Journal of Applied Psychology
1998, Vol. 83, No. 2, 234-246
Copyright 1998 by the American Psychological Association, Inc.
0021-9010/98/S3.00
Functional Relations Among Constructs in the Same Content Domain at
Different Levels of Analysis: A Typology of Composition Models
David Chan
Michigan State University and National University of Singapore
Composition models specify the functional relationships among phenomena or constructs
at different levels of analysis (e.g., individual level, team level, organizational level) that
reference essentially the same content but that are qualitatively different at different levels
(M. T. Hannan, 1971; K. H. Roberts, C. L. Hulin, & D. M. Rousseau, 1978; D. M.
Rousseau, 1985). Specifying adequate composition models is a critical component of
good multilevel research. A typology of composition models is proposed to provide a
framework for organizing, evaluating, and developing constructs and theories in multilevel
research. Five basic forms of composition are described and illustrated. Implications of
the typology are discussed.
Organizational phenomena have the properties of dy-
namic systems, with critical antecedents, processes, and
outcomes conceptualized and measured at multiple levels
of organizational analysis (e.g., individual, group, organi-
zation). Because more researchers are beginning to real-
ize that the organizational phenomenon under investiga-
tion often is inherently multilevel as opposed to occurring
at a single level or in a level vacuum, organizational stud-
ies increasingly are adopting a multilevel approach. Sev-
eral influential theoretical frameworks for multilevel re-
search have been proposed (e.g., House, Rousseau, &
Thomas-Hunt, 1995; Klein, Dansereau, & Hall, 1994;
Rousseau, 1985). Excellent discussions on important
mathematical issues related to the analysis of multilevel
data (e.g., Bliese & Halverson, in press; Ostroff, 1993a)
and analytical models for structuring multilevel data (e.g.,
Bryk & Raudenbush, 1987; McArdle & Epstein, 1987;
Meredith & Tisak, 1990; Muthen, 1994; Willett & Sayer,
1994) also are available.
However, despite the existence of broad theoretical
David Chan, Department of Psychology, Michigan State Uni-
versity, and Department of Social Work and Psychology, Na-
tional University of Singapore.
I thank Neal Schmitt, Daniel Ilgen, and Richard DeShon for
their helpful comments on the article, and i am grateful to Steve
Kozlowski for introducing me to multilevel research.
An earlier version of this article was presented at the 1997
Annual Meeting of the Academy of Management in Boston,
Massachusetts, August 1997.
Correspondence concerning this article should be addressed
to David Chan, Department of Psychology, 129 Psychology Re-
search Building, Michigan State University, East Lansing, Mich-
igan 48824-1117. Electronic mail may be sent to dchan@
pilot.msu.edu.
frameworks and methodological advances, the fundamen-
tal substantive issue of construct validation in multilevel
research has not been addressed adequately. Accompa-
nying the increased interest in multilevel research is an
increased proliferation of new constructs at multiple lev-
els. Unless we have explicit composition models to guide
the development and validation of newly proposed con-
structs in multilevel research, there is a danger of violating
the scientific principle of parsimony. Organizational re-
searchers could easily end up with a multitude of labels,
all of which purportedly refer to scientific constructs but
in reality have no incremental explanatory value. Compo-
sition models specify the functional relationships among
phenomena or constructs at different levels of analysis
(e.g., individual level, team level, organizational level)
that reference essentially the same content but that are
qualitatively different at different levels (Hannan, 1971;
Roberts, Hulin, & Rousseau, 1978; Rousseau, 1985).
Specifying functional relationships between constructs at
different levels provides a systematic framework for map-
ping the transformation across levels. The explicit trans-
formation relationships provide conceptual precision in
the target construct, which in turn aids in the derivation
of test implications for hypothesis testing. Unfortunately,
the specification of functional relationships between con-
structs has not always been adequate or even explicit in
multilevel research. This is partly because no systematic
frameworks for specifying functional relationships exist.
An adequate typology of composition models addresses
the above problems and contributes to multilevel research
in at least two important ways. First, it provides an or-
ganizing framework for existing focal constructs facilitat-
ing scientific communication in multilevel research. Re-
searchers can be more confident that they are referring to
234
COMPOSITION MODELS 235
the same construct when it is explicated according to the
same form of composition. Meaningful replications and
extensions of current findings then are possible. Apparent
contradictory findings may be reconciled, and debates
may be clarified. For example, many so-called inconsis-
tent findings simply could be a result of confusion of
terminology (i.e., comparing apples and oranges), and
the confusion may become apparent when each study lo-
cates its construct in the typology corresponding to the
composition model. Organizing existing constructs also
aids cumulation of research findings by providing a frame-
work for performing meaningful meta-analytic studies in
multilevel research.
Second, a typology provides a conceptual framework
for developing and validating new focal constructs and
multilevel theories. As described later in this article, the
typology of models could help compose new explanatory
constructs from established ones. In addition, being cogni-
zant of different models allows the researcher to consider
alternative designs, measurements, and data analyses for
testing competing hypotheses, modifying existing theories
or developing new ones, or performing a more rigorous
test of the original hypothesis. The purpose of this article
is to propose a typology of composition models.
A Typology of Composition Models
The proposed typology is concerned with elemental
composition; that is, situations in which data from a lower
level are used to establish the higher level construct. In
other words, the higher level construct is of a collective
or aggregate nature and is construed as some form of
combination of the lower level units. All lower level units
play some substantive role in composing the lower level
construct to the higher level construct, and the value of
the higher level construct is not solely determined by any
single lower level unit (i.e., each unit is used in some
way or another, such as for computing the mean level).
Note that the use of data from the lower level to establish
the higher level construct does not imply that it is neces-
sary to begin conceptualization at a level lower than the
level of the target or composed construct. The starting
level of conceptualization is dependent on the research
question. For example, a researcher may start at the group
level' with the established construct of group norms and
then move down to the individual level to collect percep-
tual data for subsequent aggregation to the group level
to establish the construct of group norms. The focus on
elemental composition is consistent with the actual con-
straints and practice in empirical multilevel research. As
noted by several researchers (e.g., Ostroff, 1993a; Roberts
et al., 1978), we often do not have global indices of the
higher level (organizational or group) variables of interest
and hence have to rely on aggregated data from the lower
level (individuals) to represent the higher level variable.
Because the focus is on elemental composition, this article
does not address the traditional issues of disaggregation
and ecological fallacies (Cronbach, 1976; Hannan, 1971;
Langbein & Lichtman, 1978).
Table 1 presents the typology of composition models.
The typology describes the basic forms composition mod-
els can take. The forms described are ideal types. The
five basic forms of composition models are: (a) additive,
(b) direct consensus, (c) referent-shift consensus, (d)
dispersion, and (e) process composition. A theory of the
focal construct in a multilevel study may contain one or
more of the five composition forms. The five forms of
composition models in Table 1 are not presented in any
specific order. The reader is cautioned against ordering
the five forms in terms of the similarity of functional
relationships between constructs. The simplistic notion of
similarity of functional relationships is not an adequate
way of representing the complexity in or guiding the de-
velopment of composition forms because there is an infi-
nite number of answers (i.e., dimensions) to the question,
"Similar with respect to what?"
As shown in Table 1, each composition model is defined
by a particular form of functional relationship specified
between constructs at different levels. Corresponding to
each form of functional relationship is a typical opera-
tional process by which the lower level construct is com-
bined to form a higher level construct. Note that the opera-
tional combination process is the typical form as opposed
to a necessary consequence of the functional relationship
specified. The column in Table 1 labeled empirical sup-
port suggests what constitutes the forms of evidence
needed to support the relevant functional relationships
and to establish that appropriate combination rules are
applied. To illustrate the forms of composition, examples
from climate research are consistently used throughout
1
In describing the typology and throughout this article, the
terms group and team are used interchangeably to refer to the
level of the collection of individuals immediately higher than
the individual level. Several researchers have distinguished
groups and teams as the polar ends on a continuum of task
interdependence (e.g., Salas, Dickinson, Converse, & Tannen-
bauem, 1992). In addition, groups are often characterized by
low role differentiation and low task differentiation, whereas
teams are characterized by high role differentiation, high task
differentiation, distributed expertise, and high levels of task in-
terdependence (Sundstrom et al., 1990). The present typology
is concerned with work groups or work teams in organizations.
The terms groups and teams are used to refer to multiple individ-
uals formed to perform some organizationally relevant task-
functions. These individuals interact, exhibit task interdepen-
dence, possess one or more shared goals, and are embedded in
a larger organizational setting (Kozlowski et al., 1994, 1996;
Salas etal., 1992).
236 CHAN
Table 1
A Typology of Composition Models
Functional relationships
Typical operational
combination Empirical support
Example from climate
research
Additive model
Higher level unit is a summation
of the lower level units
regardless of the variance
among these units
Summing or averaging lower
level scores
Validity of additive index (e.g.,
mean of lower level units)
From psychological climate to
organizational climate
(Click's T1985]
conceptualization)
Direct consensus model
Meaning of higher level
construct is in the consensus
among lower level units
Within-group agreement to
index consensus and justify
aggregation
Value of within-group
agreement index (e.g., >•„,);
validity of aggregated scores
From psychological climate to
organizational climate
(Jarnes et al.'s [1984]
conceptualization)
Referent-shift consensus model
Lower level units being
composed by consensus are
conceptually distinct though
derived from the original
individual-level units
Within-group agreement of new
referent lower level units to
index consensus and justify
aggregation
Value of within-group
agreement index (e.g., r^);
validity of aggregated scores
From psychological climate to
organizational collective
climate
Dispersion model
Meaning of higher level
construct is in the dispersion
or variance among lower level
units
Within-group variance (or its
derivative) as
operationalization of the
higher level construct
Absence of multimodality in
within-group distributions of
lower level scores; validity
of dispersion index
From psychological climate to
climate strength
Process model
Process parameters at higher
level are analogues of process
parameters at lower level
No simple algorithm; ensure
analogues exist for all critical
parameters
Nomological validity for source
and target constructs at their
respective levels to
distinguish shared core
content from level-specific
aspects
From psychological climate
development to
organizational climate
emergence
the typology. I assume that a researcher collects climate
data (i.e., individual perceptual responses on climate
questionnaires) from hundreds of individuals across many
organizations (or groups). For each composition model,
I present examples of hypotheses and explanations of how
the model can be applied to the data. Where appropriate,
examples other than climate also are presented to clarify
the composition form. The first two forms of composition
are familiar to most researchers, and I will only describe
them briefly. The remaining three forms are less familiar
even though they characterize many focal constructs ex-
amined in multilevel studies. As mentioned earlier, com-
position models are seldom made explicit in existing mul-
tilevel research. The following sections describe each of
the five composition models in the typology.
Additive Models
Additive composition models specify a straightforward
functional relationship between constructs at different lev-
els. In such models, the meaning of the higher level con-
struct is a summation of the lower level units regardless
of the variance among these units. In additive composition
models, the variance of the lower level units is of no
theoretical or operational concern for composing the lower
level construct to the higher level construct. The typical
operational combination process is a simple sum or aver-
age of the lower level scores on the lower level variable
to represent the value on the higher level variable. The
validity of the additive index (e.g., the mean) constitutes
empirical support for the composition.
In the climate example, the researcher may be interested
in relating organizational climate to organizational perfor-
mance. The researcher has an established measure of orga-
nizational performance but he or she has to develop some
conceptualization and measure of the construct of organi-
zational climate. Adopting Click's (1985) conceptualiza-
tion, the researcher stipulates that all organizations have
an organizational climate that can be described as high
or low on various dimensions regardless of the level of
COMPOSITION MODELS 237
within-organization individual-level agreement. Within-
organization agreement, according to this view, is an issue
of measurement accuracy reflecting individual-level ran-
dom error and sources of bias (see Glick, 1985, pp. 604—
605). Hence, using an additive composition model, the
researcher averages the climate perceptions of individuals
within each organization, regardless of the within-organi-
zation variance, to represent the organizational climate
variable. The organizational mean climate scores and the
organizational performance variable then are correlated,
and the validity coefficient (i.e., validity of the additive
index; in this case, the mean) provides empirical support
for the additive composition model.
Clearly, whether the relationship between organiza-
tional climate and psychological climate (i.e., individual-
level climate perceptions) is additive or some other com-
position form depends on how the construct of organiza-
tional climate is conceptualized. For example, if the level
of individual perceptual agreement within an organization
is central in the substantive definition of organizational
climate, then an additive composition model would be
inappropriate because within-group (i.e., organization)
variance among lower level units (i.e., individual percep-
tions) becomes relevant in composing the lower level con-
struct (i.e., psychological climate) to the higher level con-
struct (i.e., organizational climate). In this case, a direct
consensus composition model, described in the next sec-
tion, is appropriate.
Direct Consensus Models
Direct consensus composition is probably the most fa-
miliar and popular form of composition among multilevel
researchers. This model uses within-group consensus of
the lower level units as the functional relationship to spec-
ify how the construct conceptualized and operationalized
at the lower level is functionally isomorphic to another
form of the construct at the higher level. The typical opera-
tional combination process is using within-group agree-
ment of scores to index consensus at the lower level and
to justify aggregation of lower level scores to represent
scores at the higher level (e.g., James, Demaree, & Wolf,
1984; Kozlowski & Hattrup, 1992; Ostroff, 1993b; Os-
troff & Rothausen, 1997). This operational combination
process has two components. The first component in-
cludes an operationalization (i.e., a measure) of the con-
ceptual definition for each of the two constructs (i.e., one
at each level). For example, individual-level perceptual
responses on a climate measure are used to operationalize
psychological climate, whereas the mean of those individ-
ual responses within an organization is used to operation-
alize organizational climate. The second component spec-
ifies the manner of and precondition(s) for combining
the individual lower level measurements to represent the
higher level measurement. For example, within-group
agreement indexes such as the rwg index (James et al.,
1984) may be calculated and some cutoff level of agree-
ment is used to justify aggregation of individual re-
sponses. In this example, the aggregation procedure and
preconditions, together with the conceptual definition of
the higher level construct, determine the meaningfulness
and validity of the operationalization of the higher level
construct.
In the climate example, the researcher may follow
James (1982) to construe psychological climate as an
individual's perception or cognitive representation of the
work environment in terms of the psychological meaning
and significance to the individual. Organizational climate
simply refers to the shared assignment of meanings among
individuals within the organization. In this conceptualiza-
tion, within-group agreement among individual climate
perceptions indicates shared assignment of psychological
meaning. It is this sharedness that constitutes functional
equivalence between the climate constructs at the two
levels. Hence, the definition of organizational climate is
essentially the same as psychological climate, except that
the former refers to the shared perceptions among the
individuals. The conceptual relationship between the two
forms of the construct at different levels then drives the
manner in which the lower level construct composes to the
higher level construct. So the researcher would proceed to
check within-group agreement (group here refers to the
organization) of individual climate responses using some
agreement index (e.g., rn,s). High within-group agreement
indicates consensus and justifies aggregation of individual
climate responses to represent scores on the organiza-
tional climate variable.
In direct consensus composition, the lower level attri-
butes need not be restricted to individual perceptions.
Consensus, as indexed by within-group agreement, can
apply to individual-level attributes such as cognitive abil-
ity and styles, personality, mental representation, and be-
havioral v