Interfacing system dynamics with various soft system methodologies is currently engaging the attention of leading practitioners of system dynamics. There is a great deal of concern because of the isolation of system dynamics from other techniques and because of methodological issues in system dynamics that the field of soft OR has already begun to address. There is much benefit to be derived from a dialogue between the practitioners of system dynamics and those of soft OR (Lane, 1994). These disciplines have grown up separately, and yet they have many things in common, perhaps the most important being a concern with effective organizational intervention.

Ackoff (1979) criticized objectivity as an impossible goal in a specific situation although he admitted it could be a systemic property of scientific endeavour as a whole. He rejected the concept of optimality as being impractical because of the exclusion of esthetics and the emphasis on the utility of ends to the exclusion of means and irrelevant because of the rate of change in social and organisational systems. In this context, the focus thus changes from the hard system (technololgy based structured organisation) to the soft system (the human activity and relationships within the hard system) Bentlay(1992). It is important to note that soft system management emerged from the failure of system engineered concepts to be applied to the resolution of 'messy' people based organisational problems (Bolton & Gold, 1994).

Soft OR involves an array of tools for coping with complexity, uncertainty, and conflict. It is concerned about stakeholder participation, transparency of the process, the use of soft data, and social judgement. Models are understood not as true, solvable, objective representations of parts of the real world (ontology-based). They are accepted as subjective intellectual constructs (epistemology-based), explanatory devices that can be used to explore and understand parts of the real world (Lane, 1994). Solution and optimization are put aside, not just because they are thought to be unachievable from a practical point of view but because they are not meaningful in the contexts where soft OR seeks to operate (Checkland 1981).

A system is first of all a way of looking at the world. It is a mental construct of a whole, for which it is possible to establish a set of interrelated parts that make up the perceived whole. The system - its identity, parts, and relationships - cannot be anything else but a construct or distinction by an observer; and different observers in different contexts and with different purposes may make different distinctions. In this sense, defining a system is viewpoint-dependent (Espago, 1994). But this raises the question of the status of models in SSM. Clearly they are not would be descriptions of the world, and hence they cannot be tested by checking how well they represent the world, since this is something they do not purport to do. (Checkland, 1995).

We now come back to the basic question of establishing a dialogue between system dynamics and the various soft OR approaches. The tools of soft OR and system dynamics are both being used to try to implement the idea of learning processes. There is a rich array of approaches for systems thinking, of which system dynamics is but one. Not all problems can be addressed using system dynamics, and soft OR lacks a tool for examining the time-evolutionary behaviour of systems. Many combinations are possible as we wrestle with messes and strive for learning. Knowledge of soft OR would render more vigorous the methodological frame work of system dynamics. First, the issues arising from application problems are now straining the original framework, and second, there are unresolved weaknesses in the theory of the original (Lane, 1994).

Awareness of the strengths and weaknesses of the different systems methodologies, and of the social consequences of using each type, leads to the possibility of employing them in a pluralist or complementarist manner - each used when and where it is the most appropriate. Complementarism at the level of methodology requires a meta-methodology that respects all the other features of critical systems thinking and employs these, together with a full understanding of each individual systems approach, to describe procedures for operationalizing a pluralistic employment of methodologies in practice (Jackson, 1995). The notion 'complementarism' can be used in systems research to refer to two very different things. It can be understood as a strategy for developing the field of systems, or as a strategy for systems research. We are more concerned with the second meaning of the term. Here the idea is transferred from systems at the macro level as a field of enquiry, to the realm of systems research and practice. This idea describes a process in which the researcher thinks creatively about the context within which the problem is located, before deciding which hard, soft, or critical systems methodology/ies should be invoked.

The structural modelling tools originally developed have as a common basis the mathematical theory for directed graphs as well as the related structural modelling concepts (Harary et al, 1965, Roberts, 1976). John Warfield, as quoted by Lendaris (1980) argues that structural modelling is a methodology which "employs graphics and words in carefully defined patterns to portray the structure of a complex issue, a system, or a field of study".

The term "structure" could be defined as the way in which the component parts of the complex whole are inter-related; that which is made up of many components parts. Thus, the essence of structural modelling is one of emphasis only; that is to say that a structural model focuses on the task of selecting the components of a model and explicitly stating the interactions between them (McClean and Shepherd, 1976). Structural modelling, then, can be viewed as a wholistic process in that the user aspires to gain an overall appreciation of the system as a whole by studying a structural model of the elements which comprise the system.

Barlas (1995) considers the relationships between system dynamics and the methodologies and approaches like chaos, simulation gaming, soft systems methodology and systems thinking and comments that there are still various issues in these relatioships that need to be clarified. Hammond et al (1977), Mumpower et al (1979) contends that policy-oriented simulation models would prove useful only if they were constructed with a sophisticated understanding of the many types of judgements that are required both for model building and for model use. Anderson (1992) describes a demonstration experiment that was designed to link a simulation model with formal models of judgement. The simulation-modelling technique chosen was system dynamics, and the judgement-modelling approach selected was social judgement analysis. Models derived from social judgement analysis were attached to a system dynamics model to create a new objective function sector. In a recent paper, Rios and Schwaninger (1996) show that a combination of system dynamics and the methodology of network thinking (MNT) developed at the University of St. Gallen can help overcoming some of the limitations of both methdologies, and realizing substantial synergies between them. They call this synthesizing methodology "Intrgrative Systems Modelling". Richardson (1996) while commenting upon the problems for the future of system dynamics, states that the field is experiencing the increasing use of qualitative tools - systems archetypes, word-and-arrow diagrams under various labels (casual-loop diagrams, influence diagarams, cognitive maps), and other approaches and techniques that fall under the general rubric of qualitative systems thinking. The relationships between these qualitative practices and the quantitative core of the field of system dynamics are unclear. The question of such an interface, therefore, assumes criticality while modelling soft systems using system dynamics. Patching, as quoted by Checkland (1995) suggests that hard systems analysis addresses those parts of an enterprise that have a tangible form. Soft systems thinking, however, considers the systems that could be envisaged throughout, and, in particular, those that involve human activity.

Conventional methods and models are based on hard (quantitative, cardinally-measured) information. The problems are different in the analysis of soft, qualitative or categorically measured data. Soft modelling methodologies aim at taking into account the limitations caused by measuring variables on a non-metric scale, and try to avoid the use of non-permissible numerical operations on qualitative variables.

There is a body of literature which criticize the approach to model human phenomena in time and space for their attempts to subject human relations to numerical analysis. According to these critics, knowledge of human beings involve the apprehension of qualities, which in their very nature escape the net of numbers. Measurement is pointless at best, a hopeless distortion or obfuscation of what is really important (Kaplan, 1971). Notwithstanding these criticisms pointing to the limitations of measurement, however, there is increasing recognition that a qualitative approach need not eschew measurement. Social scientists have been more and more concerned with measuring qualities in order to grapple with complex configurations and the ambiguities inherent in human perceptions and behaviour (Leitner, Nijkamp and Wrigley, 1985) `

The problems occur at two stages in such a modelling approach. Roy and Mohapatra (1994) have earlier attempted to model the work climate of a research and development (R&D) laboratory using the system dynamics framework. First, most of the variables encountered in soft systems are measured using a quasi-quantitative framework. The problems of reliability and validity of such measurement have to be addressed. Second, the relationships among the indices have to be ascertained in a way that takes into account this quasi-quantitative measurement approach. Only thereafter could a system dynamics model of such a soft system be developed. This would also help minimize judgmental scaling error often encountered in such modelling endeavours. The question assumes criticalilty as the latent variables used in system dynamics models of soft systems are measured by such quasi-quantitative methods and as such the system dynamics model itself becomes subjective upon the validity and reliability of such measures.

The path analytic model representing the structure of relationships as implied or ascertained can be tested with the help of LISREL technique (Joreskog and Sorbom, 1989, Saris and Stronkhorst, 1984). The model incorporates unobserved (latent variables), the relation between these and observed variables and an allowance for errors of measurement in the independent and dependent latent varibles, and a causal model linking the latent variables together. It consists of two components. The first component is the measurement model, relating exogenous and endogenous latent variables to those observed variables and hypothesized to measure them and incorporating assumptions about the behaviour of unobserved measurement errors. The second component is a strucural model that represents the causal relationships between endogenous and exogenous latent variables. The latent variables are also called concepts and the observed variables are also called indicators. If a concept is directly caused or influenced by any of the other concepts, it is classified as endogenous. If a concept always acts as a cause and never as an effect, then it is exogenous, and fluctuations in the values of theses concepts are not to be explained by this model (though they may be used to explain fluctuations in the values of the endogenous concepts). Thus the direct causal effects that are of interest are located (Hayduk, 1987). In LISREL parameters of the measurement and structural model are estimated simultaneously, using full information maximum likelihood (ML) technique. Ordinary least squares imposes restrictions on correlated errors, which the LISREL 7.16 model does not impose. Thus, models with independent variables that can be fixed can be considered to directly influence the dependent variables. Also, when utilizing LISREL 7.16, measurement error can be accounted for when the measurement error of the dependent variables can be calculated. LISREL is a computer programme for estimating general linear structural equation models with the specific advantage of allowing for unmeasured hypothetical constructs or latent variables, each of which may be measured by several observed indicators. The method allows for differentiation between errors in equations (disturbances), and errors in the observed variables (measurement errors) and yields estimates for both. Thus, in LISREL, measurement concerns become integrated with model development, estimation, evaluation and interpretation (Bohrnstedt, 1983). LISREL 7.16 also allows for the examination of the fit of the model. A test statistic (t) indicates significance of the specefic coefficients, whereas goodness-of-fit can also be used simultaneously (La Du and Tanaka, 1989).

Structural Equation Models of Organizational Climate and Performance of Research Units

The data was collected from a stratified random sample of 236 RUs out of a total population of 602 RUs from different laboratories of the Council of Scientific and Industrial Research (CSIR), India. The effectiveness of RUs and other latent variables conceptualizing various dimensions of organizational climate were operationalized by observed variables or indicators measured on 5-point semantic differential scales. This data was used to develop two structural models involving these latent variables. The respondent strata consisted of RU head and the core scientists of the units (the external evaluators - both scientific as well as administrative were there only for the RU effectiveness measures) . The latent variables for RU effectiveness are R&D effectiveness (REF), user-oriented effectiveness (USE), administrative effectiveness(AEF) and recognition (REC). The rest of the ten latent variables are applied research thrust (ART), technical services (TEC), leadership quality (LSQ), supervisor contact effectiveness (SCE), innovative ethos (ETH), administrative constraints (ADC), communication (COM), research orientation (RES), conflict (CON) and research planning quality (RPQ). Figures 1 and 2 show the two structural models developed after the hypothesized models among the latent variables were run on LISREL 7.16 program. The exogenous concepts are indicated by xi ( ) and the endogenous concepts are indicated by eta ( ).

Figure 1 shows the first structural model involving the following exogenous variables - ART and TEC and the following endogenous variables - ETH, ADC, COM, RES, USE, and AEF. Figure 2 shows the second structural model involving the following exogenous variables - LSQ and SCE and the following endogenous variables - ETH, COM, CON, RPQ, REF and REC. Only the significant causal linkages are shown in the models (the t-values are ct 5 per cent significant level). The values of the structural coefficients gamma ( ) between the exogenous and the endogenous concepts and those of beta ( ) among the endogenous concepts are shown with each causal link along with their respective t-values shown within brackets. An analysis of the results indicate the following for the two models:

In the first model (figure 1), the following (gamma) values for the causal linkages from the exogenous concepts to the endogenous concepts in the hypothesized model were not found significant - from ₁ to ₆ (₁₆ ) and ₂ to ₂ (₂₂ ); and the following (beta) values for the causal linkages among the endogenous concepts in the hypothesized model were not found significant - from ₁ to ₄ (₁₄ ), ₁ to ₅ (₁₅ ), ₃ to ₅ (₃₅ ) and ₃ to ₆ (₃₆ ). These linkages, therefore, are not shown in the figure. The total Coefficient of Determination for Structural Equations was found to be 0.597 indicating that about 60 per cent of variance is captured by the model. Both the Root Mean Square Residual (RMSR) of 0.059 as compared to the average size of the S matrix (the actual observed covariances among the indicators) and the Goodness-of-Fit index (GFI) of 0.961 are within acceptable limits. The t-values of the measurement errors of all the endogenous concepts are found to be significant.

In the second model (figure 2), the following values were not found significant - from ₁ to ₁ (₁₁ ), from ₁ to ₃ (₁₃ ), and from ₁ to ₆ ( ₁₆ ); and the following values were not found significant - from ₁ to ₅ (₁₅ ), from ₂ to ₅ (₂₅ ), and from ₃ to ₄ (₃₄ ). These linkages are also not shown in the figure. The total Coefficient of Determination for structural equations was found to be 0.431 indicating that about 43 per cent of variance is captured by the model. Both the RMSR of 0.069 as compared to the average size of the S matrix and the GFI of 0.965 are within acceptable limits. The t-values of the measurement errors of all the endogenous concepts are found to be significant.

For the first model, the chi-square ( ² ) value with 13 degrees of freedom is evaluated as being equal to 34.42 (p=0.001). For the second model, the ² value with 12 degrees of freedom is evaluated as 30.15 (p=0.003). The probability level associated with a given ² statistic indicates the probability of obtaining a larger ² value, given that the hypothesized model is supported. The higher the value of p, the better is the fit. Values of p greater than 0.10 are considered as indicators of satisfactory fit (Reddy, 1992). However, the ² statistic is sensitive to sample size. In large samples, the ² test is too merciless, even models that approximate the sample covariance matrix are usually rejected. In small samples, the ² test lacks power and it forgives important misspecifications in the model. Thus, this test is more a reflection of the sample size than of the adequacy of the model (Browne and Cuddeck, 1992).

In conclusion, it is emphasized that the subjective measures of soft variables are influenced by systematic and random measurement errors. Hence, it is essential that their reliability and construct validity should be assessed before these are used in empirical studies. The validity of the system dynamics models of soft systems is thus dependent upon the construct validity and reliability of such quasi-quantitative measures. Moreover, the relationships among the latent variables or concepts developed from the indicators or the observed variables have to be ascertained in a way that takes into account this quasi-quantitative measurement approach. Only thereafter could a System Dynamics model of such a soft system be developed. This would also help minimize judgmental scaling errors often encountered in such modelling endeavours. Structural Modelling using LISREL 7.16 programme is an approach to tackle these issues and problems, and it also serves as a pre-validation exercise for the System Dynamics model.

References

Ackoff, R.L. (1979), The Future of Operational Research is Past, Journal of the Operational Research Society, Vol. 30, No. 2, pp. 93-102.

Andersen, David F. and John Rohrbaugh (1992), Some Conceptual and Technical Problems in Integrating Models of Judgement with Simulation Models, IEEE Transactions on Systems Man and Cybernetics, Vol. 22, No. 1, pp. 21-34

Barlas, Yaman (1995), Academics of System Dynamics: Its Core Definition, Topics and Terminology, System Dynamics: An International Journal Of Policy Modelling, Vol VII, No. 1, pp. 1-16.

Bentley, T. (1992), Soft Systems Methology, Management Accounting, July-August

Bohrnstedt, G.W. (1983), Measurement, in Rossi, Peter,James Wrightand Andy Anderson, eds., Handbook of Survey Research, Academic Press, New York, pp. 69-121.

Bolton, R. and Gold, J. (1994), Career Management Matching the Needs of Individuals with the Needs of the Organization, Personnel Review, Vol. 23, No. 1, pp. 6-24.

Browne, M.W. and R. Cuddeck (1992), Alternative Ways of Assessing Model Fit, Sociological Methods and Research, 21, pp. 230-258.

Checkland, P.B. (1981), Systems Thinking: Systems Practice, Chichester, UK., Wiley, .

Checkland, P.B. (1995), Model Validation in Soft Systems Practice, Systems Research, Vol. 12, No. 1, pp. 42-54.

Espago, Raul (1994), What is Systemic Thinking?, System Dynamics Review, Vol. 10, Nos. 2-3, pp. 199-212.

Hammond, K.R., J.L.Mumpower, and T.H. Smith (1977), Linking Envirnmental Models with Models of Human Judgment: A Symmetrical Decision Aid, IEEE Transactions on Systems, Man and Cybernetics, Vol. SMC-7, pp. 358-367.

Harary, F., R.Z.Norman, and D.Cartwright (1965), Structural Models: An Introduction to the Theory of Directed Graphs, New York, Wiley.

Hayduk, Leslie A.(1987), Structural Equation Modelling with LISREL: Essentials and Advances, The John Hopkins University Press, Maryland, U.S.A.

Jackson, M.C. (1995), Beyond the Fads: Systems Thinking for Managers, Systems Research, Vol. 12, No.1, pp. 25-42.

Joreskog, K. and D.Sorbom (1989), LISREL 7.16: Analysis of Linear Structured Relationship by Maximum Likelihood and Leasat Squares Method, International Educational Services, Chicago.

Kaplan, A. (1971), Measurement in Behavioural Sciences, in Franklin, B.J. and H.W. Osborne, eds., Research Methods: Issues and Insights, Wadsworth Publishing Company, Belmont, California, pp. 121-128.

La Du, T.J. and J.S. Tanaka (1989), Influence of Sample Size, Estimation Method, and Model Specification as Goodness-of-Fit Assessments in Structural Equation Models, Journal of Applied Psycology, 74, pp. 625-635.

Lane, David, C.(1994), With a Little Help from our Friends: How System Dynamics and Soft OR can Learn from Each Other, System Dynamics Review, Vol.10, Nos. 2-3, pp.101-134.

Leitner, H., P. Nijkamp and Wrigley, N. (1985), Qualitative Spatial Data Analysis: A compendium of Approaches in Nijkamp, P., H. Leitner and N. Wrigley, eds., Measuring the Unmeasurable, Martinus Nijhoff Publishers, pp. 1-28.

McLean, J.M. and P. Shepherd (1976), The Importance of Model Structure", Futures, February.

Mumpower, J.M., V. Veirs, and K.R. Hammond (1979), Scientiific Information, Social Values and Policy Formation: The Application of Simulation Models and Judgment Analysis to Denver Regional Air Pollution Problem, IEEE Transactions on System, Man and Cybernetics, Vol. SMC-9, pp. 464-476.

Reddy, S.K. (1992), Effects of Ignoring Correlated Measurement Errors in Structural Equation Models, Educational and Psychological Measurement, 52, pp. 549-570.

Richardson, George P.(1996), Problems for the Future of System Dynamics, System Dynamics Review, Vol.12, No. 2, pp. 141-157.

Rios, J.P. and Markus Schwaninger (1996), Integrative Systems Modelling: Leveraging Complementarities of Qualitative and Quantitative Methodologies, Proceedings of 1996 International System Dynamics Conference, Cambridge, MA, pp. 431-434.

Roberts, F.E.(1976), Discrete Mathematical Models, Englewood Cliffs, NJ, Prentice Hall.

Roy, Santanu and P.K.J. Mohapatra, (1994), Study of Work Climate in R&D Organizations: A System Dynamics Approach, Proceedings of the 1994 International System Dynamics Conference: Production and Operations Management, University of Stirling, Scotland, pp. 61-70.

Saris, W.E. and Stronkhorst, L. Henk (1984), Casual Modelling in Non-experimental Research: An Introduction to the LISREL Approach, Scientometric Research Foundation, Amsterdam, The Netherlands.

Tsouvalis, Costas and P.Checkland (1996), Reflecting on SSM: The Dividing Line Between Real World and Systems Thinking World, Systems Research, Vol.13, No.1, pp. 35-45.

ISDC '97 CD Sponsor