and go
directlyto the results.
an
6) -4 N(0
2(6))
(6)
296
MICHAEL
E. SOBEL
N(O,[fP(8)]2
-f(8))
n112(f(6n)
(7)
v2(8))
b(x - x) + e
2(bn_-b)
-$N(O,
lim((x -)2/n)-)
lim((x
/n)- )
Note, however, that as b approaches zero, (b)-1 becomes unbounded and the delta method cannot be applied.
To see heuristicallywhythe single-deltamethod works,
we consider a Taylor expansion off, a functionof the parameter set. Suppose thatfis twicedifferentiablein a neighborhood
of 6. A Taylor expansion off about 8 is given by
f(6M)
-f(8)
= f '(8)
(8n
6) + f"(6*)
(8n
6)2
/2
(8)
-f(8)
f '(8)
(6n
6)
(9)
ASYMPTOTIC
297
INTERVALS
CONFIDENCE
f ( ))-n -l
V(
[ft(8)]1V2(8)
(af/dl1'
If
8)
n1/2(-
df/d6s)
A N(O,1(8))
(10)
then
-f(8))
n12 (f(-n)
-f (8)
(8n
8)' (Of/O8)+ Rn
(12)
distributedrandom
This is a linear functionof a multinormally
vector; hence its distributionis normal,E(f(Ak) =f(8) and
V(f8({n)) ==n-1 [(a*f/a*)8) 1(8)
model
(c*f/c*)8)
b(x - x) + c(z
z) + e
298
MICHAEL
E. SOBEL
nl 2(bn- b)
N (?,(
j2
22))
(14)
(bc
bc)
-*
(15)
/8
Jaf,
Qfi/a81
If
n2(6n
6)
N(0,1(8))
(16)
A N(0,
(OF/68)1(6)(OF/O8)')
(17)
299
INTERVALS
CONFIDENCE
ASYMPTOTIC
(f1, . . ,f,)' now be a column vector of indirecteffects.Because thefi are sums of products of the structuralparameters,
theyare differentiablein a neighborhoodof S. Thus results(5)
and (17) may be combined, yieldingthe conclusion that
n1/2(F(
F(6))
A N(O,
CD
(OF/06)[Q(6)]-1(OF/068)')
(18)
c - (n/2)log I (B')-14B-1 l
-(1)
m'=l m=1
-mm(YM,
-Zm'1m')'
(YM-zmam)
300
MICHAEL
E.
SOBEL
AN EXAMPLE
Figure 1.
/3
?2
see J6reskog, 1973, p. 112; Theil, 1971, p. 526; and Rothenberg and
Leenders, 1964, p. 67.) Provided the indirecteffectsare differentiable,
confidence intervalsmay be obtained as before.
For models withunobserved variables,LISREL may be used to obtain
the maximum-likelihoodestimatesand the inverseoifthe informationmatrix.
(See J6reskog,1973, p. 109, for formulas.)Provided the indirecteffectsare
differentiable,
confidenceintervalsmay be obtained as before. (On the subject of indirect effectsin models with unobserved variables, see Schmidt,
1980.)
ASYMPTOTIC
CONFIDENCE
INTERVALS
301
entation(Xc),responde~nt'seducation (yl),respondent'soccupational statusin March 1962 (y2), and respondent's income, expressed in unitsof thousands,in 1961 (Ye). The resultsapply to
nonblack men with nonfarm background in the experienced
civilianlabor force,of age 35 to 44 in March 1962.6
Table 1 presentsthe coefficientsforthe structuralmodel
and Table 2 displaysthe estimatedindirecteffectsin both symbolic and numerical form.7In addition, Table 2 contains the
asymptoticstandard errors of the indirecteffects.
The standarderrorsin Table 2 were obtained as follows.
First thefi, i = 1, . . ., 10, are differentiated with respect to
the structural parameters, yielding the 10 x 12 matrix
(OF/06), with (is)thelement (af/1a8s),evaluated at the solution
6. In symbolicform, Table 3 presents this matrix of partial
derivatives;the parameter at the top of each column indicates
thatthe partial derivativeis taken withrespectto that parameter. Next the estimatedvariance-covariancematrixof 8 is premultipliedby the matrixof Table 3 and postmultipliedby its
transpose, yielding the estimated asymptoticvariance-covariance matrix of the indirect effects(Table 4).8 From here it
6 All the results are presented for the unstandardized form of the
model. However, the resultsthat have been established can be modified to
hold asymptoticallyfor the standardized solution.
7 Readers of the Alwin and Hauser paper maywonder whythe simpler
expressionsforthe indirecteffectstheyuse are not employed here. Their expressionsare a functionof both structuraland reduced-formparameters; in
general, to work with these expressions in a distributionalframeworkone
needs thejoint distributionof the reduced-formparametersand the structural parameters. Furthermore,the formulasgiven by Alwin and Hauser are
valid for only a limitedclass of recursivemodels, as the authors note.
Similarly,the expressions given by Fox (1980, p. 11) are not used because the indirecteffectshe defines are "total" indirecteffects-that is, the
indirecteffectof a variableon a subsequent variableis given as the sum of the
indirecteffectsthroughall the interveningvariablesand combinationsof the
interveningvariables.Here I preferto allow the possibilitythatusers may also
find it worthwhileto examine particularcomponents of "total" indirecteffects.See, for example, the illustrationin this section.)
8 The estimatedasymptoticcovariance matrixis 12 x 12, as opposed
to 15 x 15, because the variablesin the analysiswere,withoutloss of generality,deviated about theirmeans. This is also the reason whythe matrixof partial derivativesis 10 x 12 rather than 10 x 15.
302
MICHAEL
E. SOBEL
TABLE 1
COEFFICIENTS AND STANDARD ERRORS (IN PARENTHESES) FOR THE
ACHIEVEMENT
STRUCTURAL EQUATION MODEL OF SOCIOECONOMIC
IN FIGURE 1
Dependent Variable
Independent
Variable
Yi
Xa
8ia
0.0385
(0.0025)
0.1707
81b
(0.0156)
81C= -0.2281
(0.0176)
Xb
Xc
Y3
Y2
Yi
82a
82b
82c
821
0.1352
83a
0.0490
83b =
(0.0175)
(0.1082)
-0.4631
83c
(0.1231)
4.3767
(0.1202)
831 =
Y2
0.0114
(0.0045)
832 =
0.0712
(0.0275)
-0.0373
(0.0314)
0.1998
(0.0364)
0.0704
(0.0045)
NOTE: The sample size is 3,214; the population is nonblack,nonfarm,U.S. men in the
experienced civilianlabor force,aged 35-44 in March 1962. Symbolsare as described
in the text and Figure 1. The analysis is based on the correlation matrix and the
standard deviationsreported by Duncan, Featherman,and Duncan (1972, p. 38).
TABLE 2
INDIRECT EFFECTS AND ASYMPTOTIC STANDARD ERRORS
(IN PARENTHESES) FOR THE STRUCTURAL EQUATION MODEL
OF FIGURE 1
Dependent
Variable
Y2
Y3
Indirect
Effects
of Independent
Variables
Indirect EffectThrough
Yi Alone
Xa
11 =
Xb
f2 = 82181b =
Xc
f3 = 82181c= -0.9983
Xa
f4 = 8318ia =
Xb
f5
Xc
16
Yi
82181a =
Y2
0.1685
(0.0118)
0.7471
(0.0712)
(0.0818)
83181b
=
83181c =
0.0077
f7 = 832(82a + 82181a) =
(0.0015)
0.0341 f8
(0.0069)
-0.0456
(0.0090)
832(82b
82181b)
fA
832(82c
82181c)
il
= 832821 =
0.0214
(0.0020)
0.0560
(0.0096)
-0.1028
(0.0111)
0.3081
(0.0214)
ASYMPTOTIC
CONFIDENCE
303
INTERVALS
SYMBOLIC
TABLE 3
FORM FOR aF/la
81a
81b
81c
82a
82b
82c
821
83a
83b
83c
831
832
f1
f2
821
8ia
821
81b
f3
821
8lc
0?0
f4
831
8ia
f5
831
81b
f6
831
8lC
f7
832821
832
83281a
82a
82181a
f8
f9
832821
832
83281b
82b +
82181b
832821
832
83281C
82C +
f1o
832
82181C
821
is a simple matterto pick out the diagonal elements and compute asymptoticstandard errors,using these to create the (1 a)100% confidence intervals.
Table 2 indicatesthatwhile severalof the indirecteffects
are small, in no case would a 95 or 99 percent confidence intervalforany indirecteffectcover the value zero.9Hence it may
be assumed that the indirect effects are nonzero. Thus,
although the directeffectof father'seducation on respondent's
occupational statusis insignificant,
the indirecteffecton occupational statusthroughrespondent'seducation is greaterthan
zero. In other words, educational advantages in the familyof
orientation favorably influence occupational status by augmentingrespondent's education, which in turn has a positive
direct effect on occupational status. Similarly,although the
directeffectof the siblingvariable on income is not statistically
differentfrom zero, increasing the number of siblingsin the
familyof orientationdecreases income both by decreasing respondent's educational attainment(f6) and by decreasing respondent's educational attainmentand respondent's occupa-
tionalstatus(fe).
ASYMPTOTIC
f2
A
fi
f2
fA
f4
fA
Ae
f7
A
Ag
fio
1.3912
-2.7139
-0.2800
0.0538
-0.1671
0.0449
0.0830
-0.2579
0.0695
0.3902
50.6970
4.3676
-0.1671
2.1225
0.4554
-0.2579
3.2740
0.7039
1.7322
TABLE 4
VARIANCE-COVARIANCE MATRIX (MULTIPLIED
f3
66.9629
0.0449
0.4554
2.7148
0.0695
0.7037
4.1868
-2.3137
0.0220
0.0792
-0.1140
-0.0058
-0.0371
0.0363
-0.1389
0.4825
-0.4942
-0.0545
0.0373
0.1792
-0.6166
0.8119
0.0603
0.1819
-0.0053
0.8236
0.
-0.0
-0.0
0.
ASYMPTOTIC
CONFIDENCE
INTERVALS
305
and indirecteffects(Alwin and Hauser, 1975; Duncan, Featherman, and Duncan, 1972). There are a varietyof ways in
which such comparisons may be made; forexample, one could
assess the proportionof the totaleffectthatis direct (indirect).
the ratio of the direct(indirect)effectof interest
Alternatively,
to the relevantindirect(direct) effectmay be assessed, yielding
informationon both the relative magnitude and direction of
To implementsuch
influenceof the direct (indirect)effects.10
comparisons,
as
-82a/82181a
(82181a)-i
arl/a821 = -82a/81a 21
306
MICHAEL
E. SOBEL
definedas 63b/f5
f9.
Dependent
Variable
Indirect
Effects
of Independent
Variables
Y2
Xa
il =
Xb
f2
Y3
Indirect EffectThrough
Yi Alone
82a/82181a
0.8024
(0.1231)
82b/82181b
0.0656
(0.1452)
Xc
r3
82c/82181c
Xa
r4
83a/83181a
Xb
r5
83b/83181b
Xc
rS
83c/83181c
Yi
Y2
0.4639
(0.1318)
1.4820
(0.6860)
2.0876
(0.9631)
0.8184
(0.7298)
83a/832(82a
82181a)
83b/832(82b
82181b)
rs
83c/832(82c
82181c)
r10
r7
r8
0.5332
(0.2205)
1.2704
(0.5376)
0.3625
(0.3098)
831/832821
0.6485
(0.1451)
ASYMPTOTIC
CONFIDENCE
307
INTERVALS
TABLE 6
ASYMPTOTIC CONFIDENCE INTERVALS
FOR RATIOS OF DIRECT TO INDIRECT
EFFECTS IN A MODEL OF SOCIOECONOMIC
ACHIEVEMENT
r2
r3
r4
r5
r8
r7
r8
r9
rio
rll
r12
r13
95% CI
99% CI
(0.5827,1.0221)
(-0.2190,0.3502)
(0.2056,0.7222)
(0.1374,2.8266)
(0.1999,3.9753)
(-0.6120,2.2488)
(0.1010,0.9654)
(0.2167,2.3241)
(-0.2447,0.9697)
(0.3641,0.9329)
(0.0729,0.7107)
(0.1407,1.4397)
(-0.1723,0.6745)
(0.4848,1.1200)
(-0.3090,0.4402)
(0.1239,0.8039)
(-0.2879,3.2519)
(-0.4839,4.5724)
(-1.0645,1.8828)
(-0.0357,1.1021)
(-0.1166,2.6574)
(-0.4368,1.2164)
(0.2741,1.0229)
(-0.0280,0.8116)
(-0.0648,1.6452)
(-0.3062,0.8084)
basis of these data. Thus the indirecteffectof the siblingvariable on respondent'soccupational statusis more negativethan
the directeffect,indicatingthatadditional siblingsin the family
of orientation detract from subsequent occupational attainments predominantlyby attenuatingthe educational experiences of individualsreared in larger families.Similarly,the effectof father'soccupational status on respondent's income is
predominantlyindirect: Respondent's income is enhanced by
high-statusorigins,primarilybecause high-statusoriginsincrement subsequent educational and occupational achievements.
The widthsof the confidenceintervalsdo not appear to
allow stronginferencesabout the magnitudesof the othercomparisons in Table 6. Nevertheless,it is not unreasonable to
suggest that rl, r7, and r,0 are not greater than 1. In other
words,the directeffectsof father'sstatuson son's statusand income are not larger than the effectsthat intervene through
son's education and son's occupational status,respectively;similarly,the direct effectof respondent'seducation on income is
not larger than the indirecteffectof education on income.
On balance, the analysissuggeststhatthe indirecteffects
in the model under considerationare nonzero, but the confidence intervalsfor the ratiosin Table 6 are too large to permit
308
MICHAEL
E.
SOBEL
=M,
(A-1)
.-1
ASYMPTOTIC
CONFIDENCE
309
INTERVALS
ZjSj)' (yj
Zj8j) = TX'
1,
. . . ,
&)
m,
L(81,
(A-2)
log TI'
C - (n/2) I
j=1
(z;(Y, -
j = 1, . .
ZSj,))
.,
(A-3)
ifj7h
- T-'Zzj
+
(yj
- Zj8j)
2(nTX.)-1(Z;(yj
= Gj
ifj
Zj6j)'Zj)
(A-4)
Let S = (Si, . . . , t) ' as in the text; then combiningthe elementsof (A-4) gives
-G,
_0
0
a2L/ 6O6
o~0
0
G2
G3
0
.
GM
310
MICHAEL
Suppose
Since
1,
E.
SOBEL
. . .,
M.
ZjSj)) = 0
from(A-4) we readilyobtain
o
0
Qi
Q2
Q3
0
0
.0
? . .
6)A N(O,[I(8)]-1)
(A-5)
block-diagonalmatrix
Qi1
0
o
tMe
Q21
0
o
~o0
o Q-1
0 o
is the
. . . Q-1
D.
F., AND
BIELBY,
W. T.,
HAUSER,
R. M.
HAUSER,
R. M.
ASYMPTOTIC
BISHOP,
CONFIDENCE
Y. M.,
FIENBERG,
311
INTERVALS
AND
S. E.,
HOLLAND,
P. W.
H.
M. (ED.)
P. M., AND
DUNCAN,
0.
D.
0.
D.
0.
D.,
FEATHERMAN,
D.
L., AND
DUNCAN,
B.
1972 Socioeconomic
Backgroundand Achievement.
New York: Seminar Press.
FINNEY,
J. M.
HEISE,
D.
R.
JORESKOG,
1973 "A general method forestimatinga linear structuralequation system." In A. S. Goldberger and 0. D. Duncan
(eds.), Structural
EquationModelsin theSocial Sciences.New
York: Academic Press.
LAND,
K. C.
LEWIS-BECK,
M. S.
MALINVAUD,
E.
MILLER,
R.
1966 Simultaneous
Statistical
Inference.New York: McGraw-Hill.
312
MICHAEL
E.
SOBEL
C. R.
RAO,
P.
1980 "On decomposition of effectsin causal models." Mannheim: Zentrum fur Umfragen,Methoden und Analysen.
Unpublished manuscript.
THEIL,
H.