Personal Relationships, 6 (1999),449469. Printed in the United States of America.

Copyright 0 1999 ISSPR. 1350-4126/99$9.50

The correlational analysis of dyad-level

data in the distinguishable case

RICHARD GONZALEZ‘ AND DALE GRIFFINb

‘University of Michigan; and bUniversity of Sussex

Abstract
Many theories of interpersonal relationships distinguish between individual-levelprocesses and dyadic or
group-level processes. This suggests that two-person relationships should be studied at the level of the dyad as
well as at the level of the individual. We discuss correlational methods for dyads when each dyad contains two
different types of individuals (e.g., a husband and wife, a mother and child, or an expert and a novice). In such
dyadic interaction designs,the dyad members are said to be distinguishable. We present a method for
computing the overall correlation for distinguishable dyads, and we discuss a model for separating the
dyad-level and individual-level components of such a correlation. The computational techniques and their
interpretation are described using data from 98 heterosexual couples.

By definition, interpersonal interaction in- the dyad, or both. Statistically,correlational

volves more than one person. The simplest methods developed for independent indi-
interaction to study, and likely the most viduals are not appropriate for designs fea-
common (Bakeman & Beck, 1974),involves turing interdependent dyad members (Grif-
two people. However, despite their relative fin & Gonzalez, 1995;Kenny, 1995).Instead,
simplicity, dyadic designs present both con- the degree of relatedness, or “statistical non-
ceptual and methodological challenges independence,” between dyadic partners
(Kenny, 1996; Gonzalez & Griffin, 1997). must be taken into account when correla-
Dyadic designs,like all group designs, pose a tions are tested and interpreted.
conceptual “levels of analysis” problem Furthermore, the statistical analysis de-
(Robinson, 1950). One could test theory at pends on whether the dyadic partners are
either the level of t-heindividual, the level of exchangeable (i-e., the two members of each
dyad are drawn from the same class or cate-
gory) or distinguishable (i.e., the two mem-
This research was supported by a grant from the Na- bers of each dyad are drawn from different
tional Science Foundation (R. Gonzalez), and by classes or categories).l Two examples will
grants from the Social Sciences and Humanities Re-
search Council of Canada as well as the Leverhulme
help make this distinction clear. Research
Trust Research Fellowship (D. Griffin). We thank on romantic relationships might focus on
Tony Greenwald, David Kenny, Bill Ickes, Jeff Simp- male homosexual couples (which contain
son, and the structural equation modeling seminar at exchangeable dyad members) or heterosex-
the University of Washington for their helpful sugges- ual couples (which contain distinguishable
tions on previous drafts.
Correspondence concerning this article can be ad-
dyad members). Research on social devel-
dressed to either author: Richard Gonzalez, Depart- opment might focus on dyadic play in same-
ment of Psychology, University of Michigan, Ann Ar- sex children (exchangeable dyad members)
bor, MI 48109, or Dale Griffin, School of Cognitive or dyadic play between fathers and sons
and Computing Sciences, University of Sussex, Fal-
mer, Brighton BN1 9QH, United Kingdom. E-mail 1. Kenny (1996) uses the terms “interchangeable”
can be sent to either [email protected](R. GonzaIez) and “noninterchangeabIe” to denote this distinc-
or [email protected] (D. Griffin). tion.

449
450 R. Gonzalez and D. Griffin

(distinguishable dyad members). In the ex- 1985) that decomposes the observed over-
changeable case, each partner is drawn from all within-partner correlation into latent dy-
the same category or type (e.g., adult males), adic and individual-level components, and
and thus both scores are treated as though the generalized actor-partner model (Gon-
they are sampled from the same statistical zalez & Griffin, 1997; Kenny, 1996) that de-
distribution. In the distinguishable case, composes the observed relations into par-
each partner belongs to a different category tial path coefficients (beta weights). The
or type (e.g., fathers vs. sons), and their latter model was used by Murray et al.
scores on each variable are drawn from dis- (1996). This article will focus on the dyadic
tributions that may differ, yielding the possi- effects model and discuss the meaning and
bility that these two samples may differ in computation of couple-level and individ-
their means, their variances, or their covari- ual-level correlations. However, it must be
ances. emphasized that the choice of an underly-
The decision to treat dyadic partners as ing model is a theoretical rather than a sta-
distinguishable or exchangeable should be tistical question (Kenny, 1996).
guided by theoretical assumptions rather We now turn to a more detailed discus-
than empirical tests. Thus, even if the two sion of these overall correlations. We will
types of partners do not differ in mean, review the pairwise coding, show how to
variance, or covariance terms, they should estimate these two overall correlations, pro-
be treated as distinguishable rather than ex- vide tests of significance,compare the SEM
changeable if they are distinguished in the (structural equation modeling) approach to
relevant theory. When other variables are the pairwise approach, and provide a con-
tested, the two types of partners may well crete example. We will then review the dy-
differ. Furthermore, no power is lost by adic model that decomposes the observed
treating the partners as distinguishable as overall within-partner correlation into la-
long as the relevant parameter estimates tent dyadic and individual-level compo-
are pooled across both types of partners. nents. This multilevel model will not only
We now turn to a specific research exam- provide new insight for the overall correla-
ple that we will use throughout this article. tions, but will also be a useful model in
Murray, Holmes, and Griffin (1996) col- many research settings.
lected trust and conflict ratings from both
members of 98 heterosexual couples. One
Estimating and Testing the Overall
question of interest addressed by this study
Correlations in the Distinguishable Case
is: Across the entire sample of men and
women, do ratings of trust relate to reports Suppose a researcher collected two vari-
of conflict? In other words, is there a signifi- ables, X and I: from each individual in a
cant overall within-partner correlation be- sample of heterosexual couples. For in-
tween trust and conflict? A second question stance, as in Murray et al. (1996), a re-
of interest involves a different overall cor- searcher could collect trust ratings (variable
relation: Across the entire sample, do rat- X ) from both the husband and the wife, and
ings of an individual’s trust relate to the reports of conflict (variable Y) from both
partner’s report of conflict? In other words, the husband and the wife. This results in
is there a significant overall cross-partner four variables: x h , X,, Yh, and Y,, where
correlation between trust and conflict? the subscript h represents husband and the
The meaning of these “overall” correla- subscript w represents wife. There are a to-
tions could then be examined in light of a tal of six correlations that can be computed
model of the underlying process that gave in this situation; these correlations are
rise to it. Although the data could be mod- shown in Figure 1.
eled in a number of ways (Kenny, 1996), The correlation between x h and Yh, de-
two specific underlying models seem of noted cor(Xh, Y h ) in Figure 1, is the within-
greatest relevance: a dyadic version of the partner correlation between trust and con-
group effects model (Kenny & La Voie, flict for husbands, and the correlation
Correlational analysis of dyad-level data 45 1

Figure 1. The six possible pairwise correlations in the distinguishable case.

between Xw and Y, is the within-partner Returning to Figure 1, the correlation

correlation between trust and conflict for between x h and Y , is the cross-partner cor-
the wives. Under some conditions (de- relation between the husband’s trust and
scribed below) it is possible to pool these his wife’s report of conflict, and the correla-
two correlations to form a single index, the tion between X , and Yh is the cross-partner
overall within-partner Correlation between correlation between the wife’s trust and the
trust and conflict. husband’s report of conflict. Again, under
The purpose of the overall within-part- some conditions, it is possible to pool these
ner correlation is to index the strength of two “cross” correlations to form the overall
the linear relation between two variables, X cross-partner correlation, which has advan-
and Y; across all individuals in the sample. tages similar to that of the overall within-
With N dyads, the overall within-partner partner correlation discussed above.
correlation involves 2N scores on each vari- Computing the overall within-partner
able. The practice of computing the overall correlation between trust and conflict in the
correlation has a number of advantages exchangeable case is relatively straightfor-
over the more common practice of comput- ward, because the distributions of each part-
ing separate correlations within the two ner’s scores are by definition equivalent.
types of dyadic partners (e.g., computing With N dyads, the 2N individual scores on
the correlation between trust and conflict trust are correlated with the 2N individual
separately for men and for women). First, if scores on conflict, and the result is evaluated
the correlational structure is identical using a significance test that corrects for the
across the two types, then the overall corre- degree of intradyadic similarity on each
lation serves as a more efficient estimate of variable (see Griffin & Gonzalez, 1995, for
the true correlation between variable X and the development of this test). However,
variable Y than either of the separate esti- when each dyad is made up of two different
mates. Second, the overall correlation will types of individuals (e.g., men and women),
usually have substantially more power to additional complications arise because of
reject the null hypothesis relative to either the possibility of between-partner differ-
one of the separate correlations. The proce- ences in means, variances, and covariances.
dure described here includes an explicit test In this article we extend the pairwise corre-
of whether the correlational structure is the lational model (Griffin & Gonzalez, 1995) to
same within both classes, a step often ig- the more complicated case of distinguish-
nored by researchers looking at each type able dyads and compare the pairwise ap-
of partner separately. proach with structural equation modeling.
452 R. Gonzalez and D. Griffin

Whenever possible, we relegate statistical the two within-category population vari-

details to the Appendix, allowing us to focus ances on Y must be equal, and (c) the two
the text on the conceptual “big picture.” within-category population covariances be-
In the distinguishable case, computing tween X and Y must be equal (e.g., the
and testing the overall within-partner and population covariance between X and Y for
cross-partner correlations can be achieved women must equal the population covari-
either through the pairwise correlation ance between X and Y for men). For the
method or through structural equation overall cross-partner correlation conditions
modeling (SEM) programs. When SEM is (a) and (b) must hold, and in addition (d),
estimated through the maximum likelihood the population covariances between one
approach, the two approaches yield identi- partner’s X and the other partner’s k; must
cal parameter estimates because the pair- be equal across the two classes (e.g., the
wise approach also provides the maximum population covariance between the wife’s
likelihood estimate, and the two approaches X and the husband’s Y must equal the
provided asymptotically equal significance population covariance between the hus-
tests under the null hypothesis. For those band’s X and the wife’s Y). Methods for
comfortable with SEM programs, the SEM testing each assumption separately are pre-
approach offers substantial advantages in sented in the Appendix. Below we present
terms of flexibility and adaptability. How- a method for testing all four assumptions
ever, the pairwise correlational approach is simultaneously using SEM.
a useful tutorial method and provides a uni- These equality assumptions have con-
fied framework across the distinguishable ceptual meaning and are more important
and exchangeable cases. After discussing than mere statistical worries. First, a differ-
the assumptions common to both ap- ence in variances between the two types of
proaches, we first describe the SEM method partners might suggest that different proc-
and then describe the pairwise method, esses may be operating in each type or cate-
which does not require special software. gory. Second, a difference in variances may
lead to differences in the observed correla-
tions even when the population covariances
Assumptions of the overall within-partner,
are equivalent. Third, when the assumption
cross-partner correlations
of equal covariances is violated for the
The overall within-partner correlation rxyis within-partner comparison, the pooled cor-
a summary of the strength of the linear re- relation no longer represents the relation
lation between variables X and Y across all within either type of partner (e.g., women
partners of both types. Because the overall or men). When pooling is appropriate, the
within-partner correlation is essentially a conceptual focus should be on the general
weighted average of the two within-partner processes occurring for both partners.
correlations (e.g., the correlations between When the pooling is not appropriate, the
trust and conflict for men and women sepa- conceptual focus should be on the differ-
rately), it is necessary to examine whether ence between the types of partners. As
the components of the separate within- these arguments suggest, the tests of the
partner correlations are similar enough to assumptions have implications for the di-
be combined. For an overall within-partner rection the theory should take.
correlation to be a legitimate summary of
the linear relation between the two vari-
Using SEM to estimate the
ables across both types of dyad members,
overall correlations
the following conditions must hold: (a) the
two within-category population variances The overall within-partner correlation and
on X must be equal (e.g., the population the overall cross-partner correlation can be
variance on variable X for men equals the estimated within SEM (using covariances
population variance on X for women), (b) or raw data as input; see Footnote 4). The
Correlational analysis of dyad-level data 453

SEM model presented in Figure 1 con- lation approach, the estimation of the over-
strains the two X variables to have equal all within-partner and overall cross-partner
variances, and also constrains the two Y correlations proceeds in four steps. First,
variables to have equal variances. Of the six the data are arranged in the pairwise man-
observed covariances, the two within-part- ner (Griffin & Gonzalez, 1995). Second, the
ner covariances between variables X and Y assumptions necessary for pooling the cor-
are constrained to be equal as are the two relations across the two categories of indi-
cross-partner cross-variable covariances. viduals (i.e., equality of variances within
The resulting SEM analysis yields a x 2 with variables and equality of covariances within
4 degrees of freedom (df)that provides an and across categories) are checked. Third,
omnibus test on all four constraints. Note the overall within-partner and cross-part-
that it is also possible to test each assump- ner correlations are computed as partial
tion separately within SEM by relaxing correlations with between-class mean dif-
each assumption and comparing the differ- ferences (e.g., sex differences) partialled
ence in x2s from the full model (df = 4) to out. Fourth, the overall within-partner and
each reduced model (df= 3). overall cross-partner correlations are tested
If the omnibus x 2 test is nonsignificant, for significance by a method that takes into
then the pooled covariance between X and account the degree of nonindependence
Y is an appropriate estimate of the “overall within dyad members.
within-partner covariance,” and the signifi- In the distinguishable case, the painvise
cance test associated with this parameter correlation model requires five columns of
tests the overall correlation against the null data (see Griffin & Gonzalez, 1995, for a
hypothesis. Similarly,the pooled covariance general discussion of the pairwise model).
between an individual’s X and the partner’s As shown in Table 1, the first column (la-
Y is an estimate of the “overall cross-part- beled C ) consists of binary codes repre-
ner covariance,” and the significance test senting the “category” variable (e.g., sex of
associated with this parameter tests the subject). If the researcher decided to code
overall cross-partner correlation against wives as “1” and husbands as “2” the first
the null hypothesis.2 In the SEM output, the column would consist of “1”in the first row
overall within-partner and cross-partner and “2” in the second row. This pattern
correlations will appear as standardized co- would be repeated for each of the Ndyads in
efficients. the sample, yielding a column containing 2N
binary codes. The second column (labeled
X)contains the scores on variable X corre-
Using the pairwise approach to estimate
sponding to the dyad member represented
the overall correlations
by the category code in column one. For ex-
The painvise approach offers an alternative ample, the first woman’s score on trust is ad-
method for estimating the overall correla- jacent to the first “1”in column one. Below
tions; it provides parameter estimates that that, the first man’s score on trust is adjacent
are identical to those given by SEM under to the first “2” in column one. This pattern
maximum likelihood. In the pairwise corre- continues, yielding a total of 2N scores.
Column three is created by the pairwise
2. One possible generalization of this model is a re-
reversal of column two. For example, adja-
laxation of the variance constraints (i.e., a model cent to each person’s score on trust in col-
that has all four variances as free parameters). On umn two is placed her or his partner’s score
the surface this generalized model may appear at- on trust in column three. This “reversed”
tractive; however, it is difficult to interpret because column of scores on X is referred to as X ‘ .
the resulting parameters are standardized accord-
ing to the particular variances involved. Thus, even
Columns four and five consist of the scores
when the covariances are set equal across classes, on variable Y (e.g., conflict), which are also
the standardized solutions for each class (e.g., the coded in the “painvise” format and labeled
correlations for men and women) will be unequal. Y and Y’, respectively. Table 1 presents
454 R. Gonzalez and D. Griffin

Table 1. Symbolic representation for the pairwise data setup for two variables in the
distinguishable case

Variable
Dyad
No. C X x Y r
1 1 XI1 XlZ YI 1 YlZ
2 x
12 Xl 1 y12 Yl 1
2 1 x
21 x
22 y2, y22
2 x22 XZl y22 Y2l
3 1 x3 1 x32 y31 y32
2 x
32 x3, y32 y31
4 1 x41 x42 y41 y42
2 x42 x41 y42 y41

Note: The first subscript represents the dyad and the second subscript represents the individual. Categorization
of individuals as 1 or 2 is based on the class variable C. Primes denote the reverse coding described in the text.

these five columns in symbolic form. The served overall within-partner correlation,
reasoning behind the pairwise coding is with the direction of the bias controlled by
that interdependence is built directly into the magnitudes and directions of the two
the data matrix. By constructing such links mean differences. The overall within-
between data points, subsequent analyses partner correlation must be corrected for
become relatively straightforward. these mean differences by partialling on the
The Appendix provides statistical tests category variable.Thus, the appropriate esti-
for each of the assumptions discussed in the mate of the overall within-partner correla-
previous section. Note that the painvise ap- tion is
proach tests assumptions separately,
whereas above we described an omnibus
test for the set of assumptions in SEM.
Those concerned about inflated Type 1 er-
ror rates for the separate tests of the as-
sumptions in the painvise approach may This equation is simply the partial correla-
wish to use a Bonferroni corrected a-level. tion between variables X and X holding C
Separate tests may be desirable because the constant, denoted by r,,., (where variables
omnibus test could mask violations (e.g., an to the right of the dot have been partialled
assumption may be violated even when the out). In our ongoing example, the corrected
omnibus test is not statistically significant). estimate of the overall within-partner cor-
Given that the assumptions are met, the relation would be the partial correlation
overall within-partner correlation is com- between trust and conflict with sex of sub-
puted using all 2N individual scores in col- ject partialled out. Most statistical packages
umns two (X)and four (Y) of Table 1-or have procedures for computing partial cor-
equivalently, columns three ( X ’ ) and five relations.
( Y ) . However, the simple correlation be- The computation of the overall cross-
tween column X and column Y does not, in partner correlation proceeds in a similar
general, estimate the “true” overall within- manner. The “partial” overall cross-partner
partner correlation. Instead, if there are correlation is given by
between-category mean differences on one
or both variables (e.g., men have a higher
score on trust than do women or vice versa),
these mean differences will bias the ob-
Correlational analysis of dyad-level data 455

Again, this is simply the partial correlation Rempel, 1989) and conflict (a 5-item index
between an individual’s X (column 2) and adapted from Braiker & Kelley, 1979, in-
his or her partner’s Y (column 5, or Y’) dexing the self-rating of the frequency of
holding C constant. overt behavioral conflict). The relevant
Significancetests for the two overall cor- means, variances, covariances, and correla-
relations are presented in the Appendix. tions are presented in Table 2. The relevant
The reader may examine the symbolic form partial pairwise correlations needed for
of the significance test to develop intuition testing the overall within-partner correla-
for how interdependence influences these tion rxy and the overall cross-partner cor-
tests. Even though significance tests for the relation rXy’are presented in Table 3. Each
painvise approach are asymptotically entry in Table 3 is a partial correlation in
equivalent to the Z-tests in SEM under the that the category variable C has been par-
null hypothesis, in a single sample the two tialled out. The information needed to
tests will not be identical. A Monte Carlo compute all statistics and tests presented
simulation that assesses the performance of in this article is contained in those two
the painvise and SEM tests for rxyis given tables.
in the Appendix. The simulation suggests The SEM analysis estimating the overall
that the pairwise approach performs at within-partner and cross-partner correla-
least as well as the SEM approach with re- tions tests all assumptions simultaneously;
spect to the effective Type I error rate, with the omnibus test with 4 degrees of freedom
the SEM approach tending to be slightly was nonsignificant ( x 2 = 7.98, p = .09), in-
liberal. dicating that it was appropriate to interpret
the pooled overall correlations as summa-
ries of the relation between trust and con-
Illustrating the overall within-partner and
flict across the entire sample.
cross-partner correlations for the
In the pairwise approach, the assump-
distinguishable case
tions are tested individually. An examina-
In the following section we use data col- tion of the diagonal elements in Table 2
lected by Murray et al. (1996) to illustrate reveals that men and women had approxi-
the pairwise analysis of the overall within- mately equal variances on each variable.
partner correlation when dyadic partners The t values corresponding to the depend-
are distinguishable. Murray and colleagues ent variances test presented in the Appen-
collected data from both members of 98 dix were 1.31 for trust and 1.04 for conflict,
heterosexual couples. We chose two vari- neither of which is statistically significant at
ables from their data: trust (Holmes & a = .05 with 96 degrees of freedom. The

Table 2. Correlations, variances, and covariances from 98 couples

Trust Conflict
Female Male Female Male
Female Trust 1.351 0.440 - 0.857 - 0.443
Male Trust 0.287 1.743 - 0.365 - 0.438
Female Conflict - 0.474 - 0.178 2.421 1.025
Male Conflict - 0.269 - 0.234 0.465 2.006
Means 7.631 7.509 2.992 2.882

Note: Correlations appear below the diagonal, variances on the diagonal, and covariances above the diagonal.
The means on each variable (e.g., the mean female trust score) are presented in the last row. From Murray et
al.. 1996.
456 R. Gonzalez and D. Griffin

Table 3. Partial pairwise correlation p < .001. The SEM Z was -4.27. The over-
matrices for trust (T) and conflict (C)from all cross-partner correlation was -.218,
Murray et al. (1996) with an effective sample size of 175.0 and
a pairwise Z of -2.88. The SEM Z was
Trust and Conflict -2.67 for the identical standardized’ esti-
T T’ mate.

T‘ .284
C - .350 - .218 Separating Dyad-Level and
C’ - .218 - .350 .463 Individual-Level Effects
In contrast to the overall correlations dis-
Note: The prime denotes the “reverse” coding of the
variable as described in the text. Boldfaced values cussed above, the dyadic effects model fo-
are the partial intraclass correlations. cuses on three different questions. First,
are dyadic members similar to each other
on each variable? That is, is there signifi-
assumption that covxy for women equals cant dyad-level variance in each variable?
covxyfor men is tested next3 (again, see the Second, if there is dyad-level variance in
Appendix for details on tests of dependent each variable, are the two variables corre-
covariances). For trust and conflict, covxy lated at the dyadic level? That is, is there
equaled -.857 for women and -.438 for a significant dyad-level correlation be-
men. This difference was not statistically tween the trust ratings and the reports of
significant by the test on dependent covari- conflict? Third, within each dyad, does the
ances given in the Appendix, Z = -1.64. relative standing of the two individuals on
The final assumption tested is that the two one variable relate to their relative stand-
cross-partner covariances are equal. The as- ing on the second variable? That is, is there
sumption that cov,,’ for one partner (e.g., a significant individual-level correlation
the population covariance between between trust and conflict? These ques-
women’s trust and partner’s report of con- tions assess whether the overall within-
flict) equals covxy‘for the other partner partner correlation reflects dyad-level
(e.g., the population covariance between processes, individual-level processes, or
men’s trust and partner’s report of conflict) both.
appears to fit. The difference between the The overall within-partner correlation
trust-conflict covariances across partners serves as a summary of the relation between
was not statistically significant by the co- variables X and Y across all individuals,but
variance test presented in the Appendix, Z it does not reveal whether the relation be-
= -.33. tween variables X and Y exists at the level
Estimates of the two overall correlations of the individual, at the level of the dyad, or
are identical for the SEM and pairwise ap- both. Figure 2 presents one model of the
proaches, although the tests of significance sources of the linear relation between Xand
differ slightly. The overall within-partner Y that allows the separation of individual-
correlation (partialled on sex) between level and dyad-level effects (Kenny & La
trust and conflict was -.350. Thus, a nega- Voie, 1985). Once again, we emphasize that
tive correlation existed between trust and the appropriateness of this model is not
reports of conflict when pooling both sexes provable by statistical means; alternate
and controlling for mean differences. In the structural models (e.g., Kenny, 1996) may be
pairwise analysis, the effective sample size theoretically justified. In the dyadic effects
N*1 = 166.24, and the Z value was -4.51, model for the distinguishable case, the vari-
ance of a given observed variable is as-
3. Note that when the variances are equal across
sumed to result from three different
classes, the equality of covariance assumption can sources: variation due to category member-
be tested using either covariances or correlations. ship (which is partialled out from the model
Correlational analysis of dyad-level data 457

'i
'i /
,
--
J J 1

variance variance

Ixl El lyrJ

2
&
"\ \m A
m Shared

variance

t 'd

Figure 2. A latent variable model separating individual-level (unique) and dyad-level

(shared) effects using the pairwise approach.

shown here), variation due to dyads, and scores in a distinguishable dyadic design is
variation due to individuals within dyads. modeled as a weighted sum of the dyad-
The dyadic component of a given variable is level correlation (rd) and the individual-
the portion of that variable that is shared within-dyad correlation (ri).The dyad level
between dyadic partners; the individual correlation contributes more to the overall
component is the portion of a variable that within-partner partial correlation when the
is unshared between dyadic partners. shared variance between partners is large
As Figure 2 illustrates, in this model the (as indexed by the partial intraclass correla-
covariation of two variables is also assumed tions rxx'.cand ryy,.c),whereas the individ-
to result from the same three sources. First, ual-level correlation contributes more to
mean category differences on X and Y the overall within-partner partial correla-
(which are held constant in the pairwise tion when the shared dyadic variance is
model) give rise to one portion of the ob- small.
served overall correlation. Second, the dy-
adic (shared) portions of X and Y are re-
lated through the dyadic correlation r d . Assumptions of the dyadic model
Finally, the individual (unshared or unique)
portions of X and Y are related through the The assumptions for the distinguishable
individual-level correlation ri. After the dyadic model are identical to the assump-
class differences are partialled out, the cor- tions for the overall correlations (equal
rected overall within-partner correlation variances on X, equal variances on X
can be decomposed into the remaining two equal covariances between an individual's
parts: X and the partner's I: and equal covari-
ances between X and Y within each class).
As with the overall case, these assumptions
'xy.c = C G r d -k
can be tested either separately with the
J I - ) ; x ' c J G5 (3) tests provided in the or simultaneously
with SEM. The omnibus test in SEM is the
4-degrees-of-freedom x* test given by the
In words, the overall within-partner partial model in Figure 2 with the observed co-
correlation between all X scores and all Y variance matrix as input. Note that this is
458 R. Gonzalez and D.Grifin

the identical x 2 from the omnibus test of structural model. Instead, this model is
the assumptions for the overall correla- based on the six covariances4 presented in
tions. Table 2. Third, when using SEM the model
The dyadic effects model has one addi- in Figure 2 could be generalized to allow for
tional assumption that must be examined as separate individual-level correlations ri for
well: The intraclass correlations must both each class (i.e., separate individual-level
be significant (Kenny & La Voie, 1985).The correlations for men and women), whereas
partial intraclass correlation yields the pro- the pairwise model can only estimate a
portion of shared or dyad-level variance in pooled individual-level correlation.
each variable. Stated another way, the The output from an SEM program will
square root of the intraclass correlations provide the estimates of rd and ri, as well as
define the “reliability coefficients” for the tests of significance for these parameters.
underlying dyadic latent variables. If the in- We identified the model by setting the vari-
traclass correlations are zero, then there is ances of all latent variables to 1 (there are
no dyad-level variance, and Td is meaning- two latent dyadic variables and four latent
less. A simple asymptotic test for the partial individual variables) and requiring the two
intraclass correlation is provided in the Ap- individual-level loadings on each variable
pendix. to be equal (e.g., the individual loading on
The SEM framework offers more flexi- trust is set equal for men and women) and
bility than does the pairwise approach be- the two dyadic loadings on each variable to
cause it permits a relaxed model with dif- be equal (e.g., the dyadic loading on trust is
ferent individual-level correlations for each set equal for men and women). This para-
member. If the x2 is significant, then the meterization facilitates generalization to
pooled dyadic structural model does not fit more variables and groups of larger size, as
the data. One option available in that case discussed later in this article. Under the
is to refit the model with separate individ- present identification scheme, the unstan-
ual-level correlations, and examine the dardized coefficients are the correlations.
change in x 2 between the full and reduced For different identification schemes, one
model (see the Appendix for the necessary would need to use the standardized coeffi-
EQS syntax). cients rather than the unstandardized coef-
ficients. See Bentler (1995) for an illustra-
Using SEM to estimate the dyadic tion of different methods of identifying
model such a model; Gonzalez and Griffin (1998)
Figure 2 also illustrates the structural
model used to estimate the two underlying 4. For some simple SEM models, a correlation matrix
correlations rd and ri in the SEM approach. or a covariance matrix as input leads to identical
Sample EQS syntax and guidelines for solutions (see Bollen, 1989, or Long, 1983, for a
nontechnical discussion of this issue). However, a
AMOS are presented in the Appendix. The covariance matrix should be used as input when-
SEM approach involves three important ever there are equality constraints in the model. In
changes from the pairwise structural model. the present models, the two dyadic loadings from
First, instead of the pairwise data setup each latent variable leading to the two individual
(e.g., columns X and X ’ each with the entire scores are constrained to be equal and the two
individual loadings leading to a given variable are
2N set of observations differing only in or- constrained to be equal. Thus, the models pre-
der), the observed variables are coded sented here must be evaluated by using unstan-
separately by categories (e.g.,Xwrepresents dardized input. Note also that the individual-level
the wife’s score on trust and x h represents correlations are identical to correlated errors pro-
the husband’s score on trust, each column vided by standard SEM models; however, the para-
meterization provided here is slightly different
with N observations). Second, and follow- from usual to make the multilevel nature of the
ing from the first point, the category effect data clearer and to allow easy generalization to
is not directly partialled out from the SEM more than two variables.
Correlational analysis of dyad-level data 459

participate individually but are randomly

discuss the implications that different iden-
tification schemes have on the 2 tests. collected into dyads by the experimenter
after the data have been collected. In this
situation, as long as there is a significant
Parameter estimation and significance individual-level correlation, correlations
testing computed on “dyadic” means could be sub-
In this section we show how the dyad-level stantial, even though the dyads never ex-
isted! This is because the dyad average
and individual-level correlations can be
combines the unique qualities of the indi-
computed from the partial pairwise correla-
viduals with the “emergent properties” of
tions given in Table 3. These estimates are
the dyad. A dyad-level correlation should
identical to those given in SEM under
maximum likelihood estimation (i.e., the represent only the relation between the
shared or emergent properties of the dyad
pairwise approach also provides maximum
likelihood estimators). on two variables.
The dyad-level correlation addresses The individual-level correlation ad-
the following question: Does the similarity dresses the following question: Does the re-
sidual of an individual’s score on variable X
of the members of a dyad on X (i.e., both
tending to be high on X or both tending (beyond what would be expected by mem-
to be low on X) relate to the similarity of bership in a particular dyad) relate to the
residual of that individual’s score on vari-
the members of that dyad on Y (i.e., both
tending to be high on Y or both tending able Y (beyond what would be expected by
to be low on Y)? Following the logic de- membership in a particular dyad)? Under
veloped in Griffin and Gonzalez (1995), the dyadic model, the individual-level cor-
the dyad-level correlation in the distin- relation can be expressed as
guishable case is

Individual-level correlations represent the

Because dyad-level correlations represent degree of relation between the unique vari-
the degree of linear relation between the ance on variable X (how a person differs
shared dyadic components of each variable, from his or her partner on X) and the
there can be no dyad-level correlation with- unique variance on variable Y (how a per-
out within-dyad similarity on both vari- son differs from his or her partner on Y).
ables, as indexed by a significant intraclass The individual-level correlation can exist
correlation (or equivalently, by a significant whether or not there is intra-dyadic similar-
amount of variance in the latent variable). ity, as long as the intra-dyadic similarity is
Therefore, the partial intraclass correla- not perfect. The individual-level correlation
tions rxx,.cand nwd to be nonzero.
ryyl.c differs from the overall within-partner cor-
Note that the dyad-level correlation is relation because the latter correlation is not
quite different from the correlation be- constrained by the unique portion of an in-
tween the dyad means because the mean- dividual’s score, and instead contains (like
level correlation is not constrained by the the mean-level correlation) a mix of indi-
dyadic similarity within each variable, and vidual and dyadic effects (see Equation 3).
instead contains a mix of dyad-level and Both r d and Ti are estimated latent vari-
individual-level effects (see Griffin & Gon- able correlations and are “disattenuated”
zalez, 1995, for an extended discussion). for the proportion of dyad-level and indi-
This problem can be seen most clearly in a vidual-level variance, respectively, in each
“pseudo-dyadic’’ design where individuals variable. The disattenuation interpretation
460 R. Gonzalez and D. Griffin

is driven by the denominators of Equations tion of variance in conflict that was shared
4 and 5. The numerator of rd is the overall between the partners. The same informa-
cross-partner correlation r,,‘.,, which in- tion is obtained from the standardized path
dexes the strength of the dyadic relation in coefficients of the SEM model after they
raw score terms. When rxY’.,is 0, then rd are squared. When the two partial intraclass
must be 0. correlations are tested by the asymptotic 2
Significance tests for the distinguishable tests, both are significant: 2 = 3 . 0 6 , ~< .01
painvise case are, with one slight exception, for trust, and 2 = 5.83,~< .001,for conflict.
identical to those presented in Griffin and The relevant tests from the SEM output are
Gonzalez (1995) for the exchangeable case. 5.39 and 8.28, again statistically significant.
The asymptotic tests include those for the Recall that the two tests will not coincide
partial intraclass correlations rXx,.,and r,,‘.,, except asymptotically under the null hy-
which index the amount of dyadic or shared pothesis.
variance in each variable, the partial raw- Turning to the multilevel correlations,
score dyadic correlation rxY’.,, which in- the pairwise analysis yields a dyad-level
dexes the dyadic relation uncorrected for correlation between trust and conflict of
the degree of shared variance, and the la- -.601. This value is obtained by disattenu-
tent dyadic correlation rd. The individual- ating the overall cross-partner correlation
level correlation ri is tested using a standard rxy’ = -.218 by the amount of dyad-level
t test for Pearson correlations, and this test variance in each variable, rxx‘= .284 and ryyl

loses a degree of freedom relative to the

exchangeable case owing to the partialling = .463. In this case, rd = -.’*’
J284Jzz
=

of variable C . Details of the tests can be -.601. Note that exactly the same estimate
found in the Appendix. is obtained from SEM. Under special cir-
Although the significance tests from the cumstances the estimated latent correlation
SEM and pairwise approaches are equiva- can obtain “out of bounds” values greater
lent under the null hypothesis, in practice than 1.0.Some programs such as EQS auto-
the null hypothesis will not be exactly true matically restrict the possible values of cor-
and the significance tests from the two relations to fall between - 1and 1, and then
methods will differ. In the Appendix we the pairwise and SEM values will diverge.
present simulation data comparing the tests When tested by the pairwise significance
of significance provided by the two models.
test, rd is statistically significant,2 = -2.88,
Note that the significance test from the
p < .01 (the SEM 2 value is -3.49).
pairwise approach is slightly more well-be-
Finally, the individual-level correlation is
haved than the SEM version (see Gonzalez
obtained by correcting the difference be-
& Griffin, 1998,for a discussion of the limi-
tween the combined individual and dyadic
tations in SEM significance tests).
correlation rxyand the pure dyadic correla-
tion rxy’for the amount of unique individ-
An illustration of the dyadic effects model ual-level variance in each variable, l-rxxl
and l-ryy*.In this case, the individual-level
We use the Murray et al. (1996) data to
correlation between trust and conflict is
illustrate the distinguishable dyadic effects
-.212, yielding a t(96) = 2 . 1 2 , ~< .05 (the
model. For trust, the observed partial in-
SEM the corresponding estimate of ri is
traclass correlation of .284 indicates that
identical and the 2 value is 2.19).
about 28% of the variance in trust was
shared between dyadic partners.5 Similarly,
for conflict, the observed partial intraclass
Interpreting these results
correlation of .463 represents the propor-
How should these results be interpreted?
5. Note that for the intraclass, the percentage of vari- First, the relations across men and women
ance explained is given by r not rz. between trust and conflict were symmetric.
Correlational analysis of dyad-level data 461

That is, the relation between men’s trust members, or (b) because of the selection
and women’s report of conflict was roughly principles that led couples to end up to-
equal to the relation between women’s gether, or (c) because of some third vari-
trust and men’s report of conflict. This is an able operating on both partners. Only when
important conclusion in itself, but is also a dyad members are randomly assigned can
necessary assumption for the dyadic effects we conclude that the similarities on each
model to be appropriate. Second, the mem- variable, and the dyad-level correlation be-
bers of heterosexual romantic couples were tween each variable, reflect actual dyadic
moderately similar (at a correlational, but processes (or shared environments that
not necessarily at a mean level, because have dyad-level effects).
mean sex differences were partialled out of Once again, we stress that multilevel
this analysis) on levels of both trust and conclusions cannot be reached by the com-
conflict. The significant intraclass correla- mon practice of (a) correlating dyad means
tions indicated that men and women within on X and Y or (b) correlating individual
a couple tended to resemble each other, scores on X and E: As noted earlier, both
although the proportions of variance were the mean-level correlation r, and the over-
only moderate (28% and 46% dyad-level all within-partner correlation are weighted
variance for trust and conflict, respec- combinations of the dyad-level and the in-
tively). dividual-level correlation.
Finally, and most important, trust and
conflict were related at both the level of the
dyad and the level of the individual. This
Extensions to Larger Groups and More
situation is by no means preordained. Sub-
Variables
stantial correlations may be observed at
only one level of analysis, and in special Clearly, the two-person, two-variable prob-
cases, each of the two levels can have oppo- lem we have demonstrated here is the sim-
site signs. Recall that the dyad-level correla- plest possible setup. In the distinguishable
tion represents dyadic processes because it case, because of the multiple equality con-
indexes the extent to which the similarity straints, the pairwise approach becomes
between men and women on X (trust) re- cumbersome when larger groups or multi-
lates to the similarity between men and variate analyses are desired. The SEM ap-
women on Y (conflict). Conceptually, this proach, however, is well-suited to such ex-
suggests that the dyadic processes that give tensions. (Note that the standard SEM
rise to shared positive trust are related to approach is not easy to implement in the
those dyadic processes that reduce shared exchangeable case.) For example, imagine
perceptions of conflict. In this case, an indi- that one collected trust and conflict scores
vidual who was relatively more trusting on members of family units where each
tended to report relatively less conflict, so family consisted of a father, mother, daugh-
that both levels moved in the same direc- ter, and son, and where the four categories
tion. But it is easy to imagine situations were treated as distinguishable. An SEM
(theoretically justified) for an opposite re- model generalized from Figure 2 could ad-
sult. If the individual-level relation primar- dress such questions as: Are trust and con-
ily reflected neuroticism, for example, one flict related overall, at the family level, and
might find that the more neurotic and en- at the individual level? (See Cook, 1994, for
meshed partner was both more trusting and a similar model of family functioning using
more likely to report conflict. a latent variable approach.) With four dis-
Note the limitations provided by an ex- tinguishable family members, a step-down
isting couples design. We do not know approach to the equality constraints should
whether the significant dyad-level correla- be used: For example, after testing for a
tions came about (a) because of dyadic common overall within-individual correla-
processes that occurred between the dyad tion, the constraints might be relaxed to al-
462 R. Gonzalez and D.Griffin

low the children to have a different overall ther benefits come when multiple distin-
correlation than the adults. guishable members and/or multiple vari-
Or, as another example, one might wish ables are analyzed. However, we believe
to study whether the multilevel relations that contrasting the merits of the two ap-
between trust and conflict in couples re- proaches is not as important as illustrating
main when a third variable, such as rela- their common abilities to separate dyadic
tionship satisfaction, is held constant. This and individual-level relations.
could be accomplished by extending the We hope that analytic methods that sepa-
model presented in Figure 2 so that the cor- rate the relative contribution of dyad or
related latent dyadic variables underlying group-level processes from individual-level
trust and satisfaction were used to predict processes may encourage the development
the latent dyadic variable underlying con- of theory in interpersonal behavior. Every
flict, and the correlated individual-level step of the model-testing procedure pre-
variables underlying trust and satisfaction sented here may help to guide theory about
were used to predict the individual-level the multilevel nature of interpersonal be-
variable underlying conflict. Examples of havior. For example, the explicit testing of
this extension (and others) are given in the equality of the two cross-partner corre-
Gonzalez and Griffin (1997). lations should encourage theory-building
about when relations between variables
across classes of individuals will be equal
Summary and Conclusion
and when they are likely to be unequal.
We have presented a technique for analyz- We believe that the most important as-
ing correlational data from distinguishable pect of this model is the explicit separation
dyads. The pairwise correlation model has of dyad-level and individual-level relations.
the advantage of simplicity,both conceptu- At the moment, little theory in social psy-
ally and in the computer software required. chology has been developed about how
Simulation results presented in the Appen- dyad-level or group-level processes differ
dix show that the pairwise approach is at from individual-level processes. However,
least as accurate as SEM in protecting in the present examples we found that in
against Type I error when the null hypothe- romantic couples the relations between
sis is true. A benefit of the structural equa- trust and conflict appear at both levels. Per-
tions modeling approach is that it allows a haps empirical findings such as these will
test of whether separate individual-level promote more precise and complex multi-
correlations are needed for each class. Fur- level theories of interpersonal behavior.

References
Bakeman, R., & Beck, S. (1974). The size of informal iterative approach to the analysis of family data.
groups in public. Environment & Behavior, 6, Annals of Human Genetics, 55, 77-90.
378-390. Gonzalez, R., & Griffin, D. (1997). On the statistics of
Becker, R. A., Chambers, J. M., & Wilks, A. R. (1988). interdependence: Treating dyadic data with re-
The new S language. Pacific Grove, CA: spect. In S. Duck (Ed.), Handbook of personal re-
Wadsworth. lationships (2nd ed., pp. 271-302). New York: Wiley.
Bentler, P. M. (1995). EQS Structural Equations Pro- Gonzalez, R., & Griffin, D. (1998). Correct tests of
gram Manual. Encino, CA: Multivariate Software. significance f o r parameters in SEM. Unpublished
Bollen, K. A. (1989). Structural equations with latent manuscript, University of Michigan, Ann Arbor,
variables. New York: Wiley. and University of Sussex, Brighton, UK.
Braiker, H. B., & Kelley, H. H. (1979). Conflict in the Griffin, D., & Gonzalez, R. (1995). The correlational
development of close relationships. In R. L. Bur- analysis of dyad-level data: Models for the
gess & T. L. Huston (Eds.), Social exchange in de- exchangeable case. Psychological Bulletin, 118,
veloping relationships. New York Academic Press. 430439.
Cook, W. (1994). A structural equation model of dy- Holmes, J. G., & Rempel, J. K. (1989). Trust in close
adic relationships within the family system.Journal relationships. In C. Hendrick (Ed.),Review ofper-
of Consulting & Clinical Psycho6ogy, 62, 500-510. sonality and social psychology (pp. 187-220). New-
Eliasziw, M., & Donner, A. (1991). A generalized non- bury Park, CA: Sage.
Correlational analysis of dyad-level data 463

Kendall, M., & Stuart, A. (1966). The advanced theory Long, J. S. (1983). Confirmatory factor ana1ysis:A pref-
of statistics (2nd ed.). London: Griffin. ace to L l S R E L . Beverly Hills, CA: Sage.
Kenny, D. A. (1979). Correlation and causality. New Murray, S. L., Holmes, J. G., & Griffin, D. W. (1996). The
York: Wiley. benefits of positive illusion: Idealization and the
Kenny, D. A. (1995). Design issues in dyadic research. construction of satisfaction in close relationships.
Review of Personality and Social Psychology, 11, Journal of Personality and Social Psychology, 70,
164-184. 79-98.
Kenny, D. A. (1996). Models of non-independence in Robinson, W. S.(1950). Ecological correlations and the
dyadic research.Journal of Social and Personal Re- behavior of individuals.American Sociological Re-
lationships, 13, 279-294. view, 15, 351-357.
Kenny, D. A., & La Voie, L. (1985). Separating individ-
ual and group effects. Journal of Personality and
Social Psychology, 48, 339-348.

Appendix
This appendix summarizes several statisti- A computationally simpler method to
cal tests that are mentioned in the text. test two dependent variances was given by
Kenny (1979). One creates two new vari-
ables: a sum within dyads and a difference
Testing the Difference Between Two within dyads. That is, a new variable equal
Dependent Variances to the sum of the two scores within a dyad
To test the difference between the variance (denoted S ) and a new variable equal to the
of each individual in the dyad, select only difference of the two scores within the dyad
those cases coded with a “1”in column one (denoted D).The test for the correlation
and compare the variance of X (column between S and D (i.e., the usual t-test for
two) with that of X ’ (column three), and the Pearson correlation with N - 2 degrees
compare the variance of Y (column four) of freedom as in Equation A4 below) is
with that of Y’ (column five). We illustrate equivalent to the test for two correlated
with an example from Murray et al. (1996). variances presented in Equation A l .
The variance for the women on trust was
1.35 and the variance for the men on trust Testing the Difference Between ’Itro
was 1.74. These two samples are not inde- Dependent Covariances
pendent because the women and men are
paired as a function of their dyad. To test We present a simple asymptotic Z test that
the null hypothesis that the population vari- directly compares two dependent covari-
ance for the women equals the population ances (rather than correlations). Using the
variance for the men, one can use the t-test asymptotic property that

(Kendall & Stuart, 1966), a Z test for two

dependent covariances can be constructed.
with N - 2 degrees of freedom where r is We illustrate with an example from the
the cross-partner correlation on a given Murray et al. (1996) data. Suppose the re-
variable, V , is the sample variance for the searcher wanted to test whether the covari-
women, and V m is the sample variance for ance between a woman’s trust and her con-
the men. From Table 2, the correlation be- flict equaled the covariance between a
tween the woman’s trust and the man’s man’s trust and his conflict (i.e., the equality
trust is .287, yielding a nonsignificant test constraint necessary to justify the pooling
in the overall within-partner correlation
(1.351 - 1 . 7 4 3 ) a rxy). Let C denote the sample covariance
t= = -1.31 and V denote the sample variance. For con-
2J(1.351)(1.743)(1 - .2872) venience we denote the woman’s trust as 1,
464 R. Gonzalez and D. Griffin

the woman’s conflict as 2, the man’s trust as indexing the dependence of an individual’s
3, and the man’s conflict as 4. The necessary X on his or her partner’s Y).
variances and covariances are denoted us- Given the four elements of the partial
ing subscripts (e.g., C12 denotes the covari- pairwise correlation matrix in the distin-
ance between the woman’s trust and her guishable case (rXy,,,rxx’.crryy‘.c,and rxy?.&
conflict; VI denotes the variance of the significance testing of the overall within-
woman’s trust). The test for two dependent partner correlation proceeds in the same
covariances (e.g., C12 and C34) is given by manner as for the exchangeable case (Grif-
fin & Gonzalez, 1995).6 First, we consider
the test for the overall within-partner corre-
lation. Under the null hypothesis that pXy.,
= 0, the approximate large-sample variance
of r,,, is -1 ,where
where COrepresents the average of the two
sample covariances being tested (i.e. (C12 + N;
C34)/2). This 2 can be compared to the criti-
cal value of 1.96 for a two-tailed test at a =
.05.
Using the values from Table 2, we can
test whether the covariance between the Thus, the overall (partial) correlation rXy.,
woman’s trust and her conflict ( - 3 5 7 ) dif-
can be tested using the critical ratio 2 =
fers from the covariance between the
male’s trust and his conflict (- .438). The rxy., , or more simply z = rxy.,JN7. The
resulting test statistic is

(-.857 - (-.438))&% observed 2 can be compared to the stand-

ard table of the normal distribution. Be-
cause this test is asymptotic we ignore the
+ 2(-.64752) - 2[(.440)(1.025) loss of one degree of freedom due to par-
+ (-.443)(-.365)] tialling the class variable C.
= -1.64, Intuitively, N ; can be thought of as the
“effective sample size” for rXy.,adjusted for
which is not statistically significant by con- dependent observations, and it typically
ventional standards. ranges between N (the number of dyads)
and 2N (the number of individuals) de-
pending on the degree of nonindependence
Significance Testing for the Pairwise
within dyads (Griffin & Gonzalez, 1995).7
Approach
Thus, under the null hypothesis, when there
The significance test for the overall within- is complete dependence within dyads the
partner correlation requires four elements: overall within-partner correlation and its
the number of dyads and three indices of significance test reduce to the usual (par-
the degree of nonindependence within dy-
ads (Griffin & Gonzalez, 1995). In the dis- 6. As in our previous article dealing with the ex-
tinguishable case, the indices of noninde- changeable case, we focus on hypothesis testing
pendence include the partial intraclass rather than interval estimation. The hypothesis-
correlation on variable X (rxx,.c,indexing testing approach has the advantage of simplifying
the within-dyad similarity on X),the partial the standard error under the null hypothesis (see
the Appendix in Griffin & Gonzalez, 1995).
intraclass correlation on variable Y (rYy,.,, 7. As discussed in our earlier report (Griffin & Gon-
indexing the within-dyad similarity on Y), zalez, 1995), when the two intraclass correlations
and the partial overall cross-partner corre- differ in sign it is possible for the effective sample
lation between variables X and Y’ (rXy,.,, size to exceed 2N.
Correlational analysis of dyad-level data 465

tial) Pearson correlation using one individ- adjusted for dependent observations and
ual from each dyad because the two indi- disattenuation. When rxx’.cand/or ryyf.care
viduals are essentially identical. When there small, rd will tend to be large and may even
is complete independence within dyads on exceed 1.0. Under these circumstances, D*
both variables, the correlation and its sig- will be small and so the “inflated” value of
nificance test essentially reduce to the (par- rd will generally not be significant. Note
tial) Pearson estimates obtained using both that the p values for testing rxy’..and rd
members of each dyad. against their respective null hypotheses will
Similarly, the asymptotic variance of the always be identical in the painvise ap-
overall cross-partner correlation rxy’.cunder proach.
1 The individual-level correlation rj can be
the null hypothesis is - where N ; = computed directly from the pairwise setup
N2*
as per Equation 5, or equivalently by corre-
2N . Nr can be thought of lating the difference scores within a dyad
+ rx2yy‘.c
1+ rxx‘.cryy’.c (e.g., female minus male scores on trust cor-
as the “effective sample size” for the overall related with female minus male scores on
cross-partner correlation rxy’.cadjusted for conflict). The significance of ri can be tested
dependent observations, and under the null using the usual Pearson correlation table
hypothesis can range between N, the (or the associated t test for a Pearson corre-
number of dyads, and 2N, the number of lation
‘xy’.c
individuals. The critical ratio - -
-
Ir rJN-2

rXyf, ,/N; is tested as a 2 statistic. Note

that these tests (and those following), de- with N - 2 degrees of freedom).
spite being simplified, are asymptotically
equivalent to the 2 tests provided in an
SEM program under maximum likelihood Simulation Study on the Dyad-Level
estimation. Correlation and the Overall
The partial intraclass correlation (e.g., Within-Partner Correlation
rxx’.c)can be tested against its standard
We conducted a small simulation study on
1 - r&j.c the dyad-level correlation to compare the
error as a Z test of the null hy-
JN painvise approach to the SEM approach.
pothesis, where N is the number of dyads. The simulation was based on the algorithm
The observed 2 ratio of ( f i r x x t . c ) / ( l - described by Eliasziw and Donner (1991)
and was written in the S statistical language,
rxx’.c2) is compared to 1.96 for a two-tailed
making use of its built-in random number
test at a = .05.
generator (Becker, Chambers, & Wilks,
The significance test for rd is related to
1988). For each set of population values
the test for rxy’. presented earlier. Under listed in Table 4,500 samples of either N =
the null hypothesis that Pd = 0, the asymp- 30 or N = 200 were drawn from a multivari-
1
totic variance of rd is - where ate normal distribution with unit variances
D* and a mean difference of 5 between the
population ps for the two classes. Popula-
tion correlation parameters were chosen to
ensure a correlation matrix with full rank.
All other details parallel the previous simu-
Thus, Z = rdm.Intuitively, D* can be lation performed on the exchangeable case
thought of as the “effective sample size” (Griffin & Gonzalez, 1995).
466 R. Gonzalez and D. Griffin

Table 4. Bias and effective Type I error rates for the latent variable model for the pairwise
and SEM approaches
Null hypothesis that Pd =0 NO.
of Samples
Average Pairwise SEM Consistent
Pxx. = Pyy' Pi Estimate rd Type I Error 'Qpe I Error with Model
N = 30
.so -.5 .078 .039 .047 487
SO 0 -.010 .040 .058 497
SO .5 - .055 .043 .070 486
.75 -.5 .016 .052 .080 500
.75 0 .006 .040 .080 500
.75 .5 .002 .050 .072 500
N = 200
SO - .5 .001 .060 .050 500
SO 0 - .002 .030 .036 500
SO .5 - .004 .048 .060 500
.75 - .5 .001 .058 .064 500
.75 0 .001 .054 .060 500
.75 .5 - .001 .068 .074 500

Note: Number of dyads ( N ) was either 30 or 200. Details of the simulation are provided in the text.

Examination of Table 4 suggests that, for from the pairwise approach (see Gonzalez
the population values examined in this & Griffin, 1998, for a discussion of the ef-
simulation, the pairwise approach per- fects of parameterization on tests of signifi-
formed at least as well as the SEM ap- cance).
proach (as implemented in EQS) with re- We also performed an analogous simula-
spect to Type I error rates, even with a small tion on the overall within-partner correla-
number of dyads ( N = 30).* In all but one tion rxyto examine the Trpe I error rate of
case the SEM approach had a greater effec- the asymptotic test presented here for the
tive Type I error than did the painvise ap- pairwise distinguishable case. Again, a
proach; in 8 out of the 12 combinations of population mean difference of 5, with unit
population values, the pairwise approach variance, was introduced between the two
had an effective Type I error rate that was classes. Table 5, based on N = 30 dyads,
closer (in an absolute difference sense) to presents the average estimate over the 500
0.05 than was the SEM approach. We note runs and the effective Type I error rate. As
that a different parameterization of the shown in Table 5, the asymptotic test for the
equivalent SEM model (i.e., variances of pairwise approach performed well even
the latent variables as free parameters and with samples of 30 dyads. Again, the SEM
indicator paths fixed to 1) leads to effective approach produced effective Type I error
Type I errors that are quite close to rate rates that tended to be liberal.

8. With N = 200 there were no model-inconsistent

estimates in the sense of out-of-bounds rd. How- Instructions to Estimate the SEM Model
ever, with N = 30 and the population intraclasses in Figure 2
both set to S O , a few sample values of rd > 1 oc-
curred, and they were excluded from the summary We present EQS syntax and guidelines for
table. The last column of Table 4 shows the number the package AMOS to estimate the two
of model-consistent samples (i.e., 500 - the models described in the text. We first give
number of discarded samples). Note that when the
population intraclass correlations are high (in this instructions for estimating rxyand r,,', and
simulation, .75) all samples were model-consistent then give instructions for estimating ri and
even when N = 30. rd. All relevant correlations will appear in
Correlational analysis of dyad-level data 467

Table 5. Bias and effective Type I error rates f o r the overall within-partner correlation rXy
Pairwise SEM
Pxxq = Pyy Pxy, Average Estimate Type I Error Type I Error
-.5 - .5 - .006 .046 .068
- .5 0 .006 .060 .088
- .5 .5 - .003 .046 .064
0 - .5 - .006 .040 .064
0 0 .Ooo .050 .064
0 .5 .ooo .040 .046
.5 -.5 -.002 ,058 .082
.5 0 .007 .054 .072
.5 .5 - .001 .042 .058

Note: Number of dyads in the simulation was 30. Details of the simulation are provided in the text.

the section of the program output giving directly into AMOS, assigns names to each
“standardized” estimates. Note that the two of the six paths (making sure that paths that
models test identical constraints (e.g., yield are assumed equal receive identical names),
identical x* values), and the parameters in and gives names to the variances of each of
one model can be transformed into the pa- the four observed variables so that they can
rameters in the other model using the equa- be estimated (again, making sure to assign
tions presented in the text. the same name for variances that are as-
The first EQS syntax provides estimates sumed to be equal).
of rxyand rxy’.As displayed in Figure 1, this The second EQS syntax provides esti-
model does not contain latent variables, so mates of ri and rd with the restriction that the
we need to create latent variables that are two ri’s are equal to each other. This pro-
identical to the observed variables. This al- vides parameter estimates that are identical
lows us to place the necessary equality re- to those from the pairwise model. As men-
strictions required by the model. For exam- tioned in the text, the SEM framework has
ple, by equating a latent variable to the the advantage that it can estimate a model
wife’s trust score and a different latent vari- that permits different ri’s for each individual.
able to the husband’s trust score, one can To perform this more relaxed test, simply
set the variance of those two latent vari- delete the line “(E3,El) = (E4,E2);”. The
ables equal to each other, thus providing a model in Figure 2 can also be implemented
way to test whether the observed variance in AMOS. One simply draws the model, as-
for the wife’s trust is equal to the observed signs names to each path, making sure that
variance of the husband’s trust. paths that are assumed to be equal are as-
In the SEM package AMOS, the model signed identical names, and fixes the vari-
in Figure 1 is a little more straightforward ances of the latent variables to 1.
to implement. One simply draws Figure 1

/TITLE
MODEL I N F I G U R E 1: ESTIMATING r x y and r x y ‘
L i n e s b e g i n n i n g w i t h ! a r e comments
/SPECIFICATIONS
VARIABLES = 4; CASES = 98;
METHODS = ML;
MATRIX = COV;
/LABELS
V 1 = W i f e T r u s t ; V2 = Husband T r u s t ;
V3 = W i f e C o n f l i c t ; V4 = Husband C o n f l i c t ;
468 R. Gonzalez and D. Griffin

/EQUATIONS
V 1 = 1 F1 ; ! l a t e n t v a r i a b l e s w i t h o u t
V2 = 1 F2 ; ! e r r o r a r e d e f i n e d ,
V3 = 1 F3 ; ! i n d i c a t o r s a r e f i x e d t o 1
V4 = 1 F4 ;
/VARIANCES
F 1 = l*; ! l a t e n t variances are f r e e
F2 = l*;
F3 = l*;
F4 = l*;
/ C O V A R I ANCES
F2. F1 = .5*; ! a l l s i x p o s s i b l e c o v a r i a n c e s
F3, F 1 = .5*; ! a r e f r e e
F4, F 1 = .5*;
F3, F2 = .5*;
F4. F2 = .5*;
F4. F3 = .5*;
/CONSTRAINTS
( F l , F 1 ) = ( F 2 . F 2 ) ; ! l a t e n t X v a r i a n c e s equal
(F3, F3) = (F4, F 4 ) ; ! l a t e n t Y v a r i a n c e s equal
(F3, F 1 ) = ( F 4 . F 2 ) ; ! p o o l e d cov(X,Y)
(F4, F1) = (F3. F2); ! p o o l e d cov(X,Yprime)
/MATRIX
1.351
0.440 1.743
-0.857 - 0.365 2.421
-0.443 - 0.438 1.025 2.006
/PRINT
C O V A R I A N C E = YES;
/END

/TITLE
MODEL I N FIGURE 2 WITH INDIVIDUAL COVARIANCES SET EQUAL
L i n e s b e g i n n i n g w i t h ! a r e comments
/SPECIFICATIONS
VARIABLES = 4; CASES = 98;
METHODS = ML;
MATRIX = C O V ;
/LABELS
V 1 = W i f e T r u s t ; V2 = Husband T r u s t ;
V3 = W i f e C o n f l i c t ; V4 = Husband C o n f l i c t ;
F 1 = L a t e n t Dyad T r u s t : F2 = L a t e n t Dyad C o n f l i c t ;
E l = Err W i f e T r u s t ; E2 = E r r Hus T r u s t ;
E3 = Err W i f e Con; E4 = Err Hus Con;
/EQUATIONS
V 1 = 1* F 1 + 1* E l ;
V2 = I*F1 + 1* E2 ;
V3 = 1* F2 + 1* E3 ;
V4 = 1* F2 -k 1* E4 ;
Correlational analysis of dyad-level data 469

/VARIANCES
F1 = 1; ! l a t e n t and e r r o r v a r i a n c e s a l l f i x e d t o 1
F2 = 1;
E l = 1;
E2 = 1:
E3 = 1;
E4 = 1;
/COVARIANCES
F2. F 1 = .5*; ! e s t i m a t e dyad l e v e l cov(X.Y)
E3, E l = .5*; ! e s t i m a t e i n d i v i d u a l l e v e l cov
E4, E2 = .5*; ! e s t i m a t e i n d i v i d u a l l e v e l cov
/CONSTRAINTS
( E 3 , E l ) = (E4,E2): ! s e t two c o v ( X . Y ) ’ s equal
( V 1 , E l ) = (V2,E2); ! c o n s t r a i n e r r o r v a r i a n c e s w i t h i n
(V3,E3) = (V4,E4); ! v a r i a b l e t o be equal
(V1,Fl) = (V2,Fl); !constrain i n d i c a t o r s w i t h i n
(V3.F2) = ( V 4 , F 2 ) : ! v a r i a b l e t o be equal
/MATRIX
1.351
.440 1.743
-.857 - .365 2 . 4 2 1
-.443 - .438 1.025 2.006
/PRINT
C O V A R I A N C E = YES:
/END