Experimental Study of using Petri nets and EPC from the End-user Perspective

Table of Content

List of Figures and Tables

List of Appendixes

List of Abbreviations

1 Motivation and Task

2 Theoretical framework
2.1 Levels of representation and formalism
2.2 Focus on the End-User perspective
2.3 Research questions and resulting hypotheses

3 Experimental Design and Methodology
3.1 Basics of the Laboratory Experiment
3.2 Subject information
3.3 Procedure of the experiment
3.4 Data analysis basics

4 Data analysis and result discussion
4.1 Sub question 1: Difficulty of EPC’s OR connector
4.2 Sub question 2: Difficulty of the EPC’s event element
4.3 Sub question 3: Difficulty of identifying concurrencies in Petri nets
4.4 Sub question 4: Usefulness of the tokens in Petri nets
4.5 Sub question 5: Acceptability of the modeling language
4.6 Sub question 6: Over-all superiority of the EPC
4.7 Non-hypothesis related and non-planned results

5 Concluding discussion

Acknowledgements

Literature

Appendix

List of Figures and Tables

Figures

Fig. 1: Examples for the general agreement of a three steps approach

Fig. 2: Oberweis’ 4-layer model in contrast to Scheer’s ARIS model

Fig. 3: Basic coherences of the controlled laboratory experiment

Tables

Tab. 1: Statistical data of the participants

List of Appendixes

illustration not visible in this excerpt

1 Motivation and Task

“Neither the prestige of your topic nor the power of your instruments,
the broadness of your knowledge and the carefulness of your planning will ever substitute the origin of your imagination and the sharpness of your observations”.

Anonymous^[1]

In winter semester 2003/04 a seminar working paper about the ‘relevance of the laboratory experiment in MIS research’ was generated, based on an empirical literature study. One result was that this method is generally established within information systems in the English-speaking language area (especially in the data modeling domain and the formal versus non-formal contestation).^[2] But surprisingly practically no experimental studies could be found, using the laboratory experiment as research method in process modeling.

Based on this cognition, in summer semester 2004 the project EXPEND (Experimental Study of using Petri nets and EPC from the End-user Perspective) was started, intending to realize an experimental study just in that research environment. As a framework for the further study, a comparison of the two process modeling languages EPC and Petri nets (in particular C/E nets) from the end-user perspective was given. Another important target for the study is, in addition to the principle realization of an empirical laboratory experiment, to prove or disprove some of the most prominent hypothesis in the theoretical literature discussion between the supporters of the semiformal EPC and the formal Petri nets. Their theoretical background, justification and transfer to provable hypothesis form the following theoretical framework chapter 2.

Because using the laboratory experiment in the data modeling research provides a set of similarities, the approaches from Green, Petre and Bellamy^[3] and from Moher, Mak and Blumenthal^[4] provided a basic proceeding guideline for the study on hand. Both workgroups compared the comprehensibility of Petri nets against textual program representations within a laboratory experiment. Chapter 3 illustrates the specific experimental design used for the study on hand.

In chapter 4, the results of the data analysis and statistical tests for the particular hypothesis are illustrated and discussed.

The over-all concluding and comprising discussion forms the closing chapter 5.

2 Theoretical framework

2.1 Levels of representation and formalism

There is general agreement among information system researchers that in the business process description and modeling procedure several levels of data representation exist, which range from more business administration-related spheres to those who are related to information technologies. In the German-speaking language area authors like particularly Pohl with his “Three Dimensions of Requirements Engineering”^[5] and Scheer in his ARIS-Architecture^[6] differ between such stages and focus on their exigency. In a different context, but with similar considerations, also ANSI/SPARC’s layered model of database architecture comprises three levels, so called “schemes”. First, there is the user’s view, very close to the business realness, so called an informal level. It is followed by the conceptual and physical schemes, which become more and more close to the structures of the information system (the database).^[7]

But as long as the science agrees with those different levels of representation, the particular degree of formalism is matter of an open and lively discussion, which is part of the general discussion on formal methods.^[8] So at the moment the need and appliance as well as the usefulness of formal or informal methods and modeling languages is not well clarified.

The advocacies of formal methods argue, that the use of formal methods brings the same advantages to software and hardware design that other engineering endeavors have exploited: mathematical analysis, based on models. Formal methods can further be used to specify and model the behavior of a system and to formally verify that the system design and implementation satisfy functional and safety properties.^[9]

The adversaries whereas argue, that especially the clearness and intuitivism of informal or semiformal methods as well as their ability to integrate different views or dimensions speaks in favor for them.^[10]

The general debate about formal methods is, at least for this study, out of the scope. But the specific sub-question about the degree of formalism on the different stages of information presentation is a central point for this assay. Especially the first step out of the business administration sphere is considered, because it is a central point of interest in process modeling.

Exemplary the general agreement about the coherence between the levels of representation and the degree of formalism is displayed in the following two figures. In the left part of figure 1 Pohl’s “coordinate plane of the three dimensions of requirements engineering” is shown, and it is apparent that on the way from the “initial input”, the starting point which lies more in the business administration sphere, to the “desired output”, the aiming point seated by the information technologies, the degree of formalism increases constantly. On the right side, the basic understanding of Scheer’s ARIS-World is shown, also giving a way from the business administration environment to the sphere of information technology. And along this three-stepped path the degree of formalism increases as well as the formal requirements on the particular used modeling languages.

illustration not visible in this excerpt

Pohl: 3 Dimensions of RE Scheer: ARIS – Basis Understanding

Figure 1: Examples for the general agreement of a three steps approach^[11]

A contrary position to this engages Oberweis, who pleads for an appliance of formal methods on every stage of the analysis and configuration of business information systems. He proposes several forms of Petri nets as a general interface between the semiformal, also for a non-trained information systems end-user understandable, chiefly graphical modeling methods for the design of application-orientated business processes, and the formal machine-orientated programming languages for workflow management-systems.^[12] Petri nets provide a well-known graphical technique, often used in distributed systems and they are a fully formal technique.^[13] But, and this is the crux of the matter for the actual study, even himself did appreciate that not all requirements on modeling languages on every stage of the process can be arranged with each other in one “almighty” method. So it is hard to get the requirements for “a good analyzable and validationable method” and for “an easy comprehensible and intuitionally method” together.^[14] Oberweis’ position is illustrated in the following figure 2, and in contrast Scheer’s ARIS position, representing the opposite mindset, is shown again on the right side.

illustration not visible in this excerpt

Figure 2: Oberweis’ 4-layer model in contrast to Scheer’s ARIS model^[15]

2.2 Focus on the End-User perspective

As mentioned yet, this study concentrates on a specific cut-out – the process modeling from the end-user perspective. This cutout was selected because of two reasons: First, process modeling, especially in the ERP-environment, is an ongoing topic in MIS science. The (formal) representation and also the end-user related illustration or mapping of workflows and business processes has become a fundamental component in the implementation and later the enhancement of ERP-systems. But also in other scopes inside companies and in the public sector (for example clinical pathways for differentiated accounting in the healthcare sector^[16]), process modeling has become more and more important during the last years. Secondly, process modeling represents a sort of “interface”, so to speak it considers the borderline between the business administration sphere and the first step of representing the business processes and workflows in a specific modeling language.

This representation can be embellished whether semiformal or formal. Reconsidering the two contrary positions about this topic, the above-all research-question for the study reads as follows:

For end-user orientated process modeling, which notation method – a semiformal or a formal one – is better qualified for the specific communication between the end-user and the information systems designer?

So a first topic is to clarify the boundary between the terms “semiformal” and “formal”. System decomposition, abstraction, and information representation lead to sub-problems that can be addressed using formal methods and tools, such as mathematical modeling, control law synthesis, and control implementation verification.^[17] Methods and tools, which rely basically on mathematical formulations of the underlying problem and which provide a formal language for describing a software artifact (e.g. specifications, designs, source code) in a way that formal proofs are possible, are understood as formal methods. They need to be complemented by semiformal techniques that address the problem of system integration, mainly by modeling the system with an intuitive and user-friendly graphical representation.

As shown, both perspectives have devotees and adversaries. Because a standalone theoretical discussion, as it was lead in the last years, seems not to provide results, this EXPEND is about to clarify the discussion by arranging a laboratory experiment, also for empirical studies are an “important tool for understanding the nature and efficacy of software engineering”^[18] in the United States, and no obvious reason exists why this should be different in here.

Therefore, a concentration on explicit exponents of both “paradigms” is necessary, as reputed, generally accepted and used in practice as possible. So as formal method B/E Petri nets and as semiformal method EPCs will be considered in the following. On the basis of those two methods, the central contentious issues can be emphasized, because there is a similar contestation between the adherences of both notations.^[19]

2.3 Research questions and resulting hypotheses

Now the superior research question, which notation – the semiformal EPC or the formal B/E Petri net – is more suitable for the end-user communication, is splitted into (mostly literature based) sub questions, which in turn will be transferred into the particular research hypotheses.

Sub question 1 : Difficulty of EPC’s OR connector

EPC’s OR connector is often blamed in literature for disregarding the locality principle. Because this implicates an inconsistent semantics, a continuous formalization of the EPC would be hard or even impossible.^[20] This charge is theoretically undoubted correct, but the question is, if this inconsistence got effects on the individual comprehension of the EPC through end-user, who often do not even recognize such discrepancies? Does the semiformality restricts the applicability not only theoretical but also practical? We do not think so and thus the first hypothesis reads as follows:

h1: The disregarding of the locality principle by EPC’s OR connector has no effect on comprehending the process-model through end-users.

Sub question 2: Difficulty of the EPC’s event element

Basically the EPC consists of two elements: functions, which mirror the business activities and events, which got two assignments: they represent the actual system state and, at the same time, they act as release for the following function. The EPC concept does not provide a graphical or otherwise syntactical differentiation between both characters. But a wrong assignment can lead to misinterpretations of the process flow.^[21] The question is if this difficulty, which is also undoubted theoretically correct, actually appears empirically and shows an effect on the probands’ process understanding. Again we assume the probands not to have problems in differentiating between both states.

h2: A clear assignment between the state- and the release-character of the EPC event element is possible and will not lead to misinterpretations.

Sub question 3: Difficulty of identifying multi-level concurrencies in Petri nets

The theoretical acquisition and identification of concurrencies, especially those with multi-level branching, is fundamental for end-users to comprehend complex process chains. The EPC seems to have an advantage here, because of its syntactically simple and graphically well cognizable connectors, which enhance the appreciation particularly for untrained end-users. Petri nets whereas, which realize the concurrencies through more or less complex transitions (and from a formal point of view are superior to EPCs in this case^[22]), are considered as difficult and incomprehensible, at least for end-users.^[23] So the question is, if differences in comprehending the process chain between Petri net and EPC concurrencies, mainly those with multi-level decisions) will appear in the experiment. We think they will, and so the hypothesis is:

h3: Multi-level concurrencies will lead to interpretation and comprehending problems in the Petri net group.

Sub question 4: Usefulness of the tokens in Petri nets

Guiding tokens through the Petri net serves primarily the formal process control and can be used additionally for testing the plausibility (Possible plausibility questions could be: Terminates the net? Can all tokens get through the net or exist deadlocks?). Furthermore the tokens may have a positive effect on the process comprehension, because they indicate the current status of the net, the transitions and the actual positions of concurrencies. We assume that this visual enhancement helps end-users to orientate themselves within the process model.

h4: Using tokens in Petri nets helps improving the end-user comprehension for the process model.

Sub question 5: Acceptability of the modeling language

Wehler represents the hypothesis, that the individual acceptance of a specification is decided at first on its readability and comprehensibility and after that on its completeness – not on its formal correctness.^[24] In case we assume Petri nets to be formal correct and theoretically more sophisticated but in return less comprehensible, EPCs would be superior to Petri nets. And because in the research field of end-user communication and modeling, the acceptance of a particular method is a significant variable, the question is, if the EPC in fact is more accepted by mainly untrained end-user than Petri nets are. We think the probands will favor the EPC, and so the hypothesis reads as follows:

h5: The subjective acceptance of the EPC by end-users is significant higher than that of Petri nets.

Sub question 6: Over-all superiority of the EPC in the specific end-user domain

On the basis of the expected results of hypothesis h1 to h5, we assume the EPC group to archive significant better results in an over-all comparison of both experimental groups, than the Petri net group will. The relative simple access to this modeling language and its intuitive graphical notation will help the untrained probands in comprehending the process flow and the concurrencies, while the Petri net group will have problems with the more sophisticated notation. So we expect the Petri net group to answer the questions inferior that their competition group.

H6: The EPC group will achieve significant better results in an over-all consideration than the Petri net group.

3 Experimental Design and Methodology

3.1 Basics of the Laboratory Experiment

The study was prepared and written over a timeframe of about eight month, in which the main experiments took place on two days, for the whole examination was spitted into two parts: the EPC part and the Petri nets part. As mentioned before, the aim was to compare the reaction of mainly untrained end-users if they are confronted to the semiformal modeling method EPC and the formal language Petri net. Therefore, a classical controlled laboratory experiment was chosen as research method.^[25] The basic coherences of this empirical method are shown in the following figure:

illustration not visible in this excerpt

Assigned to the study on hand, the changing causing factor and thus the independent variable are the two different process-models (one designed with EPC and one with Petri net) shown to the probands (they form the controlled variable). The research hypotheses were given in the theoretical framework chapter. The dependent variable is the correctness of the answers concerning the process-model, given from the subjects through a multiple-choice questionnaire. Thus, the acceptability of the particular process modeling method can be operationalized.

3.2 Subject information

Fifty students from the department of jurisprudence and economics of the Johannes Gutenberg-University in Mainz participated in the study, all with a similar background, but some with more practice in information systems modeling than the others. So additionally the effect of experience to the results could be measured. The following table gives a survey of the participant’s background:

illustration not visible in this excerpt

To get the necessary number of subjects for significant results additionally to the advertising, which consists in electronic and paper-based placards, an incentive system was designed: each student received a confirmation for participating in the experiment and in addition an access authorization for the faculty’s computer pools, amounting to five Euros. The registration for participating in the study was realized per PHP web form- reply on the ISYM homepage.^[27] Out of the interested students, a homogeneous number of participating probands was selected. The selection criteria consist in whether the interested person has selected MIS as elective subject or participated in lecture MobIS before, so both groups reached a similar level of experience with modeling.

The behavior of students in an experiment is surely not the same behavior we would expect of end users in practice, because there are differences in learning style, exposure and motivation, maybe also in skills.^[28] But we believe, however, that the occurring differences between both modeling languages in the experimental treatment would also be present in a field setting.

3.3 Procedure of the experiment

The laboratory experiment last about one hour per group (it took place in a normal lecture room where the participants sit close together, so there might be a theoretical possibility the probands transcribed from each other, but the supervisor could not recognize such attempts). The participating probands did not know each other and got no further information about the experiment before, so no preparation was possible.

In a first step the subjects were given a short introduction to the study and to the syntactic and semantic of the particular modeling-method. The introduction last about 20 minutes, so that the basic understanding of the modeling technique could be well communicated and the group could be situated on a consistent and homogeneous basis.

Then each student was given an identical “experimental-set”, consisting of a multiple choice-questionnaire and a large printout of the business process, for the first group modeled with EPC, for the second group with a B/E Petri net. The used process “Supply logistics” originates from Scheer’s basics book “Wirtschaftsinformatik” 1994^[29] and was also used in the theoretical essay about the differences and problems in process modeling with EPCs and Petri-Nets from Langner, Schneider and Wehler 1997^[30]. Some additions and modifications were made, so that all situations needed for testing the hypothesis could be covered. The multiple-choice questionnaire contains of three typed of questions: statistical questions about the proband (age, gender, etc.), process-related questions (which refer to a specific situation in the process model) and non process-related questions (which aim at the individual understanding and impressions of the proband concerning the applicability of the particular modeling language).

Additionally each student got a short notepad with the most important facts to know about the modeling method, because we had to act on the assumption that not all students would remind the meaning of the modeling syntax from the introduction, while later answering the questions.^[31]

After starting the experiment, the probands got unlimited time for answering the questions, so no pressure to perform was constituted. We assumed the students to need about 40 minutes for answering the questions, but the average disposal-time was about 25 minutes. From 28 participating students in the EPC group, 25 questionnaires were valid and used for the analysis. In the Petri net group 25 students participated, all questionnaires were valid.

3.4 Data analysis basics

The analysis of the multiple choice-questionnaires was arranged with SPSS 11.5 for Windows. The scale-level of the answering-options was ordinal, the sample did not show a normal distribution (detected per non-parametric Kolmogorov-Smirnov test^[32]). So for testing the hypothesis in the data analysis and results discussion part of the study, only non-parametric tests could be used. Therefore, the non-parametric U-test by Mann and Whitney^[33] and the Spearman-Rho test for correlations^[34] were applied.

Additionally a statistical analysis (also calculated per SPSS), consisting of the average value and the standard deviation, was used for identifying trends.

The detailed analyses and values of all hypothesis tests and examinations (with complementary graphics) can be found in the appendix; in the following data analysis and discussion chapter the results are mentioned without explanatory statement.

At the end of each questionnaire a question block O (for Open block) was situated. It consists of two areas: first, a free area where the probands could note down impressions and experiences within the experiment, which are not covered through the foregoing questions; secondly, a question concerning the organization and execution of the experiment as well as another free area for annotations. Those results and annotations of the probands can be found only in the appendix, because they did not participate in the hypothesis testing and analysis.

4 Data analysis and result discussion

4.1 Sub question 1: Difficulty of EPC’s OR connector

Testing procedure

Sub question 1 was tested in a two-stage process. First, the process-related questions H1.1, H1.2, H2.1, H2.2, H3.1 and H3.2 of the EPC group were analyzed by a formal testing method; secondly the non process-related, more individual impression testing questions H7.1, H7.2 and H7.3 were analyzed statistically.

To prove or disprove the hypotheses it was necessary to find out, if the XOR- (H2.1, H2.2) and respectively the AND-related questions (H1.1, H1.2) were superior answered in average than the OR-related questions (H3.1, H3.2). Therefore, the cumulative values of the OR-related questions and those of the XOR- and AND-related questions were accumulated. Afterwards a Mann-Whitney and Wilcoxon-test for paired samples was used to check if the answers to the OR-related questions were in average below or superior to those of the XOR- and respectively the AND-related questions.

Additionally the analysis of the non process-related questions H7.1 to H7.3 was consulted for testing the hypothesis. The three questions aim at the textual distinction between the connectors and at therewith connected consequences.

Results

The results of the Mann-Whitney and Wilcoxon test showed that there is a clear and significant difference between the results of the OR-related questions compared to the XOR- and the AND-related questions. To analyze the strength of this difference, the average values and the standard deviations of the answers were calculated.

Both values of the AND-related questions indicate that nearly all probands (definitely more than 95 percent) answered the questions right. The XOR-related questions were even completely (100 percent) answered correct. But the OR-related questions were answered correct by only 79 percent of the probands, with an obvious higher standard deviation.

The evaluation of the non process-related questions backed this trend, because the only problem field (from the probands’ point of view) exists in the differentiation and assignment between the XOR- and the OR-connector.

Discussion

Obviously the hypothesis could not be verified, because the disregarding of the locality principle is not only theoretically a problem, also the untrained probands realized the discrepancy between the OR-connector and the AND-, respectively the XOR-connector, which are both semantically correct. So the semiformality of the EPC, which is often praised for its advantages especially for end-users, seems to have a disadvantage at this point.

4.2 Sub question 2: Difficulty of the EPC’s event element

Testing procedure

For testing the hypothesis, the answers to the process-related questions H5.1 and H5.2 of the EPC questionnaire were evaluated. They ask for the probands notional association of the EPC’s event module. In both questions, three out of four possible answers were false, one answer was correct. A statistical analysis and interpretation should clarify the probands’ degree of understanding the bivalent character.

Results

With clearly over 80 percent in each case, both questions were answered correct. The standard deviation is comparatively high, but this can be traced back on the minor sample. But none the less there is a definitely and strong trend, that a (definitely not marginal) minority of about 20 percent of the probands did not comprehend the bivalent character.

Discussion

Although the majority of the probands answered the questions right, the theoretical risk of misinterpretations, causing out of the missing semantical differentiation between the state and release character, is apparent. The positioned hypothesis could not be approved. The wrong assignment can lead to misinterpretations of the process flow; the probands’ process understanding is affected. So Schütte, who described the problem theoretically^[35], can be backed tendential, even if the result is not absolutely significant, because of the high standard deviation.

4.3 Sub question 3: Difficulty of identifying concurrencies in Petri nets

Testing procedure

For testing the hypothesis, a comparison of the EPC and the Petri net group questionnaires concerning the answers to H3.1, H3.2 and H3.3 (EPC group) respectively H3.4 (Petri net group) was drawn up. All three questions refer to the same situation in the particular process models, and were equal posed in the questionnaires of both groups. They ask for the probands comprehension of multi-level concurrency situations. A statistical analysis and interpretation should clarify the probands’ degree of understanding those situations.

Results

All three questions were inferior answered in the Petri net group than in the EPC group. The standard deviation is comparatively high in all three questions, which results out of the comparatively small sample, so the result must be qualified. Particularly in question H3.2 it is a significant difference between both groups, in Questions H3.1 and H3.3 it is more a (strong) trend, which is apparent but fragile. But because all three questions tend to the same direction, the result is definitely.

Discussion

Although the result is not absolutely significant, the hypothesis is backed by a strong and definitely trend. Because the theoretical acquisition and identification of concurrencies with multi-level branching by the probands was much better in the EPC group, Petri nets seems to have a disadvantage here. Especially for the end-user-environment, where mostly untrained users possibly get in touch with a modeling language for the first time, Petri nets often are considered as too difficult and incomprehensible.^[36] The results support this point of view. The EPC group archived better results, so the often praised advantages (syntactically simple, graphically well cognizable) appeared also in the experiment.

4.4 Sub question 4: Usefulness of the tokens in Petri nets

Testing procedure

For testing the hypothesis, three different approaches were used. At first, the non process-related questions H4.1 to H4.4 of the Petri net questionnaire aim for the principle understanding of the fundamental rules in C/E nets and therewith the function and appliance of the tokens. Next, question H5.3 reflects the probands individual impression of the tokens. And last, the non process-related questions H5.1, H5.2 and H5.4 are a comparison of the answers of the Petri net and the EPC group, concerning possible problems or misunderstandings of the process progress and decision points. Again, a statistical analysis and interpretation should clarify the probands’ personal impressions concerning the usefulness of the tokens in Petri nets.

Results

The analysis of questions H4.1 to H4.4 arises, that in average 80 percent of the probands answered the questions correct. But a group of about 20 percent of the probands had problems and failed. The standard deviation again is comparatively high, so the result is not absolutely significant, but definitely.

In question H5.3 clearly over 30 percent of the probands stated explicitly they would understand the process better, without using tokens.

The analysis and comparison of the three questions H5.1 EPC group (H6.1 Petri net group), H5.2 (H6.2) and H5.4 (H6.4) does not show any trend. The results of the Petri net group and the EPC group are too close together and the standard deviations are at any one time too high to get even a weak trend.

Discussion

The first result lead to the trend, that the principal understanding of the fundamental Petri net rules and the function and appliance of the tokens seems to be understood by the majority of the probands, but the 20 percent who answered false are considerable. It seems that the Petri net rules are not as easy to understand, as they were understood by all end-users when they are confronted with them for the first time.

Questions H5.3 arises the trend that many end-users seems to have problems with the tokens, although 40 percent accept they as a support for their notional process. So for the moment it is not possible to come to a clear statement.

Concerning the prove or disprove of the hypothesis, the three comparable questions did not show any trend. So in an over-all consideration of the results, the hypothesis that using tokens in Petri nets helps improving the end-user comprehension for the process model or flow can not be backed. Because even a comparatively strong group of probands states they would understand the process better without using tokens, the usefulness of tokens for the specific end-user environment must be challenged. The possible visual enhancement did not appear.

4.5 Sub question 5: Acceptability of the modeling language

Testing procedure

For testing the hypothesis, the probands individual impressions and personal opinions concerning both modeling languages, asked in the non process-related questions H7.4, H7.5, H7.8, H7.10 and H7.11 of both questionnaires, were compared between the groups. In all five questions, the probands were asked, how much they back the particular statement on a tetravalent scale. A statistical analysis and interpretation should clarify the differences between both groups.

Results

Questions H7.4 and H7.5 yield the result that the majority of the probands perceive the EPC as the more suitable and comprehensible modeling language for business process modeling (96 percent) and felt less insecure while answering the process questions (84 percent). Both results are significant. Questions H7.8 tends towards EPC. Only 12 percent of the probands would not use the EPC in a company to illustrate business processes and workflows for employees, versus 32 percent of the Petri net group. Question H7.10 shows a balanced image, in both groups about 34 percent of the probands think a description of the process in a natural language they would have better understood. The last question, H7.11 also tends towards EPC. About 80 Percent of the probands think that the method is not only appropriate for modeling and planning processes per experts, but also for illustrating processes for (untrained) employees, compared to 64 percent of the Petri net group.

Discussion

Four out of five questions back the hypothesis that the subjective acceptance of the EPC by mainly untrained end-users is significant higher than that of Petri nets. So Wehlers conclusion, that the individual acceptance of a specification is decided at first on its readability and comprehensibility and after that on its completeness, not on its formal correctness^[37], can be approved by the results. Also our own assumption, that because Petri nets may be formal accurate and theoretically more sophisticated but in return less comprehensible, the EPC would be superior in the probands point of view, turned out to be true.

4.6 Sub question 6: Over-all superiority of the EPC

Testing procedure

For testing the hypothesis of the over-all superiority of the EPC, first a Mann-Whitney U-test was made to analyze if there is a principle difference between the over-all results of the Petri net and the EPC group. Therefore, the sum values of all comparable process-related questions from both questionnaires were accumulated and compared. Successional, after realizing that there is a difference, a descriptive statistical analysis should show the strength and characteristics of the difference.

Results

The Mann-Whitney test yields the result, that there is a significant difference between both sum values (2-sidet asymptotic significance value: 0.006; a strong significance). The following descriptive statistic analysis showed, that the EPC group results a significant better amount than the Petri net group. Although the standard deviation is relatively high, which is because of the comparatively minor sample, the result is definitely.

Discussion

Because of the clear result, the hypothesis that the EPC group will achieve significant better results in the over-all consideration is fully backed by the results. The assumption, that because of the relative simple access to the EPC and its intuitive graphical notation the EPC probands will comprehend the process flow and the (multi-level) concurrencies better than the Petri net group, could be approved.

4.7 Non-hypothesis related and non-planned results

Testing procedure

To analyze possible influences of the probands’ statistical data (age, gender, study course, semesters, MIS as an elective subject, lecture MobIS attended, apprenticeship and/or internship, experiences with the modeling technique and knowledge of other modeling languages) on their answers, a non-parametric Spearman-Rho test for correlation was made. For the parameters, where a correlation was asserted, again a Mann-Whitney U-test was arranged and after that a descriptive statistical analysis should show the strength and characteristics of the correlation.

Results

The Spearman-Rho test yields in result, that a correlation between the probands’ gender and the questionnaire exists in both groups, but in each case oppositional. The negative value in the Petri net group indicates, that female probands seems to have more problems with Petri nets than their male colleagues. Otherwise the women in the EPC group archived better results. Another positive correlation, between the experience with the modeling technique and the results, appeared only in the Petri net group.

Sequencing the particular statistical data was analyzed to clarify, if the correlations are significant or just a (strong) trend. Both correlations concerning the gender and also the one concerning the experience are just under the limit of significance, so they form only a strong trend. On the basis of this awareness, a Mann-Whitney U-test should clarify, if one gender is superior over both groups in a direct comparison. But no significant correlation between the gender and the results in the over-all consideration appeared. Also the consideration of the other statistical data concerning both groups lead to no result.

Discussion

The strong trend correlation between the gender and the results of the particular questionnaires are not completely surprising. Petri nets are considered as formal, mathematical and not very user-friendly. Among psychologists exists the general agreement, that the male mindset can manage such problem situations tendential better than the female one. Otherwise, women are considered as more intuitive and adaptive, exactly the adjudicated characteristics of the EPC, and so they perform better in the experiment. But consolidated and in an over-all consideration, none of the gender is superior.

The strong trend correlation between the experience with Petri net modeling and the results in the experiment is also not astonishingly. It indicates that a longer learning-process is needful and necessary to understand and adopt Petri nets in the same way as other, more intuitive modeling techniques, such as the EPC.

[...]

---

^[1] Inscription at the Division of experimental Medicine, McGill University Montreal, Canada.

^[2] Cp. Berry, Tichy: 2003.

^[3] Cp. Green et al.: 1991.

^[4] Cp. Moher et al: 1993.

^[5] Cp. Pohl: 1994, p. 3 et sqq.

^[6] Cp. Scheer: 2001.

^[7] Cp. Tsichritzis: 1977.

^[8] Cp. for example: Hall: 1990; Bowen, Hinchey: 1995; Heitmeyer: 1998.

^[9] Cp. Butler et al.: 2002, p. 6.

^[10] Cp. v. Uthmann: 1997.

^[11] Own presentment, based upon Pohl: 1994, p. 4 and Scheer: 2001.

^[12] Cp. Oberweis: 1996.

^[13] Cp. Peterson: 1977.

^[14] Both quotes: Oberweis: 1996, p. 11.

^[15] Own presentment, based upon Oberweis: 1996, p. 174 and Scheer: 2001.

^[16] Cp. Sloane, Wagner: 2004, p. 1.

^[17] Cp. Sastry: 1998.

^[18] Berry, Tichy: 2003, p. 1.

^[19] Cp. for example: Moldt, Rodenhagen: 2000; Langner, Schneider, Wehler: 1997.

^[20] Cp. for example: Rittgen: 1999; Rittgen: 2000; Langner, Schneider, Wehler: 1997;
Chen, Scheer: 1994.

^[21] Cp. Schütte: 2001, p. 101 et sqq.

^[22] Cp. Wehler: 2000.

^[23] Cp. Moldt, Rodenhagen: 2000.

^[24] Cp. Wehler: 2000, pp. 1-2.

^[25] For details about the laboratory experiment, see for example: Roth: 1995.

^[26] MobIS : Modellierung betrieblicher Informationssysteme – Modeling of business information systems

^[27] The PHP registration web form can be found in the appendix.

^[28] Cp. Berry, Tichy: 2003, p. 1.

^[29] Cp. Scheer: 1994.

^[30] Cp. Langner, Schneider, Wehler: 1997.

^[31] All materials can be found in the appendix.

^[32] For the detailed analysis see appendix; The Kolmogorov-Smirnov test was originally proposed in the 1930's; see Kolmogorov: 1933 and Smirnov: 1936;

^[33] The Mann-Whitney U-test is one of the best-known non-parametric statistical significance tests. It is sometimes also called the Mann-Whitney-Wilcoxon test. For the origin papers see: Mann, Whitney: 1947 and Wilcoxon: 1945.

^[34] The test for Spearman's rank correlation coefficient, often denoted by the Greek letter ρ (rho), is a non-parametric measure of correlation; see Spearman: 1904.

^[35] Cp. Schütte: 2001, p. 101 et sqq.

^[36] Cp. Moldt, Rodenhagen: 2000.

^[37] Cp. Wehler: 2000, pp. 1-2.