Você está na página 1de 20

Psychological Methods Copyright 2000 by the American Psychological Association, Inc.

2000, Vol. 5, No. 2~ 155-174 1082-989X/00/$5.00 DOI: 10.1037//1082-989X.5.2,155

On the Nature and Direction of Relationships Between


Constructs and Measures
Jeffrey R. Edwards Richard P. Bagozzi
University of North Carolina at Chapel Hill and Rice University and University of Michigan
University of Michigan Business School Business School

Theory development typically focuses on relationships among theoretical con-


structs, placing little emphasis on relationships between constructs and measures. In
most cases, constructs are treated as causes of their measures. However, this causal
flow is sometimes reversed, such that measures are viewed as causes of constructs.
Procedures have been developed to identify and estimate models that specify con-
structs as causes or effects of measures. However, these procedures provide little
guidance for determining a priori whether constructs should be specified as causes
or effects of their measures. Moreover, these procedures address few of the possible
causal structures by which constructs and measures may be related. This article
develops principles for specifying the direction and structure of relationships be-
tween constructs and measures. These principles are illustrated using examples
from psychological, sociological, and organizational research.

A theory can be divided into two parts: one that measures. These relationships are of paramount im-
specifies relationships between theoretical constructs portance because they constitute an auxiliary theory
and another that describes relationships between con- that bridges the gap between abstract theoretical con-
structs and measures (Bagozzi & Phillips, 1982; Cost- structs and measurable empirical phenomena (Cost-
her, 1969). Presentations of theory often place great ner, 1969). Without this auxiliary theory, the mapping
emphasis on explaining causal relationships among of theoretical constructs onto empirical phenomena is
constructs but devote little attention to the nature and ambiguous, and theories cannot be meaningfully
direction of relationships between constructs and tested (Blalock, 1971).
The nature and direction of relationships between
constructs and measures have been discussed in the
literature on construct validity and structural equation
Jeffrey R. Edwards, Department of Management, Kenan-
modeling (Blalock, 1971; Bollen, 1989; DeVellis,
Flagler Business School, University of North Carolina at
Chapel Hill and Department of Organizational Behavior and 1991). In this literature, constructs are usually viewed
Human Resource Management, University of Michigan as causes of measures, meaning that variation in a
Business School; Richard P. Bagozzi, Department of Man- construct leads to variation in its measures (Bollen,
agement, Jones Graduate School of Management, Rice Uni- 1989). Such measures are termed reflective because
versity and Department of Marketing, University of Michi- they represent reflections, or manifestations, of a con-
gan Business School. struct (Fornell & Bookstein, 1982). Reflective mea-
We thank Richard S. Blackburn, Kenneth A. Bollen, Kyle surement underlies classical test theory (Lord &
D. Cattani, James W. Dean Jr., David A. Harrison, Novick, 1968), reliability estimation (Nunnally,
Lawrence R. James, Keith A. Markus, Edward E. Rigdon, 1978), and factor analysis (Harman, 1976; Kim &
Albert H. Segars, Larry J. Williams, and members of the
Mueller, 1978), each of which treats a measure as a
Society of Organizational Behavior for their helpful com-
function of a latent variable (i.e., construct) plus error.
ments during the development of this article.
Correspondence concerning this article should be ad- In some situations, measures are viewed as causes
dressed to Jeffrey R. Edwards, Department of Management, of constructs (Bagozzi & Fornell, 1982; Blalock,
Kenan-Flagter Business School, University of North Caro- t971; Bollen & Lennox, 1991; Heise, 1972; MacCal-
lina, Campus Box 3490, Chapel Hill, North Carolina 27599- lum & Browne, 1993). Such measures are termed for-
3490. Electronic mail may be sent to jredwards@unc.edu. mative, meaning the construct is formed or induced by

155
156 EDWARDS AND BAGOZZI

its measures (Fornell & Bookstein, 1982). 1 Formative causal direction and structure of the relationship be-
measures are commonly used for constructs con- tween constructs and measures. By applying these
ceived as composites of specific component variables, principles, researchers may work toward developing
as when socioeconomic status (SES) is defined in auxiliary theories linking constructs to measures with
terms of occupation, education, and income (Hauser the same precision and rigor used to specify relation-
& Goldberger, 1971; Marsden, 1982). ships between theoretical constructs.
Discussions of reflective and formative measures We begin by defining constructs and measures and
have focused primarily on identification and estima- addressing key philosophical issues underlying these
tion issues (Blalock, 1971; Bollen & Lennox, 1991; definitions. Next, we draw from principles of causal-
MacCallum & Browne, 1993). These issues are criti- ity to develop criteria for determining the direction of
cally important because they must be addressed be- the relationship between a construct and a measure.
fore models with reflective and formative measures We then derive hypothetical models that delinate ba-
can be empirically tested. However, little attention has sic causal structures by which a construct and a mea-
been devoted to the conditions in which measures sure might relate. Cumulatively, this discussion yields
should be specified as reflective or formative in the guidelines for specifying the relationship between a
first place. Rough guidelines can be inferred from construct and a measure in terms of (a) direction (i.e.,
examples used to illustrate reflective and formative whether a construct causes or is caused by its mea-
sures) and (b) structure (i.e., whether the relationship
measures (Bollen & Lennox, 1991; MacCallum &
is direct, indirect, spurious, or unanalyzed). We then
Browne, 1993), but these examples are open to dif-
apply these guidelines to constructs and measures
ferent interpretations and may not apply to measures
from the psychological, sociological, and organiza-
used in a particular study. Covariances among mea-
tional literatures. We conclude with recommendations
sures can help differentiate formative from reflective
for theory development, with particular emphasis on
measures, because these covariances follow a predict-
the integration of auxiliary theories relating constructs
able pattem for reflective measures but are indetermi-
to measures into substantive theories of relationships
nate for formative measures (Bollen & Lennox, 1991;
among constructs.
Bollen & Ting, 1993). However, this approach per-
mits only tentative conclusions, because imperfect re-
flective measures may yield covariances that deviate D e f i n i n g Constructs and M e a s u r e s
from the pattern expected for such measures, and for-
mative measures may exhibit covariances that happen Before we discuss relationships between constructs
to follow the pattern expected for reflective measures. and measures, we must first define these terms. We
Moreover, a particular pattem of covariances may be define a m e a s u r e as an observed score gathered
consistent with numerous causal structures (Bollen & through self-report, interview, observation, or some
Ting, 1993; Duncan, 1975; Lee & Hershberger, 1990; other means (DeVellis, 1991; Lord & Novick, 1968;
MacCallum, Wegener, Uchino, & Fabrigar, 1993). Messick, 1995). Put simply, a measure is a quantified
Hence, researchers currently have few conceptual or record, or datum, taken as an empirical analog to a
empirical criteria for determining whether measures construct. Note that a measure refers not to an instru-
should be specified as reflective or formative. This ment used to gather data or to the act of collecting
shortcoming seriously hampers tests of substantive data, but to the score generated by these procedures. A
theories because if the causal structure relating con- c o n s t r u c t is a conceptual term used to describe a phe-
structs to measures is specified incorrectly, relation-
ships among constructs cannot be meaningfully tested
(Blalock, 1971; MacCallum & Browne, 1993).
This article provides general principles for specify- 1 Blalock (1964), who is often credited with distinguish-
ing measures as effects versus causes of constructs, labeled
ing the nature and direction of the relationship be-
reflective measures effect indicators and formative mea-
tween constructs and measures. These principles are sures cause indicators. We prefer the terms reflective and
derived from logical and philosophical arguments re- formative because these terms do not imply that the rela-
garding the meaning of constructs and measures and tionship between a construct and a measure is necessarily
the nature and form of their relationship (Blalock, causal. As we later show, relationship between some con-
1971; Costner, 1969). These principles provide a structs and measures should be viewed as definitional rather
priori criteria by which researchers can specify the than causal.
RELATIONSHIPS BETWEEN CONSTRUCTS AND MEASURES 157

nomenon of theoretical interest (Cronbach & Meehl, causal. Causal language pervades discussions of re-
1955; Nunnally, 1978; Schwab, 1980). Several points flective measures. For example, Lord and Novick
regarding this definition should be elaborated. First, (1968) described measurement error as that part of an
although constructs are terms researchers literally observed variable that is not "determined by" a con-
construct, or put together (Nunnally, 1978), we intend struct (p. 531), and Nunnally (1978) characterized
that constructs refer to phenomena that are real and measurement error as the variance in a measure that is
exist apart from the awareness and interpretation of not "explained by" the true score (p. 201). Long
the researcher and the persons under study (Cook & (1983) was more explicit, stating that the arrow lead-
Campbell, 1979; Loevinger, 1957; Messick, 1981). ing from a construct to its measure represents the
For instance, psychologists develop cognitive, attitu- "causal effect" of the construct on the measure (p. 11).
dinal, and emotional constructs to describe real phe- Likewise, DeVellis (1991) asserted that a latent vari-
nomena that are experienced by people. Likewise, so- able (i.e., a construct) is "a cause of the item score"
ciologists formulate constructs such as social (p. 13). Discussions of formative measures often use
stratification, social mobility, and SES to represent causal language as well. For example, Blalock (1971)
real phenomena in social collectives. Second, al- described formative measures as indicating that a la-
though constructs refer to real phenomena, constructs tent variable is measured using "one or more of its
themselves are not real in an objective sense (Nun- causes" (p. 336). Likewise, MacCallum and Browne
nally, 1978). Rather, they are elements of scientific
(1993, p. 533) and Bedeian, Day, and Kelloway
discourse that serve as verbal surrogates for phenom-
(1997, p. 788) stated that formative measures "may be
ena of interest. Thus, when we speak of the relation-
viewed as causing" latent variables. Bollen and Len-
ship between a measure and a construct, we refer to
nox (1991) were more cautious, stating that they "do
the relationship between a measure and the phenom-
not attribute any special significance to the term cause
enon named by the construct. Third, the phenomena
other than the fact that the indicators determine the
that constructs describe can be unobservable (e.g., at-
latent variable" (p. 306), although it is unclear how
titudes) or observable (e.g., task performance). In ei-
"determine" differs from the concept of cause.
ther case, the construct itself is an abstract term that
Although discussions of formative and reflective
describes the phenomenon. Hence, subjective states
measures often use causal language, they rarely in-
are described as attitudes and behavior is described as
corporate principles of causality from the philosophy
task performance when researchers derive labels for
of science (e.g., Hume, 1946; Mill, 1886; Popper,
these phenomena. Finally, constructs differ in how
1959; Suppes, 1970). These principles are an ongoing
well they describe and assign meaning to phenomena
source of debate, although the current literature sug-
of theoretical interest. Some constructs may demon-
gests some consensus on four conditions for estab-
strate ongoing usefulness, whereas others initially
lishing causality in the social, behavioral, and man-
considered useful may be modified or abandoned as
agement sciences (Asher, 1983; Bagozzi, 1980;
knowledge accumulates. These advances may occur
Bollen, 1989; Cook & Campbell, 1979; Heise, 1975;
even when the phenomenon of interest remains un-
James, Mulaik, & Brett, 1982). First, causality re-
changed. In sum, our definition of a construct repre-
quires that the cause and the effect are distinct enti-
sents a critical realist perspective, in that we view
ties. When two variables are not distinct, their rela-
constructs as attempts to describe real phenomena, but
tionship is tautological and therefore should not be
we recognize that these phenomena cannot be known
viewed as causal. Second, causality requires associa-
directly or with complete accuracy because of mea-
tion, meaning that the cause and the effect covary.
surement error and the imperfect epistemological lens
Association is usually viewed as probabilistic rather
that a construct provides (Cook & Campbell, 1979;
than definitional, meaning the cause increases the
Delanty, 1997; Loevinger, 1957; Messick, 1981; Zu-
likelihood of the effect but does not guarantee that the
tiff, 1998).
effect will occur (Cook & Campbell, 1979; James et
al., 1982; Suppes, 1970). Third, causality requires
Causal Direction o f the Relationship Between temporal precedence, such that the cause occurs be-
Constructs and Measures fore the effect. Although causality may be nearly in-
stantaneous at the micromedial level, the cause must
Discussions of the relationship between constructs precede the effect by some minimal time interval
and measures often characterize this relationship as (Cook & Campbell, 1979). Finally, causality requires
158 EDWARDS AND BAGOZZI

the elimination of rival explanations for the presumed struct (Bollen, 1989). Alternatively, if a construct has
relationship between the cause and the effect (Cook & one or more formative measures, the relationships be-
Campbell, 1979). Of the four conditions for causality, tween the measures and the construct can be estimated
this last condition is often the most difficult to satisfy provided the construct is specified as a direct or an
in practice, particularly in nonexperimental research. indirect cause o f at least two r e f e c t i v e measures
We now apply the foregoing conditions for causality (Bollen & Davis, 1994). Although these procedures
to the relationship between constructs and measures. provide empirical evidence of association, this evi-
dence may be consistent with other models with dif-
Distinct Entities ferent causal orderings that yield different estimates
The first condition for causality stipulates that a of the relationships between constructs and measures
construct and a measure must be distinct. This condi- (Duncan, 1975; Lee & Hershberger, 1990; MacCal-
tion is satisfied by the definitions of construct and lum et al., 1993). Therefore, empirical evidence of
measure provided earlier, in which a construct refers association provides necessary but not sufficient sup-
to a phenomenon of theoretical interest and a measure port for a hypothesized causal relationship between a
is an observed score. The distinction between a con- construct and a measure. 2
struct and a measure is evident for attitudes, cogni- A second method for evaluating the association be-
tions, and other mental states or events, for which the tween a construct and a measure entails the use of
construct refers to a phenomenon within the mind of "mental experiments" (Bollen, 1989), in which a re-
a person and the measure is a recorded trace taken as searcher imagines a change in the construct and then
evidence of the construct (e.g., a number circled by a considers whether a change in the measure is likely.
respondent on a survey). Constructs that refer to be- For example, a researcher might consider whether
havior are also distinct from measures of behavior, variation in a j o b involvement construct would relate
because exhibiting a behavior is not the same as the to different scores for items such as "The most im-
score associated with that behavior. For example, the portant things that happen in life involve work"
act o f being absent from work is distinct from a score (Kanungo, 1982). Mental experiments entail the use
on the item "How many days o f work have you of disciplined imagination (Weick, 1989) to deduce
missed in the past year?" (Johns, 1994). The distinc- whether an association between a construct and a
tion between a construct and a measure breaks down measure is plausible. Ultimately, mental experiments
under operationalism, in which a construct is defined rely on speculation and "appeals to reason" that a
in terms of its measures; for example, intelligence was construct and a measure covary (cf. Nunnally, 1978,
once defined as that which the Stanford-Binet mea- p. 93). Therefore, mental experiments should be used
sures (Campbell, 1960). in conjunction with empirical procedures, acknowl-
edging that neither approach provides definitive evi-
Association dence of association between a construct and a mea-
sure.
The second condition for causality is association,
Ideally, the association between a construct and its
meaning the construct and the measure must covary.
measures should remain stable regardless of the larger
In philosophical discussions of causality, association
causal model in which the construct and measures are
usually refers to an empirical relationship between
embedded. If the association varies, then no unique
two objects. For a construct and a measure, however,
meaning can be assigned to the construct (Burt, 1976),
the researcher has direct access to the measure but not
and claims of association between the construct and
the construct (Nunnally, 1978). Hence, association
measures are tenuous. The association between a con-
between a construct and a measure must be inferred
struct and its measures is generally stable for reflec-
by using methods that do not rely on direct observa-
tion.
One such method involves the use of covariances
among multiple measures of a construct to infer the
z We refer to association net of other influences as a
relationships between the focal construct and its mea- necessary condition for causality (Bollen, 1989, p. 57). A
sures. For example, if a construct has three or more bivariate association between a construct and a measure is
reflective measures, the covariances among the mea- not a necessary condition for causality because their asso-
sures may be used to estimate the association (i.e., ciation may entail two effects of opposite sign that yield a
factor loading) between each measure and the con- null total effect.
RELATIONSHIPS BETWEEN CONSTRUCTS AND MEASURES 159

tive measures that serve as alternative (i.e., substitute) sequence in which one link in the sequence must be
indicators of a construct, provided these measures taken as given (i.e., the stimulus ~ organism link
correlate more highly with one another than with mea- must be assumed to test the effect of a construct on a
sures of other constructs. In contrast, associations of measure, whereas the organism ~ response link must
formative measures with their construct are deter- be assumed to test the effect of a measure on a con-
mined primarily by the relationships between these struct).
measures and measures of constructs that are depen- If conventional experimentation is not possible, re-
dent on the construct of interest. This point was made searchers may use mental experiments (Bollen, 1989)
by Heise (1972), who noted that a construct measured to evaluate various temporal orderings between a con-
formatively is not just a composite of its measures; struct and a measure. For example, constructs repre-
rather, "it is the composite that best predicts the de- senting well-formed attitudes presumably exist before
pendent variable in the analysis . . . . Thus, the mean- being measured by a researcher. Such constructs
ing of the latent construct is as much a function of the should therefore precede their measures, indicating a
dependent variable as it is a function of its indicators" temporal sequence consistent with reflective measure-
(p. 160). Unstable associations between formative ment. Some attitudes may be formed or changed in
measures and constructs not only create difficulties response to the act of measurement (Feldman &
for establishing causality but also obscure the mean- Lynch, 1988), which may invite the conclusion that a
ing of these constructs because their interpretation
measure precedes its construct. However, studies of
depends on the dependent variables included in a
these effects entail a temporal sequence in which pre-
given model.
sentation of an item is followed by an attitude, which
in turn is followed by an emitted score. Although the
Temporal Precedence presentation of the item precedes the formation or
change of the construct, the construct precedes its
The third condition for causality is temporal prece- associated score. This reasoning draws from our defi-
dence, or whether change in the construct precedes, nition of a measure as the score itself, not the device
accompanies, or follows change in the measure. Be- used to gather the score.
cause researchers have direct access to measures but Some researchers have used mental experiments to
not constructs, temporal precedence cannot be as- argue that formative measures precede their con-
sessed directly. However, it may be established by structs. For example, Heise (1972) argued that the
using experiments that control the timing and se- construct SES is caused by measures of education,
quence of a construct and its measures. To create a income, and occupational prestige on the basis of
sequence in which the construct precedes the mea- the premise that changes in these socioeconomic vari-
sure, researchers may apply experimental manipula- ables lead to changes in SES, but not the reverse.
tions known to influence the construct and collect Although this logic is appealing, it relies on a
scores after the manipulation. For example, studies of subtle form of operationalism in which measures of
perceived control have shown that giving participants education, income, and occupational prestige are
a switch that terminates electric shock increases per- equated with the socioeconomic phenomena they are
ceived control (Averill, 1973). If this manipulation is intended to represent. Thus, a score denoting years
followed by a self-report measure of control, one may of schooling is treated as education itself, not as an
conclude that variance in the perceived-control con- indicator of education containing errors due to im-
struct preceded variance in scores on the measure of perfect recall, coding mistakes, and so forth. If we
control. To create a sequence in which the measure assume that scores on education, income, and occu-
precedes the construct, scores may be manipulated by pational prestige contain measurement error and
providing bogus feedback to subjects, under the as- occur after the phenomena they represent (e.g., re-
sumption that variation in these scores will create porting one's years of schooling logically occurs
variation in a latent attitude or belief construct (Bin- after the last year of schooling has been completed),
ning, Zaba, & Whattam, 1986; Phillips & Lord, then these scores may be viewed as reflective mea-
1982). However, to verify that the construct has in- sures of socioeconomic constructs that in turn cause
deed changed, a valid reflective measure of the con- SES. Following this logic, it is these socioeconomic
struct must follow. Both of these experimental proce- constructs, not their measures, that cause SES
dures entail a stimulus ~ organism --~ response (Blalock, 1971).
160 EDWARDS AND BAGOZZI

Eliminating Rival Causal Explanations the presumed effect of job performance on ratings.
This effect is a special case of method variance as a
The fourth condition for causality is the elimination
nuisance factor in observed scores (Bagozzi & Yi,
of rival explanations for the presumed causal relation-
1990; Williams & Brown, 1994). Additional threats to
ship between a construct and a measure. This condi-
validity (Cook & Campbell, 1979) may be considered
tion is perhaps the most difficult to satisfy because of
to assess other rival explanations for relationships be-
the myriad third variables that may induce a relation-
tween constructs and measures.
ship between a presumed cause and effect (Cook &
A related approach to identifying rival causal ex-
Campbell, 1979). Moreover, rival causal explanations
planations for constructs and measures is to apply
must be examined with regard to specific constructs
and measures (Hauser, 1972). Hence, ruling out rival mental experiments, as discussed earlier. For ex-
causal explanations is a daunting task that cannot be ample, overall job satisfaction is often measured using
reduced to universal prescriptions. Thus, our objective items that describe satisfaction with specific job facets
(e.g., Warr, Cook, & Wall, 1979; Weiss, Dawis, En-
here is to suggest general procedures for identifying
gland, & Lofquist, 1967). Scores on these items are
rival explanations for a presumed causal relationship
between a construct and a measure and to identify then taken as direct reflective measures of overall job
broad categories of rival explanations that may apply satisfaction, as implied by the use of statistical pro-
in a variety of contexts. cedures (e.g., reliability estimation) consistent with
One approach to identifying rival causal explana- reflective measurement (Bollen & Lennox, 1991;
tions for constructs and measures is to consider threats Nunnally, 1978). However, the presumed effect of the
overall job satisfaction construct on facet satisfaction
to validity relevant to quasi-experimental research
measures may be spurious, attributable to facet satis-
(Cook & Campbell, 1979). One threat to internal va-
faction constructs. This rival explanation is based on
lidity is history, in which the relationship between the
presumed cause and effect is due to some intervening two plausible assertions: (a) Facet satisfaction mea-
event that is not part of the experimental treatment. sures are directly influenced by facet satisfaction con-
Applied to constructs and measures, history entails structs, not overall job satisfaction (Ironson, Smith,
some event that occurs between the construct and the Brannick, Gibson, & Paul, 1989), and (b) facet satis-
measure that may explain their relationship. This faction constructs affect the overall job satisfaction
threat to validity was illustrated by Brief, Butcher, and construct, meaning that one's overall attitude toward a
job arises from attitudes toward specific aspects of the
Roberson (1995), who examined job satisfaction in
job (Aldag & Brief, 1978; Ferratt, 1981). This logic
hospital employees. In general, one might assume that
variation in the job satisfaction construct would gen- can be applied to other general constructs operation-
erate variation in job satisfaction scores, indicating alized using measures of specific aspects of those con-
reflective measurement. However, Brief et al. showed structs.
that positive mood-inducing events at the time of
measurement (i.e., giving respondents small gifts, Models o f the Relationship Between Constructs
such as cookies) influenced job satisfaction scores. and Measures
These events are a historical threat to validity because The preceding discussion outlined general prin-
they occurred between the construct and the measure ciples for determining the causal direction between a
(assuming employees had some notion of their job construct and a measure. In this section, we build on
satisfaction before the study) and influenced the mea- those principles by developing formal models of the
sure irrespective of the job satisfaction construct. 3 causal structure relating constructs to measures. His-
Another threat to validity that is relevant to con-
structs and measures is instrumentation, in which the
relationship between a presumed cause and effect is
3 A subtle difference between Brief et al.'s (1995) study
due in part to variation in the instrument or method
and history as described by Cook and Campbell (1979) is
used to collect data (Cook & Campbell, 1979). For
that the positive mood-inducing events were not con-
example, job performance scores from a set of raters founded with some other treatment intended to influence job
may vary not only because job performance varies satisfaction. Nonetheless, these events aptly demonstrate
across ratees but also because raters use different rat- how conditions unrelated to the construct of interest can
ing procedures (Coovert, Craiger, & Teachout, 1997). intercede the construct and its measure and lead to errone-
Thus, rater variance provides a rival explanation for ous inferences regarding their relationship.
RELATIONSHIPS BETWEEN CONSTRUCTS AND MEASURES 161

torically, models relating constructs to measures have


focused on direct effects (Blalock, 1964; Bollen &
Lennox, 1991; Costner, 1969; MacCallum & Browne,
1993). We extend these models by applying rules for
decomposing correlations in path analysis and struc- 52
tural equation modeling (Alwin & Hauser, 1975; Fox,
1980). These rules stipulate that the correlation be-
tween two variables can be decomposed into four 83
components: (a) a direct effect, in which one variable
directly affects another; (b) an indirect effect, in Figure 1. Direct reflective model.
which the effect of one variable on another is medi-
ated by (i.e., transmitted through) one or more other focal construct or its effects (Bollen & Davis, 1994;
variables; (c) a spurious component that is due to MacCallum & Browne, 1993). However, these modi-
common or correlated causes; and (d) an unanalyzed fications should not alter the specified causal flow
component resulting from associations among prede- between the construct and its measures. If this causal
termined (i.e., exogenous) variables. Direct and indi- flow is altered, the result is a model that is empirically
rect effects can operate in either direction, depending testable but incorrectly represents the causal relation-
on which variable is specified as the cause of the ships of interest. The usefulness of testing such a
other. Applying these distinctions to the relationship model is dubious at best.
between a construct and a measure yields six basic
categories of models, shown later in Figures 1-6. Direct Reflective M o d e l
To simplify our presentation of these models, we The direct reflective model specifies direct effects
made the following general assumptions and restric- from a construct to its measures. This model is de-
tions: (a) Relationships in the models are linear, (b) picted in Figure 1, in which each of the x i measures is
constructs and measures are expressed as deviations influenced by the construct ~ and the random mea-
from their means, (c) measurement errors and distur- surement error ~i. Hence, variance in each measure is
bance terms are random and are therefore uncorre- explained by a construct common to all measures and
lated with one another and with predictors in the equa- error unique to each measure, and covariation among
t i o n that c o n t a i n s the m e a s u r e m e n t e r r o r or the measures is attributed to their common cause, ~.
disturbance term, (d) each construct has three mea- The direct reflective model is expressed by the fol-
sures, (e) models with indirect effects contain a single lowing equation:
stage o f mediation, (f) causal flow between the con-
struct and the measure is unidirectional (i.e., recur- xi = Xi~ + ~i. (1)
sive), and (g) measures cannot directly cause other
In Equation 1, the h i are factor loadings depicting the
measures (i.e., scores are inert quantitative symbols
magnitude of the effect of ~ on the x i, and 8 i represents
and therefore cannot directly influence one other). 4
random measurement error (i ranges from 1 to 3 for
We later suggest how the simplified models consid-
the model in Figure 1). Equation 1 has its roots in
ered here can be combined to represent more complex
classical test theory (Lord & Novick, 1968; Nunnally,
causal mechanisms linking constructs to measures.
The notation for all models follows conventions from
the structural equation modeling literature (e.g.,
Jtreskog & Strbom, 1996). 4 Conceivably, models could be specified in which mea-
Our discussion of these models does not address sures were the causes of other measures. For example, in a
issues of identification, and some of the models we model specifying an indirect effect of a construct on a mea-
sure, the mediating variable could, in principle, be either a
discuss are not identified as presented. However, our
construct or a measure. However, the relationship of one
goal here is to derive models that are logically pos-
measure with another is generally limited to deterministic
sible, independent of whether they are identified. This functional relationships, as when one score is mathemati-
goal reflects the premise that correct model specifi- cally transformed into another score (e.g., logarithm, square
cation should take priority over identification. Un- root) or multiple scores are summed to create scales or item
identified models can become identified by constrain- parcels. These relationships are not causal; rather, they sim-
ing parameters or by adding reflective measures of the ply entail mathematical operations on inert data.
162 EDWARDS AND BAGOZZI

1978) and underlies reliability estimation, common exposure to discrimination as measured by race, age,
factor analysis, and confirmatory factor analysis sex, and disabilities (Bollen & Lennox, 1991).
(Bollen, 1989; Harman, 1976; Kim & Mueller, 1978; In some cases, the equation relating an induced
Long, 1983; Nunnally, 1978). latent variable to its measures excludes the distur-
bance term 4, such that the latent variable is a
Direct Formative Model weighted linear function of its measures. This use of
The direct formative model specifies measures as the direct formative model underlies principal-
correlated causes of a construct. This model is de- components analysis (Kim & Mueller, 1978), canoni-
cal correlation analysis (Thompson, 1984), and partial
picted in Figure 2, which shows effects from the x i
least squares (Garthwaite, 1994; Wold, 1982), each of
measures to the construct "q. The disturbance term
which uses observed measures to create weighted lin-
represents that part of the construct "q that is not ex-
ear composites that serve as conceptual variables in
plained by the x i measures and thus may be inter-
preted as measurement error. In contrast, the x i are subsequent analyses. By omitting the disturbance
conceived as error-free causes of "q (MacCallum & term, measurement error in these composites is effec-
tively ignored.
Browne, 1993). Because causality runs from the mea-
sures to the construct, the construct -q is not presumed A second use of the direct formative model is to
to explain the variances of the x i or the covariances create block variables that summarize the effects of
among the x i. The direct formative model is repre- several variables (Coleman, 1976; Heise, 1972; Igra,
sented by the following equation: 1979; Marsden, 1982). A block variable yields a
single summary estimate of the effects of variables in
"I1 ~- ~i yiXi d- 4, (2) the block on some outcome. Variables in the block
often represent conceptually distinct causes of the out-
where the Yi are structural parameters depicting the come. For example, in a model predicting political
magnitudes of the effects of the xi on "q, and ~ rep- liberalism, Heise (1972) introduced a block variable
resents the summation of the yixi products (i ranges labeled f a m i l y socialization, which itself was a func-
from 1 to 3 for the model in Figure 2). tion of mother's liberalism, father's liberalism, and
Because the direct formative model may be unfa- other unspecified variables captured by the distur-
miliar to the reader, we describe several of its com- bance term 4. The overall effect of family socializa-
mon uses. One use of the model is to create an "in- tion on political liberalism was summarized with a
duced" latent variable that represents an aggregation "sheaf coefficient," which was calculated from coef-
of observed variables (Heise, 1972). An oft-cited ex- ficients obtained by regressing political liberalism on
ample is SES, which is typically viewed as a function mother's and father's liberalism.
of background variables such as income, education, Third, the direct formative model has been used to
and occupational prestige (Bollen & Lennox, 1991; depict the effects of an experimental manipulation on
Heise, 1972). Other induced latent variables include a latent variable (Alwin & Tessler, 1974; Bagozzi,
stress as measured by the experience of major life 1977; Blalock, 1971; Costner, 1971). For example,
events (Cohen, Cohen, Teresi, Marchi, & Velez, Costner (1971) described an experiment in which a
1990); social support as measured by number of sup- construct signifying fatigue was manipulated by with-
portive incidents (MacCallum & Browne, 1993); and holding sleep from participants. Here, degree of sleep
deprivation was taken as a measure of fatigue. If an
experiment includes measures designated as manipu-
lation checks, these measures may be viewed as re-
flective indicators of the manipulated construct
(Bagozzi, 1977; Costner, 1971).

Indirect Reflective Model

The preceding models specified relationships be-


tween a construct and its measures as direct. We now
Figure 2. Direct formative model (for simplicity, qb.. la- consider two models that capture indirect effects. One
bels on covariances among exogenous variables are omit- is the indirect reflective model (see Figure 3). This
ted). model shows that the effects of the construct ~ on its
RELATIONSHIPS BETWEEN CONSTRUCTS AND MEASURES 163

A equations that capture both versions of the indirect


reflective model can be written as follows:

= + (3)
~2 and

Yi = Xij'q* + el. (4)


1:31- ~3
In Equation 3, ~j indicates the effect of ~ on the "qT,
and in Equation 4, hij captures the effects of the "q* on
the Yi. For both models shown in Figure 3, i ranges
from 1 to 3; for the model in Figure 3A, j = 1 (and
B is therefore disregarded; see footnote 5), whereas for
the model in Figure 3B,j ranges from 1 to 3. Note that
Equation 4 accommodates cross-loadings, meaning
that each Yi may load on each "q*. However, the model
in Figure 3B depicts a special case of Equation 4 in
which each Yi loads on a single -q*, meaning that each
Yi represents its own mediating constructs. Equations
~22~"~2~ ~2 3 and 4 correspond to a second-order factor model in
which the Yi load on the first-order factors Xl*, which
in turn load on the second-order factor ~ (Rindskopf &
Rose, 1988). However, unlike most applications of
~33~'1~31~ ~3 second-order factor analysis, our emphasis here is on
cases in which Yi are mistaken as indicators of the
Figure 3. Indirect reflective model. A: Single mediating construct 6.
construct. B: Multiple mediating constructs.
An important insight regarding the relationship be-
tween ~ and the yi can be obtained by substituting the
expression for "q* from Equation 3 into Equation 4:
Yi measures are mediated by one or more latent vari-
ables, "q*. We use an asterisk to indicate that the -q* Yi = )kij('~j'~ + ~t) at- ei
are not the construct intended to be measured by the = )kij'~j6 -Jr )kij~t + ~i" (5)
Yi, which instead is labeled 6. The indirect reflective
model applies when measures are interpreted as indi- Equation 5 shows that the relationships between the Yi
cators of a construct but actually represent one or and the construct they are intended to measure (i.e., 6)
more of its effects. For example, job satisfaction has are represented by the products Xij'Y:. Therefore, any
been measured with items describing turnover intent estimate of the relationships between ~ and the Yi nec-
(e.g., "I frequently think of quitting my job"; Hack-
man & Oldham, 1980). However, theories of job sat-
isfaction and turnover indicate that turnover intent is 5 For simplicity, when a model contains a single variable
a consequence of job dissatisfaction, not the affective of a particular type, we omit the subscript for that variable.
experience of dissatisfaction itself (Locke, 1976; Mo- For example, because the model in Figure 3A contains a
bley, 1982). single -q*, we omit thej subscript in our discussion of'q* for
Figure 3 shows two versions of this model. The this model, and we do not include a subscript for this vari-
first incorporates a single mediating construct, such able in Figure 3A (the implied subscript of 1 is superfluous,
given that there is no need to index this variable). By the
that ~ influences a single "q*, which in turn affects
same token, parameters linked to a single variable of a
multiple yi .5 The second version contains multiple
particular type do not include a subscript in reference to that
mediating constructs (i.e., -q*), each of which affects variable. Thus, the k parameters in Figure 3A contain sub-
a single Yi. Note that, for both versions of the model, scripts for the Yi but not for the single -q*. When a parameter
the "q* or "q* are also affected by disturbance term ~* links two variables that do not require subscripts, the pa-
or ~*, respectively, and the Yi are influenced by mea- rameter likewise has no subscripts (as exemplified by the ~/
surement errors designated as e i. A general set of linking ~ and xl* in Figure 3A).
164 EDWARDS AND BAGOZZI

essarily confounds the relationships between ~ and the measures of depression could be recast as an indirect
"q* with the relationships between the -q* and the Yi. formative model in which supportive incidents cre-
Equation 5 also shows that variance in the y~ is attrib- ated an induced social support variable, which in turn
utable not only to the construct ~ and the measurement influenced depression. We consider two versions of
errors e~ but also to the disturbance terms ~*. The the indirect formative model: one in which the x i in-
effects of the ~* on the Yi represent additional sources fluence a single -q*, which in turn affects "q, and an-
of measurement error in the Yi, given that the ~* rep- other in which the x i influence multiple -q* (in this
resent that portion of the variance in the "q* not ex- example, three), each of which affects the "q. Note that
plained by ~ and hence signify error. 6 the lq*, -q*, and -q are also influenced by correspond-
ing disturbance terms ~*, ~*, and ~. Both versions of
Indirect Formative Model
this model can be represented by the following two
A second model that incorporates indirect effects equations:
between constructs and measures is the indirect for-
mative model, shown in Figure 4. Here, the x i have a
"1]7= E~fjiXi + ~ (6)
i
direct impact on one or more "q*, which in turn influ-
ence the construct of interest, "q. This model applies ~1 = E13j'q* + ~" (7)
when formative measures assigned to a construct ac- J

tually represent one of its causes. For example, Mac- In Equation 6, the ~/ji depict the magnitudes of the
Callum and Browne (1993, p. 538) showed that a effects of the x i on the -q*, and E is the summation of
model using supportive incidents as direct formative the "Yjixi products over i (i = 1 to 3 for both examples
shown in Figure 4). In Equation 7, the 13i indicate the
magnitudes of the effects of the "q* on "q, and ~ is the
A
summation of the 13fq* products o v e r j (j = 1 for the
model in Figure 4A and is therefore omitted; j = 1 to
3 for the model in Figure 4B). The relationships be-
tween the construct of interest, "q, and its measures x i
may be further clarified by substituting the expression
for "q* in Equation 6 into Equation 7:

= EE13j j, x i .
j i j
(8)
Equation 8 reveals that the relationship between the xi
and the intended construct, 0, is represented by the
summation of the 13fYji across all ~1". Thus, an esti-
B mate of the effect of xi on ~1 confounds the effects of
x i on the ~1" with the effects of the 11" on O. In addi-
tion, variance in ~1 is attributable not only to its mea-
sures x~ and the disturbance ~ but also to the distur-
b a n c e s ~*. Thus, m e a s u r e m e n t error in ~1 is
represented by ~ as well as the products 13j~* associ-
ated with each of the ~1".

6 The interpretation of the 4" as error is based on the


assumption that the Yl are intended to measure ~ rather than
Indirect formative model (for simplicity, qb.. la-
F i g u r e 4. the "q*. If yl were intended to measure "q*, measurement
bels on covariances among exogenous variables are omit- error would be represented solely by the e i and the 4" would
ted). A: Single mediating construct. B: Multiple mediating be interpreted simply as that part of the true score variance
constructs. in the -q* not explained by ~.
RELATIONSHIPS BETWEEN CONSTRUCTS AND MEASURES 165

Spurious Model on "q. Assuming the 8; and ~ are independent, a gen-


eral expression for the covariance of the construct -q
The preceding models were based on the premise
with its ith measure x i can be written as
that, either directly or indirectly, the construct caused
its measures or the measures caused the construct. An Or(Xi, "I]) = E E h i j f ~ j k ~ t k . (11)
alternative model depicts the relationships between j k
the construct and measures as spurious, due to the
In Equation 11, i indicates the measure under consid-
influence of one or more common causes (Blalock,
eration, j indexes the linkage between the measure
1971; Costner, 1969). Figure 5 displays two versions
and a particular ~*, k indexes the linkage between a
of this model. In the first version, covariances be-
particular ~* and "q, and ~bjkrefers to the variances and
tween the construct of interest, "q, and its intended
covariances of ~* and 6~' (separate subscripts for ~*
measures, x i are due to a single common cause, ~*.
are required because some elements of the covariance
The disturbance g indicates that the relationship be-
between x i and "q involve different 6*). Equation 11
tween 6" and "q is imperfect, such that some portion of
allows for cross-loadings, meaning that each of the
the variance in "q is due to forces other than 6*. In the
X i measures can be assigned to each of the ~* con-
second version, covariation between "q and each x i is
structs. If each item is assigned to a single ~* con-
due to a separate common cause, 6" (J ranges from 1
struct, as in the model shown in Figure 5B, Equation
to 3). Equations corresponding to both versions of this
11 simplifies to
model can be written as follows:

X i = )kij6~ "4" ~i (9) o) = (12)


I
k
"q = ~ / j ~ * + 4. (10)
J In Equation 12, i again indicates the measure in ques-
tion, j represents the construct to which the measure is
In Equation 9, hq designates the effects of the ~* on assigned, and k indexes each of the 6" constructs. For
the x i, and in Equation 10, "yj indicates the effect of ~* illustration, we apply this equation to the covariance
between x~ and "q in Figure 5B, which yields the fol-
A lowing:

O'(Xl, 3]) = ~kll(l)ll~l + ~kll~b12~2 + )kll~b13~3. (13)


1
The first term in Equation 13 represents that part of
82 the covariance between x~ and "q that is due to their
common cause, ~*, whereas the latter two terms rep-
resent parts of the covariance between xl and -q that is
53 due to correlated causes (i.e., ~* and ~', ~* and 6~',
respectively). This expression shows that the relation-
ship between x~ and "q confounds the relationship of
x 1 with ~*, the relationship of each ~ ' with "q, and the
correlations among the ~'.
If a single common cause 6" is involved (as in
B Figure 5A), Equation 11 further simplifies to

O.(Xi, oq) : kifJ)~t. (14)


1
Hence, the covariance between x I and -q in Figure 5A
52 is simply hld~. This expression shows that the rela-
tionship between x~ and ~q confounds the relationship
of x I with 6" with the relationship between 6" and r I.
53 A noteworthy special case of the spurious model
with multiple common causes (Figure 5B) treats the x i
Spurious model (for simplicity, ~b.. labels on
F i g u r e 5. as error-free measures of the 6", such that the h 0 are
covariances among exogenous variables are omitted). A: unity and the variances of the 8i are zero. In this case,
Single common cause. B: Multiple common causes. the x; and 6" are equivalent, and Equation 9 reduces to
166 EDWARDS AND BAGOZZI

A B

C D

Figure 6. Unanalyzed model (for simplicity, ~b..labels on covariances among exogenous variables are omitted). A: Constructs
and measures exogenous. B: Exogenous measures with correlated construct. C: Exogenous measures with correlated measures.
D: Exogenous construct with correlated construct.

x i = ~*. Substituting x i for ~j* in Equation 10 (and (b) the measures are exogenous and correlated with a
modifying the subscript on ~j to correspond to the Xi) construct that causes the focal construct, (c) the mea-
yields sures are exogenous and correlated with other mea-
sures that cause the construct, or (d) the focal con-
"q = E ~ l i X i d- ~. (15) struct is exogenous and correlated with another
i
construct that causes the measures. 7 These distinc-
Equation 15 is identical to the direct formative model tions yield four versions of the unanalyzed model,
as specified in Equation 2. Hence, the direct formative shown in Figure 6. For each model, we presume the
model may be considered a special case of the spuri- researcher does not realize that the unanalyzed mea-
ous indicator model in which the x i are viewed as sures are indicators of the focal construct.
perfect (i.e., error-free) indicators of the ~*.

Unanalyzed Model
7 For reasons stated earlier, we excluded models in which
Finally, the relationships between a construct and measures cause measures (for the unanalyzed case, this
its measures can be unanalyzed. This model applies would entail a model in which a construct is correlated with
when (a) the construct and measures are exogenous, a set of measures, which in turn cause other measures).
RELATIONSHIPS BETWEEN CONSTRUCTS AND MEASURES 167

The model in Figure 6A specifies that the construct (r(xi, "11): ~d(bih'Yh. (17)
and its measures are exogenous, meaning their rela- h
tionships are explained by forces outside the model.
In Equation 17, x i represents the measure of interest,
This model applies when the researcher is unwilling
and h indexes the covariance between x i and each x~
or unable to specify the relationship between a con-
and the effect of each x~' on "q. This equation shows
struct and its measures. Hence, the relationships be-
that the covariance between the x i and "q confounds
tween the construct and its measures are represented
the covariance between the x i and each x~' and the
by the covariances ~bi~ (here, the subscripts i and
effect of each x~' on "q.
identify the covariance between each x i and ~, thereby
Finally, the model in Figure 6D indicates that the
disregarding the covariances among the xi). Because
unanalyzed x i are considered correlational measures
all variables are exogenous, the model does not at-
of ~ but are actually reflective measures of ~*, which
tempt to explain variance in the x i or ~. Thus, the
signifies a construct correlated with ~. This model
model cannot distinguish true score variance from
applies when a measure reflects a construct that is
measurement error in the x i, nor can the model iden-
correlated with, but conceptually distinct from, the
tify residual variance in ~.
construct of interest. The covariance between each
The model in Figure 6B indicates that the unana-
measure and the intended construct can be written as
lyzed xi measures are correlated with a construct, ~*,
follows:
that causes the construct of interest, "q. Here, the x i are
mistakenly interpreted as correlational measures of "q, O'(Xi, ~) = ~.~ki, (18)
when in fact they represent correlational measures of
~*, which in turn causes "q. The covariance between where ~b~. is the covariance of ~ and ~*, and h i is the
each measure and the intended construct can be ex- loading of xi on ~*. Thus, the relationship of ~ with a
pressed as given x i confounds the covariance between ~ and ~*
and the loading from ~* to x i.
cr(x i, "q) = ",td~i¢., (16) Summary and Extensions

where ~/signifies the effect of ~* on "q, and ~bi~. rep- The preceding models are summarized in Table 1,
resents the covariance between the ith measure and which classifies the models according to whether (a)
the construct ~*. Hence, the covariance between the xi the measure refers to the construct itself or to a cause
and "q confounds the covariance between the x i and ~* or an effect of the construct and (b) the measure is
with the effect of ~* on "q. reflective or formative with respect to the construct to
According to the model in Figure 6C, the x i are which the measure is directly linked (i.e., a causal
viewed as correlational measures of "q but in fact are construct, the construct of interest, or an effect con-
correlated with a second set of measures, x* (where h struct, moving left to right across Table 1). As a gen-
ranges from 1 to 3), which in turn are formative mea- eral rule, if a measure describes the inherent attributes
sures of "q. The covariance between each xi measure of a construct, the relationship between the construct
and the intended construct is represented as follows: and the measure should be considered direct, whereas

Table 1
Classification o f M o d e l s Relating Constructs to Measures
Referent of measure
Inherent attributes
Type of measure Cause of construct of construct Effect of construct
Reflective Spurious model Direct reflective Indirect reflective
model model
Formative Indirect formative Direct formative Unanalyzed model
model model
Note. The columns indicate the referent of the measure, meaning whether the proximal construct (i.e.,
the construct to which the measure directly refers) is a cause of the construct of interest, the construct of
interest itself, or an effect of the construct of interest. The rows indicate whether the measure is reflective
or formative, meaning whether causality runs from or toward the proximal construct, respectively.
168 EDWARDS AND BAGOZZI

if the measure refers to a cause or an effect of the event measures can be viewed as causes of life stress,
construct, their relationship is indirect, spurious, or corresponding to the direct formative model.
unanalyzed. The preceding controversy can be resolved by ap-
Naturally, the models in Table 1 do not exhaust all plying principles developed in this article to the SRRS
possible causal structures. For example, the indirect (see Table 2). Specifically, the life stress construct
reflective and indirect formative models can be elabo- (i.e., change in life pattern) and the SRRS are distinct
rated by including additional mediators of the rela- entities, given that life change is distinct from ratings
tionship between the construct and its measures. The of events that may prompt change. Association is
models can also be combined to form hybrid models. plausible, although change in life pattern may occur
For instance, the direct reflective and direct formative for reasons other than the 43 events that constitute the
models can be combined to depict reciprocal causa- SRRS. Temporal precedence is likewise difficult to
tion between a construct and its measures (Bollen, establish, given that SRRS scores should logically fol-
1989) or to incorporate both reflective and formative low the occurrence of their corresponding events but
measures of a single construct (JOreskog & Gold- can follow or precede the change in life pattern these
berger, 1975). Thus, the models in Table 1 represent events presumably create. These ambiguities are
conceptual building blocks that can be used to expli- symptomatic of a rival causal explanation for the re-
cate the relationship between constructs and mea- lationship between life change and the SRRS in which
sures. the occurrence of individual life events causes change
in life pattern and produces scores on the SRRS. The
Illustrative Applications association between event occurrence and the SRRS
scores is imperfect because of reporting biases and
The preceding discussion developed principles for errors in recall (Raphael, Cloitre, & Dohrenwend,
specifying the direction and structure of relationships 1991). On the basis of this reasoning, the relationship
between constructs and measures. For expository pur- between life change and the SRRS would follow the
poses, we applied selected aspects of these principles spurious model with multiple common causes (Figure
to various constructs and measures (e.g., SES, in- 5B), in which the common causes are individual life
volvement, satisfaction, performance). We now more events.
fully apply these principles to constructs and mea-
sures characteristic of psychological, sociological, Organizational Commitment
and organizational research.
Mowday, Steers, and Porter (1979) defined organi-
Life Stress zational commitment as "the relative strength of an
individual's identification with and involvement in a
Following Selye (1956), stress researchers often particular organization" (p. 226). To measure this
conceptualize life stress as the degree of change re- construct, Mowday et al. developed the 15-item Or-
quired by major life events (Dohrenwend & Dohren- ganizational Commitment Questionnaire (OCQ). Al-
wend, 1981). The most widely used measure of life though Mowday et al. stated that commitment may
events is the 43-item Social Readjustment Rating consist of multiple factors, they evaluated the OCQ
Scale (SRRS), developed by Holmes and Rahe (1967) using internal consistency reliability estimation and
to measure "the amount and duration of change in common factor analysis and concluded that the OCQ
one's accustomed pattern of life" (p. 213). Studies of represents "a single common underlying construct"
the factor structure and the internal consistency reli- (p. 232). Subsequent studies have treated the OCQ as
ability of the SRRS (e.g., Kipper & Furcon, 1981; Lei a unidimensional measure and have routinely reported
& Skinner, 1980) imply that the relationship between internal consistency reliability estimates (Mathieu &
the SRRS and life stress follows the direct reflective Zajac, 1990). Hence, the development and application
model. However, Turner and Wheaton (1995) argued of the OCQ have implicitly followed the direct reflec-
that life event measures such as the SRRS should not tive model.
exhibit internal consistency because items in these Closer scrutiny of the OCQ suggests that the direct
measures are not alternative indicators of a single un- reflective model may be inappropriate (see Table 2).
derlying construct but instead describe distinct events For illustration, we focus on OCQ Items 5 ("I find that
a person might experience. On the basis of similar my values and the organization's values are very simi-
reasoning, Cohen et al. (1990) suggested that life lar"), 14 ("For me this is the best of all possible or-
RELATIONSHIPS BETWEEN CONSTRUCTS AND MEASURES 169

i
r..) N

-!
E
c,
E E r..)

i
o .=
II
°~ CY
0
©
b~

8
II
o o

©
170 EDWARDS AND BAGOZZI

ganizations for which to work"), and 2 ("I talk up this measure suggests a different model. Scores on items
organization to my friends as a great place to work describing social interaction are distinct from those
for"). Scores on these items are distinct from the com- interactions, and higher scores should be associated
mitment construct, which appears to represent an at- with the increased occurrence of interactions. How-
titude within the mind of the respondent (assuming ever, social interaction scores should follow the inter-
"involvement" refers to the importance ascribed to the actions they describe, given that the scores reflect
organization; cf. Kanungo, 1982). The commitment events that have already occurred. This temporal or-
construct is likely to precede the OCQ item scores, dering is the opposite of that implied by the direct
assuming respondents have some notion of their com- formative model and indicates a rival causal explana-
mitment before the OCQ is administered. Association tion in which constructs signifying forms of social
is plausible in that higher commitment should logi- interaction generate scores on their corresponding
cally be associated with higher scores on the three items. Inspection of Doney and Cannon's items sug-
illustrative OCQ items. However, the relationship be- gests that these constructs may include informal con-
tween the commitment construct and the OCQ scores versations ("talk about family, sports or other per-
may follow a causal structure more complex than the sonal interests," "talk about common interests beside
direct reflective model. In particular, Item 5 may be work") and informal meetings (i.e., "meet away from
viewed as a direct reflective measure of a cause of the workplace," "meet over breakfast, lunch or din-
commitment, in that congruence between personal ner," "attend entertainment events [sports, theater,
and organizational values may create a sense of iden- etc.]," "get together primarily to have fun," "get to-
tification with the organization (Meglino, Ravlin, & gether with other family members"). Effects of these
Adkins, 1989). Thus, the relationship between Item 5 constructs on their measures are presumably direct
and commitment may follow the spurious model (Fig- because the items describe the inherent attributes of
ure 5A). Conversely, Items 14 and 2 may be consid- their constructs. This reasoning yields two direct re-
ered direct reflective measures of evaluative and be- flective models, one for informal conversations and
havioral consequences of commitment, respectively, another for informal meetings. The informal-meetings
in that identification with an organization should cre- construct can be subdivided into constructs that de-
ate positive appraisals of that organization and prompt scribe specific types of meetings, thereby reflecting
the person to speak positively about the organization. the distinctions among the items (e.g., an entertain-
Thus, the relationship of commitment with Items 14 ment event may not entail a meal or contact with
and 2 follows the indirect reflective model. Combin- family members). If this degree of specificity is not
ing these models indicates that the relationship be- desired, items describing specific types of meetings
tween organizational commitment and OCQ Items 5, can be replaced with general items akin to "meet away
14, and 2 is captured not by the direct reflective model from the workplace," yielding additional direct reflec-
but rather by a combined spurious and indirect reflec- tive measures of informal meetings. Note that this
tive model. Similar reasoning can be applied to the respecified model does not entail a social interaction
remaining 12 OCQ items to develop a full model of construct, because this term merely serves as a label
their relationship with the commitment construct. for a category that encompasses different ways of in-
teracting informally. These forms of interaction can
Social Interaction be investigated jointly to examine the general phe-
nomenon of social interaction.
In a study of buyer-seller relationships, Doney and
Cannon (1997) conceptualized social interaction as
S u m m a r y and Implications
contact between a buyer and a seller in an informal
environment conducive to enhanced information This article has derived models that describe rela-
flow. To measure social interaction, Doney and Can- tionships between constructs and measures and has
non used seven items describing various forms of so- provided guidelines for choosing among the models
cial interaction (e.g., meeting away from the work- on the basis of principles for determining the causal
place, talking about common interests). Because the direction and structure of relationships between con-
forms of interaction differed, Doney and Cannon con- structs and measures. The models derived in this ar-
sidered this measure formative, thereby adopting the ticle go beyond simple direct effects specified in most
direct formative model. measurement models to include indirect, spurious,
Further evaluation of Doney and Cannon's (1997) and unanalyzed associations between constructs and
RELATIONSHIPS BETWEEN CONSTRUCTS AND MEASURES 171

measures. Guidelines for choosing among the models structural relationships between constructs. As
apply conditions for causality from the philosophy of pointed out long ago by Blalock (1971), "many so-
science literature, thereby supplementing the com- called 'background variables' such as race, sex, oc-
monsense assertions often used to specify causal di- cupation of father, religion, community, region, and
rection between constructs and measures. Throughout even 'age' are basically crude indicators of stimuli
the article, we have underscored the premise that the that are thought to influence behavior" (p. 337). Con-
relationship between a construct and its measures can- sequently, these background variables should be
not be specified without reference to the conceptual specified not as error-free formative measures but as
underpinnings of the construct and measures. Thus, imperfect reflective measures of constructs that cause
rather than generating narrow prescriptions, this ar- other constructs.
ticle has attempted to provide a way of thinking about Second, items that describe distinct facets of a gen-
linkages between constructs and measures that will eral construct are often misspecified as direct reflec-
shed light on the nature and direction of their rela- tive measures of that construct. This tendency is il-
tionship. lustrated by the m e a s u r e m e n t of overall job
We hope the guidelines offered in this article will satisfaction by using items that describe satisfaction
help researchers specify relationships between con- with job facets (Warr et al., 1979; Weiss et al., 1967).
structs and measures with clarity and precision, In general, facet measures should be respecified as
thereby enhancing the quality and rigor of auxiliary reflective measures of facet constructs, which in turn
theories linking constructs to measures. As auxiliary (a) influence a general construct, as in the relationship
theories are improved, researchers can better rely on between facet satisfaction and overall job satisfaction,
these theories to drive analytical decisions, as is or (b) are influenced by a superordinate construct, as
widely recommended for tests of substantive theories in second-order factor models. In the former case, the
(MacCallum, Roznowski, & Necowitz, 1992; Runkel relationship between the facet measures and the gen-
& McGrath, 1972). Moreover, stronger auxiliary eral construct would follow the spurious model shown
theories will help reduce misspecifications of relation- in Figure 5B. In the latter case, the relationship be-
ships between constructs and measures. Such mis- tween the facet measures and the superordinate con-
specification can lead to gross inaccuracies in tests of struct would follow the indirect reflective model shown
relationships among substantive constructs (MacCal- in Figure 3B. In neither case would facet measures serve
lum & Browne, 1993), inviting erroneous conclusions as direct reflective measures of the general construct.
and thwarting the accumulation of knowledge. Thus, Third, researchers should refrain from invoking the
by improving auxiliary theories linking constructs to direct formative model as an explanation for low re-
measures, tests of substantive theories relating con- liability estimates. Computerized searches for studies
structs to one another will likewise improve. citing two recent articles on formative measurement
On the basis of our experience with the guidelines (Bollen & Lennox, 1991; MacCallum & Browne,
developed in this article, we believe their application 1993) revealed several instances in which these ar-
will yield three general conclusions, each of which ticles were used to construct post hoc explanations for
has important substantive ramifications. First, many low reliability estimates (e.g., Adkins, 1995; Jarley,
instances of the direct formative model can be re- Fiorito, & Delaney, 1997; Judge & Bretz, 1994;
specified as the spurious model, as illustrated by our Thomas & Ravlin, 1995). As we have argued, the
discussion of the measurement of SES. Researchers specification of measures as formative or reflective
have noted that measures in the direct formative should be based on a priori conceptual criteria, not on
model are assumed to contain no measurement error post hoc empirical evidence, particularly when that
(Heise, 1972; MacCallum & Browne, 1993), an as- evidence consists of a low reliability estimate. Fur-
sumption that is untenable in most situations. This thermore, if measures are specified as formative, their
assumption can be relaxed by incorporating constructs validity must still be established. It is bad practice to
that each cause their own reflective measures and col- report a low reliability estimate, claim that one's mea-
lectively cause the construct of interest. To identify all sures are formative, and do nothing more.
parameters in this model, direct reflective measures
can be added to the construct of interest. This respeci- Conclusion
fled model would provide estimates of relationships This article has advanced guidelines for specifying
between each construct and its measures as well as the causal direction and the causal structure of the
172 EDWARDS AND BAGOZZI

relationship between constructs and measures. By ap- plaining the biasing effects of performance cues in terms
plying these guidelines, researchers can develop aux- of cognitive categorization. Academy of Management
iliary theories linking constructs to measures with the Journal, 29, 521-535.
rigor and clarity used to specify relationships among Blalock, H. M. (1964). Causal inferences in nonexperimen-
theoretical constructs. Thus, we advocate a holistic tal research. Chapel Hill: University of North Carolina
construal (Bagozzi & Phillips, 1982) in which auxil- Press.
iary and substantive theories are developed jointly to Blalock, H.M. (1971). Causal models involving unob-
enhance our understanding of relationships among served variables in stimulus-response situations. In H. M.
theoretical constructs and the mapping of these con- Blalock (Ed.), Causal models in the social sciences (pp.
structs onto the empirical world. 335-347). Chicago: Aldine.
Bollen, K. A. (1989). Structural equations with latent vari-
References ables. New York: Wiley.
Bollen, K. A., & Davis, W. R. 0994, August). Causal in-
Adkins, C. L. (1995). Previous work experience and orga- dicator models: Identification, estimation, and testing.
nizational socialization: A longitudinal examination. Paper presented at the 1993 American Sociological As-
Academy of Management Journal, 38, 839-862. sociation Convention, Miami, FL.
Aldag, R. J., & Brief, A. P. (1978). Examination of alterna- Bollen, K., & Lennox, R. (1991). Conventional wisdom on
tive models of job satisfaction. Human Relations, 31, measurement: A structural equation perspective. Psycho-
91-98. logical Bulletin, 110, 305-314.
Alwin, D. F., & Hanser, R. M. (1975). The decomposition
Bollen, K.A., & Ting, K. (1993). Confirmatory tetrad
of effects in path analysis. American Sociological Re-
analysis. In P. V. Marsden (Ed.), Sociological methodol-
view, 40, 37-47.
ogy 1993 (pp. 147-175). Washington, DC: American So-
Alwin, D. F., & Tessler, R. C. (1974). Causal models, un-
ciological Association.
observed variables, and experimental data. American
Brief, A. P., Butcher, A. H., & Roberson, L. (1995). Cook-
Journal of Sociology, 80, 58-86.
ies, disposition, and job attitudes: The effects of positive
Asher, H. B. (1983). Causal modeling. Newbury Park, CA:
mood-inducing events and negative affectivity on job sat-
Sage.
isfaction in a field experiment. Organizational Behavior
Averill, J. R. (1973). Personal control over aversive stimuli
and Human Decision Processes, 62, 55-62.
and its relationship to stress. Psychological Bulletin, 80,
Burt, R.S. (1976). Interpretational confounding of unob-
286-303.
served variables in structural equation models. Sociologi-
Bagozzi, R. P. (1977). Structural equation models in experi-
cal Methods and Research, 5, 3-52.
mental research. Journal of Marketing Research, 14,
209-226. Campbell, D.T. (1960). Recommendations for APA t6st
Bagozzi, R.P. (1980). Causal models in marketing. New standards regarding construct, trait, or discriminant va-
lidity. American Psychologist, 15, 546-553.
York: Wiley.
Bagozzi, R. P., & Fornell, C. (1982). Theoretical concepts, Cohen, P., Cohen, J., Teresi, J., Marchi, M., & Velez, C. N.
measurements, and meaning. In C. Fornell (Ed.), A sec- (1990). Problems in the measurement of latent variables
ond generation of multivariate analysis: Vol. 2. Measure- in structural equations causal models. Applied Psycho-
ment and evaluation (pp. 24-38). New York: Praeger. logical Measurement, 14, 183-196.
Bagozzi, R. P., & Phillips, L. W. (1982). Representing and Coleman, J. S. (1976). Regression analysis for the compari-
testing organizational theories: A holistic construal. Ad- son of school and home effects. Social Science Research,
ministrative Science Quarterly, 27, 459-489. 5, 1-20.
Bagozzi, R. P., & Yi, Y. (1990). Assessing method variance Cook, T. D., & Campbell, D. T. (1979). Quasi-experimenta-
in multitrait-multimethod matrices: The case of self- tion: Design and analysis issues for field settings. Bos-
reported affect and perceptions at work. Journal of Ap- ton: Houghton Mifflin.
plied Psychology, 75, 547-560. Coovert, M. D., Craiger, J.P., & Teachout, M. S. (1997).
Bedeian, A. G., Day, D. V., & Kelloway, E. K. (1997). Cor- Effectiveness of the direct product versus confirmatory
recting for measurement error attenuation in structural factor model for reflecting the structure of multimethod-
equation models: Some important reminders. Education- multirater job performance data. Journal of Applied Psy-
al and Psychological Measurement, 57, 785-799. chology, 82, 271-280.
Binning, J. F., Zaba, A.J., & Whattam, J. C. (1986). Ex- Costner, H. L. (1969). Theory, deduction, and the rules of
RELATIONSHIPS BETWEEN CONSTRUCTS AND MEASURES 173

correspondence. American Journal of Sociology, 75, ment Rating Scale. Journal of Psychosomatic Research,
245-263. 11, 213-218.
Costner, H. L. (1971). Utilizing causal models to discover Hume, D. (1946). Inquiries concerning the human under-
flaws in experiments. Sociometry, 34, 398-410. standing. London: Oxford University Press.
Cronbach, L. J., & Meehl, P. C. (1955). Construct validity Igra, A. (1979). On forming variable set composites to sum-
in psychological tests. Psychological Bulletin, 52, 281- marize a block recursive model. Social Science Research,
302. 8, 253-264.
Delanty, G. (1997). Social science: Beyond constructivism Ironson, G.H., Smith, P.C., Brannick, M.T., Gibson,
and realism. Minneapolis: University of Minnesota Press. W. M., & Paul, K. B. (1989). Construction of a Job in
DeVellis, R.F. (1991). Scale development: Theories and General scale: A comparison of global, composite, and
applications. Newbury Park, CA: Sage. specific measures. Journal of Applied Psychology, 74,
Dohrenwend, B. S., & Dohrenwend, B. P. (1981). Stressful 193-200.
life events and their contexts. New York: Neale Watson. James, L. R., Mutaik, S. A., & Brett, J. M. (1982). Causal
Doney, P. M., & Cannon, J. P. (1997). An examination of analysis: Assumptions, models and data. Beverly Hills,
the nature of trust in buyer-seller relationships. Journal CA: Sage.
of Marketing, 61, 35-51. Jarley, P., Fiorito, J., & Delaney, J. T. (1997). A structural
Duncan, O. D. (1975). Introduction to structural equation contingency approach to bureaucracy and democracy in
models. New York: Academic Press. U.S. national unions. Academy of Management Journal,
40, 831-861.
Feldman, J. M., & Lynch, J. G., Jr. (1988). Self-generated
Johns, G. (1994). How often were you absent? A review of
validity and other effects of measurement of belief, atti-
the use of self-reported absence data. Journal of Applied
tude, intention, and behavior. Journal of Applied Psy-
Psychology, 79, 574-591.
chology, 73, 421-435.
Jtreskog, K. G., & Goldberger, A. S. (1975). Estimation of
Ferratt, T. W. (1981 ). Overall job satisfaction: Is it a linear
a model with multiple indicators and multiple causes of a
function of facet satisfaction? Human Relations, 34, 463-
single latent variable. Journal of the American Statistical
473.
Association, 10, 631-639.
Fornell, C., & Bookstein, F. L. (1982). Two structural equa-
Jtreskog, K. G., & Strbom, D. (1996). LISREL 8: User's
tion models: LISREL and PLS applied to consumer exit-
reference guide. Chicago: Scientific Software Interna-
voice theory. Journal of Marketing Research, 19, 440-
tional.
452.
Judge, T. A., & Bretz, R. D. (1994). Political influence be-
Fox, J. (1980). Effect analysis in structural equation models. havior and career success. Journal of Management, 20,
Sociological Methods and Research, 9, 3-28. 43-65.
Garthwaite, P. H. (1994). An interpretation of partial least Kanungo, R. N. (1982). Measurement of job and work in-
squares. Journal of the American Statistical Association, volvement. Journal of Applied Psychology, 67, 341-349.
89, 122-127. Kim, J. O., & Mueller, C. W. (1978). Factor analysis. Bev-
Hackman, J. R., & Oldham, G.R. (1980). Work redesign. erly Hills, CA: Sage.
Reading, MA: Addison-Wesley. Kipper, D. A., & Furcon, J. (1981). Factors underlying the
Harman, H.H. (1976). Modern factor analysis (3rd ed.). Social Readjustment Rating Scale. Personality and Indi-
Chicago: University of Chicago Press. vidual Differences, 2, 31-36.
Hauser, R.M. (1972). Disaggregating a social-psycho- Lee, S., & Hershberger, S. (1990). A simple rule for gen-
logical model of educational attainment. Social Science erating equivalent models in covariance structure model-
Research, I, 159-188. ing. Multivariate Behavioral Research, 25, 313-334.
Hauser, R. M., & Goldberger, A. S. (1971). The treatment Lei, H., & Skinner, H. A. (1980). A psychometric study of
of unobservable variables in path analysis. In H. L. Cost- life events and social readjustment. Journal of Psychoso-
ner (Ed.), Sociological methodology 1971 (pp. 81-117). matic Research, 24, 57-65.
San Francisco: Jossey-Bass. Locke, E. A. (1976). The nature and causes of job satisfac-
Heise, D. R. (1972). Employing nominal variables, induced tion. In M. Dunnette (Ed.), Handbook of industrial and
variables, and block variables in path analysis. Sociologi- organizational psychology (pp. 1297-1350). Chicago:
cal Methods & Research, 1, 147-173. Rand McNally.
Heise, D. R. (1975). Causal analysis. New York: Wiley. Loevinger, J. (1957). Objective tests as instruments of psy-
Holmes, T. H., & Rahe, R. H. (1967). The Social Readjust- chological theory. Psychological Reports, 3, 635-694.
174 EDWARDS AND BAGOZZI

Long, J. S. (1983). Confirmatory factor analysis: A preface Rindskopf, D., & Rose, T. (1988). Some theory and appli-
to L1SREL. Beverly Hills, CA: Sage. cation of confirmatory second-order factor analysis. Mul-
Lord, F. M., & Novick, M. R. (1968). Statistical theories of tivariate Behavioral Research, 23, 51-67.
mental test scores. Reading, MA: Addison-Wesley. Runkel, P. J., & McGrath, J. E. (1972). Research on human
MacCallum, R. C., & Browne, M.W. (1993). The use of behavior: A systematic guide to method. New York: Holt.
causal indicators in covariance structure models: Some Schwab, D. P. (1980). Construct validity in organizational
practical issues. Psychological Bulletin, 114, 533-541. behavior. In L. L. Cummings & B. M. Staw (Eds.), Re-
MacCallum, R.C., Roznowski, M., & Necowitz, L.B. search in organizational behavior (Vol. 2, pp. 3-43).
(1992). Model modification in covariance structure Greenwich, CT: JAI Press.
analysis: The problem of capitalization on chance. Psy- Selye, H. (1956). The stress of life. New York: McGraw-
chological Bulletin, 111, 490-504. Hill.
MacCallum, R. C., Wegener, D. T., Uchino, B. N., & Fab- Suppes, P. (1970). A probabilistic theory of causation. Am-
rigar, L. R. (1993). The problem of equivalent models in sterdam: North-Holland.
applications of covariance structure analysis. Psychologi- Thomas, D. C., & Ravlin, E. C. (1995). Responses of em-
cal Bulletin, 114, 185-199. ployees to cultural adaptation by a foreign manager. Jour-
Marsden, P. V. (1982). A note on block variables in multi- nal of Applied Psychology, 80, 133-146.
equation models. Social Science Research, 11, 127-140. Thompson, B. (1984). Canonical correlation analysis: Uses
Mathieu, J. E., & Zajac, D. M. (1990). A review and meta- and interpretation. Newbury Park, CA: Sage.
analysis of the antecedents, correlates, and consequences Turner, R.J., & Wheaton, B. (1995). Checklist measure-
of organizational commitment. Psychological Bulletin, ment of stressful life events. In S. Cohen, R. C. Kessler,
108, 171-194. & L. U. Gordon (Eds.), Measuring stress: A guide for
Meglino, B. M., Ravlin, E. C., & Adkins, C. L. (1989). A health and social scientists (pp. 29-58). Oxford, En-
work values approach to corporate culture: A field test of gland: Oxford University Press.
the value congruence process and its relationship to in-
Warr, P. B., Cook, J. D., & Wall, T. D. (1979). Scales for
dividual outcomes. Journal of Applied Psychology, 74, the measurement of some work attitudes and aspects of
424-434.
psychological well-being. Journal of Occupational Psy-
Messick, S. (1981). Constructs and their vicissitudes in ed-
chology, 52, 129-148.
ucational and psychological measurement. Psychological
Weick, K.A. (1989). Theory construction as disciplined
Bulletin, 89, 575-588.
imagination. Academy of Management Review, 14, 516-
Messick, S. (1995). Validity of psychological assessment.
531.
American Psychologist, 50, 741-749.
Weiss, D.J., Dawis, R.W., England, G. W., & Lofquist,
Mill, J. S. (1886). A system of logic, ratiocinative and in-
L.H. (1967). Manual for the Minnesota Satisfaction
ductive (8th ed.). London: Oxford University Press.
Questionnaire. Minneapolis: University of Minnesota,
Mobley, W. H. (1982). Employee turnover: Causes, conse-
Industrial Relations Center.
quences, and control. Reading, MA: Addison-Wesley.
Williams, L. J., & Brown, B. K. (1994). Method variance in
Mowday, R. T., Steers, R. M., & Porter, L. W. (1979). The
organizational behavior and human resources research:
measurement of organizational commitment. Journal of
Effects on correlations, path coefficients, and hypothesis
Vocational Behavior, 14, 224-247.
testing. Organizational Behavior and Human Decision
Nunnally, J. C. (1978). Psychometric theory (2nd ed). New
Processes, 57, 185-209.
York: McGraw-Hill.
Wold, H. (1982). Soft modeling: The basic design and some
Phillips, J. S., & Lord, R. G. (1982). Schematic information
processing and perceptions of leadership in problem- extensions. In K. G. Jtreskog & H. Wold (Eds.), Systems
solving groups. Journal of Applied Psychology, 67, 486-
under indirect observation (Part 2, pp. 1-54). Amster-
dam: North-Holland.
492.
Popper, K. R. (1959). The logic of scientific discovery. New Zuriff, G. (1998). Against metaphysical social construction-
York: Basic Books. ism in psychology. Behavior and Philosophy, 26, 5-28.
Raphael, K. G., Cloitre, M., & Dohrenwend, B. P. (1991).
Problems of recall and misclassification with checklist Received October 26, 1998
methods of measuring stressful life events. Health Psy- Revision received September 2, 1999
chology, 10, 62-74. Accepted November 5, 1999 •