Accuracy in Modeling with Extended Entity Relationship and Object Oriented Data Models

Transcription

1 Accuracy in Modeling with Extended Entity Relationship and Object Oriented Data Models Douglas B. Bock Southern Illinois University at Edwardsville Terence Ryan Indiana University at South Bend This research compares success in developing conceptual data models for the extended entity relationship (EER) model and Kroenke s object oriented model. A laboratory study was used to evaluate model correctness for 38 subjects divided into two equally sized groups where each group was trained in one of the modeling methods. Modeling correctness is measured in terms of eight different facets of a conceptual data model: (1) entities/objects, (2) attribute/property identifiers, (3) categories, (4) unary one-one relationships, (5) binary one-many relationships, (6) binary many-many relationships, (7) ternary one-many-many relationships, and (8) ternary many-many-many relationships. The EER model provided significantly improved performance for the attribute/ property identifier, unary one-one relationship, and binary many-many relationship facets. A number of research areas are concerned with human factors and database technology. These include information elicitation, query formulation, and data modeling. This paper focuses on data modeling and extends prior data modeling research by investigating the degree of model correctness achievable by entry-level information system professions for two methods commonly used to develop conceptual data models. Conceptual data models primarily use graphical symbols to represent data and data relationships in an implementation independent manner. Model correctness is important for the creation of properly working information systems. A conceptual data model can be incorrect in several ways, for example, the model may fail to identify a class of data items to be represented, or the model may fail to capture fully the semantic meanings of individual data items, or the model may fail to specify correct relationships among data items. Each situation can result in an information system that is inadequate with respect to various system user information requirements. The implications of modeling error underscore the importance of understanding the impact of various conceptual data models on model correctness. One of the most widespread conceptual data modeling methods is the entity-relationship (ER) model first proposed by Chen (1976). The ER model is classified as a semantic data model (Peckham and Maryanski, 1988) because it enables the capture of important meanings about data and relationships among data. Semantic data models alleviate some of the inadequacies of the three major classical data models: the relational, hierarchical, and network models (see Kent, 1979 for a description of record based data model limitations). The ER model provides for the graphical representation of entities (i.e., real or abstract things about which data are collected), entity attributes (properties of entities), and relationships between entities. An extension of the ER model by Elmasri et al. (1985), termed the extended entity-relationship (EER) model, includes the ability to model entity categories (class-subclass). Over the past decade, the object oriented (OO) model has grown in popularity as an alternative conceptual data modeling approach. OO modeling methods evolved from concepts put forth by Smith and Smith (1977) and Hammer and McLeod Manuscript originally submitted October 21, 1992; Revised April 28, 1993; Accepted September 8, 1993 for publication. 30 Vol. 4, No. 4

2 (1981) and are also classified as semantic modeling methods. These original works prescribed a definitional approach to data modeling using a set of extensive descriptive clauses to describe and define data. Recently Kroenke (1991), Coad and Yourdon (1991), and others have introduced various graphical extensions of the OO modeling approach. Object oriented designers represent the world in terms of objects, i.e., real or abstract things about which data are collected. Like entities, objects have properties and there can exist relationships between objects. Example objects include customer orders, invoices, and management reports. Unlike entities, an object may be composed of successively smaller objects, or an object may be a component of more than one larger object. Although both the EER and OO modeling approaches are classified as semantic modeling methods, there are significant differences between the two. The EER approach tasks the designer to identify and build a model out of basic components (entities, attributes and relationships) that, in aggregate, represent system user views about which one desires to store data. The EER method involves view decomposition and view integration, and is very much a building-block approach to modeling. For example, the CUSTOMER ORDER view is composed of the entities CUSTOMER, ORDER, and PROD- UCT, along with various entity attributes and relationships among these three entities. Other views may share some of these entities and attributes, and require the definition of additional relationships. The OO approach allows the modeling task to be accomplished at the object level; thus, the CUSTOMER ORDER would be modeled as an object. While OO modeling also requires a decomposition and integration approach to modeling, i.e. the CUSTOMER ORDER object would undoubtedly have both CUSTOMER and PRODUCT objects as properties, the OO model may be superior to other modeling methods in terms of model correctness since the representation developed by OO modelers is purported to bear a closer semblance to reality. Recent literature has focused on the promise held by the OO approach to improve modeling success and systems development productivity (Wirfs-Brock and Johnson, 1990). Additionally, OO approaches include the ability to model methods (sometimes termed services ) as properties of an object. Methods are the rules or procedures that operate on the data properties of an object. This research compares the EER and OO approaches in a controlled laboratory experiment. Success in modeling by entry-level information system professionals is measured by the degree of model correctness. We proceed as follows. Part two discusses related research. Part three provides brief explanations of the two conceptual modeling approaches. The research method is detailed in part four, and the experimental results are given in part five. The concluding section includes a discussion of results, limitations of the study, and recommendations for future research. Related Research Figure 1: Research Model Recent human-factors information systems (IS) research can be mapped to a research model attributed to Jenkins (1982). Figure 1 gives an adaptation of the Jenkins model by Batra et al. (1990) that aids in framing conceptual data modeling research. Jenkins research framework identifies three classes of variables that may interact: system, decision maker, and task. These three variable classes have a causal effect on the dependent variable performance. Batra et al. replace the system variable with the data model variable. They also substitute the variable human for the decision maker variable in Jenkins model. In this research, performance is operationalized by model correctness. This measures how well the human designer applies the data model approach to the modeling task in developing an abstract representation of reality. Early research by Durding et al. (1977) investigated how people organize data by having subjects organize word sets. Durding found that the semantic structure of an application interacts with the data structure and affects how people organize data. This suggests that model correctness is, in part, a function of the structure of data in an application, and that a data model which provides for ease of use in modeling an application may provide a more correct result. Brosey and Schneiderman (1978) used instance diagrams to investigate comprehension, problem solving situation, and memorization tasks as variables in experiments comparing the relational and hierarchical models. Subjects were classed as programmers or non-programmers. Programmers performed best overall, while the hierarchical model was easiest for the non-programmer group to use. Hoffer (1982) conducted an experiment in which novice subjects developed a conceptual model of an application. Fall

3 Figure 2: EER Solution to the Experimental Modeling Task Subjects were free to select the model approach they preferred (none were suggested by the researcher). While a number of modeling methods were used, subjects most often selected a process flow structure as the model of choice. An important finding was that subjects have difficulty specifying data relationships. This indicates that training is probably important as a factor affecting modeling success. Three other studies report mixed results. Juhn and Naumann (1985) found that semantic data models are superior in relationship existence-finding and cardinality-finding tasks, while the relational model is superior for modeling entity identifiers (key attributes). Shoval and Even-Chaime (1987), however, concluded that the relational model is superior for most tasks. Batra and Davis (1989) reported the use of a relational representation to be favored by both novices and experts who had knowledge of other modeling methods. In investigating some of the reported inconsistencies, Batra et al. (1990) compared model correctness for end-user model builders using either the relational and EER model. The EER model produced significantly better results for representing binary one-many, binary many-many, and ternary onemany-many relationships. There was no significant difference between the relational and EER models in the ability of end-users to represent unary one-one and ternary manymany-many relationships, entity identifiers, or in perceived ease-of-use. Most recently, Palvia et al. (1992) examined novice enduser comprehension of databases represented by three conceptual data models: the ER model, data structure diagrams, and Kroenke s OO model. Novice subjects answered querytype questions about a fairly simple and well-known CUS- TOMER-ORDER-PRODUCT operational model. In general, comprehension was best for the OO model over the other two, although the differences diminished with increases in computing experience. Although the results of previous research are mixed, several findings can be extracted. First, there is no clear evidence that any one data model is dominant, however, it is apparent that, in addition to the data model used, task characteristics and human characteristics can affect model correctness. Second, there is a need to improve the systematic study of the variables that affect model correctness. The Jenkins research model provides a framework for classifying prior research and for identifying variables that affect data model correctness, but this research model is inadequate as a theory to explain model learning and knowledge transfer to new problem domains. In the concluding section we briefly comment on the appropriateness of the theory of transfer, learning and inference from psychology as a basis for improving data modeling research (Gick and Holyoak, 1987). Third, the contradictory results reported, to some extent, may be explained by apparent threats to internal and/or external validity in the various studies. The difficulty in eliminating all threats in an experiment underscores the need to replicate research in order to develop a preponderance of evidence as to the factors most critical to conceptual data modeling success. Fourth, while the semantic data modeling methods appear to provide increased model correctness over classical models, there is a need to extend the examination of modeling approaches to include the OO model. Conceptual Modeling Methods Extended Entity Relationship Model The EER data model was selected as a basis for compari- 32 Vol. 4, No. 4

4 son due to the wide-spread use of the model. Research results for the EER data model enable a direct comparison to earlier research findings. Since the EER data model is well-known, we provide a very brief description of the model. Figure 2 gives an EER model that includes unary one-one, binary onemany, binary many-many, ternary one-many-many, and ternary many-many-many relationships, as well as a category (class-subclass) example. Figure 2 is from Batra et al. (1990) with minor adaptations. The figure uses the EER notation found in McFadden and Hoffer (1991), and gives a solution to the modeling problem found in Appendix A. Entities are denoted by rectangular boxes. Entity attributes may be modeled in ovals and attached to entities by lines, or simply listed close to the entity box and attached to the entity box by lines as is done in Figure 2. Identifier (key) attributes are underlined. Relationships among entities are named and shown in diamond symbols. The lines connecting relationship and entity symbols have a crow s foot symbol to represent a many connectivity, while the absence of a crow s foot indicates a one connectivity; thus the DEPART- MENT entity has many EMPLOYEEs, but each EMPLOYEE belongs to a single DEPARTMENT. Class-subclass categorization is depicted by the ISA hexagon symbol. Object Oriented Model Several different OO model approaches exist. The general limitations of laboratory experiments with respect to the availability of subjects and other resources makes it infeasible to examine all possible OO approaches; therefore, it was necessary to select one approach from among those available. Two approaches were considered. These are Kroenke s (1991) OO model and the Coad and Yourdon (1991) object oriented analysis (OOA) model. Their selection was motivated by the fact that each is well known among information systems and data modeling professionals. The Coad and Yourdon OOA model combines process analysis and conceptual data modeling in a single approach. The symbolism in their diagrammatical method is different, but the resulting representation is quite similar to that of Kroenke in terms of modeling objects and object properties. Coad and Yourdon depict relationships among objects explicitly by various types of arrows and connecting lines, while Kroenke represents relationships implicitly in two ways: (1) by depicting objects as properties of other related objects and, (2) through the use of association objects. The Coad and Yourdon OOA enables the ability to model methods that operate on an object s data properties while the Kroenke OO model does not; however, since the ability to achieve a correct model of object methods was not a focus of this experiment, this feature was deemed unimportant in choosing an OO approach for this research. We found the Coad and Yourdon OOA diagrams more difficult to interpret (partly because they contain more information) than those in the Kroenke OO model. The OOA diagrams do not appear to capture the concept of object encapsulation as well as does Kroenke s OO model. More importantly, Kroenke s approach does not combine data modeling with process modeling. The inclusion of process modeling in this experiment could serve as a confounding variable. Further, the process modeling feature of the OOA approach is not directly comparable to the modeling approached used with the EER model. Hence, we selected Kroenke s OO model for comparison to the EER model. Figure 3 gives an OO solution to the modeling problem in Appendix A based on Kroenke s model. Kroenke defines a semantic object as a named collection of properties that sufficiently describe a distinct identity. Further, a property is like an attribute, but can also be another object (1991, chapter 4). Diagrammatically, an object is denoted by a rectangular box. Named properties of objects are listed inside the box. Kroenke does not address the concept of identifiers of objects. In order to add this dimension to the method, the protocol of underlining identifier properties was followed. Since an object may be a property of another object, this enables the specification of relationships among objects. Note that where the relationship is many in the EER sense, the multi-valued object contained within another object is so indicated with the initials mv for multi-valued as in the EMPLOYEE object that is contained within the DEPART- MENT object. Kroenke s OO model requires designers to recognize the association or nesting of objects wherein one object can be thought to contain a second object, and the second object can, in turn, be thought to contain the first. These association objects enable one to model binary many-many, and higherorder ternary degree relationships among objects, as well as properties that are shared by more than one object (intersection data in the EER sense). An example association object is the BUYS-FROM object that represents the binary many-many relationship between the DEPARTMENT and VENDOR objects, and the property Date_Last_Meeting that is shared by the DEPARTMENT and VENDOR objects. Research Method Research Design A laboratory experiment was conducted to compare relative model correctness for the two models. The experimental design is a derivation of the classical post-test-only control group design with 38 subjects randomly assigned to two groups of equal size, an EER and an OO group. Each group received separate treatments, the treatments being instruction in EER versus OO modeling. The 38 observations provide error probabilities of α=.05 and ß =.20 that a significant difference in treatment effects will not be detected if significant differences are present. The research model in Figure 1 identifies the dependent, Fall

5 Figure 3: Object Oriented Solution to the Experimental Modeling Task independent, and control variables. As noted earlier, the dependent variable is model correctness. Model correctness is defined as the degree to which a conceptual data model developed by a subject captures the semantics about the data as they are represented in a natural language, textual description of the application task. Model correctness is measured at the facet level following the protocol developed by Batra et al. (1990). Examples of facets include correctly identifying and modeling entities or objects, identifier attributes or identifier properties of entities or objects, and relationships among entities or objects. For relationships, model correctness at the facet level is evaluated in terms of both the degree and connectivity of the relationship. Table 1 gives the facets measured in this research. As Batra et al. point out, it is not reasonable to combine facet scores to achieve an overall correctness measure because of a lack of construct validity for such a measure; therefore, the analysis is at the individual facet level. Control variables are human and task. The task variable is held constant by using a single experimental task (see Appendix A). In a study of this type there are many human characteristics which we seek to control or randomize out of the study. Examples include intelligence, motivation to perform well, and field dependence. A primary factor of concern here is the application of individual cognitive problem-solving schemata (Gick and Holyoak, 1987) based on learned knowledge. The experiment controlled the development of individual schemata with respect to the data model learned, but not with respect to a subject s overall general schemata of problem-solving. We relied on randomization of assignment to experimental groups to ameliorate these latter effects. The population of interest focuses on individuals with the skill needed to obtain positions as entry-level IS professionals. This research population is appropriate since entrylevel IS professionals can be expected to participate as modelbuilder team members during the design of information systems. Subjects were drawn from Management Information System (MIS) students enrolled in a senior/graduate MIS course on Database Modeling and Design. The senior students were in their last term of work completing a baccalaureate 34 Vol. 4, No. 4

6 EER Model OO Model 1. Entity Object Experience Level EER OO Total 2. Identifier attribute Identifier property 3. Category (class-subclass Category (class-subclass of entity) of object) 4. Unary 1:1 relationship Unary 1:1 relationship as an object property 5. Binary 1:M relationship Binary 1:M relationship as an object property 6. Binary M:N relationship Binary M:N relationship as an association object 7. Ternary 1:M:N relationship Ternary 1:M:N relationship as an association object 8. Ternary M:N:O relationship Ternary M:N:O relationship as an association object Table 1: Facets of Model Correctness degree in MIS and were in the process of interviewing for entry-level positions in the field. All MIS course work with exception of the senior-level project design course was completed prior to participation in the experiment. Similarly, the graduate students were generally in their last year of study and had some experience in the IS field. As such, these subjects serve as good surrogates for entry-level IS professionals. Most subjects possessed some degree of database modeling experience through exposure in previous MIS courses or through work experience. In order to control this potential confound, the database modeling experience variable is treated as a covariate in the statistical analysis. Subjects completed a questionnaire to measure their prior IS education and IS work experience. This enabled a classification of subjects with respect to database modeling experience. Additional homogeneity of subject groups was achieved by randomly assigning subjects to the two experimental groups on a stratified basis such that each group had equal numbers of undergraduate and graduate students. Table 2 summarizes the subject profile of the two experimental groups. Hypotheses Eight hypotheses are tested. Stated in the null form, there is no significant difference in model correctness for entrylevel IS professionals in developing a conceptual data model using an EER versus OO model, where model correctness is measured for eight different facets: H 1 : entities or objects; H 2 : attribute or property identifiers; H 3 : class-subclass (category of) entities or properties; H 4 : unary one-one relationships; H 5 : 1. No Computing Experience Word Processing, LOTUS, Package Experience 3. Programming, Command Language Database Design Using a PC DBMS Previous Formal Database Course Expert: Formal Database Course and Work Experience Number of Graduate Subjects/Group Number of Undergraduate Subjects/Group Table 2: Subject Profile binary one-many relationships; H 6 : binary many-many relationships; H 7 : ternary one-many-many relationships; and H 8 : ternary many-many-many relationships. Although there are claims in the literature for the superiority of the OO modeling approach (e.g., see Kroenke, 1991, chapter 4), such claims lack a solid cognitive theory basis. This experiment assumes no directionality in the alternative hypotheses. Training and Conduct of the Experiment Each group of subjects received a comparable block of instruction in the assigned model. The block of instruction is a standardized eight-hour curriculum developed in a prior pilot study and includes: TOPIC Time Basic concepts about data, entities (or objects), 3/4 hour relationships, and intersection data. Diagrammatical modeling with the EER (or OO) 3/4 hour approach. Applying EER (or OO) modeling to simple and 3-1/2 hours complex data relationships with examples of unary, binary, and ternary relationships. Data subclasses, superclasses, and generalization. 1 hour Practice modeling exercises. 2 hours TOTAL INSTRUCTION: 8 hours Instruction was presented separately to the two groups. In each case, the groups received four two-hour blocks over Fall

7 ERROR CLASSIFICATION ITEM INCORRECT MEDIUM ERROR MINOR ERROR Entity or Object o Missing o Extra entity o Represented as an attribute or as a relationship Identifier o Missing o Not underlined o Identifier is different from that specified in the task description Relationship o Missing o Incorrect connectivity o Unary relationo Incorrect Degree o Unary relationship ship modeled as modeled by categories attribute (EER) Category of o Missing o Incorrect representation Entity or Object by using relationship symbol (EER only) or by failing to indicate by the OR symbol (OO only) Table 3: Grading Scheme several days. The same instructor presented all instruction for both groups and followed a planned outline. Examples of modeling problems and subsequent modeling exercises were equivalent for both groups. There was no mortality in the experimental groups. Subjects were graded and received credit for the work as part of their database design class. This helped to ensure the high level of motivation that was exhibited by subjects. Following the eight hours of instruction, subjects completed the modeling exercise given in Appendix A. This exercise is the same problem used by Batra et al. (1990) with minor modifications to add intersection data for the two ternary relationships, and enables a comparison to the earlier research. Subjects were allowed an unlimited amount of time to complete the experimental task. Grading Scheme The grading scheme follows the protocol used by Batra et al. (1990) and is given in Table 3. The score for an individual facet varies from zero points for completely incorrect to one point for completely correct. This protocol specifies two classes of intermediate errors, medium and minor. Scores awarded for facets with medium or minor errors are 0.50 and 0.75 points, respectively. Two experts graded each solution. For errors not described by the grading scheme, the experts independently subjectively evaluated the errors according to the effect on the conceptual data model s ability to represent accurately the semantic meaning of the data. Results The independent scores for each facet for each subject were averaged for the two raters and converted to a percentage score. The inter-rater reliability for the scores was 0.96 (Cronbach s alpha statistic). Table 4 gives the comparison results. Note that there was no significant difference in the time required for subjects in either group to complete the modeling task. The experiment-wise statistical significance level was set at α =.05. An analysis of variance for each facet with subject database modeling experience as a covariate was conducted. The covariate was not significant for any of the facets. Following this, multiple matched-pairs t-tests were used to compare model correctness for the two groups of subjects on the eight facets (hypotheses). In the case of multiple dependent comparisons (same subjects), it is appropriate to reduce the significance level used for an individual comparison in accordance with the Bonferroni inequality; therefore the significance level selected for an individual t-test was set at α =.006 to achieve an experiment-wise significance level 36 Vol. 4, No. 4

8 Time Required to Complete the Experimental Task: EER Group OO Group Mean Time in Minutes Standard Deviation Range in Minutes 30 to to 96 Facet EER Group OO Group Mean Std. Dev. Mean Std. Dev. t-value p-value Entity/Object 98% 4% 96% 10% NS Identifier p<.001 Category NS Unary one-one p<.001 Binary one-many NS Binary many-many See Note Below Ternary one-many-many NS Ternary many-many-many NS NS=Not Significant Note: A statistical comparison for the binary many-many facet is not possible due to no variance for the EER group, however, there is a clear significant difference in model correctness between the EER and OO groups for this facet. Table 4: Comparison of the EER and OO Model Correctness Results of α =.047. There was no significant difference in model correctness for five of the facets: entity/object, category, binary onemany, ternary one-many-many, and ternary many-many-many. The EER model provided significantly better model correctness on each of the three remaining facets: identifier attribute or property, unary one-one, and binary many-many. Few previous studies have cited differences in modeling the identifier facet. Indeed, in comparing the EER and relational modeling methods, Batra (1989) found the ability to model identifiers correctly to be quite high (above 90% correctness). Apparently the ability to recognize correctly an identifier for both the EER and relational models goes handin-hand with the ability to recognize an entity. Conversely, subjects in the OO modeling group of this experiment were very successful (96% correct) in modeling objects, but exhibited less success (80% correct) in selecting appropriate identifier properties for the objects. This was particularly true for objects representing categories (subclasses) of other objects; in this case, the ENGINEER and SECRETARY objects that are sub-classes of the EMPLOYEE object. Model correctness for subjects using the EER model for the unary one-one relationship was very high (96% correct) while subjects using the OO model performed significantly poorer (64% correct). The OO model required designers to model this unary one-one relationship as an object like the one termed MARRIAGE in Figure 3. The most common mistake was to model the Date-married data item as a simple property of the EMPLOYEE object, thereby failing to capture the semantic meaning of the unary one-one relationship. The third significant difference was the higher model correctness for the binary many-many relationship for the EER group (100% correct) over the OO group (63% correct). This finding was surprising since it is generally recognized that binary relationships are among the easiest to model. In this problem, the binary relationship between the VENDOR and DEPARTMENT entities/objects was associated with an intersection data item, Date-Last-Meeting, that represents the date on which the last meeting between representatives of a vendor and a department took place. The OO model required designers to represent the relationship as an association object like the one termed BUYS-FROM in Figure 3. The most common mistake was a failure to model this association Fall

9 object. Rather subjects tended to model the Date-Last-Meeting data item as a property of the VENDOR object, thereby failing to capture the semantics of the association of the data item with the DEPARTMENT object. Finally, although there was no significant difference for modeling ternary one-many-many and ternary many-manymany relationships for the two approaches, we note the low model correctness percentage associated with both the EER and OO model. The ternary many-many-many relationship was modeled correctly 79% and 72% of the time for the EER and OO models, respectively. The most common mistake was an inability to recognize the degree of the relationship as ternary. Model correctness for the ternary one-many-many relationship was worse with correctness percentages of 47% and 44% for the EER and OO models, respectively. In addition to failing to recognize the degree of the relationship, when subjects did model the degree of the relationship correctly, they tended to model incorrectly the connectivity of the relationship. Conclusions and Future Directions for Research Before drawing extensive conclusions, it is important to note the limitations of this particular research. The findings are generally applicable to the ER model and its various extensions (like the EER model used here). However, generalization of expected model correctness with other OO models that are diagrammatically different from the Kroenke model should be applied with caution. For example, subject modeling ability with the Coad and Yourdon OOA model may differ significantly from the Kroenke model examined in this research due to the alternate symbols that Coad and Yourdon use to represent object relationships diagrammatically. An additional reason for conservatively interpreting the OO model results is that there are additional features associated with OO approaches that were not examined in this experiment. The importance of object encapsulation was not directly tested. Likewise the importance of the ability to model methods that represent operations that can be performed on objects was not explored. Given the superior results for the EER model over Kroenke s OO model for three of the eight facets, designers should tend to favor the EER model. While the EER findings exhibit a high degree of generalizability, alternative ways of symbolically diagraming OO models may provide improvements in OO model correctness. Neither the EER nor Kroenke s OO model enabled a high degree of success in representing ternary relationships. While the poorer model correctness results for ternary relationships are consistent with those reported throughout the literature, we again urge caution with respect to this result. There is always the possibility that X hours of instruction in modeling ternary relationships is an inadequate treatment. In this research subjects received about 1-1/2 hours of instruction on ternary relationships. It appears reasonable to assume that better correctness scores could be obtained through increases in training. Currently we are conducting an experiment which focuses on techniques that may improve ternary relationship modeling correctness. This same reservation may hold for the unary one-one and binary many-many relationship findings. In general, subjects exhibited problems modeling these all three different types of association objects, i.e. those representing unary, binary, and ternary associations among objects. Another problem that requires investigation is the identification of the type of errors that occur in converting the EER model, or a model derived with any alternative conceptual data modeling approach, to an implementation model. This question is relevant to all implementation models, be they relational, hierarchical, network, or more recent OO implementations. This may be important because errors made in the conceptual data model can affect design implementation correctness during the conversion process. A significant weakness of research in conceptual data modeling is the lack of a sufficient theoretical foundation that would enable a fuller understanding of the learning process that surrounds conceptual data modeling. While the framework in Figure 1 can guide research in terms of classifying prior research, it does not provide this requisite theoretical foundation. Consider the limitations of this research. This experiment investigated the question of which model is better, but an important and unanswered question is why one model is superior to another. This why question is critical to the development of improved instructional methods used in teaching conceptual data modeling principles. A reasonable goal is to improve model correctness across the board for all types of model builders and for all types of data models. We believe the application of the theory of transfer, learning and inference from the field of psychology will provide the requisite theoretical foundation for such research. Our current research efforts focus on the application of this theory to improving our understanding of the conceptual data model process, and the transfer of model knowledge to new modeling domains. The interested reader is referred to Gick and Holyoak (1987) for a thorough description of the theory and a detailed discussion of related psychological research. Finally, there is a need to extend this type of laboratory research to a field setting. This will enable an examination of modeling skill for a subject pool with broader experiences in data modeling and information systems design. Appendix A. Experimental Modeling Task (from Batra et al.) Projects Inc. is an engineering firm with approximately Vol. 4, No. 4

10 employees. A database is required to keep track of all employees, their skills, projects assigned, and departments worked in. Every employee has a unique number assigned by the firm. It is required to store his/her name and date-of-birth. If an employee is currently married to another employee of Projects Inc., then it is required to store the date of marriage and who is married to whom. However, no record of marriage need be maintained if the spouse of an employee is not an employee of the firm. Each employee is given a job title (e.g., engineer, secretary, foreman, etc.). We are interested in collecting more data which is specific to the following types: engineer and secretary. The relevant data to be recorded for engineers is the type of degree (e.g., electrical, mechanical, civil, etc.) and for secretaries is their typing speeds. An employee does only one type of job at any given time and we need to retain information material for only the current job for an employee. There are 11 different departments, and each has a unique name. An employee can report to only one department. Each department has a phone number. To procure various kinds of equipment, each department deals with many vendors. A vendor typically supplies equipment to many departments. It is required to store the name and address of each vendor, and the date of last meeting between a department and a vendor. Many employees can work on a project. An employee can work in many projects (e.g., Southwest Refinery, California Petrochemicals, etc.), but can be assigned to only one project in a given city. It is necessary to track the dates on which an employee works on a project in a given city. For each city, we are interested in its city name, state name, and population. An employee can have many skills (e.g., preparing material requisitions, checking drawings, etc.), but he/she may use only a given set of skills on a particular project. (For example, an employee MURPHY may prepare requisitions (a skill) for Southwest Refinery project, and prepare requisitions as well as check drawings for California Petrochemicals.) An employee uses each skill that he/she possesses in at least one project. It is necessary to keep track of the number of hours that an employee uses each skill in a project. Each skill is assigned a number. A short description is required to be stored for each skill. Projects are distinguished by project numbers. It is required to store the estimated cost of each project. References Batra, D. (1989). The Relational Model Versus The Extended Entity Relationship Model: A Comparison of Representations Developed by Autonomous Users. Unpublished Doctoral Dissertation, Indiana University, Bloomington, Ind. Batra, D.; & J.G. Davis. (1989). Conceptual database design by novice and expert database designers. Proceedings of the Tenth International Conference on Information Systems, Boston, Mass, pp Batra, D.; J.A. Hoffer; & R.P. Bostrom. (1990). Comparing representations with relational and EER model. Communications of the ACM 33(2): Brosey, M.; & B. Schneiderman. (1978). Two experimental comparisons of relational and hierarchical database models. International Journal of Man- Machine Studies 10: Chen, P.P. (1976). The entity-relationship model--toward a unified view of data. ACM Transactions on Database Systems 1(1):9-36. Coad, P.; & E. Yourdon. (1991). Object-Oriented Analysis, 2nd Ed., Yourdon Press. Durding, B.M.; C.A. Becker; & J.D. Gould. (1977). Data organization. Human Factors 19(1):1-14. Elmasri, R.; J. Weeldreyer; & A. Hevner. (1985). The category concept: An extension to the entity-relationship model. Data Knowledge Engineering 1(11): Gick, M.L.; & K.J. Holyoak. (1987). The cognitive basis of knowledge transfer. In Transfer of Learning: Contemporary Research, S.M. Cormier & J.D. Hagman, Eds. Academic Press, pp Hammer, M.; & D. McLeod. (1981). Database description with SDM: A semantic database model. ACM Transactions on Database Systems 6(3): Hoffer, J.A. (1982). An empirical investigation into individual differences in database models. Proceedings of the Third International Conference on Information Systems, Ann Arbor, Mich., pp Jenkins, A.M. (1982). MIS Decision Variables and Decision Making Performance, UMI Research Press, Ann Arbor, Mich. Jung, S.; & J.D. Naumann. (1985). The effectiveness of data representation characteristics on user validation. Proceedings of the Sixth International Conference on Information Systems, Indianapolis, Ind., pp Kent, W. (1979). Limitations of the record based information models. ACM Transactions on Database Systems 4(1): Kroenke, D. (1991). Database Management, 4th Ed., MacMillan Publishing Company. McFadden, F.R.; & J.A. Hoffer. (1991). Database Management Systems, 3rd Ed., The Benjamin/Cummings Publishing Company. Palvia, P.C.; C. Liao; & P. To. (1992). The impact of conceptual data models on end-user performance. Journal of Database Management 3(4):4-15. Peckham, J.; & F. Maryanski. (1988). Semantic data models. ACM Computing Surveys 20(3): Shoval, P.; & M. Even-Chaime. (1987). Database scheme design: An experimental comparison between normalization and information analysis. Database 18(3): Smith, J.M.; & D.C.P. Smith. (1977). Database abstractions: Aggregation and generalization. ACM Transactions on Database Systems 2(2): Wirfs-Brock, R.J.; & R.E. Johnson. (1990). Surveying current research in object-oriented design. Communications of the ACM 33(9): Douglas B. Bock is Associate Professor of Management Information Systems in the School of Business, Southern Illinois University at Edwardsville. He teaches in the areas of database design and information resource management and conducts research on improved methods of conceptual data modeling and software engineering productivity measurement. Terence F. Ryan is Assistant Professor of Management Information Systems in the Division of Business & Economics, Indiana University at South Bend. He teaches systems analysis and design and focuses his research on the effects of information systems development on organizational work systems. Fall

11 Related Content Inclusion Dependencies Laura C. Rivero, Jorge H. Doorn and Viviana E. Ferraggine (2001). Developing Quality Complex Database Systems: Practices, Techniques and Technologies (pp ). Reverse Engineering from an XML Document into an Extended DTD Graph Herbert Shiu and Joseph Fong (2008). Journal of Database Management (pp ). NoSQL Database Phenomenon (2018). Bridging Relational and NoSQL Databases (pp ). Ex Ante Evaluations of Alternate Data Structures for End User Queries: Theory and Experimental Test Paul L. Bowen, Fiona H. Rohde and Jay Basford (2004). Journal of Database Management (pp ). Case Base Management Systems: Providing Database Support to Case-based Reasoners Radha Mahapatra and Arun Sen (1994). Journal of Database Management (pp ).