Edwards 2001 A

ORGANIZATIONAL
Edwards / MULTIDIMENSIONAL
RESEARCH CONSTRUCTS
METHODS
Multidimensional Constructs in
Organizational Behavior Research:
An Integrative Analytical Framework
JEFFREY R. EDWARDS
University of North Carolina
Multidimensional constructs are widely used to represent several distinct dimen-

sions as a single theoretical concept. The utility of multidimensional constructs
relative to their dimensions has generated considerable debate, and this debate
creates a dilemma for researchers who want the breadth and comprehensiveness
of multidimensional constructs and the precision and clarity of their dimensions.
To address this dilemma, this article presents an integrative analytical framework
that incorporates multidimensional constructs and their dimensions, using struc-
tural equation modeling with latent variables. This framework permits the study of
broad questions regarding multidimensional constructs along with specific ques-
tions concerning the dimensions of these constructs. The framework also provides
tests of issues underlying the multidimensional construct debate, thereby allowing
researchers to address these issues on a study-by-study basis. The framework is il-
lustrated using data from studies of the effects of personality on responses to con-
flict and the effects of work attitudes on employee adaptation.
Multidimensional constructs are pervasive in organizational behavior (OB) research.

A construct is multidimensional when it refers to several distinct but related dimen-
sions treated as a single theoretical concept (Law, Wong, & Mobley, 1998). Examples
of multidimensional constructs include overall job satisfaction conceptualized as sat-
isfaction with multiple job facets (Smith, Kendall, & Hulin, 1969; Warr, Cook, & Wall,
1979), overall job performance viewed as the aggregation of performance on various
job criteria (Murphy & Shiarella, 1997), and broad personality traits that comprise
specific personality dimensions (Digman, 1990; McCrae & Costa, 1992). Multidi-
mensional constructs may be distinguished from unidimensional constructs, which
refer to a single theoretical concept (Hattie, 1985), and from multiple dimensions
regarded as distinct but related concepts rather than a single overall concept.
Author’s Note: I thank Kyle D. Cattani, Lawrence R. James, Albert H. Segars, and five anonymous re-
viewers for their helpful comments during the development of this article. I am particularly indebted to Al-
bert Maydeu-Olivares for his advice and statistical programs on model identification. Correspondence con-
cerning this article should be addressed to Jeffrey R. Edwards, Kenan-Flagler Business School, University
of North Carolina, Chapel Hill, NC 27599-3490; phone: (919) 962-3144; e-mail: [email protected].
Organizational Research Methods, Vol. 4 No. 2, April 2001 144-192
© 2001 Sage Publications, Inc.
144
Edwards / MULTIDIMENSIONAL CONSTRUCTS 145
The utility of multidimensional constructs has generated considerable debate in the

OB literature. Advocates of multidimensional constructs argue that such constructs
provide holistic representations of complex phenomena, allow researchers to match
broad predictors with broad outcomes, and increase explained variance (Hanisch,
Hulin, & Roznowski, 1998; Hulin, 1991; Ones & Viswesvaran, 1996; Roznowski &
Hanisch, 1990). Critics contend that multidimensional constructs are conceptually
ambiguous, explain less variance than explained by their dimensions taken collec-
tively, and confound relationships between their dimensions and other constructs
(Gerbing & Anderson, 1988; Hattie, 1985; Johns, 1998; Paunonen, Rothstein, & Jack-
son, 1999; Schneider, Hough, & Dunnette, 1996). This debate has been ongoing for
decades (e.g., Cattell & Tsujioka, 1964; Humphreys, 1962; Schmidt & Kaplan, 1971)
and shows little sign of abating.
The multidimensional construct debate presents a dilemma for OB researchers who
want the breadth and comprehensiveness of multidimensional constructs and the clar-
ity and precision of the dimensions that constitute the construct. These apparently con-
flicting objectives cannot be achieved if a researcher adopts one side of the debate.
Moreover, criticisms underlying the debate are often characterized as necessary evils,
but many are matters of degree that can be assessed empirically. For example, the
degree to which a multidimensional construct captures variance in its dimensions can
be assessed empirically, and the variance explained by a multidimensional construct
can be statistically compared with that explained by its dimensions. Unfortunately,
methods that allow these comparisons have received little attention in the multidimen-
sional construct debate.
This article presents an integrative analytical framework for assessing the utility of
multidimensional constructs in OB research. The framework is integrative in that it
combines multidimensional constructs and their dimensions within a single analytical
approach. The framework incorporates different types of multidimensional constructs
and allows researchers to directly assess a range of assumptions regarding the relation-
ships between a multidimensional construct and its dimensions, causes, and effects.
The framework is illustrated using data from studies concerning the effects of person-
ality on responses to conflict (Moberg, 1998) and the effects of work attitudes on
employee adaptation (Hanisch & Hulin, 1991).
Types of Multidimensional Constructs

Multidimensional constructs can be distinguished in various ways. Perhaps the
most basic distinction concerns the direction of the relationship between the construct
and its dimensions (Law & Wong, 1999; Ones & Viswesvaran, 1996; Schneider et al.,
1996). If the relationships flow from the construct to its dimensions, the construct may
be termed superordinate because it represents a general concept that is manifested by
specific dimensions. If the relationships flow from the dimensions to the construct, the
construct may be termed aggregate because it combines or aggregates specific dimen-
sions into a general concept. The following discussion elaborates this basic distinc-
tion, discusses different ways in which superordinate and aggregate constructs may be
operationalized, and provides examples from OB research.
Before proceeding, it is important to clarify the nature of the relationships between
a multidimensional construct and its dimensions. Because a multidimensional con-
struct is conceptualized in terms of its dimensions, it does not exist separately from its
146 ORGANIZATIONAL RESEARCH METHODS
dimensions. Therefore, the relationships between a multidimensional construct and its

dimensions are not causal forces linking separate conceptual entities, but instead rep-
resent associations between a general concept and the dimensions that represent or
constitute the construct (Law et al., 1998). If a multidimensional construct were
replaced by a conceptually analogous construct conceived as distinct from its dimen-
sions, then relationships between this construct and the dimensions may be causal. For
example, if overall job satisfaction were defined as a general affective orientation
toward the job rather than a composite of satisfaction with job facets (Ironson, Smith,
Brannick, Gibson, & Paul, 1989), then it would be meaningful to examine causal rela-
tionships between overall job satisfaction and satisfaction with job facets.
Superordinate Construct
As noted previously, a superordinate construct is a general concept that is mani-

fested by its dimensions. The dimensions of a superordinate construct are analogous to
reflective measures, which are observed variables that serve as manifest indicators of
an underlying construct (Bollen & Lennox, 1991; Edwards & Bagozzi, 2000). How-
ever, whereas reflective measures are observed variables, the dimensions of a
superordinate construct are themselves constructs that function as specific manifesta-
tions of a more general construct.
Superordinate constructs are common in research on personality. For example, the
five-factor model of personality (Digman, 1990) comprises five broad personality traits
manifested by 30 specific personality facets (McCrae & Costa, 1992). Some investiga-
tors further cast these five traits as indicators of two broader personality dispositions,
one manifested by agreeableness, conscientiousness, and emotional stability and the
other by extraversion and intellect (Digman, 1997). Other examples of superordinate
constructs include general work values manifested by preferences for specific aspects
of work (Bolton, 1980; Pryor, 1987); leader-member exchange reflected by affect, loy-
alty, contribution, and professional respect (Liden & Maslyn, 1998); work withdrawal
manifested by absenteeism, lateness, leaving early, and escapist drinking (Hanisch
et al., 1998); and psychological climate indicated by job characteristics, leader attrib-
utes, role stress, and work group relationships (L. A. James & James, 1989).
Superordinate constructs are often operationalized by summing scores on their
dimensions. Although this approach is widespread, it disregards measurement error
and fails to capture differences in the relationships between the construct and its
dimensions. These problems are avoided when a superordinate construct is specified
as a first-order factor and dimension scores are treated as observed variables (Hanisch &
Hulin, 1991). However, this approach confounds random measurement error with
dimension specificity (i.e., systematic variance in each dimension not captured by the
superordinate construct) and ignores the relationships between each dimension and its
measures. These limitations are overcome by second-order factor models that treat the
superordinate construct as a second-order factor, its dimensions as first-order factors,
and measures of the dimensions as observed variables (Bagozzi & Edwards, 1998;
Hull, Lehn, & Tedlie, 1991; Hunter & Gerbing, 1982; Rindskopf & Rose, 1988).
Given these advantages, the framework presented in this article operationalizes
superordinate constructs using second-order factor models.
Aggregate Construct
Unlike a superordinate construct, an aggregate construct is a composite of its

dimensions, meaning the dimensions combine to produce the construct (Law et al.,
1998). The dimensions of an aggregate construct are analogous to formative measures,
which form or induce a construct (Bollen & Lennox, 1991; Edwards & Bagozzi,
2000). However, whereas formative measures are observed variables, the dimensions
of an aggregate construct are themselves constructs conceived as specific components
of the general construct they collectively constitute.
Aggregate constructs are widespread in OB research. For example, overall job sat-
isfaction has been conceptualized as a composite of satisfaction with specific job fac-
ets, such as pay, promotions, supervision, coworkers, and the work itself (Locke, 1976;
Smith et al., 1969; Warr et al., 1979). Similarly, job performance has been viewed as
the combination of performance on specific tasks (Murphy & Shiarella, 1997). Other
examples of aggregate constructs include role stress conceived as the combination of
role ambiguity, role conflict, and role overload (Bedeian, Burke, & Moffett, 1988;
Parasuraman, Greenhaus, & Granrose, 1992); organizational commitment treated as
the aggregation of commitment to the work group, supervisor, and top management
(Hunt & Morgan, 1994); and job perceptions as a composite of job challenge, job
autonomy, and job importance (James & Tetrick, 1986).
Aggregate constructs are typically operationalized by summing scores on their
dimensions, such that the dimensions are assigned equal weight. Occasionally, dimen-
sions are assigned empirically derived weights obtained from principal components
analysis or factor analysis, which calculate weights based on correlations among the
dimensions (Kim & Mueller, 1978). In some instances, dimension weights are esti-
mated by specifying the dimensions as formative indicators of the construct in a struc-
tural equation model (Bollen & Lennox, 1991; MacCallum & Browne, 1993). To iden-
tify the weights, the construct must be specified as a direct or indirect cause of at least
two observed variables (MacCallum & Browne, 1993). Hence, the dimension weights
are influenced not only by the correlations among the dimensions, but also by the rela-
tionships between the dimensions and the variables caused by the construct. A residual
term may be added to the model, such that the construct becomes a weighted compos-
ite of its dimensions plus random error and other unspecified variables (Bollen &
Lennox, 1991; Heise, 1972; MacCallum & Browne, 1993). Each of these approaches
treats the dimensions of the aggregate construct as observed variables, thereby disre-
garding error in the dimension measures. This limitation can be overcome by specify-
ing the dimensions as latent variables and their measures as manifest variables, as
demonstrated later in this article.
Other Types of Multidimensional Constructs
Although most multidimensional constructs are either superordinate or aggregate,

other types of multidimensional constructs may be considered. Some of these con-
structs combine features of superordinate and aggregate constructs. For example, a
multidimensional construct may have reflective and formative dimensions, analogous
to multiple indicator/multiple cause (MIMIC) models in structural equation modeling
(Jöreskog & Goldberger, 1975). This type of construct is illustrated by organizational
commitment defined by dimensions that include facets of commitment (e.g., accep-

tance of organizational goals and values) and manifestations of commitment (e.g.,
desire to maintain membership in the organization; Mowday, Steers, & Porter, 1979).
Other multidimensional constructs have relationships with their dimensions that are
more complex than the simple linear relationships linking superordinate and aggregate
constructs to their dimensions. For instance, person-organization fit has been defined
as the absolute or squared difference between person and organization dimensions
(Edwards, 1991; Kristof-Brown, 1996), and the motivating potential of jobs has been
defined as a multiplicative composite of core job dimensions (Hackman & Oldham,
1980). Likewise, personality dimensions have been ranked to define personality pro-
files (Holland, 1985) or dichotomized and cross-classified to define personality
typologies (Myers & McCaulley, 1985). The present article focuses on superordinate
and aggregate constructs because they are prevalent in OB research (Law & Wong,
1999) and provide a foundation for understanding other multidimensional constructs
that relate to their dimensions in more complex ways.
The Multidimensional Construct Debate

As noted previously, the utility of multidimensional constructs has been debated for
decades. Although this debate covers a wide range of issues, these issues can be dis-
tilled into five key points. This section summarizes and integrates these points, with
emphasis on the recent OB literature.
Theoretical Utility
Advocates of multidimensional constructs have argued that such constructs are

more theoretically useful than their dimensions. This argument stipulates that theories
should be general and that general theories require general constructs that combine
specific dimensions (Hanisch et al., 1998). For instance, Roznowski and Hanisch
(1990) disparaged specific behaviors as “theoretically sterile” (p. 361) and argued that
only aggregates of heterogeneous behaviors provide a complete understanding of
behavior in organizations. Likewise, Ones and Viswesvaran (1996) contended that
broad personality constructs (i.e., the Big Five) are more basic than specific personal-
ity dimensions and are therefore more useful for theory development. Humphreys
(1962) advocated a hierarchical model of human abilities in which broad abilities are
superordinate constructs, and specific abilities are subordinate constructs located
lower in the hierarchy “in the relatively unimportant position that their size and gener-
ality warrant” (p. 482).
Critics have questioned the theoretical utility of multidimensional constructs on the
grounds that such constructs are conceptually ambiguous (Cronbach, Gleser, Nanda,
& Rajaratnam, 1972; Hattie, 1985; Hunter & Gerbing, 1982; McIver & Carmines,
1981). This ambiguity occurs because variation in a multidimensional construct may
imply variation in any or all of its dimensions. Consequently, theories that explain the
relationship between a multidimensional construct and other variables are difficult to
develop, because different explanations may apply to different dimensions of the con-
struct (Johns, 1998). As an alternative, critics advocate theoretical models that relate
each dimension of the construct to other variables within a general nomological net-
work (Paunonen et al., 1999; Schneider et al., 1996). Such models accommodate dif-
ferences in relationships involving the dimensions of the construct, which critics con-
sider important for theory development and refinement (Johns, 1998; Paunonen et al.,
1999; Schneider et al., 1996).
The foregoing debate partly reflects ideological differences regarding the value of
theories that are broad versus specific. Both of these properties are desirable (Weick,
1979), yet the debate frames them as antithetical. This dilemma may be ameliorated by
developing theories that incorporate multidimensional constructs along with their
dimensions. Such theories can be used to explain how the construct and its dimensions
relate to one another and to other relevant variables, thereby addressing questions that
are broad and specific. Alternately, theories may treat conceptually related dimensions
as a set, such that they collectively represent a general concept. For example, job per-
formance may be conceived not as a multidimensional construct, but rather as a set of
performance dimensions. Such theories would allow researchers to investigate spe-
cific questions for each dimension individually along with general questions for the
dimensions collectively. These approaches for modeling multidimensional constructs
and their dimensions are illustrated later in this article.
Matching Levels of Abstraction
Multidimensional constructs have been recommended for matching general pre-

dictors with general outcomes. For example, researchers have asserted that many
important outcomes in OB research (e.g., job performance) are factorially complex
and therefore require predictors that are also factorially complex (Hogan & Roberts,
1996; Ones & Viswesvaran, 1996). Similarly, researchers have argued that general
attitudes should be matched with general behavioral outcomes that combine specific
behaviors, as when work withdrawal is represented by lateness, absenteeism, and
other unfavorable job behaviors (Hanisch & Hulin, 1990, 1991; Roznowski &
Hanisch, 1990).
Critics of multidimensional constructs have argued that general predictors and out-
comes should be matched not by combining specific dimensions into a single con-
struct, but instead by treating dimensions collectively as a set. For instance, Schneider
et al. (1996) and Paunonen et al. (1999) asserted that the relationship between person-
ality and overall job performance should be examined by linking multiple personality
dimensions to multiple performance dimensions. This approach echoes Nunnally’s
(1978) advice that “instead of building factorial complexity into a particular test, it is
far better to meet the factorial complexity by combining tests in a battery by multiple
regression, in which case tests would be selected to measure different factors that are
thought to be important” (p. 268).
Advocates and critics of multidimensional constructs agree that predictors and out-
comes should be at the same level of abstraction (Fisher, 1980; Schmidt & Kaplan,
1971). At issue is whether general constructs should be represented by combining
multiple dimensions into a single concept or by treating dimensions as a set. An appar-
ent advantage of combining dimensions is that the association between two general
constructs can be indexed by a single quantity (e.g., a correlation coefficient). How-
ever, a single index of association can also be obtained for sets of dimensions using set
correlation (Cohen, 1982) or multivariate regression (Dwyer, 1983). Moreover, most
methods for combining dimensions are special cases of methods that treat dimensions
as sets, such that the former can be statistically compared with the latter. These points
are elaborated within the framework presented in this article.
Reliability
Another source of debate concerns the internal consistency reliability of measures

of multidimensional constructs created by summing dimension scores. Critics note
that the dimensions of a multidimensional construct are necessarily heterogeneous
because they represent different facets or manifestations of the construct. As dimen-
sion heterogeneity increases, correlations among the dimensions decrease, which in
turn reduces the reliability of summed dimension scores. When sums of dimension
scores exhibit adequate reliabilities in practice, critics point out that these sums
often contain numerous items and therefore can attain high reliabilities in spite of
dimension heterogeneity (Mershon & Gorsuch, 1988; Paunonen et al., 1999). Some
critics have further argued that reliability estimation applies only to measures that are
unidimensional (Gerbing & Anderson, 1988; Hunter & Gerbing, 1982), thereby rais-
ing fundamental questions regarding the meaning and assessment of reliability for
measures of multidimensional constructs.
Advocates acknowledge that the dimensions of multidimensional constructs are
often heterogeneous (Roznowski & Hanisch, 1990). However, rather than lamenting
the effects of dimension heterogeneity on reliability, advocates have argued that reli-
ability estimates based on internal consistency are irrelevant for measures of multidi-
mensional constructs (Hanisch et al., 1998). As an alternative, some researchers (e.g.,
Ones & Viswesvaran, 1996) have applied formulas for estimating the reliability of lin-
ear composites (Nunnally, 1978), which incorporate the reliabilities of dimension
measures and can produce acceptable composite reliabilities even when dimensions
are uncorrelated (Aston, 1998).
Ultimately, reliability is an empirical matter that varies across studies. When a mul-
tidimensional construct is operationalized by summing dimension scores, reliability
may be estimated using formulas appropriate for linear composites (Nunnally, 1978)
or latent variables (Jöreskog, 1971), depending on whether the construct is aggregate
or superordinate, respectively. However, as explained earlier, multidimensional con-
structs and their dimensions are better treated as latent variables in structural equation
models. This approach corrects for measurement error in the construct and its dimen-
sions, which effectively renders the reliability debate moot. However, just as it is
important to assess the strength of the relationship between a construct and its mea-
sures, it is important to assess the strength of the relationship between a multidimen-
sional construct and its dimensions. Procedures for assessing these relationships are
demonstrated later in this article.
Construct Validity
Advocates of multidimensional constructs have criticized the construct validity of

dimension measures, arguing that such measures are dominated by specific variance
that should be considered invalid. For example, Ones and Viswesvaran (1996) con-
tended that measures of personality dimensions are “construct deficient” (p. 622)
because they contain excessive specific dimension variance that should be regarded as
measurement error. Likewise, Humphreys (1970) argued that variance specific to

dimensions represents noise and that only variance common to all dimensions is valid.
To reduce the effects of specific dimension variance, advocates of multidimensional
constructs have recommended summing numerous heterogeneous dimension mea-
sures (Humphreys, 1970; Roznowski & Hanisch, 1990) or using dimensions as indica-
tors of a general factor (Hanisch et al., 1998; Hulin, 1991). Both of these approaches
emphasize variance common to the dimensions and treat specific dimension variance
and random error variance as measurement error (Humphreys, 1970).
Critics of multidimensional constructs have defended the construct validity of
dimension measures. For instance, Schneider et al. (1996) challenged claims by Ones
and Viswesvaran (1996) that dimension specificity should be treated as measurement
error, arguing instead that specificity is valid precisely because it is not measurement
error. Accordingly, high specificity should be interpreted not as invalid construct vari-
ance, but instead as valid dimension variance that is not captured by the multidimen-
sional construct. A similar argument was advanced by Blau (1998), who criticized
superordinate work withdrawal and job withdrawal constructs (Hanisch & Hulin,
1991) on the grounds that these constructs explained little variance in their dimen-
sions, thereby failing to capture dimension specificity.
The evaluation of construct validity begins by identifying the construct of interest
(Nunnally, 1978; Schwab, 1980). For advocates of multidimensional constructs, the
construct of interest is general, and any variance specific to its dimensions is therefore
invalid. For critics, the dimensions themselves are of interest, and the multidimen-
sional construct serves primarily as an organizing category or label for the dimensions.
This difference in perspective underlies the interpretation of dimension specificity as
invalid or valid. An integrative perspective would treat multidimensional constructs
and their dimensions as theoretical constructs, each with its own claim to validity. Evi-
dence for the validity of the multidimensional construct would include the strength of
its relationships with its dimensions and the degree to which it captures relationships
between its dimensions and other theoretically relevant constructs. Evidence for the
construct validity of the dimensions would include the strength of their relationships
with their measures and the degree to which each dimension exhibits relationships
with other constructs not captured by the multidimensional construct. Methods for
assessing these aspects of construct validity are illustrated later in this article.
Criterion-Related Validity
Finally, advocates of multidimensional constructs have argued that such constructs

have higher criterion-related validity than their dimensions. In most cases, these argu-
ments pertain to multidimensional constructs as predictors. For example, Ones and
Viswesvaran (1996) reported that, across a range of criteria, broad personality traits
exhibited higher predictive validity than their dimensions. Less frequently, these argu-
ments are applied to multidimensional constructs as criteria. For instance, in a study of
the effects of attitudes on employee adaptation, Roznowski and Hanisch (1990) found
that attitudes correlated more highly with a sum of adaptive behaviors than with the
individual behaviors constituting the sum.
Critics have countered that, although multidimensional constructs often have
higher criterion-related validity than most of their dimensions, they frequently have
lower criterion-related validity than at least one dimension (Aston, 1998; Paunonen
et al., 1999; Schneider et al., 1996). Critics have further pointed out that a multidimen-
sional construct cannot have higher criterion-related validity than an optimally
weighted linear combination of its dimensions, as when the dimensions are used as
predictors in multiple regression analysis (Goldberg, 1993; Moberg, 1998; Paunonen
et al., 1999; Schneider et al., 1996). At a more fundamental level, some critics have
questioned the value of maximizing criterion-related validity as an end unto itself,
arguing instead that the goal of empirical research is to obtain parameter estimates that
are accurate, regardless of their magnitude (Johns, 1998).
The debate over criterion-related validity has two important limitations. First,
although many studies have compared multidimensional constructs with their dimen-
sions taken individually, such comparisons are irrelevant because critics recommend
using the dimensions collectively. Therefore, comparisons of criterion-related validity
should pit the multidimensional construct against its dimensions as a set. Some
researchers have argued that such comparisons are pointless because the construct
cannot explain more variance than that explained by its dimensions taken jointly
(Paunonen, 1998; Schneider et al., 1996). However, the relevant question is whether
the increase in explained variance is worth the degrees of freedom consumed by
replacing a multidimensional construct with its dimensions, a question that has
received little attention in the criterion-related validity debate. Second, comparisons
between multidimensional constructs and their dimensions taken collectively have
been limited to aggregate constructs as predictors, perhaps due to the simplicity of
comparing the bivariate correlation for an aggregate construct with the multiple corre-
lation for its dimensions. Alternative procedures are required for aggregate constructs
as outcomes and for superordinate constructs as predictors or outcomes. These proce-
dures are demonstrated later in this article.
Summary and Implications of the Debate
In sum, the multidimensional construct debate has raised issues of theoretical util-
ity, levels of abstraction, reliability, construct validity, and criterion-related validity.
Advocates of multidimensional constructs endorse generality, breadth, and simplicity,
whereas critics promote specificity, precision, and accuracy. Given that both sets of
objectives are laudable, researchers would be better served by an integrative approach
than by admonitions to adopt one side of the debate. Moreover, many issues underly-
ing the debate refer to empirical matters that vary across studies, but methods for quan-
tifying these issues have received little attention. The following section presents an
integrative framework that incorporates multidimensional constructs along with their
dimensions, quantifies issues that underlie the multidimensional construct debate, and
seeks to serve the interests of advocates and critics of multidimensional constructs.
An Integrative Analytical Framework

The framework developed here is based on four principles. First, multidimensional
constructs and their dimensions should be included in the same model. Doing so
allows tests of broad questions associated with multidimensional constructs along
with specific questions concerning the dimensions of the construct. Second, models
should incorporate assumptions regarding the direction of the relationships between
the construct and its dimensions, thereby capturing the distinction between
superordinate and aggregate constructs. Third, analyses should examine the strength
and variability of the relationships between the multidimensional construct and its
dimensions. These analyses have important implications regarding the interpretation
of the construct, given that its meaning arises from its relationships with its dimen-
sions. Fourth, analyses should quantify issues underlying the multidimensional con-
struct debate, such that the relative merits of multidimensional constructs and their
dimensions can be assessed on a study-by-study basis. Model specification differs
depending on whether the multidimensional construct is superordinate or aggregate
and whether the construct is a cause or an effect of constructs other than its dimensions.
Combining these distinctions yields the four types of models that are discussed below.
All models use standard notation from the structural equation modeling literature
(e.g., Jöreskog & Sörbom, 1996).
Superordinate Construct as a Cause
As explained earlier, a superordinate construct is best viewed as a second-order fac-

tor with its dimensions as first-order factors. Three second-order factor models are
considered here, representing different degrees of variability in the relationships
between the construct and its dimensions. The most restrictive model treats the dimen-
sions as parallel, meaning they have equal loadings and equal residual variances. This
model embodies two assumptions: (a) each dimension manifests the superordinate
construct to the same degree, such that a unit change in the construct leads to the same
degree of change in each dimension; and (b) the quality of each dimension as an indi-
cator of the superordinate construct is the same. These assumptions reflect the premise
that distinctions among the dimensions can be disregarded, as implied by procedures
that collapse the dimensions of a superordinate construct into a single score. An equa-
tion that captures this model is
ηi = γξ* + ζ. (1)
In this equation, ξ* is the superordinate construct, and the ηi are its dimensions (a su-
perscript asterisk is used to differentiate a multidimensional construct from constructs
treated as its dimensions, causes, and effects). Note that each ηi has the same loading
on ξ* (i.e., γ) and the same residual (ζ), which in turn implies that the residual variances
for the ηi are equal.
A less restrictive model treats the dimensions as tau equivalent, such that the dimen-
sions have equal loadings but different residual variances. This model captures the
assumption that a unit change in the construct leads to the same change in each dimen-
sion, but the construct explains different amounts of variance in each dimension. This
assumption implies that each dimension represents the superordinate construct to the
same degree, although some dimensions may do so with greater precision, as indicated
by lower residual variance. An equation for this model is
ηi = γξ* + ζi. (2)

Note that the relationship of ξ* with each of its dimensions is represented by a single
loading (i.e., γ) but a different residual (i.e., the ζi).
Finally, the least restrictive model considered here treats the dimensions as conge-
neric, meaning that their loadings and residual variances are free to vary. An equation
for this model is
ηi = γiξ* + ζi. (3)
Here, the relationship between ξ* and each of its dimensions is represented by a differ-
ent loading (i.e., the γi) and residual (i.e., the ζi). This equation corresponds to a stan-
dard second-order factor model (Rindskopf & Rose, 1988).
Constructs caused by the superordinate construct are specified as additional ηi, and
incorporating these effects into the preceding models adds equations that follow Equa-
tion 3. As such, distinctions between the ηi as dimensions versus effects of the
superordinate construct are strictly matters of interpretation. However, comparisons
among the parallel, tau equivalent, and congeneric models pertain only to the ηi treated
as dimensions of the construct. The effects of ξ* on the ηi as effects should be unre-
stricted, such that each ηi has a unique coefficient and residual. Moreover, causal paths
and correlated residuals may be included for the ηi as effects but not the ηi as dimen-
sions of ξ*, given that the superordinate construct is considered the only source of
covariation among its dimensions.1
As noted earlier, the framework presented here provides tests of relationships for
multidimensional constructs and their dimensions simultaneously. For a superordinate
construct as a cause, relationships between the construct and its effects are represented
by the γi in Equation 3, whereas relationships between the dimensions and effects of
the construct are captured by the pairwise products of the γi for the dimensions and
effects, each multiplied by the variance of the construct. For example, if the construct
is standardized and has three dimensions, the relationship between the first dimension
of the construct (i.e., η1) and the first effect of the construct (labeled η4, given that η1,
η2, and η3 are dimensions of the construct) is represented by the loading for η1 times
the coefficient for η4, or γ1γ4 (when the construct is standardized, its variance is unity
and can be disregarded). Other relationships between the dimensions and effects of the
construct can be similarly derived. Thus, using a superordinate construct as a cause
treats the relationships between the dimensions and effects of the construct as spuri-
ous, due to the common cause ξ*.
Aggregate Construct as a Cause
As explained earlier, an aggregate construct is best viewed as a function of latent

variables representing its dimensions. Here we consider four models that incorporate
different degrees of variability in the relationships between the construct and its
dimensions. The most restrictive of these models treats the aggregate construct as a
simple sum of its dimensions. The corresponding equation is
η* = Σξi. (4)
Here, the aggregate construct is represented by η*, its dimensions are represented by
the ξi, and the summation is over all dimensions of the construct.
A somewhat less restrictive model uses weights specified by the researcher. For
example, the correlations among the ξi may be entered into a principal components
analysis to obtain weights that maximize the variance of η*, subject to the restriction
that sum of the squared weights equals unity (Kim & Mueller, 1978). Alternately,
weights may be derived rationally to represent the judged importance of the dimen-
sions (Murphy & Shiarella, 1997; Schneider et al., 1996). An equation containing such
weights is
η* = Σwiξi. (5)
In this equation, the dimension weights are represented by the wi, each of which is as-
signed to a particular dimension.
Neither of the preceding models estimates dimension weights within the model
itself. A model that directly estimates dimension weights is captured by the following
equation:
η* = Σγiξi. (6)
In this model, the dimension weights are represented by the γi. These weights are anal-
ogous to those derived from principal components analysis, except that the criterion
for deriving the γi is not to maximize the variance of the aggregate construct, but in-
stead to reproduce the covariances among the dimensions and effects of the construct.
Because the model in Equation 6 includes weights as free parameters, it is less restric-
tive than the models in Equations 4 and 5.
The three preceding models treat the aggregate construct as an exact linear combi-
nation of its dimensions. A less restrictive model introduces a residual for the aggre-
gate construct:
η* = Σγiξi + ζ*. (7)
In Equation 7, ζ* represents aspects of η* not captured by its dimensions. Thus, Equa-

tion 7 allows the relationships between the dimensions and the construct to vary empir-
ically (as in Equation 6) and further stipulates that the variance of η* is no longer solely
a function of the variance of the linear combination Σγiξi.
Relationships between the aggregate construct and its effects are captured by equa-
tions of the following form:
ηj = βjη* + ζj. (8)
In Equation 8, the ηj represent the effects of the aggregate construct η*, and the ζj are
residuals for each ηj.
Like a superordinate construct, an aggregate construct represents the relationships

between its dimensions and effects as pairwise products of the dimension loadings
with the coefficients relating the construct to its effects. These products can be
obtained by substituting the equation for the appropriate aggregate construct into
Equation 8. Specifically, the relationships between the dimensions and effects of the
construct are represented by γiβj for Equations 6 and 7, wiβj for Equation 5, and βj for
Equation 4 due to the implied loading of unity on each dimension. These products
show that using an aggregate construct as a cause treats the relationships between the
dimensions and effects of the construct as indirect, mediated by the construct η*.
Superordinate Construct as an Effect
We now turn to models that treat a multidimensional construct as an effect, starting

with a superordinate construct. As before, the superordinate construct is specified as a
second-order factor with its dimensions as first-order factors. Three second-order fac-
tor models are again considered, corresponding to the parallel, tau equivalent, and con-
generic models discussed previously. To reiterate, the parallel model is the most
restrictive of the three models because it specifies that each dimension represents the
construct to the same degree and with the same precision. For a superordinate con-
struct as an effect, the parallel model is represented by
ηj = βη* + ζ. (9)
In Equation 9, η* is the superordinate construct and the ηj are its dimensions. Note that
each ηj has the same loading on η* (i.e., β) and the same residual (ζ), which in turn im-
plies that the residual variances are equal.
The less restrictive tau equivalent model is captured by the following equation:
ηj = βη* + ζj. (10)
Here, the relationship of η* with each of its dimensions is represented by a single load-
ing (i.e., β) and a different residual (i.e., the ζj).
Finally, the congeneric model is depicted by the following equation:
ηj = βjη* + ζj. (11)
In this equation, the relationship of η* with each of its dimensions is captured by a dif-
ferent loading (i.e., the βj) and residual (i.e., the ζj). This equation corresponds to a
standard second-order factor model but specifies the second-order factor as endoge-
nous (i.e., η*) rather than its usual specification as exogenous (i.e., ξ*).
Adding the causes of the superordinate constructs to the preceding models intro-
duces a set of equations that treat η* as dependent on one or more ξi and a residual:
η* = Σγiξi + ζ*. (12)

Note that Equation 12 is identical to Equation 7 for an aggregate construct as a cause.

However, whereas the ξi are interpreted as dimensions of an aggregate construct η* in
Equation 7, they are viewed as causes of the superordinate construct η* in Equation 12.
Moreover, the constraints applied to Equation 7 to obtain the restricted aggregate
cause models in Equations 4, 5, and 6 should not be imposed on Equation 12, which in-
stead should freely estimate the parameters linking the superordinate construct to its
causes and the residual not explained by these causes.
As with aggregate cause models, superordinate effect models represent the rela-
tionships between the causes and dimensions of the construct as the pairwise products
of the coefficients on the causes with the loadings on the dimensions, or γiβj. Thus,
using a superordinate construct as an effect treats the relationships between the causes
and dimensions of the construct as indirect, mediated by the construct η*.
Aggregate Construct as an Effect
Finally, we turn to models that specify an aggregate construct as an effect. For these
models, the dimensions and causes of the aggregate construct are exogenous variables,
and their variances and covariances should be freely estimated (MacCallum &
Browne, 1993). However, doing so renders the parameters linking the construct to its
dimensions and causes unidentified. Nonetheless, if the parameters linking the con-
struct to its dimensions are fixed to a priori values (e.g., unit weights or principal com-
ponent weights), the parameters linking the construct to its causes become functions of
the variances and covariances of the causes and dimensions of the construct. For illus-
tration, consider a model in which an aggregate construct is a sum of three dimensions
(ξ1, ξ2, and ξ3) and has one cause (ξ4). An equation relating the construct to its cause is
η* = γ4ξ4 + ζ*. (13)
Assuming ξ4 and ζ* are independent, covariance algebra yields the following solution
for γ4:
γ4 = C(η*, ξ4)/φ44, (14)
where C(.) is a covariance operator, and φ44 is the variance of ξ4. By substituting the
sum ξ1 + ξ2 + ξ3 for η* and applying covariance algebra, Equation 14 may be rewritten
as
γ4 = φ14/φ44 + φ24/φ44 + φ34/φ44. (15)
Equation 15 shows that γ4 is a function of the variance of ξ4 and its covariances with ξ1,
ξ2, and ξ3. Hence, γ4 may be constrained to the expression in Equation 15. Next, the
variance of ζ* (i.e., ψ*) may be written by solving Equation 13 for ζ* and taking vari-
ances:
ψ* = V(η* – γ4ξ4), (16)

where V(.) is a variance operator. Substituting Equation 15 for γ4 and ξ1 + ξ2 + ξ3 for η*

and applying covariance algebra yields
ψ * = φ 11 + φ 22 + φ 33 + 2φ 12 + 2φ 13 + 2φ 23 (17)
− (φ 14
2
+ φ 224 + φ 234 + 2φ 14φ 24 + 2φ 14φ 34 + 2φ 24φ 34 ) / φ 44 .
Thus, ψ* is a function of the variances and covariances of ξ1, ξ2, ξ3, and ξ4 and therefore
may be constrained to the expression shown in Equation 17.2
If η* is specified as a weighted sum of the ξi (e.g., w1ξ1 + w2ξ2 + w3ξ3), the foregoing
approach should be adapted to incorporate rules for calculating variances and
covariances of weighted sums of random variables. Doing so yields the following
solutions for γ4 and V(ζ*):
γ4 = w1φ14/φ44 + w2φ24/φ44 + w3φ34/φ44 , (18)
ψ * = w12φ 11 + w 22φ 22 + w 32φ 33 + 2 w1w 2φ 12 + 2 w1w 3φ 13 + 2 w 2 w 3φ 23 (19)

− (w12φ 14
2
+ w 22φ 224 + w 32φ 234 + 2 w1w 2φ 14φ 24 + 2 w1w 3φ 14φ 34 + 2 w 2 w 3φ 24φ 34 ) / φ 44 .
Hence, if η* is a weighted sum, γ4 and ψ* should be constrained to the expressions

shown in Equations 18 and 19.
The procedure described above becomes exceedingly complicated as more causes
are added to the model. Fortunately, the same results may be obtained by respecifying
the dimensions of the construct as endogenous variables. This approach may be illus-
trated by first substituting the sum ξ1 + ξ2 + ξ3 for η* in Equation 13, which yields
ξ 1 + ξ 2 + ξ 3 = γ 4 ξ4 + ζ * . (20)
Because ξ1, ξ2, and ξ3 are now endogenous, they are rewritten as η1, η2, and η3, which
yields
η1 + η2 + η3 = γ4ξ4 + ζ*. (21)
Equation 21 may be expanded into a set of three equations representing a multivariate

structural model, as follows:
η1 = γ1ξ4 + ζ1
η2 = γ2ξ4 + ζ2
η3 = γ3ξ4 + ζ3. (22)

In this model, the γi and the variances and covariances of the ζi are freely estimated.
The correspondence between Equations 21 and 22 can be seen by summing the equa-
tions in Equation 22, which yields
η1 + η2 + η3 = (γ1 + γ2 + γ3)ξ4 + ζ1 + ζ2 + ζ3. (23)
Hence, γ4 and ζ* in Equation 21 equal γ1 + γ2 + γ3 and ζ1 + ζ2 + ζ3 in Equation 22, respec-

tively, and the variance of ζ* equals the variance of the sum ζ1 + ζ2 + ζ3. It can be shown
that these expressions for γ4 and the variance of ζ* are algebraically equivalent to Equa-
tions 15 and 17, respectively. An estimate of γ4 can be obtained by estimating the
model represented by the combination of Equations 13 and 22 (i.e., ξ4 as a cause of η1,
η2, η3, and η*) and constraining γ4 to γ1 + γ2 + γ3. Estimates of the variance of ζ* and its
covariances with the ζi may be obtained by constraining these parameters to values in-
dicated by rules for calculating variances and covariances of linear combinations of
random variables. For the present example, the variance of ζ* equals the sum of the
variances of ζ1, ζ2, and ζ3 plus twice their covariances (i.e., ψ11 + ψ22 + ψ33 + 2ψ12 + 2ψ13
+ 2ψ 23), and the covariances of ζ * with each ζ i equal the variance of that ζ i plus
its covariances with the other two ζi (e.g., the covariance of ζ* with ζ1 equals ψ11 + ψ21
+ ψ31).
The procedure shown above may also be applied when the dimensions constituting
η* are assigned different weights. Again assuming η* has three dimensions, a differen-
tially weighted aggregate construct may be written as
η* = w1ξ1 + w2ξ2 + w3ξ3. (24)
The weights in Equation 24 may be incorporated into Equation 22 by multiplying both

sides of each equation by the appropriate weight, which yields
w1η1 = w1γ1ξ4 + w1ζ1
w2η2 = w2γ2ξ4 + w2ζ2
w3η3 = w3γ3ξ4 + w3ζ3. (25)
Thus, the effect of ξ4 on the aggregate construct η* in Equation 24 equals w1γ1 + w2γ2 +
w3γ3, and the variance of ζ* equals the variance of the weighted sum w1ζ1 + w2ζ2 + w3ζ3.
As before, estimates of these quantities may be obtained by incorporating η* and ζ*
into the model, constraining γ4 to w1γ1 + w2γ2 + w3γ3, and constraining the variance ζ*
and its covariances with the ζi to values indicated by rules for calculating variances and
covariances of weighted linear combinations of random variables. Again, this ap-
proach yields results that are identical to those provided by the more complicated pro-
cedure corresponding to Equations 18 and 19.
Next, we consider how using an aggregate construct as an effect represents the rela-
tionships between the causes and dimensions of the construct. If the dimensions are
modeled as exogenous variables, then their relationships with the causes of the con-
struct are represented as covariances in the φ matrix. If the dimensions are recast as
endogenous variables, then their relationships with the causes are represented as direct
effects, corresponding to the γi in Equation 22. In either case, the relationships between
the causes and dimensions of the construct are direct effects that bypass the construct
itself. Thus, whereas other multidimensional construct models constrain relationships
between the dimensions of the construct and its causes or effects, an aggregate effect
model conceals the relationships between the causes and dimensions of the construct.
These relationships are concealed because they are effectively summed into compos-
ites such as γ4 in Equation 21. These composites provide no information regarding the
individual relationships between the causes and dimensions of the construct, because a
given value of a composite can be produced by an infinite number of combinations of
the γi.
Model Estimation and Evaluation

To apply the preceding framework, various matters of model estimation and evalua-
tion must be addressed. Regarding model estimation, it is necessary to establish that
the model is identified and to properly specify the relationships between the multidi-
mensional construct and its dimensions. Regarding model evaluation, it is important to
assess model fit, the direction and magnitude of relationships between constructs, and
issues underlying the multidimensional construct debate. Issues of theoretical utility
and level of abstraction are conceptual rather than statistical and therefore should be
addressed as part of the substantive interpretation of the model. Reliability is not an
issue of debate when a multidimensional construct and its dimensions are treated as
latent variables that contain no measurement error. However, an analogous issue con-
cerns the strength of the relationships between the construct and its dimensions, as
these relationships indicate how well a superordinate construct is represented by its
dimensions and, likewise, how well an aggregate construct captures variance in its
dimensions. Issues of construct validity underlying the debate focus on the utility of
dimension specificities, which can be assessed by estimating relationships between
the dimensions of the construct and its causes and effects after taking the construct into
account (Hull et al., 1991). Finally, issues of criterion-related validity can be addressed
by comparing the explained variance associated with the construct with that associated
with its dimensions. Procedures for addressing these matters of model estimation and
evaluation are discussed below.
Model Estimation
Identification. Prior to estimation, it is necessary to establish that all parameters in a

model are identified. The following discussion focuses on identification issues partic-
ular to multidimensional construct models. For this discussion, it is assumed that a
scale has been set for each dimension, cause, and effect of the construct, and that the
parameters of the measurement model for these variables (i.e., item loadings and mea-
surement error variances) are identified. All identification rules set forth in this discus-
sion were verified using procedures outlined by Bekker, Merckens, and Wansbeek
(1994).
First, it is necessary to set a scale for the multidimensional construct. This may be
accomplished by fixing a path leading to or from the construct to unity or by fixing the
variance of the construct to unity, thereby standardizing the construct.3 To conduct sta-
tistical tests involving the multidimensional construct, it is useful to obtain standard
errors for paths leading to and from the construct, and these standard errors are not
available for fixed paths. Therefore, it is usually preferable to set the scale of the con-
struct by fixing its variance. If the construct is endogenous (i.e., a superordinate con-
struct as an effect or an aggregate construct as a cause or effect), the variance of the
construct can be fixed directly in some structural equation modeling programs (e.g.,
RAMONA; Browne & Mels, 1992), whereas in LISREL (Jöreskog & Sörbom, 1996)
it is necessary to write an equation for the variance of the construct, set that equation to
unity, solve for one parameter in the equation, and constrain that parameter to the
expression indicated by the equation.
Second, it is necessary to ensure that parameters in the model are identified. For
superordinate cause models, all parameters are identified if no causal paths are
included among the ηi and the residuals for the ηi (i.e., the ζi) are uncorrelated, as is
customary for second-order factor models (Rindskopf & Rose, 1988). These restric-
tions make substantive sense for the dimensions of the construct, because presumably
the construct is the only systematic source of covariance among the dimensions. How-
ever, these restrictions make less sense for the effects of the construct, which may
covary for reasons other than sharing the superordinate construct as a cause (e.g.,
effects may influence one another, omitted variables embodied in residuals may corre-
late or influence more than one effect). Fortunately, the model remains identified if, for
each pair of effects, (a) the residuals are allowed to correlate or (b) a causal path is
included between the effects, provided the relationships among the effects as a set
remain recursive.
Aggregate cause models and superordinate effect models have the same basic struc-
ture and therefore raise similar identification issues. For both types of models, the con-
struct must have paths leading to at least two endogenous variables (Bollen & Davis,
1994; MacCallum & Browne, 1993). This criterion is satisfied if a superordinate effect
has at least two dimensions or an aggregate cause has at least two effects. The model
remains identified if the residual variances are freed for all endogenous variables
(including the superordinate or aggregate construct), the covariances among all resid-
uals are fixed, and the model contains no causal relationships among endogenous vari-
ables other than those emanating from the multidimensional construct. This specifica-
tion makes sense for a superordinate construct as an effect because the construct is
considered the only systematic source of covariation among its dimensions, and a
residual for the construct must be estimated to capture the variance in the construct not
explained by its causes. However, for an aggregate construct as a cause, it may be
desirable to estimate causal paths among the effects of the construct or correlations
among the residuals of the effects. These parameters are identified if the residual for
the aggregate construct is fixed (i.e., the construct is an exact linear combination of its
dimensions, as in Equations 4, 5, and 6) and, for each pair of effects, (a) the residuals
are allowed to correlate or (b) a causal path is included between the effects and the rela-
tionships among the effects as a set remain recursive. If the residual for the aggregate
construct is freed, as in Equation 7, then at least one path or covariance among the
effect residuals must be constrained to achieve identification.
For an aggregate effect model, estimating the variances and covariances of the
dimensions and causes of the construct renders the parameters linking the construct to
its dimensions and causes unidentified. As noted previously, this dilemma can be over-
come by specifying a priori weights linking the construct to its dimensions, respecifying
the dimensions as endogenous variables, and imposing constraints on the paths from
the causes to the construct and on the variance and covariances of the residual of the
construct. Structural models of this type are saturated and are therefore just identified.
If effects of the construct are added to the model, then the parameters linking the con-
struct to its dimensions may be estimated, provided the model conforms to rules for
identification of models with an aggregate construct as a cause.
Specifying relationships between the multidimensional construct and its dimen-

sions. As noted previously, the proposed framework provides tests of the variability of
the relationships between a multidimensional construct and its dimensions. For
superordinate constructs, these tests are performed by imposing constraints analogous
to those used to compare congeneric, tau equivalent, and parallel measurement models
(Jöreskog & Sörbom, 1996). For congeneric models, no constraints are imposed on the
dimension loadings or residual variances. For tau equivalent models, the dimension
loadings are set equal to one another. For parallel models, the dimension loadings are
set equal to one another, and the residual variances are set equal to one another. These
models are nested and can be compared using chi-square difference tests.
Constraints for aggregate constructs represent different approaches to combine
dimensions to form the construct. Summing the dimensions is tantamount to assigning
them equal weight. To incorporate unit weights, one weight should be set to unity, the
remaining weights should be set equal to that weight, and the variance of the construct
should be freed. If the construct is standardized by fixing its variance to unity, then the
dimension weights may be set equal to one another without fixing any of the weights to
a particular value. In either case, the substantive meaning of the construct is the same.
For principal component weights, covariances among the dimensions should be
obtained from a confirmatory factor analysis and submitted to a principal components
analysis from which one component is extracted. The resulting component weights
can be used to impose proportional constraints on the paths from the dimensions to the
construct. For example, if the principal component weights for three dimensions are .5,
.6, and .7, the second path should be constrained to 1.2 times the first, and the third path
should be constrained to 1.4 times the first. If the scale for the aggregate construct is set
by fixing the first path to the obtained principal component weight, then these con-
straints will reproduce the full set of component weights (with this approach, the vari-
ance of the aggregate construct should be freed, and the resulting estimate will equal
the eigenvalue for the principal component). If the aggregate construct is scaled by fix-
ing its variance to unity, then one path should be freed, and the remaining paths should
be constrained proportionally to that path. For models with equal weights or principal
component weights, the residual for the aggregate construct should be fixed to zero, as
implied by Equations 4 and 5. If the aggregate construct is a cause, then all paths from
the dimensions to the construct may be freed (as in Equations 6 and 7), and the result-
ing model may be compared with models with equal or principal component weights
using chi-square difference tests.
Model Evaluation
Model fit. Model fit should be assessed using indices recommended in the structural
equation modeling literature (Gerbing & Anderson, 1993), such as the comparative fit
index (CFI; Bentler, 1990) and the root mean squared error of approximation
(RMSEA; Steiger, 1990). The CFI represents the increase in fit of the target model
over a null model in which all variables are uncorrelated (Bentler, 1990). The CFI has
an expected value of 1.00 when the estimated model is true in the population, and val-
ues of .95 or higher indicate adequate fit (Hu & Bentler, 1999). The RMSEA estimates
the discrepancy per degree of freedom between the original and reproduced
covariance matrices in the population. Values up to .05 indicate close fit, and values up
to .08 represent reasonable errors of approximation in the population (Browne &
Cudeck, 1993). Point estimates of the RMSEA may be supplemented by confidence
intervals to obtain tests of close fit, as indicated by whether the confidence interval
includes .05 (MacCallum, Browne, & Sugawara, 1996).
Assessments of model fit should be supplemented by comparisons with alternative
models (Anderson & Gerbing, 1988; MacCallum, Roznowski, & Necowitz, 1992).
For a superordinate construct, the parallel, tau equivalent, and congeneric models may
be compared with one another. For an aggregate construct, models with equal or prin-
cipal component dimension loadings may be compared with models that freely esti-
mate these loadings. In addition, models that include a multidimensional construct
should be compared with models that treat the construct as a set of related dimensions,
thereby yielding information directly relevant to the multidimensional construct
debate. Sets of dimensions should be modeled using multivariate structural models
that differ according to whether the multidimensional construct is a cause or effect. If
the construct is a cause, its dimensions are treated as exogenous variables, its effects
are treated as endogenous variables, and each dimension is specified as a direct cause
of each effect. If the construct is an effect, its dimensions are treated as endogenous
variables, its causes are treated as exogenous variables, and a direct effect is included
linking each cause to each dimension. For both models, correlations should be
included among the exogenous variables and among residuals for the endogenous
variables, with the exception of endogenous variables connected by causal paths.
Relationships between the multidimensional construct and its dimensions. Proce-
dures for assessing the overall relationship between a multidimensional construct and
its dimensions differ according to whether the construct is aggregate or superordinate.
For an aggregate construct, the relationship can be assessed with the adequacy coeffi-
cient (R a2 ), which is used in canonical correlation analysis to assess the relationship
4
between a set of variables and their associated canonical variate (Thompson, 1984).
2
R a is calculated by summing the squared correlations between the construct and its
dimensions and dividing by the number of dimensions. The relationship between a
superordinate construct and its dimensions can be assessed with the total coefficient of
determination or multivariate R2 (here labeled R m2 ), which represents the proportion of
generalized variance in a set of dependent variables explained by one or more inde-
pendent variables (Bollen, 1989; Cohen, 1982; Jöreskog & Sörbom, 1996). R m2 is cal-
culated by taking the determinant of the covariance matrix of the multidimensional
construct and its dimensions, dividing this quantity by the variance of the multidimen-
sional construct times the determinant of the covariance matrix of the dimensions, and
subtracting the resulting quantity from unity.
Relationships for dimension specificities. Relationships between dimension

specificities and the causes and effects of the construct can be assessed by incorporat-
ing specificities as latent variables in the model. This approach is straightforward for a
superordinate construct but is quite complicated for an aggregate construct for which
dimension specificities are represented by creating variables that equal the portion of
each dimension that is independent of the construct. A simpler approach that yields
equivalent results is to treat relationships for dimension specificities as direct effects
for the dimensions after taking into account the multidimensional construct. When the
construct is a cause, these direct effects represent the incremental variance explained
by each dimension after controlling for the construct. When the construct is an effect,
direct effects to the dimensions indicate whether the causes in the model relate to
aspects of the dimensions that are distinct from the construct. Individual direct effects
can be tested using modification indices (Jöreskog & Sörbom, 1996), which follow a
chi-square distribution with 1 degree of freedom and indicate the expected improve-
ment in model fit if a constrained parameter is freed.5 Alternately, a direct effect can be
tested by adding it to the model and testing whether it differs from zero (or, analo-
gously, whether the chi-square for the model is reduced). Multiple direct effects can be
tested by adding them to the model and conducting chi-square difference tests, pro-
vided the augmented model is identified.
The foregoing tests do not apply to aggregate effect models, because these models
already include direct effects from the causes to the dimensions of the construct.
Instead, dimension specificities may be assessed by comparing the R2 for the aggregate
construct to the R m2 for its dimensions, and a confidence interval for the differ-
ence between R2 and R m2 can be obtained using the bootstrap (Efron & Tibshirani,
1993). Alternately, dimension specificity can be reframed as information provided by
the dimensions beyond that obtained from the aggregate construct. Using an aggregate
construct as an effect implies that a single coefficient adequately represents the effect
of each cause on all dimensions of the construct. If the coefficients relating each cause
to the dimensions are indeed equal, then no information is lost by using the aggregate
construct. However, if the coefficients differ from one another such that the effects of
each cause vary across dimensions, then the aggregate construct is concealing poten-
tially useful information. This issue may be examined by imposing equality con-
straints on the coefficients from each cause to the dimensions and testing the increase
in chi-square.
Comparing criterion-related validity for the construct and its dimensions. Differ-
ences in criterion-related validity for a multidimensional construct and its dimensions
can be assessed usingR m2 . When the multidimensional construct is a cause, theR m2 from
the model with the construct may be compared with the R m2 from a multivariate struc-
tural model using the dimensions as correlated causes. Because the former model is
nested in the latter, the difference in R m2 for these models can be assessed using a
2
chi-square difference test. When the multidimensional construct is an effect, the R for
2
the model with the construct can be compared with the R m for a multivariate structural
model using the dimensions as effects with correlated residuals. If the construct is
superordinate, the former model is nested in the latter, and the difference in R m2 for the
two models model can be assessed using a chi-square difference test. If the construct is
aggregate, the R for the aggregate construct may be compared with the R m2 for its
2
dimensions as a set, as explained earlier. It should be noted that the foregoing tests are
equivalent to omnibus tests for all dimension specificities, given that any increase in
criterion-related validity for the dimensions is attributable to aspects of the dimensions
not shared with the construct.
Empirical Illustrations
The proposed framework is illustrated using data from two studies: one that exam-
ined personality as a multidimensional cause of responses to conflict (Moberg, 1998)
and another that examined employee adaptation as a multidimensional effect of job
dissatisfaction (Hanisch & Hulin, 1991). For both studies, superordinate and aggre-
gate construct models were examined. Analyses were based on covariance matrices
for measures of the dimensions of the construct and its causes or effects, derived from
information reported in published articles. To facilitate interpretation, covariances for
each study were based on measures converted to a common metric, and all
superordinate constructs were scaled by fixing their variances to unity.6 Measurement
error was incorporated by using each measure as a single indicator of a latent variable
with its loading set to unity and the variance of its measurement error set to one minus
the reported reliability of the measure multiplied by the variance of the measure.
Because a single indicator measurement model was used, fit indices refer specifically
to the structural model relating the multidimensional construct to its dimensions and
causes or effects. Models were estimated using LISREL 8.30 (Jöreskog & Sörbom,
1996) and RAMONA (Browne & Mels, 1992).
Multidimensional Construct as a Cause
Moberg (1998) collected data from 249 managers and supervisors who completed
the 240-item NEO-PI-R (Costa & McCrae, 1992) and the 30-item Organizational
Communication and Conflict Instrument (OCCI; Putnam & Wilson, 1982). The
NEO-PI-R contains 48 items for each of the Big Five personality traits, and items for
each trait can be scored on six dimensions, each measured with eight items. The OCCI
measures four responses to interpersonal conflict, using 5 to 12 items for each
response. The present analyses focused on extraversion as a cause of avoidance, seek-
ing resolution, and exerting control in conflict situations (correlations among residuals
of these outcomes were included in all models). Data used for these analyses are
reported in Table 1.
Superordinate cause models. Results for extraversion as a superordinate cause are

reported in Figure 1 and Table 2. The parallel, tau equivalent, and congeneric models
did not fit the data well, as evidenced by CFI values ranging from .584 to .681 and
RMSEA values ranging from .150 to .155 (for each model, the 90% confidence inter-
val for the RMSEA was well above .05, thereby rejecting the hypothesis of close fit;
MacCallum et al., 1996). Chi-square difference tests indicated that the congeneric
model fit better than the tau equivalent model (∆χ2(5) = 28.171, p < .001), which in turn
166
Table 1
Means, Standard Deviations, Reliabilities, and Correlations for Measures of Extraversion and Responses to Conflict (N = 249)
M SD α 1 2 3 4 5 6 7 8 9
Extraversion dimensions
1. Warmth 2.888 0.500 .750 0.250 0.182 0.062 0.077 0.070 0.172 –0.018 0.054 –0.045
2. Gregariousness 2.100 0.688 .810 .530 0.473 0.122 0.098 0.188 0.159 –0.039 –0.004 –0.033
3. Assertiveness 2.225 0.538 .750 .230 .330 0.289 0.124 0.072 0.076 –0.116 0.078 0.147
4. Activity 2.438 0.550 .700 .280 .260 .420 0.303 0.070 0.133 –0.079 0.087 –0.004
5. Excitement seeking 2.125 0.638 .670 .220 .430 .210 .200 0.406 0.100 –0.014 –0.004 0.039
6. Positive emotions 2.600 0.563 .750 .610 .410 .250 .430 .280 0.316 –0.044 0.068 –0.015
Responses to conflict
7. Avoidance 4.717 0.717 .870 –.050 –.080 –.300 –.200 –.030 –.110 0.514 –0.113 –0.079
8. Resolution 2.800 0.633 .810 .170 –.010 .230 .250 –.010 .190 –.250 0.401 0.048
9. Control 4.300 0.686 .740 –.130 –.070 .400 –.010 .090 –.040 –.160 .110 0.470
Source. Adapted from Moberg (1998).
Note. Table entries labeled α are Cronbach’s alpha. Table entries below the diagonal are correlations, and those on and above the diagonal (in italics) are vari-
ances and covariances used for analysis. Correlations greater than .124 in absolute magnitude are significantly different from zero (p < .05).
Figure 1: Superordinate Cause Models

**p < .01.
2
fit better than the parallel model (∆χ (5) = 29.198, p < .001). Thus, the congeneric
model provided the best fit of the three models, although none of the models fit well in
an absolute sense.
For the parallel, tau equivalent, and congeneric models, extraversion exhibited
strong multivariate relationships with its dimensions, withR m2 values of .840, .849, and
.891, respectively. For the parallel model, the R2 linking extraversion to each dimen-
sion was .462 (by construction, this value was the same for all dimensions). For the tau
equivalent model, dimension R2 values ranged from .347 for gregariousness to .608 for
warmth, whereas for the congeneric model, dimension R2 values ranged from .236 for
assertiveness to .757 for positive emotions.7 Thus, the congeneric model indicated that
the tau equivalent and parallel models concealed considerable variability in the rela-
tionships between extraversion and its dimensions. Omnibus significance tests of this
168
Table 2
Superordinate Cause Models for Extraversion (N = 249)
Extraversion Dimensions
Responses Excitement Positive
2 2
to Conflict Warmth Gregariousness Assertiveness Activity Seeking Emotions Rm χ df CFI RMSEA
Parallel model
Avoidance –0.053** –0.053** –0.053** –0.053** –0.053** –0.053** .046
Resolution 0.049** 0.049** 0.049** 0.049** 0.049** 0.049** .054
Control 0.015 0.015 0.015 0.015 0.015 0.015 .004 236.144** 34 .584 .150
Tau equivalent model
Avoidance –0.051** –0.051** –0.051** –0.051** –0.051** –0.051** .044
Resolution 0.054** 0.054** 0.054** 0.054** 0.054** 0.054** .066
Control 0.009 0.009 0.009 0.009 0.009 0.009 .002 206.946** 29 .633 .153
Congeneric model
Avoidance –0.044* –0.052* –0.027* –0.033* –0.030* –0.050* .028
Resolution 0.052** 0.062** 0.032** 0.039** 0.036** 0.060** .055
Control –0.008 –0.009 –0.005 –0.006 –0.005 –0.009 .001 178.775** 24 .681 .155
Note. CFI = comparative fit index, RMSEA = root mean squared error of approximation. For the extraversion dimensions, table entries are unstandardized spuri-
ous relationships between the extraversion dimensions and responses to conflict, calculated as the product of each dimension loading on extraversion with the
path from extraversion to each response to conflict.
*p < .05. **p < .01.
variability are provided by the chi-square difference tests reported earlier, which indi-
cate that the loadings and residuals for the extraversion dimensions differed signifi-
cantly from one another.
All three models indicated that extraversion was negatively related to avoidance,
positively related to resolution, and unrelated to control. R m2 values linking
extraversion to its outcomes were .092, .101, and .081, respectively, for the parallel,
tau equivalent, and congeneric models. The three models produced R2 values for indi-
vidual outcomes ranging from .032 to .055 for avoidance, .061 to .078 for resolution,
and .001 to .005 for control. Relationships between individual dimensions and out-
comes (see Table 2) were necessarily constant across dimensions for the parallel and
tau equivalent models, whereas the congeneric model indicated that the relationships
with outcomes were strongest for gregariousness, followed by positive emotions and
warmth.
Relationships for dimension specificities were examined using modification indi-
ces for parameters directly linking the extraversion dimensions to the outcomes. To
control for Type I error, the nominal p value of .05 was divided by the number of modi-
fication indices examined (i.e., 18), yielding a critical p value of .00278 and corre-
sponding chi-square of 8.498. For all three models, this criterion indicated a negative
effect of the gregariousness specificity on resolution, a negative effect of the warmth
specificity on control, and a positive effect of the assertiveness specificity on control.
Aggregate cause models. Results for extraversion as an aggregate cause are
reported in Figure 2 and Table 3. Models with equal loadings and principal component
loadings did not fit the data well, with CFI values of .812 and .805 and RMSEA values
of .145 and .148, respectively (for both models, 90% confidence intervals for the
RMSEA excluded .05). In contrast, the model that freely estimated the dimension
loadings fit the data reasonably well, producing a CFI of .945 and a RMSEA of .099
(the lower bound of the 90% confidence interval was .064, falling between the .05 cri-
terion of close fit and the .08 value indicating reasonable errors of approximation;
Browne & Cudeck, 1993). Chi-square difference tests confirmed that this model fit
better than the models with equal loadings (∆χ2(5) = 69.940, p < .001) or principal
2
component loadings (∆χ (5) = 73.033, p < .001). The same fit was produced by a
model that freed the residual on extraversion, as would be expected given that free-
ing this residual required fixing one of the residual covariances for the responses to
conflict. However, the estimated variance of the extraversion residual was nega-
tive, indicating an inadmissible solution. Therefore, the model was reestimated using
RAMONA, which ensures that all estimated variances are nonnegative. The resulting
estimate of the extraversion residual was zero, which essentially rendered this
model equivalent to the model in which the extraversion residual was fixed to zero.
Therefore, the model that freed the extraversion residual was excluded from further
consideration.
For the models with equal loadings and principal component loadings, relation-
ships between extraversion and its dimensions were moderate, as evidenced by R a2 val-
ues of .552 and .548, respectively. Squared correlations between extraversion and its
dimensions (which are analogous to squared structure coefficients in canonical corre-
lation analysis; Thompson, 1984) ranged from .429 for assertiveness to .677 for gre-
gariousness for the equal loadings model and from .368 for assertiveness to .759 for
gregariousness for the principal components loadings model. Relationships between
Figure 2: Aggregate Cause Models

**p < .01.
extraversion and its dimensions were lower and more varied for the model that esti-
mated the dimension loadings, as indicated by an R a2 value of .116, a squared correla-
tion of .626 for assertiveness, and squared correlations smaller than .028 for the other
five dimensions. Chi-square difference tests comparing this model with the models
with equal or principal component loadings indicated that, although the latter models
produced higher R a2 values, they concealed substantial variability in the dimension
loadings. On the other hand, the model that estimated the dimension loadings trans-
formed extraversion into a construct that was dominated by assertiveness and captured
little of the remaining five dimensions.
The models with equal loadings and principal component loadings indicated that
extraversion was negatively related to avoidance, positively related to resolution, and
Table 3
Aggregate Cause Models for Extraversion (N = 249)
to Conflict Warmth Gregariousness Assertiveness Activity Seeking Emotions 2
Rm χ2 df CFI RMSEA
Equal loadings model

Avoidance –0.063** –0.063** –0.063** –0.063** –0.063** –0.063** .044
Resolution 0.056** 0.056** 0.056** 0.056** 0.056** 0.056** .048
Control 0.016 0.016 0.016 0.016 0.016 0.016 .004 106.414** 15 .812 .145
Principal component loadings model
Avoidance –0.051** –0.080** –0.041** –0.044** –0.052** –0.058** .035
Resolution 0.044** 0.069** 0.036** 0.038** 0.045** 0.050** .037
Control 0.007 0.011 0.006 0.006 0.007 0.008 .001 109.507** 15 .805 .148
Estimated loadings model
Avoidance 0.245 0.172 –0.640** 0.302* –0.065 –0.258 .132
Resolution –0.152 –0.106 0.395** –0.186 0.040 0.159 .069
Control –0.442 –0.311 1.153** –0.544* 0.117 0.465 .550 36.474** 10 .945 .099
Estimated loadings model with residual
Avoidance 0.245 0.172 –0.640** 0.302* –0.065 –0.258 .132
Resolution –0.152 –0.106 0.395** –0.186 0.040 0.159 .069
Control –0.442 –0.311 1.153** –0.544* 0.117 0.465 .550 36.474** 10 .945 .099
Note. CFI = comparative fit index, RMSEA = root mean squared error of approximation. For the extraversion dimensions, table entries are unstandardized indirect
effects of the extraversion dimensions on the responses to conflict, calculated as the product of each dimension loading on extraversion with the path from
extraversion to each response to conflict.
*p < .05. **p < .01.
171
unrelated to control. R m2 values for the two models were .071 and .056, respectively,
and the models produced R2 values for individual outcomes of .044 and .035 for avoid-
ance, .048 and .037 for resolution, and .004 and .001 for control. In contrast, the model
that estimated the dimension loadings produced an R m2 of .609 and R2 values of .132,
.069, and .550 for avoidance, resolution, and control, respectively. Relationships
between dimensions and outcomes (see Table 3) were similar for the models with
equal loadings and principal component loadings, indicating that all dimensions were
negatively related to avoidance and positively related to resolution. A markedly dif-
ferent pattern emerged for the models that estimated dimension loadings, indicating
that assertiveness was negatively related to avoidance and positively related to reso-
lution and control, and that activity was positively related to avoidance and nega-
tively related to control.
Relationships for dimension specificities were examined using modification indi-
ces for parameters directly linking the extraversion dimensions to the outcomes, again
using a critical chi-square value of 8.498. For the models with equal loadings and prin-
cipal component loadings, this criterion indicated a positive effect of the assertiveness
specificity on control and negative effects of the warmth specificity on control and
the gregariousness specificity on resolution. For the model with estimated loadings,
modification indices pointed to positive effects of the activity and positive emotions
specificities on resolution, a positive effect of the assertiveness specificity on con-
trol, and negative effects of the warmth, activity, and positive emotions specificities on
control.8
Multivariate structural model. Results for the multivariate structural model are
reported in Table 4. Because single indicators were used for all latent variables,
the model was saturated and therefore fit the data perfectly. The fit of this model rela-
tive to the superordinate and aggregate cause models can be tested using the
chi-square statistics for those models, which are compared with a value of zero (i.e.,
the chi-square for the multivariate structural model). 9 These tests indicate that
the multivariate structural model fit the data better than any of the superordinate
cause and aggregate cause models.
The multivariate structural model yielded an R m2 of .731 for the relationships
between the extraversion dimensions and outcomes. This value is substantially higher
than values for all models except the aggregate cause model with estimated dimension
loadings, which produced anR m2 of .609. R2 values linking the extraversion dimensions
to avoidance, resolution, and control for the multivariate structural model were .158,
.224, and .622, respectively. These values were notably higher than corresponding val-
ues from the superordinate and aggregate cause models.
The multivariate structural model indicated that three of the six extraversion dimen-
sions were related to the responses to conflict. In particular, gregariousness was nega-
tively related to resolution, activity was negatively related to control, and assertiveness
was negatively related to avoidance and positively related to resolution and control.
These results indicate that the relationships for warmth, excitement seeking, and posi-
tive emotions found for the superordinate and aggregate cause models were artifacts of
the constraints imposed by these models on the loadings of the extraversion dimen-
sions. These differences between the multivariate structural model and the
superordinate and aggregate cause models were foreshadowed by the analyses of
dimension specificities reported earlier, which indicated that the superordinate and
Table 4
Multivariate Structural Model for Extraversion (N = 249)
to Conflict Warmth Gregariousness Assertiveness Activity Seeking Emotions 2
Rm χ2 df CFI RMSEA
Avoidance 0.302 –0.025 –0.539** –0.045 0.133 –0.246 .158

Resolution 0.566 –0.406* 0.300* 0.284 0.008 –0.110 .224
Control –0.577 –0.314 1.207** –0.764** 0.192 0.527 .622 0.000 0 1.000 .000
Note. CFI = comparative fit index, RMSEA = root mean squared error of approximation. For the extraversion dimensions, table entries are unstandardized coeffi-
cients linking each extraversion dimension to each response to conflict.
*p < .05. **p < .01.
173
aggregate cause models overestimated some relationships and underestimated other

relationships between the extraversion dimensions and responses to conflict.
Summary. The preceding results provide little support for extraversion as a super-
ordinate construct. Although extraversion exhibited moderate to strong relationships
with its dimensions, it distorted the relationships between the extraversion dimensions
and outcomes, exaggerating relationships with avoidance and resolution and conceal-
ing relationships with control. In addition, the criterion-related validity of the extra-
version construct was much lower than its dimensions as a set. Moreover, relationships
revealed by treating the extraversion dimensions as a set were theoretically meaning-
ful, as exemplified by the positive relationship between assertiveness and control.
Given that relationships with outcomes varied across the extraversion dimensions, it
appears that the level of abstraction embodied by the extraversion construct was too
broad. Thus, on the grounds of theoretical utility, level of abstraction, dimension speci-
ficity, and criterion-related validity, extraversion as a cause of responses to conflict is
better viewed as a set of related dimensions than as a superordinate construct.
Results also do not support extraversion as an aggregate construct with dimension
loadings that are equal or proportional to principal component weights. For these mod-
els, relationships between extraversion and its dimensions were reasonably strong but
criterion-related validities were low, particularly for control as an outcome. These
models also concealed considerable variability in the effects of the extraversion
dimensions, suggesting that the extraversion construct was too broad relative to its out-
comes. Models that freely estimated dimension loadings produced higher criterion-
related validities and better represented the effects of the extraversion dimensions on
the outcomes. However, these models effectively reduced the extraversion construct to
assertiveness, indicating that the construct was too broad and had little theoretical util-
ity beyond the assertiveness dimension taken separately. Thus, the aggregate cause
models were inferior to the multivariate structural model, although the particular
shortcomings of the aggregate cause models differed depending on how the relation-
ships between the aggregate construct and its dimensions were specified.
Multidimensional Construct as an Effect
Hanisch and Hulin (1991) obtained data from 348 university staff members who
completed measures of work, pay, and coworker satisfaction from the Job Descriptive
Index (Smith et al., 1969), a measure of health satisfaction from the Retirement
Descriptive Index (Smith et al., 1969), and measures of five dimensions of employee
adaptation, including unfavorable job behavior, lateness, absenteeism, turnover intent,
and desire to retire (Roznowski & Hanisch, 1990). Satisfaction measures contained 9
to 19 items, and adaptation measures contained three to seven items. The present anal-
yses examined health, work, coworker, and pay satisfaction as causes of adaptation.
Descriptive statistics and correlations for the measures analyzed are reported in
Table 5.
Superordinate effect models. Results for adaptation as a superordinate effect are
reported in Figure 3 and Table 6. The parallel, tau equivalent, and congeneric models
did not fit the data well, with CFI values from .342 to .793 and RMSEA values from
.106 to .153 (for all three models, 90% confidence intervals for the RMSEA excluded
Table 5
Means, Standard Deviations, Reliabilities, and Correlations for Measures of Satisfaction and Adaptation (N = 348)
M SD α 1 2 3 4 5 6 7 8 9
Satisfaction
1. Health satisfaction 5.620 1.349 .800 1.820 0.348 0.283 0.149 –0.083 –0.027 –0.122 –0.157 –0.180
2. Work satisfaction 5.951 1.033 .850 .250 1.068 0.612 0.408 0.048 –0.072 –0.075 –0.326 –0.322
3. Coworker satisfaction 5.812 1.234 .900 .170 .480 1.522 0.663 –0.038 –0.012 –0.037 –0.403 –0.275
4. Pay satisfaction 4.298 1.580 .820 .070 .250 .340 2.496 0.061 –0.284 –0.076 –0.313 0.000
Adaptation dimensions
5. Unfavorable job behaviors 2.760 0.770 .620 –.080 .060 –.040 .050 0.593 0.146 0.121 0.099 0.009
6. Lateness 2.075 0.998 .510 –.020 –.070 –.010 –.180 .190 0.995 0.266 0.163 0.044
7. Absenteeism 1.758 0.605 .530 –.150 –.120 –.050 –.080 .260 .440 0.366 0.162 0.088
8. Turnover intentions 2.143 1.167 .540 –.100 –.270 –.280 –.170 .110 .140 .230 1.361 0.325
9. Desire to retire 4.423 1.114 .820 –.120 –.280 –.200 .000 .010 .040 .130 .250 1.242
Source. Adapted from Hanisch and Hulin (1991).
Note. Table entries labeled α are Cronbach’s alpha. Table entries below the diagonal are correlations, and those on or above the diagonal (in italics) are variances
and covariances used for analysis. Correlations greater than .105 in absolute magnitude are significantly different from zero (p < .05).
175
Figure 3. Superordinate Effect Models

**p < .01.
.05). In addition, the congeneric model produced a small negative error variance for
absenteeism, although this estimate did not differ significantly from zero. Therefore,
this model was reestimated using RAMONA, which produced an estimate of zero for
the absenteeism error variance and little change in the other parameter estimates.
Chi-square difference tests indicated that the congeneric model fit better than the tau
equivalent model (∆χ2(4) = 20.569, p < .001), which in turn fit substantially better than
2
the parallel model (∆χ (4) = 152.262, p < .001). Thus, although the congeneric model
produced the best fit of the three models, none of the models fit the data well.
For the parallel, tau equivalent, and congeneric models, R m2 values relating adapta-
tion to its dimensions were .615, .852, and 1.000, respectively. The R m2 value of 1.000
for the congeneric model was attributable to the relationship between adaptation and
absenteeism, for which the residual variance of zero implied a univariate R2 of 1.000.
The congeneric model also produced an R2 of .624 for lateness and R2 values of less
than .230 for the remaining adaptation dimensions. For the tau equivalent model,
dimension R2 values ranged from .139 for desire to retire to .813 for absenteeism,
Table 6
Superordinate Effect Models for Adaptation (N = 348)
Satisfaction
2 2
Adaptation Dimensions Health Work Coworker Pay R χ df CFI RMSEA
Parallel model
Unfavorable job behaviors –0.046 –0.102* –0.041 –0.009 .048
Lateness –0.046 –0.102* –0.041 –0.009 .048
Absenteeism –0.046 –0.102* –0.041 –0.009 .048
Turnover intentions –0.046 –0.102* –0.041 –0.009 .048
Desire to retire –0.046 –0.102* –0.041 –0.009 .048 270.038** 29 .342 .153
Tau equivalent model
Unfavorable job behaviors –0.056* –0.063 –0.013 –0.023 .035
Lateness –0.056* –0.063 –0.013 –0.023 .034
Absenteeism –0.056* –0.063 –0.013 –0.023 .083
Turnover intentions –0.056* –0.063 –0.013 –0.023 .022
Desire to retire –0.056* –0.063 –0.013 –0.023 .014 117.776** 25 .747 .107
Congeneric model
Unfavorable job behaviors –0.031 –0.039 –0.006 –0.024 .014
Lateness –0.070 –0.088 –0.013 –0.055 .053
Absenteeism –0.055* –0.070 –0.010 –0.043 .086
Turnover intentions –0.051 –0.064 –0.009 –0.040 .019
Desire to retire –0.028 –0.036 –0.005 –0.022 .004 97.207** 21 .793 .102
Note. CFI = comparative fit index, RMSEA = root mean squared error of approximation. For the satisfaction facets, table entries are unstandardized indirect effects
of the satisfaction facets on the dimensions of adaptation, calculated as the product of the path from each satisfaction facet to adaptation with the loading of each
dimension on adaptation.
*p < .05. **p < .01.
177
whereas for the parallel model the R2 for each adaptation dimension was .242. Hence,
as parameters relating adaptation to its dimensions were relaxed, these relationships
because increasingly variable to the point that adaptation became isomorphic with
absenteeism.
The parallel model indicated that adaptation was negatively related to work satis-
faction, whereas the tau equivalent and congeneric models indicated that adaptation
was negatively related to health satisfaction. R2 values for adaptation produced by the
three models were .200, .102, and .081, respectively. Relationships between satisfac-
tion and the adaptation dimensions (see Table 6) were necessarily constant for the par-
allel and tau equivalent models, with the former model indicating that all dimensions
were negatively related to work satisfaction and the latter model indicating that all
dimensions were negatively related to health satisfaction. In contrast, the congeneric
model yielded a single negative relationship between health satisfaction and absentee-
ism, suggesting that the negative relationships for the other four dimensions in the tau
equivalent model were artifacts of the equality constraint imposed across the five
dimensions.
For the three superordinate effect models, relationships for dimension specificities
were examined using modification indices for parameters directly linking satisfaction
to the adaptation dimensions. The nominal p value of .05 was divided by the number of
modification indices examined (i.e., 20), yielding a critical p value of .0025 and corre-
sponding chi-square of 9.140. For the parallel model, this criterion indicated a positive
effect of work satisfaction on the unfavorable job behavior specificity, negative effects
of work satisfaction on the turnover intent and desire to retire specificities, a negative
effect of coworker satisfaction on the turnover intent and desire to retire specificities,
and a negative effect of pay satisfaction on the turnover intent specificity. Similar
results were obtained for the tau equivalent model, although modification indices did
not reach significance for the effect of coworker satisfaction on the desire to retire
specificity or the effect of pay satisfaction on the turnover intent specificity. For the con-
generic model, work satisfaction and pay satisfaction were negatively related to the
specificities for turnover intent and desire to retire. Thus, although the results of
the specificity analyses differed somewhat for the three models, each model indicated
that dissatisfaction with work and pay may relate aspects of turnover intent and desire
to retire not captured by the adaptation construct.
Aggregate effect models. Results from analyses using adaptation as an aggregate
effect are shown in Figure 4 and Table 7. Because aggregate effect models do not
impose constraints on the relationships between the adaptation dimensions and out-
comes, they yield equivalent fit to the data and therefore cannot be compared using fit
statistics or chi-square difference tests. Nonetheless, these models can be evaluated
using criteria relevant to the multidimensional construct debate, as discussed below.
Overall, the aggregate effect models yielded modest relationships between adapta-
tion and its dimensions, with R a2 values of .449 and .365 for the equal loadings and
principal component loadings models, respectively. The equal loadings model yielded
squared correlations between adaptation and its dimensions ranging from .264 for
unfavorable job behavior to .627 for absenteeism. Squared dimension correlations for
the principal component loadings model ranged from .064 for unfavorable job behav-
ior to .704 for desire to retire. The increased variability in dimension correlations for
Figure 4. Aggregate Effect Models

**p < .01.
the principal component loadings model paralleled the differential weights assigned to
the adaptation dimensions, as shown in Figure 4.
The aggregate effects models indicated that adaptation was negatively related to
work satisfaction and unrelated to health, coworker, and pay satisfaction. R2 values
were .134 and .193, respectively, for the equal loadings model and principal compo-
nent loadings model. Relationships between satisfaction and the adaptation dimen-
sions (see Table 7) differed for the two models due to scaling differences produced by
the dimension weights used to derive the adaptation construct (e.g., for the equal load-
ings model, the five coefficients linking health satisfaction to the adaptation dimen-
sions in Table 7 sum to –.087, which equals the coefficient linking health satisfaction
to adaptation in Figure 4).
Relationships for dimension specificities were examined by comparing the R2 for
adaptation with the R m2 for its dimensions. Whereas the equal loadings and principal
component loading models produced R2 values of .134 and .193, the R m2 for the adapta-
tion dimensions was .535. Information provided by the dimensions beyond that
yielded by the adaptation construct was further assessed by imposing equality con-
straints on the coefficients linking each satisfaction facet to the five adaptation dimen-
180
Table 7
Aggregate Effect Models for Adaptation (N = 348)
Satisfaction
2 2
Equal loadings model

Unfavorable job behaviors –0.031 0.051* –0.036 0.016 .048
Lateness –0.002 –0.032 0.043 –0.066** .096
Absenteeism –0.030* –0.026 0.012 –0.012 .075
Turnover intentions –0.008 –0.090* –0.071* –0.023 .214
Desire to retire –0.016 –0.133** –0.041 0.041* .134 0.000 0 1.000 .000
Principal component loadings model
Lateness –0.001 –0.017 0.022 –0.034** .096
Absenteeism –0.013* –0.012 0.006 –0.005 .075
Desire to retire –0.025 –0.205** –0.063 0.063* .134 0.000 0 1.000 .000
Note. CFI = comparative fit index, RMSEA = root mean squared error of approximation. For the satisfaction facets, table entries are weighted direct effects of
the satisfaction facets on the dimensions of adaptation (these effects differ from one another and from the multivariate structural model only due to scaling
differences).
*p < .05. **p < .01.
sions. As explained earlier, if the adaptation construct adequately represents the

effects of satisfaction on the adaptation dimensions, then a single coefficient should
adequately summarize the effects of each satisfaction facet on all five adaptation
dimensions. Imposing these constraints significantly worsened the fit of the model
(∆χ2(16) = 73.262, p < .001). Modification indices were examined to identify which
satisfaction facets were most severely affected by the equality constraints, again using
a critical chi-square of 9.140. Results indicated that the equality constraints concealed
the greatest degree of variability in the effects of work satisfaction and coworker satis-
faction, followed by pay satisfaction. Thus, using adaptation as an aggregate construct
may have concealed meaningful variability in the effects of these three facets of satis-
faction on the adaptation dimensions.
Multivariate structural model. Results for the multivariate structural model are
reported in Table 8. Given that single indicators were used for all latent variables, the
model was saturated and fit the data perfectly. Consequently, the fit of the model is nec-
essarily the same as that of the aggregate effect models. However, the fit of the
multivariate structural model can be compared with that of the superordinate effect
models using the chi-square statistics for those models (see Table 6). These tests indi-
cate that the multivariate structural model fit the data better than the superordinate
effect models.
The multivariate structural model produced anR m2 of .535 for the effects of the satis-
faction facets on the five adaptation dimensions as a set. This value is markedly higher
than R2 values for adaptation produced by the superordinate and aggregate effect mod-
els. R2 values for the individual adaptation dimensions ranged from .048 for unfavor-
able job behavior to .214 for turnover intent. It is worth noting that the R2 for turnover
intent as a single dimension was higher than the R2 for the adaptation construct pro-
duced by any of the superordinate or aggregate cause models.
The multivariate structural model indicated that all five adaptation dimensions
were related to various facets of satisfaction. In particular, unfavorable job behavior
was positively related to work satisfaction, lateness was negatively related to pay satis-
faction, absenteeism was negatively related to health satisfaction, turnover intent was
negatively related to work satisfaction and coworker satisfaction, and desire to retire
was negatively related to work satisfaction and positively related to pay satisfaction.
This pattern of relationships is consistent with the analyses of dimension specificities
for the superordinate and aggregate effect models, which indicated that the relation-
ships between the satisfaction facets and adaptation dimensions were more variable
than implied by the adaptation construct itself.
Summary. The preceding analyses indicate that adaptation is better viewed as a set
of related dimensions than as a superordinate effect. Adaptation was strongly related
to its dimensions, but this was primarily due to a relationship with absenteeism that
reached unity for the congeneric model. Criterion-related validities for adaptation pro-
duced by the superordinate effect models were much smaller than the criterion-related
validity yielded by the multivariate structural model. In addition, the superordinate
effect models oversimplified the effects of the satisfaction facets on the adaptation
dimensions, as suggested by the dimension specificities and confirmed by the
multivariate structural model. The variable effects on the adaptation dimensions, com-
bined with the higher criterion-related validity of the multivariate structural model,
182
Table 8
Multivariate Structural Model for Adaptation (N = 348)
Satisfaction
2 2

Lateness –0.004 –0.076 0.102 –0.156** .096
Absenteeism –0.072* –0.062 0.030 –0.029 .075
Desire to retire –0.039 –0.318** –0.098 0.098* .134 0.000 0 1.000 .000
Note. CFI = comparative fit index, RMSEA = root mean squared error of approximation. For the satisfaction facets, table entries are unstandardized coefficients
linking each satisfaction facet to each dimension of adaptation.
*p < .05. **p < .01.
indicate that the level of abstraction represented by the adaptation construct was too
broad. Finally, the theoretical utility of adaptation as a superordinate effect is dubious,
given that relaxing the constraints on the dimension loadings effectively reduced adap-
tation to absenteeism, and differences among the effects on the adaptation dimensions
were conceptually meaningful (e.g., the negative effect of work satisfaction on desire
to retire implies that people want to cease work they dislike, whereas the positive effect
of pay satisfaction on desire to retire suggests that people with greater financial secu-
rity are more likely to stop working). Thus, the superordinate effect models fell short
of the multivariate structural model in terms of theoretical utility, level of abstraction,
dimension specificity, and criterion-related validity.
The foregoing analyses also provide little support for adaptation as an aggregate
effect. Relationships between adaptation and its dimensions were modest for the
aggregate effect models, although some dimensions exhibited much stronger relation-
ships than others, even when the dimensions were assigned equal weight. The aggre-
gate effect models also produced much smaller criterion-related validities than the
multivariate structural model and concealed substantial variability in the effects of the
satisfaction facets on the adaptation dimensions. In conjunction, these results indicate
that the level of abstraction for adaptation as an aggregate effect is too broad. The theo-
retical utility of adaptation as an aggregate construct is also suspect, given that it con-
cealed meaningful variability in the effects of satisfaction facets on adaptation dimen-
sions. Although this variability was embedded within the aggregate effect models, it
was evident only when the coefficients for the individual adaptation dimensions were
examined, thereby revealing what the aggregate construct had concealed. Hence, the
aggregate effect models were inferior to the multivariate structural model on all points
underlying the multidimensional construct debate.
Discussion
The framework presented in this article incorporates multidimensional constructs
and their dimensions into a single analytical approach. This framework permits the
investigation of broad questions regarding multidimensional constructs along with
specific questions pertaining to the dimensions of these constructs. The framework
also provides tests relevant to issues underlying the ongoing debate over the utility of
multidimensional constructs, thereby allowing researchers to address these issues on a
study-by-study basis. Thus, the framework provides a holistic approach for research
on multidimensional constructs and their dimensions, causes, and effects.
Applications of the Framework
Results of the illustrative applications of the framework are summarized in Table 9,

which evaluates each model analyzed on criteria pertaining to the multidimensional
construct debate. Overall, the models fared reasonably well regarding the strength
of the relationships between the multidimensional construct and its dimensions.
However, on issues of theoretical utility, level of abstraction, construct validity, and
criterion-related validity, the multidimensional construct models were inferior to
multivariate structural models that used the dimensions as a set. Hence, for each of the
examples considered here, a multidimensional construct was better represented as a
set of related dimensions than as a single latent variable.
Table 9
184
Summary of Model Comparisons on Issues Underlying the Multidimensional Construct Debate

Theoretical Matching Levels Relationships Between Construct
Model Utility of Abstraction Construct and Dimensions Validity Criterion Validity
Superordinate Construct concealed Construct too broad for Strong relationships be- Specificities indicated Construct explained
cause meaningful differences outcomes, as indicated tween the construct and that construct distorted much less variance
in effects of dimensions by variable effects of di- dimensions for all models; effects of dimensions on than its dimensions as
mensions on outcomes relationships varied across outcomes a set
dimensions
Aggregate Information unique to Construct too broad for Relationships between Specificities indicated For models with con-
cause dimensions concealed outcomes for models construct and dimensions that construct distorted strained loadings, con-
by models with con- with constrained load- were moderate for models effects of dimensions on struct explained little
strained loadings; mod- ings; construct narrowed with constrained loadings outcomes, particularly for variance; for models
els with estimated load- to match outcomes for but weak and variable for models with constrained with estimated load-
ings reduced construct models with estimated models with estimated loadings ings, construct ex-
to a single dimension loadings loadings plained moderate vari-
ance but not as much
as that explained by its
dimensions as a set
Superordinate Information unique to Construct too broad for Relationships between Specificities indicated Much less variance ex-
effect dimensions concealed causes; for models with construct and dimensions that construct distorted plained in the construct
by all models; models free residual varainces, modest for parallel model or concealed effects of than in its dimensions
with free residual vari- construct narrowed to but strong for tau equiva- satisfaction on several di- as a set
ances reduced the con- capture relationship be- lent and congeneric mod- mensions
struct to a single di- tween one cause and els, due to large relation-
mension one dimension ship with absenteeism
Aggregate Construct concealed Construct too broad for Relationships between Substantial variation in Much less variance ex-
effect meaningful differences causes, as indicated by construct and dimensions effects on the dimen- plained in the construct
in effects on dimen- variable effects of causes were moderate but vari- sions concealed by using than in its dimensions
sions on dimensions able the construct as a set
Although the results reported here did not support the use of multidimensional con-
structs, other studies may produce different results. Indeed, a central premise of the
proposed framework is that the merits of multidimensional constructs and their dimen-
sions should be compared on a study-by-study basis. Nonetheless, these results exem-
plify those from other applications of the framework to personality traits, work with-
drawal, and other multidimensional constructs. Taken together, these results suggest
that support for multidimensional constructs will be the exception rather than the rule.
There are several reasons why multidimensional constructs are likely to perform
worse than their dimensions. First, a multidimensional construct comprises dimen-
sions that are necessarily distinct from one another. If the dimensions were not distinct,
then the construct would be unidimensional rather than multidimensional. Presum-
ably, the distinctions between the dimensions are conceptually meaningful, such that
the dimensions represent different aspects of a general concept or are expected to
relate differently to other variables. Consequently, the mere act of defining a multidi-
mensional construct prompts researchers to identify distinct dimensions that contain
more information than can be captured by single latent variable.
Second, most multidimensional construct models are more constrained than
multivariate structural models that treat the dimensions as a set. In general, constrained
models are inferior to unconstrained models in terms of model fit, explained variance,
and information regarding relationships among variables. However, constrained mod-
els are superior to unconstrained models in terms of parsimony. Thus, the advantages
of using models that treat dimensions as a set should be weighed against the loss of par-
simony inherent in such models. This tradeoff is built into statistical tests that incorpo-
rate differences in degrees of freedom between constrained and unconstrained models.
Although experience to date suggests that the benefits of parsimony provided by mul-
tidimensional construct models are not worth the costs, the relative magnitudes of
these benefits and costs will vary across studies.
Third, as constructs in the field of OB are refined, distinctions that were previously
overlooked often become increasingly clear and compelling. This tendency is exem-
plified by research on job characteristics (Hackman & Oldham, 1980; Hulin & Blood,
1968; Turner & Lawrence, 1965), job stress (Beehr & Newman, 1978; Cooper & Mar-
shall, 1976; Edwards, 1992; Schuler, 1980), and organizational commitment (Allen &
Meyer, 1990; Meyer, Allen, & Smith, 1993; Mowday et al., 1979; O’Reilly & Chatman,
1986), each of which has drawn progressively finer distinctions within constructs once
treated as unidimensional. As constructs become more differentiated, information
specific to construct dimensions becomes increasingly relevant, and multidimensional
construct models become less useful than multivariate structural models that treat con-
struct dimension as a set.
The use of multivariate structural models to represent multidimensional constructs
may generate resistance among OB researchers, given that these models contain no
direct vestige of the construct itself. Consequently, using these models may seem tan-
tamount to abandoning the study of multidimensional constructs. This is not the case.
Rather, a multivariate structural model represents a multidimensional construct as a set
of dimensions, and hypotheses regarding the causes and effects of the construct can be
tested using multivariate procedures, as illustrated earlier. Thus, a multidimensional
construct can be represented directly using a multidimensional construct model or
indirectly using a multivariate structural model. In either case, questions involving the
meaning, causes, and effects of the multidimensional construct can be investigated.
Extensions to the Framework
Although the framework presented in this article applies to multidimensional con-

structs commonly used in OB research, several extensions to the framework may be
considered. First, multidimensional construct models may be elaborated to include
direct relationships between the dimensions of the construct and its causes and effects.
These relationships may correspond to a priori hypotheses regarding relationships
involving the dimensions that are independent of the construct itself. Alternately, these
relationships may be added in response to tests of dimension specificities using modi-
fication indices. However, such models are exploratory, and their results should be
considered tentative, pending cross-validation. Models that add relationships for the
dimensions of the construct raise additional issues of identification, and general rules
that apply to these models are as follows. For a superordinate cause model, (a) no effect
may be caused by all dimensions, and (b) the total number of effects added must not
exceed [p2 – 2q + p(2q – 3)]/2, where p is the number of dimensions, and q is the num-
ber of effects. For an aggregate cause model, (a) no dimension may cause all effects,
(b) no effect may be caused by all dimensions, and (c) the total number of effects added
must not exceed ( p – 1)(q – 1), where p is the number of dimensions, and q is the num-
ber of effects. For both of these models, it is assumed that each effect is connected to
each other effect through a correlated residual or causal path, provided the relationship
among the effects remains recursive. For a superordinate effect model, (a) no cause
may affect all dimensions, and (b) the total number of effects added must not exceed
p(q – 1) + q(q – 3)/2, where p is the number of causes, and q is the number of dimen-
sions. No identification rules are provided for aggregate effect models because these
models already include all direct effects from the causes to the dimensions of the
construct.
Second, models may be tested in which multidimensional constructs are causes and
effects of other constructs. These models may be derived by starting with a model con-
sidered in this article and adding paths connecting the multidimensional construct to
additional causes and effects. Most of these models will follow the form of an aggre-
gate cause or superordinate effect model, although some of the latent variables these
models depict as construct dimensions will be recast as causes or effects. Models may
also be derived in which multidimensional constructs are causes and effects of one
another. Again, these models may be derived from those considered here by respecify-
ing a cause or effect as a superordinate or aggregate construct.
Third, multidimensional constructs may be compared with unidimensional con-
structs at the same level of abstraction (Rushton, Brainerd, & Pressley, 1983). These
comparisons can be used to assess whether the dimensions of the multidimensional
construct adequately capture the general concept of interest. For example, overall job
satisfaction as a multidimensional construct may be compared with a unidimensional
construct measured with items that describe general affective reactions to the job (e.g.,
“Overall, I am satisfied with my job”). The relationship between the multidimensional
and unidimensional job satisfaction constructs would indicate how well the dimen-
sions of the multidimensional construct capture the range of job facets that produce
feelings of overall job satisfaction (Ferratt, 1981; Ironson et al., 1989; Scarpello &
Campbell, 1983). In addition, relationships between the satisfaction dimensions and
the unidimensional satisfaction construct may be used to examine the effects of job
facet satisfaction on overall job satisfaction.
Finally, the framework may be used to rigorously evaluate measures formed by

summing heterogeneous items. These measures are widespread in OB research, yet
the methodological issues raised by these measures are rarely discussed. Typically,
these measures combine items intended to capture different aspects of a general con-
cept. Measures constructed in this manner are special cases of the aggregate construct
model in which the dimensions are represented by single indicators, error in the mea-
surement of the dimensions is disregarded, the dimensions are assigned equal weights,
and the residual of the construct is fixed to zero. This model is highly restrictive, and
the assumptions embedded in the model may not withstand empirical scrutiny, as dem-
onstrated in this article. Indeed, the prevalence of such measures is curious, given that
most researchers staunchly reject individual items that are heterogeneous (e.g., the
classic double-barreled item). Even when such items are avoided, problems of item
heterogeneity emerge when items that describe different aspects of a concept are com-
bined into a measure, because such measures simply move problems of heterogeneity
from the item level to the scale level. The implications of using such measures may be
investigated using aggregate construct models that test the constraints implied by sum-
ming heterogeneous items. Moreover, these models may be elaborated by using multi-
ple indicators of each dimension and treating the dimensions and the aggregate con-
struct as latent variables. Indeed, if the dimensions represented by heterogeneous
items are meaningful and worthwhile, then perhaps each dimension should be mea-
sured with multiple items rather than a single item.
Conclusion
Multidimensional constructs are widespread in OB research, yet there is little con-
sensus regarding the merits of these constructs relative to their dimensions. This lack
of consensus presents a dilemma for OB researchers who want the breadth and gener-
ality of multidimensional constructs along with the clarity and precision provided by
the dimensions of such constructs. This article has presented an integrative analytical
framework that combines multidimensional constructs and their dimensions and pro-
vides tests of issues underlying the multidimensional construct debate. By applying
this framework, researchers may obtain a better understanding of the complexities
underlying multidimensional constructs and draw firmer conclusions regarding ques-
tions that motivate the use of such constructs.
Notes
1. If two ηi as effects are not connected by a causal path, then allowing their residuals to cor-
relate admits the possibility that causes of the ηi excluded from the model may be correlated
with one another. If two ηi as effects are connected by a causal path (i.e., one ηi is a cause of the
other), then the residuals for the two ηi should not be allowed to correlate, because doing so in-
troduces a correlation between one ηi and the residual for the other ηi, and the parameters in the
equations for the two ηi are no longer identified.
2. If the estimated model includes fixed paths of unity from ξ1, ξ2, and ξ3 to η*, the variance of
η will be overestimated by φ11 + φ22 + φ33 + 2(φ12 + φ13 + φ23) + 2γ4(φ14 + φ24 + φ34), and the R2 re-
*
ported for η* will be incorrect. This problem can be solved by constraining the paths from ξ1, ξ2,
and ξ 3 to η* to zero. Although doing so may seem inconsistent with the definition of η* as
ξ1 + ξ2 + ξ3, this definition is incorporated through the constraints placed on γ4 and ψ*.
3. Throughout this discussion, parameters are fixed to values that facilitate model interpreta-
tion. The models discussed here will remain identified if parameters are fixed to other values.
4. It may seem reasonable to assess the relationship between an aggregate construct and its
dimensions by regressing the construct on its dimensions. However, this approach necessarily
yields an R2 of 1.00 for models that specify the construct as an exact linear combination of its di-
mensions, regardless of the dimension weights. In contrast, Ra2 assesses the degree to which the
construct captures the total variance of its dimensions. Therefore, Ra2 is more appropriate in the
present context.
5. Modification indices are used here not to guide specification searches (MacCallum,
Roznowski, & Necowitz, 1992), but instead to identify sources of misfit due to the omission of
direct effects for dimension specificities. This use of modification indices is consistent with the
framework presented in this article, which emphasizes comparing multiple a priori models, not
deriving models empirically.
6. In the Moberg (1998) study, all items used 5-point response scales, but measures were
formed by summing different numbers of items, producing measures with different metrics. By
dividing each measure by its number of items (i.e., transforming the measures from item sums to
item averages), all measures were returned to a common 5-point scale. A somewhat more com-
plicated procedure was required for the Hanisch and Hulin (1991) data because the satisfaction
and unfavorable job behavior items used 4-point scales, whereas the lateness, absenteeism, turn-
over intentions, and desire to retire items used 7-point scales. Therefore, after each measure was
divided by its number of items, the 4-point scales were converted to 7-point scales based on the
formula X7 = (X4 – 1)*2 + 1, where X4 represents a 1-4 scale and X7 represents a 1-7 scale. Be-
cause adding and subtracting constants has no affect on variances and covariances, this transfor-
mation amounted to doubling the standard deviations of the 4-point scales.
7. Dimension R2 values may be calculated from information reported in Figure 1 by dividing
the variance explained in each dimension by the total variance for the dimension. The former
quantity may be obtained by squaring the dimension loading, and the latter quantity may be ob-
tained by adding the squared dimension loading to the residual variance for the dimension. To il-
lustrate using results for positive emotions from the congeneric model, the squared loading is
.4242 = .180, and the total variance is .180 + .058 = .238. Dividing .180 by .238 yields .757,
which equals the reported R2 for positive emotions.
8. Although the expected parameter change statistic was negative for the effect of the asser-
tiveness specificity on control, this parameter was positive when added to the model.
9. If multiple indicators had been used for the dimensions or effects in the model, the
chi-square for the first-order factor model would have been nonzero, and chi-square difference
tests comparing this model with the other three models would be conducted in the usual manner.
References
Allen, N. J., & Meyer, J. P. (1990). The measurement and antecedents of affective, continuance,
and normative commitment to the organization. Journal of Occupational Psychology, 63,
1-18.
Anderson, J. C., & Gerbing, D. W. (1988). Structural equation modeling in practice: A review
and recommended two step approach. Psychological Bulletin, 103, 411-423.
Aston, M. C. (1998). Personality and job performance: The importance of narrow traits. Journal
of Organizational Behavior, 19, 289-303.
Bagozzi, R. P., & Edwards, J. R. (1998). A general approach to construct validation in organiza-
tional research: Application to the measurement of work values. Organizational Research
Methods, 1, 45-87.
Bedeian, A. G., Burke, B. G., & Moffett, R. G. (1988). Outcomes of work-family conflict among
married male and female professionals. Journal of Management, 14, 475-491.
Beehr, T. A., & Newman, J. E. (1978). Job stress, employee health, and organizational effective-
ness: A facet analysis, model and literature review. Personnel Psychology, 31, 665-699.
Bekker, P. A., Merckens, A., & Wansbeek, T. J. (1994). Identification, equivalent models and
computer algebra. New York: Academic Press.
Bentler, P. M. (1990). Comparative fit indexes in structural models. Psychological Bulletin, 107,
238-246.
Blau, G. (1998). On the aggregation of individual withdrawal behaviors into larger multi-item
constructs. Journal of Organizational Behavior, 19, 437-451.
Bollen, K., & Lennox, R. (1991). Conventional wisdom on measurement: A structural equation
perspective. Psychological Bulletin, 110, 305-314.
Bollen, K. A. (1989). Structural equations with latent variables. New York: Wiley.
Bollen, K. A., & Davis, W. R. (1994). Causal indicator models: Identification, estimation, and
testing. Paper presented at the American Sociological Association Convention, Miami, FL.
Bolton, B. (1980). Second-order dimensions of the Work Values Inventory (WVI). Journal of
Vocational Behavior, 17, 33-40.
Browne, M. W., & Cudeck, R. (1993). Alternative ways of assessing model fit. In K. A. Bollen &
J. S. Long (Eds.), Testing structural equation models (pp. 136-162). Newbury Park, CA:
Sage.
Browne, M. W., & Mels, G. (1992). RAMONA user’s guide. Columbus: Department of Psychol-
ogy, Ohio State University.
Cattell, R. B., & Tsujioka, B. (1964). The importance of factor-trueness and validity, versus ho-
mogeneity and orthogonality, in test scales. Educational and Psychological Measurement,
24, 3-30.
Cohen, J. (1982). Set correlation as a general multivariate data-analytic method. Multivariate
Behavioral Research, 17, 301-341.
Cooper, C. L., & Marshall, J. (1976). Occupational sources of stress: Review of literature relat-
ing to coronary heart disease and mental ill health. Journal of Occupational Psychology, 49,
11-28.
Costa, P. T., Jr., & McCrae, R. R. (1992). Manual for the Revised NEO Personality Inventory
(NEO-PIR) and the NEO Five-Factor Inventory (NEO-FFI). Odessa, FL: Psychological
Assessment Resources.
Cronbach, L. J., Gleser, G., Nanda, H., & Rajaratnam, N. (1972). The dependability of behav-
ioral measurements: Theory of generalizability for scores and profiles. New York: Wiley.
Digman, J. M. (1990). Personality structure: Emergence of the five-factor model. Annual Re-
view of Psychology, 41, 417-440.
Digman, J. M. (1997). Higher-order factors of the Big Five. Journal of Personality and Social
Psychology, 73, 1246-1256.
Dwyer, J. H. (1983). Statistical models for the social and behavioral sciences. New York: Ox-
ford University Press.
Edwards, J. R. (1991). Person-job fit: A conceptual integration, literature review, and method-
ological critique. In C. L. Cooper & I. T. Robertson (Eds.), International review of indus-
trial and organizational psychology (Vol. 6, pp. 283-357). New York: Wiley.
Edwards, J. R. (1992). A cybernetic theory of stress, coping, and well-being in organizations.
Academy of Management Review, 17, 238-274.
Edwards, J. R., & Bagozzi, R. P. (2000). On the nature and direction of the relationship between
constructs and measures. Psychological Methods, 5, 155-174.
Efron, B., & Tibshirani, R. (1993). An introduction to the bootstrap. New York: Chapman &
Hall.
Ferratt, T. W. (1981). Overall job satisfaction: Is it a linear function of facet satisfaction? Human
Relations, 34, 463-473.
Fisher, C. D. (1980). On the dubious wisdom of expecting job satisfaction to correlate with per-
formance. Academy of Management Review, 5, 607-612.
Gerbing, D. W., & Anderson, J. C. (1988). An updated paradigm for scale development incorpo-
rating unidimensionality and its assessment. Journal of Marketing Research, 25, 186-192.
Gerbing, D. W., & Anderson, J. C. (1993). Monte Carlo evaluations of goodness of fit indices for
structural equation models. In K. A. Bollen & J. S. Long (Eds.), Testing structural equation
models (pp. 40-65). Newbury Park, CA: Sage.
Goldberg, L. R. (1993). The structure of personality traits: Vertical and horizontal aspects. In
D. C. Funder, R. D. Parke, C. Tomlinson-Keasey, & K. Widaman (Eds.), Studying lives
through time: Personality and development (pp. 169-188). Washington, DC: American
Psychological Association.
Hackman, J. R., & Oldham, G. R. (1980). Work redesign. Reading, MA: Addison-Wesley.
Hanisch, K. A., & Hulin, C. L. (1990). Job attitudes and organizational withdrawal: An exami-
nation of retirement and other voluntary withdrawal behaviors. Journal of Vocational Be-
havior, 37, 60-78.
Hanisch, K. A., & Hulin, C. L. (1991). General attitudes and organizational withdrawal: An
evaluation of a causal model. Journal of Vocational Behavior, 39, 110-128.
Hanisch, K. A., Hulin, C. L., & Roznowski, M. (1998). The importance of individuals’ reper-
toires of behaviors: The scientific appropriateness of studying multiple behaviors and gen-
eral attitudes. Journal of Organizational Behavior, 19, 463-480.
Hattie, J. (1985). Methodology review: Assessing unidimensionality of tests and items. Applied
Psychological Measurement, 9, 139-164.
Heise, D. R. (1972). Employing nominal variables, induced variables, and block variables in
path analysis. Sociological Methods and Research, 1, 147-173.
Hogan, J., & Roberts, B. W. (1996). Issues and non-issues in the fidelity-bandwidth trade-off.
Journal of Organizational Behavior, 17, 627-637.
Holland, J. L. (1985). Making vocational choices (2nd ed.). Englewood Cliffs, NJ: Prentice
Hall.
Hu, L., & Bentler, P. M. (1999). Cutoff criteria for fit indexes in covariance structure analysis:
Conventional criteria versus new alternatives. Structural Equation Modeling, 6, 1-55.
Hulin, C. L. (1991). Adaptation, persistence, and commitment in organizations. In M. D.
Dunnette & L. M. Hough (Eds.), Handbook of industrial and organizational psychology
(2nd ed., pp. 445-505). Palo Alto, CA: Consulting Psychologists Press.
Hulin, C. L., & Blood, M. R. (1968). Job enlargement, individual differences, and worker re-
sponses. Psychological Bulletin, 69, 41-55.
Hull, J. G., Lehn, D. A., & Tedlie, J. C. (1991). A general approach to testing multifaceted per-
sonality constructs. Journal of Personality and Social Psychology, 61, 932-945.
Humphreys, L. G. (1962). The organization of human abilities. American Psychologist, 17,
475-483.
Humphreys, L. G. (1970). A skeptical look at the pure factor test. In C. E. Lunneborg (Ed.), Cur-
rent problems and techniques in multivariate psychology: Proceedings of a conference hon-
oring professor Paul Horst (pp. 23-32). Seattle: University of Washington.
Hunt, S. D., & Morgan, R. M. (1994). Organizational commitment: One of many commitments
or key mediating construct? Academy of Management Journal, 37, 1568-1587.
Hunter, J. E., & Gerbing, D. W. (1982). Unidimensional measurement, second order factor anal-
ysis, and causal models. In B. M. Staw & L. L. Cummings (Eds.), Research in organiza-
tional behavior (pp. 267-320). Greenwich, CT: JAI Press.
Ironson, G. H., Smith, P. C., Brannick, M. T., Gibson, W. M., & Paul, K. B. (1989). Construction
of a job in general scale: A comparison of global, composite, and specific measures. Jour-
nal of Applied Psychology, 74, 193-200.
James, L. A., & James, L. R. (1989). Integrating work environment perceptions: Explorations
into the measurement of meaning. Journal of Applied Psychology, 74, 739-751.
James, L. R., & Tetrick, L. E. (1986). Confirmatory analytic tests of three causal models relating
job perceptions to job satisfaction. Journal of Applied Psychology, 71, 77-82.
Johns, G. (1998). Aggregation or aggravation? The relative merits of a broad withdrawal con-
struct. Journal of Organizational Behavior, 19, 453-462.
Jöreskog, K. G. (1971). Statistical analysis of sets of congeneric tests. Psychometrika, 36,
109-133.
Jöreskog, K. G., & Goldberger, A. S. (1975). Estimation of a model with multiple indicators and
multiple causes of a single latent variable. Journal of the American Statistical Association,
10, 631-639.
Jöreskog, K. G., & Sörbom, D. (1996). LISREL 8: User’s reference guide. Chicago: Scientific
Software International.
Kim, J. O., & Mueller, C. W. (1978). Factor analysis. Beverly Hills, CA: Sage.
Kristof-Brown, A. L. (1996). Person-organization fit: An integrative review of its conceptual-
ization, measurement, and implications. Personnel Psychology, 49, 1-49.
Law, K. S., & Wong, C. S. (1999). Multidimensional constructs in structural equation analysis:
An illustration using the job perception and job satisfaction constructs. Journal of Manage-
ment, 25, 143-160.
Law, K. S., Wong, C. S., & Mobley, W. H. (1998). Toward a taxonomy of multidimensional con-
structs. Academy of Management Review, 23, 741-755.
Liden, R. C., & Maslyn, J. M. (1998). Multidimensionality of leader-member exchange: An em-
pirical assessment through scale development. Journal of Management, 24, 43-72.
Locke, E. A. (1976). The nature and causes of job satisfaction. In M. Dunnette (Ed.), Handbook
of industrial and organizational psychology (pp. 1297-1350). Chicago: Rand McNally.
MacCallum, R., & Browne, M. W. (1993). The use of causal indicators in covariance structure
models: Some practical issues. Psychological Bulletin, 114, 533-541.
MacCallum, R., Browne, M. W., & Sugawara, H. M. (1996). Power analysis and determination
of sample size for covariance structure modeling. Psychological Methods, 1, 130-149.
MacCallum, R., Roznowski, M., & Necowitz, L. B. (1992). Model modification in covariance
structure analysis: The problem of capitalization on chance. Psychological Bulletin, 111,
490-504.
McCrae, R. R., & Costa, P. T., Jr. (1992). Discriminant validity of the NEO-PIR facet scales. Ed-
ucational and Psychological Measurement, 52, 229-237.
McIver, J. P., & Carmines, E. G. (1981). Unidimensional scaling. Beverly Hills, CA: Sage.
Mershon, B., & Gorsuch, R. L. (1988). Number of factors in the personality sphere: Does in-
crease in factors increase predictability of real-life criteria? Journal of Personality and So-
cial Psychology, 55, 675-680.
Meyer, J. P., Allen, N. J., & Smith, C. A. (1993). Commitment to organizations and occupations:
Extension and test of a three-component conceptualization. Journal of Applied Psychology,
78, 538-551.
Moberg, P. J. (1998). Predicting conflict strategy with personality traits: Incremental validity
and the five factor model. International Journal of Conflict Management, 9, 258-285.
Mowday, R. T., Steers, R. M., & Porter, L. W. (1979). The measurement of organizational com-
mitment. Journal of Vocational Behavior, 14, 224-247.
Murphy, K. R., & Shiarella, A. H. (1997). Implications of the multidimensional nature of job
performance for the validity of selection tests: Multivariate frameworks for studying test
validity. Personnel Psychology, 50, 823-854.
Myers, I. B., & McCaulley, M. (1985). Manual: A guide to the development and use of the
Myers-Briggs Type Indicator. Palo Alto, CA: Consulting Psychologists Press.
Nunnally, J. C. (1978). Psychometric theory (2nd ed.). New York: McGraw-Hill.
Ones, D. S., & Viswesvaran, C. (1996). Bandwidth-fidelity dilemma in personality measure-
ment for personnel selection. Journal of Organizational Behavior, 17, 609-626.
O’Reilly, C., III, & Chatman, J. (1986). Organizational commitment and psychological attach-
ment: The effects of compliance, identification, and internalization on prosocial behavior.
Journal of Applied Psychology, 71, 492-499.
Parasuraman, S., Greenhaus, J. H., & Granrose, C. S. (1992). Role stressors, social support, and
well-being among two-career couples. Journal of Organizational Behavior, 13, 339-356.
Paunonen, S. V. (1998). Hierarchical organization of personality and prediction of behavior.
Journal of Personality and Social Psychology, 74, 538-556.
Paunonen, S. V., Rothstein, M. G., & Jackson, D. N. (1999). Narrow reasoning about the use of
broad personality measures for personnel selection. Journal of Organizational Behavior,
20, 389-405.
Pryor, R.G.L. (1987). Differences among differences: In search of general work preference di-
mensions. Journal of Applied Psychology, 72, 426-433.
Putnam, L. L., & Wilson, C. (1982). Communicative strategies in organizational conflict: Reli-
ability and validity of a measurement scale. In M. Burgoon (Ed.), Communication yearbook
(Vol. 6, pp. 629-652). Newbury Park, CA: Sage.
Rindskopf, D., & Rose, T. (1988). Some theory and applications of confirmatory second-order
factor analysis. Multivariate Behavioral Research, 23, 51-67.
Roznowski, M., & Hanisch, K. A. (1990). Building systematic heterogeneity into work attitudes
and behavior measures. Journal of Vocational Behavior, 36, 361-375.
Rushton, J. P., Brainerd, C. J., & Pressley, M. (1983). Behavioral development and construct va-
lidity: The principle of aggregation. Psychological Bulletin, 94, 18-38.
Scarpello, V., & Campbell, J. P. (1983). Job satisfaction: Are all the parts there? Personnel Psy-
chology, 36, 577-600.
Schmidt, F. L., & Kaplan, L. B. (1971). Composite vs. multiple criteria: A review and resolution
of the controversy. Personnel Psychology, 24, 419-434.
Schneider, R. J., Hough, L. M., & Dunnette, M. D. (1996). Broadsided by broad traits: How to
sink science in five dimensions or less. Journal of Organizational Behavior, 17, 639-655.
Schuler, R. S. (1980). Definition and conceptualization of stress in organizations. Organiza-
tional Behavior and Human Performance, 25, 184-215.
Schwab, D. P. (1980). Construct validity in organizational behavior. In L. L. Cummings & B. M.
Staw (Eds.), Research in organizational behavior (Vol. 2, pp. 3-43). Greenwich, CT.: JAI
Press.
Smith, P. C., Kendall, L., & Hulin, C. L. (1969). The measurement of satisfaction in work and re-
tirement. Chicago: Rand McNally.
Steiger, J. H. (1990). Structural model evaluation and modification: An interval estimation ap-
proach. Multivariate Behavioral Research, 25, 173-180.
Thompson, B. (1984). Canonical correlation analysis: Uses and interpretation. Newbury Park,
CA: Sage.
Turner, A. N., & Lawrence, P. R. (1965). Industrial jobs and the worker: An investigation of re-
sponses to task attributes. Boston: Harvard University Press.
Warr, P. B., Cook, J. D., & Wall, T. D. (1979). Scales for the measurement of some work attitudes
and aspects of psychological well-being. Journal of Occupational Psychology, 52,
129-148.
Weick, K. E. (1979). The social psychology of organizing. New York: McGraw-Hill.
Jeffrey R. Edwards (Ph.D., Carnegie Mellon University) is the Belk Distinguished Professor of Management
at the Kenan-Flagler Business School at the University of North Carolina. He studies stress, coping, and
well-being, person-organization fit, and methodological issues. His work has appeared in the Academy of
Management Review, the Academy of Management Journal, the Journal of Applied Psychology, Organiza-
tional Behavior and Human Decision Processes, Personnel Psychology, Psychological Methods, and Orga-
nizational Research Methods. He is past chair of the Research Methods Division of the Academy of Manage-
ment.

Edwards 2001 A

Uploaded by

Copyright:

Available Formats

Edwards 2001 A

Uploaded by

Document Information

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

Edwards 2001 A

Uploaded by

Copyright:

Available Formats

ORGANIZATIONAL

Multidimensional constructs are widely used to represent several distinct dimen-

Multidimensional constructs are pervasive in organizational behavior (OB) research.

The utility of multidimensional constructs has generated considerable debate in the

Types of Multidimensional Constructs

dimensions. Therefore, the relationships between a multidimensional construct and its

As noted previously, a superordinate construct is a general concept that is mani-

Unlike a superordinate construct, an aggregate construct is a composite of its

Other Types of Multidimensional Constructs

Although most multidimensional constructs are either superordinate or aggregate,

commitment defined by dimensions that include facets of commitment (e.g., accep-

The Multidimensional Construct Debate

Advocates of multidimensional constructs have argued that such constructs are

Matching Levels of Abstraction

Multidimensional constructs have been recommended for matching general pre-

Another source of debate concerns the internal consistency reliability of measures

Advocates of multidimensional constructs have criticized the construct validity of

measurement error. Likewise, Humphreys (1970) argued that variance specific to

Finally, advocates of multidimensional constructs have argued that such constructs

Summary and Implications of the Debate

An Integrative Analytical Framework

Superordinate Construct as a Cause

As explained earlier, a superordinate construct is best viewed as a second-order fac-

ηi = γξ* + ζi. (2)

ηi = γiξ* + ζi. (3)

Aggregate Construct as a Cause

As explained earlier, an aggregate construct is best viewed as a function of latent

η* = Σγiξi + ζ*. (7)

In Equation 7, ζ* represents aspects of η* not captured by its dimensions. Thus, Equa-

ηj = βjη* + ζj. (8)

Like a superordinate construct, an aggregate construct represents the relationships

Superordinate Construct as an Effect

We now turn to models that treat a multidimensional construct as an effect, starting

ηj = βη* + ζj. (10)

ηj = βjη* + ζj. (11)

η* = Σγiξi + ζ*. (12)

Note that Equation 12 is identical to Equation 7 for an aggregate construct as a cause.

Aggregate Construct as an Effect

η* = γ4ξ4 + ζ*. (13)

γ4 = C(η*, ξ4)/φ44, (14)

γ4 = φ14/φ44 + φ24/φ44 + φ34/φ44. (15)

ψ* = V(η* – γ4ξ4), (16)

where V(.) is a variance operator. Substituting Equation 15 for γ4 and ξ1 + ξ2 + ξ3 for η*

γ4 = w1φ14/φ44 + w2φ24/φ44 + w3φ34/φ44 , (18)

ψ * = w12φ 11 + w 22φ 22 + w 32φ 33 + 2 w1w 2φ 12 + 2 w1w 3φ 13 + 2 w 2 w 3φ 23 (19)

Hence, if η* is a weighted sum, γ4 and ψ* should be constrained to the expressions

η1 + η2 + η3 = γ4ξ4 + ζ*. (21)

Equation 21 may be expanded into a set of three equations representing a multivariate

η3 = γ3ξ4 + ζ3. (22)

η1 + η2 + η3 = (γ1 + γ2 + γ3)ξ4 + ζ1 + ζ2 + ζ3. (23)

Hence, γ4 and ζ* in Equation 21 equal γ1 + γ2 + γ3 and ζ1 + ζ2 + ζ3 in Equation 22, respec-

η* = w1ξ1 + w2ξ2 + w3ξ3. (24)

The weights in Equation 24 may be incorporated into Equation 22 by multiplying both

w1η1 = w1γ1ξ4 + w1ζ1

w2η2 = w2γ2ξ4 + w2ζ2

w3η3 = w3γ3ξ4 + w3ζ3. (25)

Model Estimation and Evaluation

Identification. Prior to estimation, it is necessary to establish that all parameters in a

Specifying relationships between the multidimensional construct and its dimen-

Relationships for dimension specificities. Relationships between dimension