Upload
uyha-indigo-beisek
View
213
Download
0
Embed Size (px)
Citation preview
7/30/2019 78164802
1/29
RESEARCH NOTE
ARESEARCH NOTE ON REPRESENTING PARTWHOLE
RELATIONS IN CONCEPTUAL MODELING1
Gove N. Allen
Marriot School of Management, Brigham Young University, 783 Tanner Building,
Provo, UT 84602 U.S.A. {[email protected]}
Salvatore T. March
Owen Graduate School of Management, Vanderbilt University,
Nashville, TN 37203 U.S.A. {[email protected]}
Empirical research is an important methodology for the study of conceptual modeling practices. The recently
published article Representing PartWhole Relations in Conceptual Modeling: An Empirical Evaluation
(Shanks et al. 2008) uses the lens of ontology to study a relatively sophisticated aspect of conceptual modeling
practice, the representation of aggregation and composition. It contends that some analysts argue that a
composite should be represented as a relationship while others argue that a composite should be represented
as an entity. We find no evidence of such a dispute in the data modeling literature. We observe that composites
are objects. By definition, all object-types should be represented as entities. Therefore, using the relationship
construct to represent composites should not be seen as a viable alternative. Additionally, we found significant
conceptual and methodological issues within the study that call its conclusions into question. As a way to offer
insight into the requisite methodological procedures for research in this area, we conducted two experiments
that both explicate and address the issues raised. Our results call into question the utility of using ontology
as a foundation for conceptual modeling practice. Furthermore, they suggest a contrary but at least equally
plausible explanation for the results reported by Shanks et al. In conducting this work we hope to encourage
dialogue that will be beneficial for future endeavors aimed at identifying, developing, and evaluating
appropriate foundations for the discipline of conceptual modeling.
Keywords: Conceptual modeling, empirical research, ontology, information systems development, compo-
sition, UML, entityrelationship model
Introduction1
Academic research in the field of information systems haslong been interested in examining the relationship between
conceptual models (grammars, methods, and scripts) and
human performance in tasks related to information system
development and use (Batra et al. 1990; Bodart et al. 2001;
Burton-Jones and Meso 2006; Kim and March 1995). Oneimportant question deals with identifying a sound theoretical
basis for the development and assessment of conceptual
modeling grammars and practices (Wand and Weber 2002).
Bunges (1977) ontology has been suggested as such a theo-
retical basis (Wand et al. 1995; Wand et al. 1999; Wand and
Weber 1993). Several articles have reported experiments that
study the effects of conformance to Bunges ontology on
human performance in various tasks (Bodart et al. 2001;
1Dale L. Goodhue was the accepting senior editor for this paper. Sandeep
Purao served as the associate editor.
The appendices for this paper are located in the Online Supplements
section of theMIS Quarterlys website (http://www.misq.org).
MIS Quarterly Vol. 36 No. 3, pp. 945-964/September 2012 945
7/30/2019 78164802
2/29
Allen & March/Representing PartWhole Relations in Conceptual Meaning
Bowen et al. 2006, 2009; Gemino and Wand 2005; Shanks et
al. 2008.
Establishing that a specific ontological foundation consis-
tently leads to superior human performance in conceptual
modeling tasks would be a significant contribution to our
field. Shanks et al. (2008) draw this conclusion, contending
that
our results add strength to a growing body of empi-
rical work that supports the usefulness of ontological
theories, especially Bunges (1977) theory, as a
means of predicting the strengths and weaknesses of
conceptual modeling models and practices (p. 570).
However, we found a number of apparent problems with the
Shanks et al. study, in both its conceptual underpinnings and
in the execution of the empirical tests that cast considerable
doubt on such a conclusion.
Articles in MIS Quarterly are often the starting point for
further research. Because reliance on results with problematic
conceptualization and methodology will have detrimental
effects on the development of a cumulative research tradition
within the discipline (Straub 2008), we feel it is important to
discuss these problems, to recast and retest the assertions of
Shanks et al., and to provide an exemplar of careful work in
this difficult area of research.
While we focus on what we see as critical concerns with the
study reported by Shanks et al., it is important to note the
significant contributions of the work. First, this is among avery small number of studies aimed at empirically testing the
utility of a well-established philosophical ontology as the
basis for conceptual modeling. Empirical studies of this
nature are crucial. Second, it presents an innovate framework
for data analysis that combines a quantitative assessment of
subjects answers with a qualitative assessment of explana-
tions for their answers and their interpretations of the modeled
domain. It presents a detailed description of procedures used
in this qualitative data analysis. Analysis and categorization
of verbal protocols is a difficult and time-consuming process,
but one that can yield significant insight into subjects
reasoning processes.
In undertaking this work, our intent is to encourage dialogue
that will be beneficial for future endeavors as our field works
toward identifying, developing, and evaluating appropriate
foundations for the discipline of conceptual modeling. We
identify and explicate two concerns with respect to the con-
ceptualization and five with respect to the execution (method-
ology) of the study presented in Shanks et al. Further, we
conducted two experiments that validate our concerns and
demonstrate how such research should be conducted.
With respect to conceptualization, we have two fundamental
concerns. First, the study is structured as an empirical evalua-
tion of alternative conceptual-modeling representations of
partwhole relations (p. 554) contending that, in practice,
some analysts argue [that] composites should be represented
as a relationship (p. 553) rather than as entities.2 We find no
support for this contention in the literature. Composition is a
relatively sophisticated construct available in the Unified
Modeling Language (UML). It is used when a modeler deems
it important to represent the fact that a relationship has a
specialized meaning, that is, the association of part objects
that constitute a composite object as opposed to a simple
association (Fowler 2003). Representing the composite object
itself as a relationship violates the basic definitions of the
UML and generally accepted conceptual modeling practice
(Carlis and Maguire 2001; Keet 2008; Kent 1978).
Second, and more importantly, the study appeals to Bunges
ontology as the theoretical foundation for predicting dif-
ferences in human performance; but it does not conform to
Bunges ontology in the conceptualization or design of the
experiment. The experiment uses two conceptual models
(diagrams). The first, termed ontologically clear, is argued to
be consistent with Bunges ontology with respect to the
representation of composite objects. The second, termed
ontologically unclear, is argued to be inconsistent with
Bunges ontology in this respect. The underlying proposition
is that the ontologically clear model will result in superior
user performance over the ontologically unclear model
because of its conformance to Bunges ontology. However,
upon close scrutiny and as discussed in detail below, neither
of these models conforms to Bunges ontology with respect to
the definition of composite objects.
We have five concerns with respect to the studys execution.
First, it does not successfully operationalize the independent
variable, ontological clarity. Appealing to the theory of
ontological clarity (Wand and Weber 1993), the study con-
tends that the ontologically unclear model lacks ontological
clarity because it exhibits construct overload. Specifically, it
contends that the UML relationship construct is used to
explicitly represent associations and to implicitly representcomposites (Shanks et al. 2008, p. 561). Association and
composition are two distinct ontological constructs; however,
2Although the precise terms for the diamonds and rectangles in an ER
diagram are relationship types and entity types, to be consistent with Shanks
et al. and for ease in reading, we will refer to them as relationships and
entities. We will use the terms relationship instance and entity instance to
refer to members of those sets represented in a diagram.
946 MIS Quarterly Vol. 36 No. 3/September 2012
7/30/2019 78164802
3/29
Allen & March/Representing PartWhole Relations in Conceptual Meaning
UML has two graphical symbols to represent relationships:
a labeled arc and a diamond (Rumbaugh et al. 2005). The
ontologically unclear model exclusively uses labeled arcs to
explicitly represent association relationships and exclusively
uses diamonds to implicitly represent composite objects.
Thus it does not overload either graphical symbol.
Second is a concern with the development of the experimental
materials. The study describes a procedure for generating the
ontologically unclear model from the ontologically clear
model by transforming composites into ternary relationships
intended to implicitly represent those composites. It does not,
however, follow that procedure. As a result there is an incon-
sistent mapping between the models with respect to the
representation of composite objects. Furthermore, the ternary
relationships in the ontologically unclear model need not be
interpreted as representing composites at all. They are more
reasonably interpreted as representing relationships.
Third is a concern that the treatments are confounded. The
ontologically unclear model uses binary and ternary relation-
ships while the ontologically clear model uses only binary
relationships. The semantics of ternary relationships are
significantly more difficult to understand than are the seman-
tics of binary relationships, especially for novice users (Topi
and Ramesh 2002). Inclusion of ternary relationships in the
ontologically unclear treatment but not in the ontologically
clear treatment represents a substantial confound, making it
impossible to determine the source of observed performance
differences. We examine the strength of this confound in our
first experiment. Moreover the questions that make up the
experimental task systematically favor the ontologically cleardiagram in a manner that is independent from the experi-
mental treatment, further strengthening the confounding
effects.
Fourth is a concern that subjects were instructed that the
diagrams they were asked to interpret were standard UML
diagrams; however, the instructions they were given about
interpreting the diagrams were not consistent with the defini-
tions of the UML constructs used, specifically with respect to
the definitions of ternary relationship constraints. Although
this may not be a concern for novices who are unfamiliar with
UML, many of the subjects had modeling experience and a
number had explicit experience with UML. Such subjectslikely experienced cognitive dissonance with respect to the
nonstandard definitions.
Fifth is a concern with the analysis of participants perfor-
mance. Correct responses are reported for only two of the
questions used in the experimental task. In one of these
questions, subjects answers were incorrectly scored for one
of the experimental treatments.
Each of these concerns is assessed and explained in the
following sections. Our goal in this endeavor is to sharpen
the understanding of researchers with respect to the difficulty
and the care required in working in this challenging but
important area of inquiry. We hope to encourage future
research in the development and evaluation of appropriate
theoretical underpinnings for conceptual modeling and to help
improve the conceptualization and execution of empirical
research in this area. Moreover, we seek to open a dialog
among researchers in developing and using appropriate
theoretical foundations for conceptual modeling and in
developing and using appropriate experimental protocols in
conceptual modeling research.
Foundations in Conceptual ModelingPractice
Shanks et al. base their study on the contention that in con-
ceptual models composite things are sometimes represented
as entities (classes) and sometimes represented via relation-
ships (associations) between the components of the com-
posite (p. 554). However, doing so violates the basic defi-
nitions of both the entityrelationship (ER) grammar (Chen
1976) and the Unified Modeling Language (UML) grammar
(Rumbaugh et al. 2005). In both grammars, entities (classes)
are defined to represent collections of objects (instances) and
relationships are defined to represent collections of associa-
tions among objects. An instance of a relationship is the
association of one instance of each related entity.
However, the definitions of conceptual modeling grammars
and the way in which those grammars are appropriated in
practice may vary widely (Feldman and Pentland 2003). The
study presented by Shanks et al. would be important if,
indeed, a significant proportion of analysts argue composites
should be represented as a relationship or association (p.
553), independent of the fact that doing so violates the defini-
tion of the relationship construct. Shanks et al. contend that
widely used conceptual modeling/database textbooks,
including Elmasri and Navathe (2004) and Teorey et al.
(2006), show a composite represented implicitly via a rela-
tionship or association construct (p. 555). We disagree with
this interpretation of the conceptual models presented in these
textbooks, as do the authors of these textbooks.3 The strength
of a conceptual model is that domain semantics are repre-
sented explicitly. As discussed above, the entities in a con-
ceptual model represent collections, or types of objects; the
3Personal communications with the authors of these textbooks between
May 12, 2008, and June 30, 2008.
MIS Quarterly Vol. 36 No. 3/September 2012 947
7/30/2019 78164802
4/29
Allen & March/Representing PartWhole Relations in Conceptual Meaning
relationships represent collections or types of associations
among them; object types not explicitly represented are not
part of the conceptual model (Carlis and Maguire 2001).
Figures 1 and 2 are the examples used by Shanks et al. to
motivate their study. A detailed description of the syntax
used in Figure 1 is found in Appendix A, Section 1; a detailed
description of the syntax used in Figure 2 is found in Appen-
dix A, Section 5. Figure 1 is a portion of a conceptual model
taken from Elmasri and Navathe (2004, p. 102). It shows a
committee relationship between a faculty member and a
graduate student. Shanks et al. argue that a composite entity
committee is represented implicitly by the so labeled
diamond. They argue that Elmasri and Navathe clearly
intend the committee to be a composite entity that has
faculty entities and student entities as components (p.
555). While readers of a conceptual modeling diagram will
use their existing domain knowledge to make sense of a
diagram, if the designer of a diagram intends to convey some
meaning, it should be done explicitly rather than implicitly
(Carlis and Maguire 2001).
Our interpretation of the diagram is that the committee rela-
tionship indicates that there is an association between
individual graduate students and individual faculty members
by virtue of their mutual participation in a committee, even
though the committee itself is notrepresented in the model.
If Elmasri and Navathe intended to represent the committee
object they would have done so using an entity. The label
committee on the relationship simply stands in place of
another, potentially more descriptive label such as advises/
is advised by representing the role played by each entity inthe relationship. This interpretation is more consistent with
the way in which graduate committees are typically con-
ceptualized. That is, an instance of the committee relationship
indicates that a faculty member participates in a committee
that advises a student; only faculty members can be members
of a committee; the graduate student is not a member of (or
any other kind of component of) a committee. Furthermore,
a relationship instance is defined and identified by the com-
bination of the related entity instances (Elmasri and Navathe
2004). Hence, a committee relationship instance, as repre-
sented in Figure1, has exactly one faculty member and exactly
one graduate student. The cardinality constraints, M and N in
the diagram, indicate that one faculty member may be asso-ciated with many graduate students and one graduate student
may be associated with many faculty members.
Using the name committee for the relationship may evoke
reasoning processes related to academic committee objects
with which the reader is familiar; however, all that can be
known from the diagram is that faculty members and graduate
students are associated with each other and that committee
is deemed an appropriate, although potentially misleading,
name for that relationship. If the common meaning of aca-
demic committees is assumed (i.e., each student has one
committee and each faculty member can serve on many
committees), then Figure 1 does allow for reasoningabout
committees. The members of a students committee are those
faculty members associated with that student through the
committee relationship; however, the committee itself is not
represented in the diagram.
Shanks et al. use a second motivating example (Figure 2),
reproduced from Teorey et al. (2006, p. 91), and describe it as
follows: Engineers are divided into groups for certain
projects. Each group has a leader. By this, Shanks et al.
contend that the association with the aggregation symbol in
Figure 2 implicitly represents a group composite that has
engineers as components (p. 555). Here the use of the
language divided into groups may suggest that groups are
composite entities made up of engineers; however, thatcomposite is not represented in the diagram. The relationship
shows that some engineers are related to others by virtue of
either being a group leader of other engineers or being led by
another engineer who is a group leader.
Moreover, the purpose of that figure is not to explicate the
semantics of composition but rather to illustrate rules for
transforming recursive one-to-many relationships into SQL
for implementation (Teorey et al. 2006). The use of the
diamond was not intended to represent a composition (aggre-
gation) relationship, but rather an association relationship as
in the corresponding ER diagram from the prior page in the
textbook. If the intention were to implicitly represent an
aggregation relationship as Shanks et al. contend, then the
interpretation would be as follows: The UML notation for
aggregation (the open diamond at one end of an association
arc) indicates that members of the class touching the diamond
are composed of members of the class at the other end of the
association arc. Thus the semantics of the diagram do not
indicate thatgroups are composed of engineers; rather they
indicate that engineers are composed of engineers. This is
neither congruent with the English language understanding of
what an engineer is nor is it valid UML syntax. The UML
does not allow aggregation hierarchies to be circular as in the
recursive relationship of Figure 2; an object may not directlyor indirectly be a part of itself (Rumbaugh et al. 2005, p.
164). Hence, Figure 2 cannot be interpreted to implicitly
represent any composite.4
4In a personal communication on June 10, 2008, Tom Nadeau (Teorey et al.)
confirmed that the intended meaning of the diagram was simple association,
not aggregation. He also confirmed that if the group composite object were
to be modeled, an entity would have been used to explicitly represent it.
948 MIS Quarterly Vol. 36 No. 3/September 2012
7/30/2019 78164802
5/29
Allen & March/Representing PartWhole Relations in Conceptual Meaning
Figure 1. Portion of an ER Diagram that Shankset al. Argue Uses a Relationship to ImplicitlyRepresent a Committee Object (Source: Elmasri/
Navathe FUNDAMENTALS OF DATABASE SYSTEMS, p. 102, Figure
4.9, An EER Schema for a University Database, 2007, 2004
Shamkant B. Navanthe & Ramez A. Elmasri. Reproduced by
permission Pearson Education, Inc.)
Figure 2. Portion of a UML Diagram that Shankset al. Argue Uses an Association to Implicitly
Represent a Composite Group Object (Source:Database Modeling and Design: Logical Design (4th ed.), T. Teory, S.
Lightstone, and T. Nadeau, p. 91. Copyright 2006, Elsevier)
Therefore, neither of these textbook examples uses the rela-
tionship construct to represent composites; in both examples
the relationship construct represents association. Thus neither
example serves as support for the claim that some textbook
authors use relationships to represent composition. As dis-
cussed above, such a claim is inconsistent with the definitions
of both the ER and the UML grammars.
Two related concepts have been discussed in the conceptual
modeling literature, reification and association classes (El-
masri and Navathe 2004; Halpin and Morgan 2008; Hansen
and Hansen 1996; Rumbaugh et al. 2005). Reification, also
termed objectification, refers to a process of transforming
phenomena initially represented as a relationship into an
entity or class. Association classes combine the properties of
entities and relationships into a single construct. Neither,
however, infers that an ordinary relationship can or should be
used to represent a composite entity. In fact, these constructs
were introduced because the relationship construct alone is
insufficient to express phenomena that have characteristics of
both entities and relationships, specifically the ability to
participate in additional relationships.
Finally, Shanks et al. use ternary relationships in the ontologi-
cally unclear treatment. A ternary relationship associates
three entities. As with any relationship, an instance of a ter-
nary relationship associates exactly one instance of each of
the related entities. However, the semantics of ternary rela-
tionship constraints are complex and frequently misunder-
stood (Topi and Ramesh 2002). Furthermore, there are
significant differences in the definitions of ternary relation-
ship cardinality constraints between UML and a number of
extended ER model proposals. Rather than using UML,
Shanks et al. use one of these extended ER proposals in the
ontologically unclear treatment. Details of the relevant
definitions are presented in Appendix A.
Theoretical Foundations
Shanks et al. indicate that they are testing the theory of onto-
logical clarity posed by Wand and Weber (1993). This theory
asserts that when the constructs of a conceptual modeling
grammar exist in a bijective correspondence with the con-
structs of an ontology, scripts expressed in that grammar will
better communicate meaning to users than will scripts formed
using grammars with ontological mappings that are either
surjective orinjective in either direction. In a bijective corre-
spondence, each construct in the conceptual modeling gram-
mar maps to exactly one construct in the ontology and each
construct in the ontology maps to exactly one construct in the
grammar. In a surjective correspondence, multiple constructs
in the conceptual modeling grammar map to the same con-
struct in the ontology or multiple constructs in the ontology
map to the same construct in the conceptual modeling gram-
FACULTY
COMMITTEE
M
N
GRAD-STUDENT
Engineer 0..*
is-led-by
is-group-leader-of
1
MIS Quarterly Vol. 36 No. 3/September 2012 949
7/30/2019 78164802
6/29
Allen & March/Representing PartWhole Relations in Conceptual Meaning
mar. In an injective correspondence, there may be constructs
in the conceptual modeling grammar that have no corre-
sponding construct in the ontology or constructs in the on-
tology that have no corresponding construct in the ontology.
The specific aspect of ontological clarity tested by Shanks et
al. is the existence of a surjective correspondence where one
construct in a conceptual modeling grammar is used to
represent multiple constructs in an ontology. This type of
surjective relationship is commonly termed construct over-
load (Burton-Jones et al. 2009).
The authors select the ontology of Mario Bunge (1977) as the
referent ontology for their study. This presents problems for
conceptual modeling because Bunges ontology is intended to
represent concrete objects (things) that possess substantial
properties. Concrete objects exist objectively in space and
time, independent of human interpretation and ascription of
meaning. Conceptual models, however, must frequentlyrepresent conceptualobjects and attributes that exist subjec-
tively, are based exclusively on human interpretation, beliefs
and collective agreement, and convey ascribed meaning (Kent
1978; Searle 1995, 2006). These are explicitly excluded from
Bunges ontology (Allen and March 2006a; Wyssusek 2006).
In this regard Wand et al. (1999) deviate from Bunges
ontology by asserting that the notion of a concrete thing
applies to anythingperceivedas aspecific objectby someone,
whether it exists in physical reality or only in someones
mind (p. 497). In our opinion, this deviation severely
reduces any guidance a modeler would obtain from Bunges
ontology in the identification and classification of phenomena
to be represented in a conceptual model. In this deviation,phenomena are represented as the modeler perceives to be
appropriate, not as they exist in what Bunge (p. 157) calls
real existence. A complete assessment of Bunges ontology
as a foundation for conceptual modeling is beyond the scope
of this paper and is the subject of ongoing research activities.
Hence, we restrict our discussion to the ontological issues
surrounding the diagrams used in the experimental treatments.
Problems with the Experimental
Treatments
Bunges definition of a composite is quite different from the
notion of composition in the UML. None of the composites
in the ontologically clear UML diagram (Figure 3) conforms
to the definition of composition in Bunges ontology.
Appendix A contains a detailed review of the syntax used in
Figure 3. Bunge (p. 29) states that an individual is com-
posite [if and only if] it is composed of individuals other than
itself and the null individual. The composite Team vio-
lates this definition because it permits a Team to be composed
of one Team Leader and a minimum of zero Team Members.
That is, a Team Leader composed with no individuals other
than itself (i.e., having no team members) is represented to be
a Team. In Bunges ontology, a Team Leader composed with
no other individual is the Team Leader, not a Team.
Thus defining a team as the composition of a leader and some
number of members but specifying that a team can exist with-
out any members violates Bunges definition of composition.
It is ontologically equivalent to defining a water molecule as
being composed of oxygen and hydrogen but specifying that
a water molecule can exist without any hydrogen. This
violation could have been eliminated by specifying that a
team requires a team leader and at least one member. Similar
violations exist in the definitions of Project and Project
Plana Project can be composed of zero Phases (i.e.,
composed of nothing); a Project Plan can be composed ofzero Budgets and zero Scopes (i.e., composed of nothing).
Although Purchase Requisition requires its component parts,
the parts cannot exist without the whole, as indicated by the
use of the solid diamond. In Bunges ontology, nothing is
created or destroyed; composite concrete objects are com-
posed from existingconcrete objects (p. 35). The components
of a composite must have independent existence from the
composite itself. Thus the ontologically clear conceptual
model is not consistent with Bunges ontological definition of
reality. To be consistent with Bunges ontology, each part
must have existence prior to being composed. Hence, we
conclude that Team, Project, Project Plan, and PurchaseRequisition as represented in the ontologically clear diagram
are not composites as defined by Bunges ontology.
Finally, although the ontologically unclear diagram (Figure 4)
is purported to violate Bunges ontology because it uses ter-
nary relationships to implicitly represent composites, it need
not be interpreted as representing composites at all (see Ap-
pendix A for a detailed review of the syntax used in Figure 4).
It is wholly consistent with the predominant application of
Bunges ontology to conceptual modeling to view these ter-
nary relationships as representing mutual properties of distinct
things rather than as representing composites (Wand et al.
1999). Thus the ternary relationships in the ontologicallyunclear diagram need not be interpreted as a departure from
ontological clarity. Given these problems with the instan-
tiation of ontological concepts in the ontologically clear and
ontologically unclear treatments (Figures 3 and 4, respec-
tively), we argue that no conclusions should be drawn from
this study regarding the suitability of Bunges ontology as a
foundation for conceptual modeling.
950 MIS Quarterly Vol. 36 No. 3/September 2012
7/30/2019 78164802
7/29
Allen & March/Representing PartWhole Relations in Conceptual Meaning
Department
Employee
Team MemberTeam Leader
Team
Client
Project
Phase Consumable
Project Plan
Budget
Scope
Requisition Line Requisition Header
Purchase
Requisition
Supplier
Require
Has
Requests
Ordered Via
Sent to
Responsible For0..* 0..*
0..*
0..*
0..*
0..*
1
0..1
0..1
0..1
0..*
1
1
1
1 0..*
1..* 1
1..*
1..*1..*
0..*
1
1
1
UML Class Diagram 1
"ET Technologies"
- Client #- Address
- Contact Phone #
- Project Plan #
- Project Plan Version #
- Consumable #
- Consumable Description
- Department #- Department Name
- Employee #- Employee Name- DOB- Internal Phone #- Skill Level
- Qualification Level
- Phase #
- Phase Description- Phase Goal- Work Hours Required- Priority Level
- Project #
- Project Description- Required Completion Date
- Actual Completion Date
- Purchase Requisition #
- Quantity
- Purchase Header #- Date of Requisition- Total Value of Requisition
- Supplier #- Supplier Name- ABN- ACN- Address- Contact Phone #
- Budget #
- Budget Description- Budget Version #
- Team #- Team Name- Team Size- Date Formed
Key Deliverable
Belong To
Assigned To
0..*
1
0..*
- Deliverable #- Deliverable Name- Date Due
1 1
- Scope #- Scope Description- Scope Version #
Department
Employee
Team MemberTeam Leader
Team
Client
Project
Phase Consumable
Project Plan
Budget
Scope
Requisition Line Requisition Header
Purchase
Requisition
Supplier
Require
Has
Requests
Ordered Via
Sent to
Responsible For0..* 0..*
0..*
0..*
0..*
0..*
1
0..1
0..1
0..1
0..*
1
1
1
1 0..*
1..* 1
1..*
1..*1..*
0..*
1
1
1
UML Class Diagram 1
"ET Technologies"
- Client #- Address
- Contact Phone #
- Project Plan #
- Project Plan Version #
- Consumable #
- Consumable Description
- Department #- Department Name
- Employee #- Employee Name- DOB- Internal Phone #- Skill Level
- Qualification Level
- Phase #
- Phase Description- Phase Goal- Work Hours Required- Priority Level
- Project #
- Project Description- Required Completion Date
- Actual Completion Date
- Purchase Requisition #
- Quantity
- Purchase Header #- Date of Requisition- Total Value of Requisition
- Supplier #- Supplier Name- ABN- ACN- Address- Contact Phone #
- Budget #
- Budget Description- Budget Version #
- Team #- Team Name- Team Size- Date Formed
Key Deliverable
Belong To
Assigned To
0..*
1
0..*
- Deliverable #- Deliverable Name- Date Due
1 1
- Scope #- Scope Description- Scope Version #
Figure 3. Experimental Materials, Ontologically Clear Diagram (Source: Figure 5 from Shanks et al.
2008, p. 561)
Construct Operationalization
As discussed above, the theory of ontological clarity predicts
that users of conceptual models will better understand the
domain semantics represented by the diagrams when there is
a bijective mapping between constructs in the modeling
grammar and constructs in the ontology than when there is a
surjective or injective mapping between them. When a bijec-
tive mapping exists in a given diagram, the diagram is said to
be ontologically clear.
Shanks et al. argue that using the relationship construct to
represent both associations among objects and to implicitly
represent composite objects results in construct overload, a
surjective mapping from the ontology to the conceptual
model. While we agree that using the same construct to
represent both associations and composite objects would
result in construct overload, the study does not permit such an
assessment. In each experimental treatment, associations are
represented with one graphical symbol while objects are
represented with another. That is, with respect to the repre-
sentation of association and composite objects there is no
construct overload in either treatment. In the ontologically
clear diagram (Figure 3), objects are represented using the
rectangular class symbol, association relationships are
represented using labeled arcs, and aggregation/composition
relationships are represented using small diamonds on the
composite object side of the relationship. Shanks et al. state
that the ontologically unclear diagram (Figure 4) was pro-
duced from the ontologically clear diagram by converting
each aggregate/composite class to a diamond, that is, a
UML association (p. 561).
On the surface, this appears to operationalize different levels
of ontological clarity by creating construct overload; the rela-
tionship construct is intended to convey two different meanings.
MIS Quarterly Vol. 36 No. 3/September 2012 951
7/30/2019 78164802
8/29
Allen & March/Representing PartWhole Relations in Conceptual Meaning
Department
Employee
Team MemberTeam Leader
Client
Project
Consumable
Budget
Scope
Requisition Line Requisition Header
Supplier
Requests
Ordered Via
1..*
0..*
1..*
0..1
0..*
1
0..*
1..*
1 0..*
11..*
0..*
1
1
Project Plan
Phase
PurchaseRequisition
Team
UML Class Diagram 2"ET Technologies"
- Client #
- Address- Contact Phone #
- Consumable #
- Consumable Description
- Department #- Department Name
- Employee #
- Employee Name
- DOB- Internal Phone #- Skill Level
- Qualification Level
- Project #
- Project Description- Required Completion Date
- Actual Completion Date
- Consumable Type #- Quantity
- Purchase Header #
- Date of Requisition- Total Value of Requisition
- Supplier #
- Supplier Name- ABN- ACN
- Address- Contact Phone #
Key Deliverable
0..*
0..*
- Deliverable #
- Deliverable Name- Date Due
0..*
Belong To
- Budget #- Budget Description- Budget Version #
- Scope #- Scope Description- Scope Version #
Figure 4. Experimental Materials, Ontologically Unclear Diagram (Source: Figure 6 from Shanks et al.
p. 562)
However, in UML a diamond is defined to represent an
association, not a composite object. Hence, the diamonds in
Figure 4 should be interpreted as representing ternary rela-
tionships (associations), even though the authors intend them
to communicate the existence of composite objects. More
importantly, all and only ternary relationships intended to
communicate the existence of composites are represented
using the diamond symbol. All binary relationships (associa-
tions) are represented using labeled arcs. Consequently we
see no construct overload in Figure 4; rectangles are used to
represent objects; labeled arcs are used to represent binary
relationships; diamonds are used to represent ternary rela-
tionships intended to communicate the existence of composite
objects.
We conclude that there are significant problems with the
operationalization of the independent variable, construct
overload.
Experimental Implementation
Shanks et al. indicate that they produced the ontologically
unclear diagram (Figure 4) from the ontologically clear
diagram by converting each aggregate/composite class to a
diamond (p. 561). Figure 3 contains four aggregate/com-
posite classes: Team, Project, Project Plan, and Purchase
Requisition. If the stated procedure had been followed, then
each of these would appear as a diamond in Figure 4. How-
ever, only Team, Project Plan, and Purchase Requisition ap-
pear as diamonds. Project has inexplicably been represented
as a UML class (a rectangle). Conversely, each diamond in
Figure 4 should correspond to an aggregate/ composite class
in Figure 3. Again, one of the four is not mapped according
to the specified procedure. The Phase relationship in Figure 4
corresponds to a class in Figure 3 that is neither aggregate nor
composite. However, if ternary relationships are intended to
implicitly represent composites then, according to the rea-
952 MIS Quarterly Vol. 36 No. 3/September 2012
7/30/2019 78164802
9/29
Allen & March/Representing PartWhole Relations in Conceptual Meaning
soning of Shanks et al., this change has the effect of asserting
that projects are composed of phases in the ontologically clear
diagram, but that phases are composed of projects (and con-
sumables and key deliverables) in the ontologically unclear
diagram. This is problematic itself; however, the implications
for the studys validity are deeper than the introduction of
counterintuitive domain semantics in one treatment.
Recall that Shanks et al. set out to study differences in user
performance when composites are represented explicitly as
classes as compared to implicitly as relationships. To effect
this examination, they implement the ontologically clear
treatment where the four composite entities are represented as
classes (Figure 3). Their experimental protocol calls for a
second, ontologically unclear treatment, analogous to the first,
where the four composites are represented implicitly by using
relationships (Figure 4). However, in Figure 4, Project and
Phase are represented in a way that is inconsistent with the
experimental protocol. As described by Shanks et al., sub-
jects answering questions involving projects should be
examining a class in one treatment and a relationship in the
other. This difference is the basis for drawing inferences
about differences in subject performance. However, because
Project is presented as a UML class in both diagrams, no such
inferences can be made. Moreover, because Phase is repre-
sented as a relationship in Figure 4 even though it is not a
composite in Figure 3, differences in subject performance will
undoubtedly be introduced that cannot be attributed to repre-
senting a composite explicitly as a class versus implicitly as
a relationship.
These errors in the experimental materials might not be
critical if Project and Phase were of only peripheral impor-
tance to the experimental task. However, each of the 11
questions refers directly to either Project or to Phase and most
(6 of 11) refer to both (see Appendix B for the questions used
in the Shanks et al. study). Accordingly, performance on
every part of the experimental task is influenced by this error
in the experimental materials.
Confounds in the ExperimentalTreatment
Two factors confound the experimental treatments: the
imbalanced use of ternary relationships and the experimental
task favoring one diagram in a manner distinct from the
experimental treatment. The ontologically unclear treatment
(Figure 4), with which subjects had inferior performance, uses
ternary relationships while the ontologically clear treatment
(Figure 3) does not. That is, even if the ontologically unclear
treatment represented construct overload, the effects of this
factor would be confounded by the use of ternary relation-
ships. The ontologically unclear treatment would differ from
the ontologically clear treatment in that it would include both
construct overload and ternary relationships while the onto-
logically clear treatment contained neither. The semantics of
ternary relationships are more difficult to understand than the
semantics of binary relationships, especially for novices
(Batra et al. 1990; Rumbaugh et al. 2005; Topi and Ramesh
2002). Thus, even without any ontology-based differences in
the diagrams, one would expect Figure 4 to result in lower
participant performance. The qualitative analysis presented
by Shanks et al. suggests that confusion around the ternary
relationship construct was fundamental to differences in
observed performance. Reported subject statements (p. 568)
included the ternary relationship is what confuses me and
I cant understand why consumables and deliverables are
related (via a ternary relationship). This conjecture is
supported by our first experimental study presented below.
Moreover the questions that make up the experimental task
systematically favor the ontologically clear diagram in a
manner that is independent from the intended experimental
treatment (ontological clarity), further strengthening the
confounding effects. Several of the questions in the experi-
mental task are more difficult to answer in the ontologically
unclear diagram than in the ontologically clear diagram.
Other questions make statements that are incompatible with
the semantics expressed in ontologically unclear diagram,
likely causing confusion among subjects in this treatment.
Details are discussed in Appendix B.
Deviations from UML Definitions
Shanks et al. indicate that they made a single deviation from
standard UML class diagram notation: To avoid having
cluttered diagrams, we placed attributes beside class boxes
rather than in them. Otherwise we have used standard UML
notation (p. 560). However, the experimental materials
exhibit two additional notational deviations that may have had
significant, unintended influence on the studys results.
First is the positioning of the names of the ternary rela-
tionships in Figure 4. In ER notation, the names of relation-
ships appear inside the diamond; however, according to therules of UML the name of the association (if any) is shown
near the diamond not in the diamond (Rumbaugh et al. 2005,
p. 470).5 Of the symbols used in Figure 4, the only construct
5We use Rumbaugh et al. (2005) as the authoritative reference for the UML
because its authors are the creators of the UML and because it is the source
cited by Shanks et al.
MIS Quarterly Vol. 36 No. 3/September 2012 953
7/30/2019 78164802
10/29
Allen & March/Representing PartWhole Relations in Conceptual Meaning
allowed to have writing inside its borders is the rectangular
class symbol. Placing the names of associations inside the
symbol borders could have the effect of making them seem
more like classes. This, in conjunction with using nouns to
name the ternary relationships, encourages the misreading of
the relationships as if they were classes. Such a misreading
is likely to lead to the erroneous interpretation of the car-
dinality constraints, which could cause subjects to conclude
that an instance of a ternary relationship can exist without an
instance ofeach of the related classes (e.g., a phase rela-
tionship instance without a related key deliverable). This is
clearly prohibited by the definition of a relationship and doing
so would result in incorrect answers for several questions.
The second problematic departure from standard UML
notation pertains to multiplicity (cardinality constraints). For
ternary relationships (associations), Rumbaugh et al. state,
the multiplicity on an association end represents the potential
number of values at the end when the values at the other [two]
ends are fixed (p. 470). Clarifying this definition they con-
tinue, for example, given a ternary association among classes
(A, B, C), the multiplicity of the C end states how many C
objects may appear in association with a particular pair of A
and B objects. If the multiplicity of this association is (many,
many, one),6 then for each possible (A, B) pair there is a
unique C value (p. 472).
Consider the Team ternary association from Figure 4. If the
multiplicities are read according to the UML definition, the
0..* near the Project class indicates that each possible
combination of a team member and a team leader need not be
associated with any project, but could be associated with
several, which seems a reasonable constraint. However,
consider the 1..* near the Team Leader class. This notation
signifies that each possible combination of team member and
project must be associated with at least one team leader.
Accordingly, every team member must be associated with
every project according to the UML definition.
However, Shanks et al. describe in footnote 7 (p. 560) that
subjects were taught to interpret the multiplicity symbols of
ternary relationships differently. According to the instruction
that subjects received, they should have read the 0..* near
Project to indicate that any given project might not participate
in the Team association or might participate in it many times.
Likewise, the 1..* near Team Leader should be interpreted
to mean that each Team Leader must participate in the Team
association at least once but can participate in it multiple
times. This definition conforms to the notion of participation
constraints, as specified in some variants of the ER model
(Liddle et al. 1993); however, it differs significantly from the
definition of cardinality constraints in UML (see Ap-
pendix A).
The effects of using nonstandard definition of cardinality
constraints depends on subjects experience with UML.
Subjects with extensive UML experience were likely to be
confused by the nonstandard notation and perhaps apply the
standard definitions to the model, resulting in incorrect
answers for some of the questions. While Shanks et al.
disclose that most of their subjects had prior modeling
experience, including experience with UML, neither the
number of subjects having UML experience nor the level of
that experience is reported. Accordingly, the severity of any
concern resulting from this second departure from the UML
standard is uncertain.
Analysis of Performance
Ascribing a numerical representation to subjects performance
is fundamental to analyzing the results of any experiment.
The validity of the statistical analysis is only as strong as the
scoring of participant performance on the experimental task.
In the study conducted by Shanks et al., the experimental task
required participants to answer 11 questions by referencing
the UML class diagram they received (each participant
received either Figure 3 or Figure 4). Although all 11 ques-
tions are presented, answers were provided for only questions
2 and 10. As discussed below, question 2 was scored incor-rectly for the ontologically clear treatment. Question 2 reads
as follows:
A team leader has resigned. Does the model allow
the team to continue to work on the project without
him?
Shanks et al. indicate that the correct answer for the onto-
logically clear treatment (Figure 3) is possible and that the
correct answer for the ontologically unclear treatment
(Figure 4) is not possible. We agree that the correct answer
for the ontologically unclear treatment is not possible. As
discussed above, the existence of an instance of a relationship
requires exactly one instance of each related entity; hence, a
team cannot exist without its leader. However, that is also the
correct answer for the ontologically clear treatment.
Explaining their rationale for claiming that the correct answer
for the ontologically clear model is possible, Shanks et al.
state that because team leader is weakly aggregated in the
6The designation (many, many, one) for the relationship (A, B, C) indicates
that the cardinality constraint is 0..* on the A end, 0..* the B end, and 1..1 on
the C end.
954 MIS Quarterly Vol. 36 No. 3/September 2012
7/30/2019 78164802
11/29
Allen & March/Representing PartWhole Relations in Conceptual Meaning
ontologically clear diagram, the team can still exist if the
leader resigns (p. 565). This statement is inconsistent with
the definition of aggregation in the UML grammar. Shanks
et al. assert that weak aggregation allows a team to exist
without all of its constituents.7 However, aggregation does
not indicate that the aggregate can exist without its parts; but
rather, that a partmay exist independently from the
aggregate (Rumbaugh et al. 2005, p. 164). Rumbaugh et al.
further explain that constraints such as existence dependency
are specified by the multiplicity, not the aggregation marker
(p. 166). Thus, to answer the question of whether the team
can exist without its leader, the multiplicity of the aggregate
association must be examined. On the Team side of the
relationship the multiplicity (1..*) indicates that a team leader
may be associated with one or more teams; the multiplicity (1)
on the Team Leader side indicates that a team must be
associated with exactly one team leader.8
Accordingly, a team must have a leader to exist and thecorrect answer for question 2 in the ontologically clear treat-
ment (Figure 3) is not possible. Because of this misunder-
standing of the UML notation for aggregation, subjects
responses for question 2 in the ontologically clear treatment
were scored incorrectly. According to Figure 3, the team
cannot exist without its leader.
Without more information about how the other nine questions
were scored, it is not possible to determine if this is an
isolated or a pervasive problem. However, with such prob-
lems in the scoring of participant answers and the afore-
mentioned confounds, the statistical analysis of the experi-
mental results involving subjects performance cannot be
interpreted as supporting the hypotheses.
New Experimental Evidence
We argue that the significant statistical results reported by
Shanks et al. could be explained equally well by the asym-
metric use of ternary relationships in their experimental
materials, as opposed to Shanks et al.s contention they were
caused by construct overload. To investigate this, we con-
ducted two experiments. The first experiment tested the
conjecture that users of conceptual modeling diagrams better
understand domain semantics when expressed using binary
relationships than when expressed using ternary relationships.
Our results, described in more detail below, clearly support
this conjecture.
The second experiment further explicates the effects of con-
founding construct overload and conceptual modeling con-
structs and demonstrates requisite methodological procedures
to test the effects of construct overload while considering the
effects of differences in conceptual modeling constructs. As
described below, the results of this study show that construct
overload does not have a significant effect on subjects
performance, at least in the context studied.
Subjects in both experiments were university students
majoring in Information Systems at a large private university.
Each had received several months training on UML including
binary and ternary associations, association classes, andaggregationthe main concepts used in the experimental
treatments. Neither of the authors was involved in the
training. In all, 33 subjects participated in the first experiment
and 82 participated in the second experiment. There was no
overlap of subjects in the two experiments. In each experi-
ment, subjects were randomly assigned to treatments.
The two treatments used for experiment 1 are illustrated in
Figure 5. Subjects were told that, due to public health con-
cerns, Wasatch Pork Distributors (WPD) had been required to
track which pigs were involved in purchases made by each of
its customers and that two different diagrams were proposed
by competing consultants. Each subject was asked to answer
nine questions about the semantics conveyed by each diagram
(order of presentation was randomized). The decision of
casting subjects in the role of evaluating proposed diagrams
was made to allow subjects to identify what they perceived to
be errors in the diagrams without feeling the need to reconcile
the diagram to the given description of the business domain.
The diagram proposed by Consultant 1 contained a ternary
relationship while the diagram proposed by Consultant 2 con-
tained three binary relationships. Each diagram expresses
similar, though not identical, domain semantics. Both al-
lowed WPD to track which customers purchased portions ofwhich pigs.
Subjects overall performance for the diagram containing the
ternary relationship was significantly worse than subjects
performance for the binary-only diagram (n = 33; p = 0.0002;
within subjects analysis). Subjects performed almost 50 per-
cent better with the binary treatment than they did with the
ternary treatment, correctly answering, on average, 6.6 as com-
7The term weak aggregation is not expressly defined in the UML, although
it may reasonably be used to distinguish aggregation from a stronger from
of aggregation, called composition (Rumbaugh et al. 2005, p. 164). In
general, the term used to distinguish aggregation from its stronger form
(composition) is shared aggregation (Rumbaugh et al. 2005, p. 164).
8When a single integer is used to express multiplicity, it stands for both the
minimum and the maximum value (Rumbaugh et al. 2005, p. 466).
MIS Quarterly Vol. 36 No. 3/September 2012 955
7/30/2019 78164802
12/29
7/30/2019 78164802
13/29
Allen & March/Representing PartWhole Relations in Conceptual Meaning
Ontological Clarity
UML Syntactic Construct Studied Overloaded Not Overloaded
Association Class Figure 7a Figure 9a
Relationship Figure 9b Figure 7b
Figure 6. Experimental Treatment Summary
a. Association Class Overloaded
Association class notation is used to represent a simple association
(i.e., Works For and Contracts With) as well as a composite (i.e.,
Team and Assignment)
b. Association Class Not Overloaded
Labeled arc notation is used to represent a simple association (i.e.,
Works For and Contract With) while aggregation is used to
represent a composite (i.e., Team and Assignment).
Figure 7. Experiment 2 (Form 1): Collaborative Auditing Incorporated
Each experimental form individually replicates the Shanks etal. study illustrating the methodological difficulties they
encountered. These can be overcome simply by a combined
analysis of the two forms. In the first experimental form
(Figure 7), two UML 2.0 class diagrams were presented that
represent a common underlying business domain, each pur-
ported to have been developed by a different consultant.
Figure 7a exhibits construct overload by using association
classes to represent both composition and regular association.
Figure 7b avoids construct overload by using labeled arcs torepresent regular association and the aggregation notation (the
open diamond at one end of a labeled arc) to represent com-
position. Each subject answered the same set of 11 questions
(shown with answers in Figure 8) for each representation. In
all cases the answers to the questions were the same for both
representations, thereby eliminating the possibility of perfor-
mance differences arising as a result of unbalanced counter-
intuitive domain semantics. Because each subject answered
MIS Quarterly Vol. 36 No. 3/September 2012 957
7/30/2019 78164802
14/29
Allen & March/Representing PartWhole Relations in Conceptual Meaning
Questions Answers
1. An account rep is concerned that client employees who have more experience than auditors might unduly influence
assignment findings. Can she list all teams for a given client where the client employee has a hire date that is earlier than
the auditors?
Yes
2. A team is assigned to perform five tasks. Does the model make provisions for more than one auditor to work on some of
those assignments?
No
3. A client entered a contract which will begin in five months time. CAI wants to create a team with an auditor now, and give
the client the next five months to select an appropriate client employee. Is this possible?
No
4. A contract as a number of assignments that can be performed simultaneously. Is it possible for a single client employee to
work on two concurrent assignments?
Yes
5. A client has supplied CAI with the list of employees that will be used for an audit; however, no decision has been made
about which auditor will be assigned. Does the model allow CAI to record the names of the client employees?
No
6. An auditor wishes to check her skill appropriateness for an upcoming contract by checking the tasks to which she will be
assigned. At this point, the client employees have been selected but they have not been paired with auditors to form
teams. Does the model allow assignments to be made before auditors are paired with client employees?
No
7. An auditor and a client employee have been formed into a team on a contract that has been running for 3 months. Upon
reviewing the contract details, they feel that the current set of assignments is inadequate to fulfill the contract. Can they
produce a list of tasks which are planned for the current contract but which have not yet been assigned?
No
8. Because of financial difficulties, a client has requested a reduction in the scope of a contract. CAI is willing to renegotiate
the contract scope. Can each team list the risks of their remaining assignments?
Yes
9. The Account Rep and client on a given contract are working to determine the set of tasks that will be completed to meet
the contract terms. Does the model allow CAI to establish teams before any assignments are made?
No
10. CAI has hired a number of new auditors. Does the model allow CAI to track auditors who are not currently a part of any
team?
Yes
11. A client has requested CAI to perform work that is not included in the task list. Does the model allow CAI to assign a team
to perform that work without creating any new tasks?
No
Figure 8. Collaborative Auditing Incorporated Experimental Questions
the questions for a diagram with construct overload and for adiagram without construct overload, we balance subjects
ability across treatments, reducing exogenous variability in
subject responses. We randomized the order in which sub-
jects received the two treatments to eliminate any systematic
ordering bias.
However, this experimental form suffers some of the same
methodological difficulties as the Shanks et al. study. As
illustrated in Figure 6, it varies two factors, the UML con-
struct studied and construct overload. That is, although we do
not use ternary relationships, our study uses association
classes in only one treatmentthe one with construct over-
load. If subjects perform worse on that treatment, we cannottell if the performance difference is due to construct overload
or to the complexities of the association class notation. Our
experience in teaching UML suggests that association classes,
like ternary relationships, are difficult to understand. To
correct for this methodological concern, we concurrently
administered a second form of the experiment, randomly
assigning subjects to one form or to the other.
In this form (Figure 9), again two UML 2.0 Class Diagramswere presented that represent the same underlying business
domain, each purported to have been developed by a different
consultant. Figure 9a is similar to Figure 7a in that both make
use of association classes; however, Figure 9a avoids con-
struct overload by using association classes to represent
composition and labeled arcs to represent simple associations.
Figure 9b exhibits construct overload by using labeled arcs to
represent both simple association and to represent composi-
tion. Again, as illustrated in Figure 6, this form varies two
factors, the UML construct studied and construct overload.
However, in this form, the more complex construct, associa-
tion class, is not overloaded while the simpler construct,
association, is overloaded. Combining the results from bothforms enables us to differentiate the effects of construct over-
load from the effects of association classes. The complete
experimental design is summarized in Figure 10.
A total of 82 subjects participated in this experiment, each
randomly assigned to one of the two experimental forms.
Each subject answered each of the 11 questions for each of
958 MIS Quarterly Vol. 36 No. 3/September 2012
7/30/2019 78164802
15/29
Allen & March/Representing PartWhole Relations in Conceptual Meaning
a. Labeled Arc Not Overloaded
Labeled arc notation is used to represent a simple association (i.e.,
Works For and Contracts With) while association class notation is
used to represent composites (i.e., Team and Assignment)
b. Labeled Arc Overloaded
Labeled arc notation is used to represent simple association (i.e.,
Works For and Contracts With) as well as to represent composition
(i.e., Team and Assignment)
Figure 9. Experiment 2 (Form 2): Collaborative Auditing Incorporated
the two diagrams in that experimental form. A correct answer
was given a value of 1; an incorrect answer was given a value
of 0. To facilitate a within-subject analysis, each subject
received a score of 1, 0, or 1 for each of the 11 questions. A
score of 1 indicates that, for a particular question, the subject
produced the correct answer under conditions of no overload
and the incorrect answer under conditions of overload. A
score of 1 indicates the opposite. A score of 0 indicates no
treatment effect (the subject answered the question correctly
or incorrectly for both treatments).
If the two experimental forms are evaluated independently,
the effect of construct overload is confounded by the degreeto which subjects understood association classes. Table 1
illustrates the results of doing so. In the first experimental
form, all significant performance differences favored the treat-
ment with no overload and no association classes (Figure 7b
over Figure 7a). However, in the second experimental form
all significant performance differences favored the treatment
with overload and no association classes (Figure 9b over
Figure 9a). That is, if the first experimental form were used,
and the effects of association classes were ignored, then the
analysis would suggest that construct overload is detrimental
to human performance. However, if the second experimental
form were used, and the effects of association classes were
ignored, then the analysis would suggest that construct over-
load is beneficial to human performance. Only by combining
the experimental forms can the analysis separate the effects of
construct overload from the effects of association classes.
The results of the combined analysis are summarized in
Table 2. Each of the 11 questions is analyzed in two ways.
First, combining the results from the treatments with constructoverload (Figure 7a and Figure 9b) compared to the
treatments with no construct overload (Figure 7b and Figure
9a), we performed a within-subjects (paired observations)
analysis, using the scoring calculation described above.
Second, we performed two between-subjects analyses using
a rank sum mean difference scoring, and isolating the effects
of association classes.
MIS Quarterly Vol. 36 No. 3/September 2012 959
7/30/2019 78164802
16/29
Allen & March/Representing PartWhole Relations in Conceptual Meaning
Figure 10. Experimental Analysis Summary
Table 1. Independent Analysis of Experimental Forms
Experimental Form 1 Experimental Form 2
# n Mean Difference P-Value Mean Difference P-Value
1 41 0.12 0.267 0.00 1.000
2 41 0.37 < 0.001* -0.37 0.003*
3 41 0.29 0.004* -0.15 0.210
4 41 0.02 1.000 -0.05 0.688
5 41 0.12 0.227 -.05 0.688
6 41 -0.07 0.509 0.07 0.549
7 41 0.10 0.388 -0.20 0.008*
8 41 0.00 1.000 -0.10 0.289
9 41 0.27 0.013* -0.24 0.013*
10 41 -0.02 1.000 0.07 0.549
11 41 -0.12 0.180 0.15 0.238*Significant at = 0.05.
Note: Mean differences are calculated as performance on treatment without overload minus performance on treatment with overload.
Composition
Association Class
Composition
Aggregation
Composition
Association Class
Composition
Labeled Arc
Composition
Labeled Arc
Composition
Association Class
Composition
Labeled Arc
Composition
Labeled Arc
Figure 7a: n = 41 Figure 7b: n = 41
Figure 9b: n = 41 Figure 9a: n = 41
n = 82n = 82
Within-Subject Analysis
Between-Subject Analysis
Experimental
Form 1
Figure 7
Experimental
Form 2
Figure 9
960 MIS Quarterly Vol. 36 No. 3/September 2012
7/30/2019 78164802
17/29
Allen & March/Representing PartWhole Relations in Conceptual Meaning
Table 2. Results of Statistical Tests for Experiment 2
# N
Within Subjects Analysis
Between Subjects Analysis
Figure 7a vs. Figure 9a Figure 7b vs. Figure 9b
Mean
Difference* p-value
Rank Sum
Mean
Difference* p-value
Rank Sum
Mean
Difference* p-Value1 82 0.061 0.332 8 0.030 -3 0.341
2 82 0.000 1.000 3 0.514 -3 0.401
3 82 0.073 0.377 4 0.384 2 0.615
4 82 -0.012 1.000 -3 0.375 2 0.509
5 82 0.037 0.629 -1 0.770 4 0.313
6 82 0.000 1.000 0 1.000 0 1.000
7 82 -0.049 0.503 0 1.000 -4 0.313
8 82 -0.049 0.455 -2 0.586 -2 0.541
9 82 0.012 1.000 4 0.295 -3 0.515
10 82 0.024 0.832 -1 0.792 3 0.457
11 82 0.012 1.000 -4 0.221 5 0.228
*Null hypothesis: population mean difference = 0. Negative mean difference indicates that raw mean score on the diagram with construct overload
as higher than its counterpart for that question.
Note: Question 1 is not included in the performance analysis. It was included in the experimental treatment to force subjects using the overloaded
diagrams from Figures 7a and 9b to engage the parts of the diagrams that overloaded constructs. Because of the unbalanced complexity this
introduces in Figure 7a, performance differences cannot be attributed to construct overload for this question.
The within-subjects analysis yielded no significant perfor-
mance differences for any question. The p-values ranged
from 0.332 to 1.000 (Table 2, Within-Subjects Analysis).
This provides compelling evidence that, at least in this con-
text, construct overload is not predictive of user performance.
Beyond the within-subject analysis, our experimental design
allows us to conduct two meaningful between-subjects
analyses to further examine any effect of construct overload
while isolating the confounding effects of variations in UML
syntax, specifically association classes. By comparing the
performance of subjects on the overload treatment in experi-
mental form 1 (Figure 7a) with the performance of subjects on
the no-overload treatment of experimental form 2 (Figure 9a)
we can examine the effect of construct overload with only
minor variation in UML syntax. Both treatments used asso-
ciation classes. Likewise, we can compare performance on
Figure 7b with that of Figure 9b to get another look at the
effect of construct overload with very little variation in UMLsyntax. In this comparison, neither treatment uses association
classes. In brief, none of the performance differences is signi-
ficant even at an alpha of 0.20 (Table 2, Between-Subjects
Analysis).
In summary, if construct overload is predictive of human per-
formance in problem solving tasks using conceptual models,
we would expect to see a preponderance of statistically
significant results among the more than 30 tests (question
comparisons) conducted. In fact, we observe none. This
finding strongly suggests that construct overload is not a
salient predictor of human performance in this context.
Hence, based on our own experimental evidence, we argue
that the results presented by Shanks et al. are explained
simply and parsimoniously by the complexities of ternary
relationships and unbalanced counterintuitive domain seman-
tics rather than by any differences in construct overload.
Conclusions
Based on the above analysis of conceptual underpinnings,
experimental procedures, and data analysis, we argue that
there is insufficient evidence upon which to base the conclu-
sions reported in Shanks et al., that their results add strengthto a growing body of empirical work that supports the
usefulness of ontological theories, especially Bunges (1977)
theory, as a means of predicting the strengths and weaknesses
of conceptual modeling grammars and practices (p. 570).
Our own experimental results lead us to the opposite conclu-
sion. Moreover, describing the theoretical underpinning for
their study, Shanks et al. assert
MIS Quarterly Vol. 36 No. 3/September 2012 961
7/30/2019 78164802
18/29
Allen & March/Representing PartWhole Relations in Conceptual Meaning
The theory of ontological clarity is also not a con-
tingency theory. In other words, according to the
theory, instances of construct overload, construct
redundancy, construct deficit, and construct excess
will always undermine users understanding of
conceptual models (p. 557).
Given that we have provided compelling evidence that
construct overload does not result in inferior human perfor-
mance in the use of conceptual models, the theory of
ontological clarity has been falsified (Popper 1963) and must
either be discarded or modified to account for the new
experimental findings. To the extent that prior studies rely on
or support the theory of ontological clarity, they should be
reexamined to determine the conditions under which the
theory is predictive. For example, while we have falsified the
theory in the context of construct overload, Bodart et al.
(2001) argue support in the context of construct excess. All
such studies that have reported support for the predictions ofthe theory of ontological expressiveness should be reviewed
in detail for clues about the proper disposition of the theory.
We again emphasize the importance of experimental research
on conceptual modeling and we appreciate the difficulty of
formulating research questions and appropriately developing
and executing experimental procedures. Our critical com-
ments are offered in the spirit of improving subsequent
research in this area. The difficulties we describe can be
addressed in future studies. If Bunges definition of compo-
sition is to be evaluated as a guide to conceptual modelers in
their representation of phenomena, then that definition must
be applied precisely and accurately and not confounded with
differences in conceptual modeling constructs. Otherwise the
experiment cannot be construed as providing evidence that
supports the use of the ontology as a means of predicting the
strengths and weaknesses of conceptual modeling models and
practices (Shanks et al. 2008, p. 570).
However, we see no compelling evidence to indicate that
Bunges ontology holds significant promise as an underlying
foundation for conceptual modeling. With no construct to
represent conceptual objects or events, we contend that
Bunges ontology suffers significant construct deficit with
respect to the representation of business domains. This is byno means a criticism of Bunges ontology. Bunge developed
his ontology for the expressed purpose of representing objec-
tive, scientific knowledge. It was not developed to represent
human commerce or phenomena that occur in human
industry. It was developed to represent concrete objects and
only concrete objects, objects that exist in space and time,
independent of human knowledge or intentions. We render
no judgment on its value for the representation of such
objects. However, we assert that it was not intended to be
used as a referent ontology for conceptual modeling of busi-
ness systems and using it in that context is a misappropriation.
We contend that social ontology, epistemology, cognitive
theories, and theories of human memory structures and
communication are a more appropriate basis for the evaluation
of conceptual modeling grammars and practices (Allen and
March 2006b; Gemino and Wand 2005). Searle (1995, 2006),
for example, recognizes the significance of socially con-
structed objects such as corporations, contracts, agreements,
commitments, money, and intellectual property in the
execution of human endeavor. Such objects are socially
defined and created by performative acts of social entities.
They depend on human recognition and beliefs (episte-
mology) for their existence. That is, they exist only because
people believe (or agree) they exist. Such objects are expli-
citly excluded from Bunges ontology, yet they are crucial
components of business domains and must be included in
conceptual models.
The fundamental question for researchers, elegantly expli-
cated by Wand and Weber (2002), is how can we model such
domains to better facilitate our developing, implementing,
using, and maintaining more valuable information systems?
To answer this question we must first establish a concep-
tualization for information systems within business contexts.
A fundamental role of an information system is that of a
communication mechanism, providing a medium in which not
only facts but intentions, obligations, commands, and inter-
pretations of facts (knowledge and meaning) are recorded and
disseminated within an organization (Kent 1978). Informa-
tion systems play an active role in business processes (March
and Allen 2007), participating in the gathering of business
intelligence and the execution of business activities based
upon the interpretation of the information gathered. These are
cognitive tasks. Conceptual modeling grammars and prac-
tices should attempt to leverage human information pro-
cessing capabilities rather than merely follow philosophical
conceptualizations of existence.
For example, humans have an innate competency for pro-
cessing events. Human memory for events and past experi-
ences is psychologically and physiologically different fromhuman memory for facts and concepts (Nyberg 1998; Tulving
1983, 2002). Moreover, events are fundamental to narrative
thinking (Robinson and Hawpe 1986) and to the represen-
tation of causality (Pillemer 1998; Ramesh and Browne
1999). Both are principal processes in human sense-making
(Gee 1985). Furthermore, humans use this narrative/event
processing competency as a powerful tool for verbal and
written communication (Orr 1990).
962 MIS Quarterly Vol. 36 No. 3/September 2012
7/30/2019 78164802
19/29
Allen & March/Representing PartWhole Relations in Conceptual Meaning
Ontologically, events are categorically distinct from objects
(Bunge 1977; Davidson 1980). Hence, it can be argued that
using the UML class construct to represent both events and
objects will result in construct overload. However, studies in
psychology (e.g., Wakslak et al. 2006; Zacks et al. 2007;
Zacks and Tversky 2001) indicate that people conceptualize
events and objects in a similar manner. That is, people
ascribe identification and attributes to events and recall them
in the same manner in which they recall physical objects.
They conceptualize events as having parts (sub-events) and
events constitute a significant component of human memory
structures. Using the class construct to represent both events
and objects can be viewed as an abstraction mechanism used
to manage the complexity inherent in the modeled domain.
This has been recognized by a number of researchers (e.g.,
Geerts and McCarthy 2002) and empirical research has
indicated that representing events as entities is beneficial to
human performance in certain information processing tasks
(Allen and March 2006b). We encourage future researchefforts aimed at gaining an understanding of the effects of
using such abstraction mechanisms in conceptual modeling
grammars and practices.
We hope that our assessment of the work reported by Shanks
et al. and the discussion of alternative foundational disciplines
for conceptual modeling will stimulate discussion and will
lead to the design and execution of additional experiments to
test the implications of such proposed foundations.
Acknowledgments
We thank Gordon B. Davis, John V. Carlis, and Toby Teorey for
their insightful comments and suggestions on an earlier version of
this paper. We also thank Sham Navathe and Tom Nadeau for
clarifying the intended meaning of data models used in their
textbooks, Stephen W. Liddle for clarifying the meaning of the
cardinality constraints for ternary relationships in the UML, and
Mario Bunge for clarifying the definition of concrete objects in his
ontology. Finally, we thank the senior editor, associate editor, and
four anonymous reviewers for helpful comments and suggestions.
References
Allen, G. N., and March, S. T. 2006a. A Critical Assessment of
the BungeWandWeber Ontology for Conceptual Modeling,
in Proceedings of the 16th Annual Workshop on Information
Technologies and Systems, Milwaukee WI, pp. 25-30.
Allen, G. N., and March, S. T. 2006b. The Effects of State-Based
and Event-Based Data Representation on User Performance in
Query Formulation Tasks, MIS Quarterly (30:2), pp. 269-290.
Batra, D., Hoffer, J. A., and Bostrom, R. P. 1990. Comparing
Representations with Relational and EER Models, Com-
munications of the ACM(33:2), pp. 126-139.
Bodart, F., Patel, A., Sim, M., and Weber, R. 2001. Should
Optional Properties Be Used in Conceptual Modeling? A Theory
and Three Empirical Tests, Information Systems Research
(12:4), pp. 384-405.Bowen, P. L., OFarrell, R. A., and Rohde, F. H. 2006. Analysis
of Competing Data Structures: Does Ontological Clarity Produce
Better End User Query Performance,Journal of the AIS(7:1),
Article 22.
Bowen, P. L., OFarrell, R. A., and Rohde, F. H. 2009. An
Empirical Investigation of End User Query Development: The
Effects of Improved Model Expressiveness versus Complexity,
Information Systems Research (20:4), pp. 565-584.
Bunge, M. 1977. Treatise on Basic Philosophy: Volume 3:
Ontology I: The Furniture of the World, Boston: Reidel.
Burton-Jones, A., and Meso, P. 2006. Conceptualizing Systems
for Understanding: An Empirical Test of Decomposition
Principles in Object-Oriented Analysis, Information SystemsResearch (17:1), pp. 38-60.
Burton-Jones, A., Wand, Y., and Weber, R. 2009 Guidelines for
Empirical Evaluations of Conceptual Modeling Grammars,
Journal of the AIS(10:6), Article 1.
Carlis, J., and Maguire, J. 2001. Mastering Data Modeling: A
User-Driven Approach, Boston: Addison-Wesley.
Chen, P. P-S. 1976. The Entity-Relationship ModelToward a
Unified View of Data,ACM Transactions on Database Systems
(1:1), pp. 9-36.
Davidson, D. 1980. Essays on Actions and Events, New York:
Oxford University Press.
Elmasri, R., and Navathe, S. B. 2004. Fundamentals of Database
Systems (4
th
ed.), Boston: Pearson/Addison-Wesley.Feldman, M. S., and Pentland, B. T. 2003. Reconceptualizing
Organizational Routines as a Source of Flexibility and Change,
Administrative Science Quarterly (48:1), pp. 94-118.
Fowler, M. 2003. UML Distilled: A Brief Guide to the Standard
Object Modeling Language, Boston: Addison-Wesley.
Gee, J. P. 1985. The Narrativization of Experience in the Oral
Style,Journal of Education (167), pp. 9-35.
Geerts, G. L., and McCarthy, W. E. 2002. An Ontological
Analysis of the Economic Primitives of the Extended-REA
Enterprise Information Architecture,International Journal of
Accounting Information Systems (3:1), pp. 1-16.
Gemino, A., and Wand, Y. 2005. Complexity and Clarity in
Conceptual Modeling: Comparison of Mandatory and Optional
Properties,Data and Knowledge Engineering(55), pp. 301-326.
Halpin, T., and Morgan, T. 2008. Information Modeling and
Relational Databases (2nd ed.), Burlington, MA: Morgan
Kaufmann Publishers.
Hansen, G., and Hansen, J. 1996. Database Management and
Design (2nd ed.), Upper Saddle River, NJ: Prentice-Hall.
Keet, C. M. 2008. Representing and Reasoning Over a Taxonomy
of PartWhole Relations,Applied Ontology (3), pp. 91-110.
MIS Quarterly Vol. 36 No. 3/September 2012 963
7/30/2019 78164802
20/29
Allen & March/Representing PartWhole Relations in Conceptual Meaning
Kent, W. 1978. Data and Reality. Amsterdam: Elsevier Science
Limited.
Kim, Y. G., and March, S. T. 1995. Comparing Data Modeling
Formalisms, Communications of the ACM(38:6), pp.103-115.
Liddle, S., Embley, D., and Woodfield, S. 1993. Cardinality
Constraints in Semantic Data Models, Data and Knowledge
Engineering, (11), pp. 235-270.March, S. T., and Allen, G. 2007. Ontological Foundations for