Beruflich Dokumente
Kultur Dokumente
Madonna Rehabilitation Hospital University of NebraskaLincoln Rhode Island School of Design / Brown University University of Western Ontario University of Bern University of Bern McGill University Stanford University University of Central Florida University of Central Florida University of Central Florida U.S. Army Research Institute University of North Texas
Modeling Complex Systems is Volume 52 in the series CURRENT THEORY AND RESEARCH IN MOTIVATION 2007 by the University of Nebraska Press All rights reserved Manufactured in the United States of America International Standard Book Number ISBN: 978-0-8032-1387-6 (Clothbound)
The Library of Congress has cataloged this serial publication as follows: Nebraska Symposium on Motivation. Nebraska Symposium on Motivation. [Papers] v. [1]1953 Lincoln, University of Nebraska Press. v. illus., diagrs. 22cm. annual. Vol. 1 issued by the symposium under its earlier name: Current Theory and Research in Motivation. Symposia sponsored by the Dept. of Psychology of the University of Nebraska. 1. Motivation (Psychology) BF683.N4 159.4082 53-11655 Library of Congress
Preface
The volume editors for this 52nd volume of the Nebraska Symposium on Motivation are Bill Shuart, Will Spaulding, and Jeffrey Poland. The volume editors coordinated the symposium that led to this volume, including selecting and inviting the contributors and coordinating all aspects of editing. My thanks to our contributors for excellent presentations and chapters. This symposium series is supported by funds provided by the chancellor of the University of NebraskaLincoln, Harvey Perlman, and by funds donated in memory of Professor Harry K. Wolfe to the University of Nebraska Foundation by the late Professor Cora L. Friedline. We are extremely grateful for the chancellors generous support of the symposium series and for the University of Nebraska Foundations support via the Friedline bequest. This symposium volume, like those in the recent past, is dedicated to the memory of Professor Wolfe, who brought psychology to the University of Nebraska. Richard A. Dienstbier Series Editor
Contents
ix Bill Shuart, Will Spaulding, and Jeffrey Poland Richard W. J. Neufeld Wolfgang Tschacher and Zeno Kupper Susanne P. Lajoie Introduction
1 85 123
Composition and Uses of Formal Clinical Cognitive Science A Dynamics-Oriented Approach to Psychopathology Developing Computer-Based Learning Environments Based on Complex Performance Models Technology for Building Intelligent Systems: From Psychology to Engineering Fostering Team Effectiveness in Organizations: Toward an Integrative Theoretical Framework Constructive Complexity and Human Change Processes Editors Postscript: Modeling Complex Processes in a Rehabilitation Application
145
Mark A. Musen
185
Eduardo Salas, Kevin C. Stagl, C. Shawn Burke, and Gerald F. Goodwin Michael J. Mahoney
245 275
x modeling complex systems sorts of variables involved, together with their interactions. I have sought, second, to show that we possess, now, many sound and useful concepts and techniques to translate these complexities into productive experimental research. I suppose the moral is: be not afraid of complexity. If motivation is indeed complex, then let us nd the means to cope with it. (Vinacke, 1962, pp. 4243) Through the representation of a few very simple psychological concepts, in a rudimentary mathematical way, a good deal of complexity can be generated. . . . Let us now proceed to generate complexity from simplicity. (Burke, 1966, pp. 4950) Although disorder may be experienced and expressed in highly patterned processes of human activity, it is diverse, individually unique, and systemic; we shall advance in our attempts at conceptualization and classication only as we are willing to embrace the limits of symbol systems to capture human uniqueness and the ultimate ineffability of complex system dynamics. (Mahoney, in this volume, pp. 265266) The contributions to this volume of the Symposium describe contemporary approaches to the modeling of complex psychological and behavioral processes, ranging from molecular to molar phenomena. Although the contributions reect a range of theoretical and epistemic perspectives, they all explicitly or implicitly incorporate complex frameworks of dynamic, system-like relations involving perception, learning, concept formation, emotion, motivation, intention, behavior, and the social context in which behavior occurs. One special feature of all the contributions from this particularly distinguished group of theorist-practitioners is an emphasis on practical applications of the conceptual frameworks in which they work. This reects an important idea in the zeitgeist of the contemporary scientic community, that of translational research. Translational research is a process of translating the principles and truths that emerge from basic science into practical applications. The complexity of the processes captured in the contributors models enhances the models applicability to the complexities of clinical practice, industry, and education. To consolidate the relevance of application
xi Introduction and translational research, this volume ends with a volume editors postscript, describing a practical model for the complex processes of rehabilitation, as manifest in rehabilitation services currently evolving in Nebraska. Translational research demands, not just practical application, but continuity with theory and basic science. This converges with the historic role of the Nebraska Symposium as a prominent (and now the oldest sustained) forum for psychological theory. All the contributions in this volume emphasize the theoretical basis of application and the necessity of logical and conceptual continuity in understanding complex processes. In the rst contribution Richard W. J. Neufeld discusses the advantages of formal mathematical theory for illuminating relations between variables as they interact in experimental science. He applies these advantages to the clinical practice of assessing cognitive impairments. Decrying a continuing overreliance in much psychological research on statistical analyses associated with Fisher and Pearson, Dr. Neufeld asserts that formal mathematical modeling of cognitive processes will, ultimately, lead to greater theoretical clarity about normal and abnormal cognition and better clinical-assessment techniques. It is noteworthy that, while the tradition of mathematical modeling in psychology has a long and honored past, the increasing availability of powerful computational tools (e.g., computers and analytic software) supports the kind of sophisticated modeling in the hospital or clinic that was impractical in earlier decades. At a more general level, Dr. Neufeld characterizes his approach as a novel form of construct validity, one based on the inherent mathematical properties of the cognitive processes he studies. In this sense his contribution is a sophisticated exemplar of the use of complex modeling to achieve traditional theoretical goals of experimental and clinical psychology, as articulated by such historical gures as Lee Cronbach and Paul Meehl. In the next contribution, Wolfgang Tschacher and Zeno Kupper provide a synthesis of dynamic systems theory and current cognitive science. Inspired by the historic role of Gestalt psychology in the evolution of cognitive science, their discussion invites us into the heart of psychologys theoretical legacy. Drs. Tschacher and Kupper then apply their perspective and methods to the complex realm of psychopathology. They present data sets and analyses from recent
xii modeling complex systems research with people diagnosed with schizophrenia and demonstrate the importance of tracking individuals with multiple measurements over time in order to detect oscillations or trajectories in rehabilitation and recovery that would be missed in the typical cross-sectional approach. Using a time-series analysis, they identify unique patterns or dimensions of intrasubject characteristics that have complex but meaningful interrelations. Returning to theoretical principles, they show how complex, dynamic formulations can be translated into useful clinical instruments and methods. Finally, in a tribute to the Nebraska Symposiums historic focus, their contribution culminates with a characterization of motivation as identical to the ongoing action of complex human cognitive processes operating to order and simplify a complex world. A second exemplar of complex modeling to achieve traditional goals is provided by Suzanne P. Lajoie. Dr. Lajoie uses theoretically grounded performance modeling in the development of computerbased intelligent tutoring systems designed to help learners master the complexities of real-world endeavors. Learning how experts go about problem solving and decision making through cognitivetask analysis is an important aspect in the process of developing an effective tutoring system. Dr. Lajoie highlights the importance of discerning experts relevant dimensions of expertise (e.g., self-monitoring), as expressed in a specic context, in developing effective models. She also emphasizes the importance of other variables, e.g., emotional, motivational, and social, and she describes strategies for determining what to model, whom or what should serve as the model, and how to model the content and/or process. She then translates these principles into design considerations for effective educational technology. The next contribution extends application of complex modeling from education to knowledge management. Mark A. Musen discusses past and current efforts to develop computer applications to support decision making and data representation in health care. Dr. Musens theory base is not psychology or neuroscience but articial intelligence. His technology is the technology of computer engineering. Nevertheless, he envisions a future role for psychology in the development of articially intelligent systems to manage our already enormous and rapidly expanding knowledge base. It is noteworthy in this regard that psychology has drawn from engi-
xiii Introduction neering as much as vice versa, from radar-inspired signal detection models of perception to band-lter models of attention to computer models of executive cognition. Herbert Simon, in his 1994 Nebraska Symposium on Motivation contribution, cited articial intelligence as a promising model for human cognition. The impact of complex models for knowledge management may be in psychologys future rather than its past. In the next contribution, Eduardo Salas, Kevin C. Stagl, C. Shawn Burke, and Gerald F. Goodwin scrutinize complex processes associated with small groups of people brought together for common purposes. They advance the science of teams by providing a detailed review of representative models of team performance in organizations and other naturalistic settings generated over the past quarter century. In their review, Drs. Salas, Stagl, Burke, and Goodwin nd the invocation of input-process-output (ipo) models, consonant with the general systems framework that has inuenced many areas in the social sciences during the past several decades, to be a key commonality among these models. There is greater diversity among models with respect to emphasis on internal team processes versus greater attention to the inuence of external, contextual factors. The authors conclude that both inuences are important and that, consequently, more sophisticated modeling techniques are needed to successfully deal with the resulting dynamic complexity, particularly in naturalistic settings. Drs. Salas, Stagl, Burke, and Goodwin then turn to a description and elaboration of a new and unique multilevel integrative framework for understanding team functioning. This new model is distinctive in the importance that it attaches to individual team members cognition as an important moderating variable as well as group decision making, shared mental models, and external factors. Michael J. Mahoneys contribution is a nuanced interlacing of several kinds of models, including verbal metaphor, narrative, photography, and poetry. Dr. Mahoney discuses various perspectives on complexity theory and its precursors in philosophy and science, including current theoretical frameworks such as dynamic systems theory, complexity studies, and chaos theory, placing them in the context of the history of ideas. He describes and elaborates on constructivism, an integrative framework and family of theories. Dr. Mahoneys contribution includes two appendices. The rst provides
xiv modeling complex systems a synopsis of important aspects of human change from the perspective of constructivism. The second provides rich and provocative perspectives for incorporation in the practice of counseling, psychotherapy, coaching, and other educational pursuits.
Complexity, Systems, and the Nebraska Symposium on Motivation: A Brief History of Ideas
The perspectives reected in this volume are exemplars of an evolving set of conceptual frameworks that inuenced thinking in many areas of science during the second half of the 20th century. These frameworks are most generally associated with general systems theory, and addressing complexity is one of their key common features. Having recently celebrated a half century of the Nebraska Symposium on Motivation, as this volumes editors we saw a useful purpose in reviewing the more than 300 individual contributions that constitute the previous volumes, to identify ideas that anticipate or shape the approaches to complexity that we nd in contemporary work. We found a richness of such ideas, so many that only a few can be highlighted here. The remainder of our introduction to this volume is a review of ve especially resonant contributions from volumes past: Heider (1960), Walker (1964), Leeper (1965), Newcomb (1953), and Barker (1960). We selected contributions that, in addition to showing the nascent ideas about complexity and systems theory discussed in this volumes contributions, have clear relevance to practical application and translational research and especially to our own particular interests in physical medicine and psychiatric rehabilitation. One common characteristic of systems theories is an organizational scheme that orders specic mechanisms and processes according to their respective complexity. Two terms from classic learning theory, molar and molecular, serve to dene the poles of these schemes. Processes and mechanisms are molar rather than molecular to the degree that they represent the integrated interaction of multiple components. Psychology itself reects this type of ordering, ranging as it does from theories of neuronal activity to neuropsychology to the psychology of social cognition and interpersonal behavior. The historical Nebraska Symposium contributions reviewed here follow
xv Introduction a molecular to molar rather than a chronological sequence. Interestingly, contributions from common theoretical perspectives can differ with respect to the molecular-molar dimension. Heider (1960) and Barker (1960) integrate classic Gestalt principles with subsequent theories, but Heider addresses comparatively molecular expressions of key processes, while Barker is at the other end of the continuum, addressing spontaneous organizational processes at the level of a human community. Other contributions to be reviewed address system organizational processes at the cognitive level (Walker, 1964), the emotional level (Leeper, 1965), and at the social/interpersonal level (Newcomb, 1953). We have included extended excerpts from the original contributions, in an effort to preserve the style and tone of the original presentation. Also, we hope to provide enough of the language in sufcient detail for readers to draw their own conclusions about the relation of the historic ideas to those of the present volume. However, these passages cannot fully convey the logic or the eloquence of the source materials. The reader is encouraged to consult the full chapters in areas of particular interest.
xvi modeling complex systems ideas presented have direct relevance to current models of rehabilitation and crucial aspects of participation and quality of life, each intimately associated with the behavior settings available to the individual, the quality of his or her social networks, and other variables that will be discussed in the nal section of this introduction. According to Klein (1959): Fritz Heiders work . . . has had over the years a signicant, if relatively unobtrusive, impact on some of the most important theorists of our time, notably Kurt Koffka, Kurt Lewin, and Egon Brunswik. More recently, Heiders inuence has been detectable in perception theory, for example, in the work of James Gibson. Still, his writings cannot be called popular (p. v). Klein continues: Heiders emphasis on the macrophysics of things (in contrast to the reductionist emphasis on microphysics), the important distinction he develops between those parts of the environment which mediate (medium) and those which are mediated (thing), his analysis of how we may distinguish behavioral events attributable to the structure of the environment and those attributable to the structure of the perceptual systemall of these merit close study. . . . Throughout the papers the composition of a unit whether spatial, temporal, or causalis of central importance to Heiders distinction between thing and medium. The dening properties of a unit, therefore, come in for extended and penetrating analysis. . . . Heider has made [an attempt] to penetrate the essential nature of the concept of structure. The general macrostructures which he describes may apply to their subjective counterparts in ego organization. . . . A unique feature of Heiders approach is his attempt to fathom environmental structure not from the response side from the inside outward, as it were, as is common in psychological theoriesbut from the outside inward, that is, by specifying the architectural rules of the extrapersonal world of physical object and event units. The result, then, is an extraordinarily fresh confrontation of the external structures which are assumed but never specied in psychoanalytic notions of reality testing and adaptation. (p. vii)
xvii Introduction In his Symposium contribution Heider provides an overview of four thought models or schemata characterizing the Gestalt tradition at the time. Heider terms the initial model the classic Gestalt theory. The model is based on the work of investigators such as Wertheimer, Khler, and Koffka. Describing this model, Heider highlights a number of aspects of this perspective that bring to mind such concepts as patterning and perceptual organizing processes that seem quite consistent with concepts used today. Heider begins by emphasizing the Gestalt concept of good gure, advanced by Wertheimer, who applied it to visual processing: This principle states that the perceived gure tends to be as good as the stimulus pattern will allow, or as Koffka says: Psychological organization will always be as good as the prevailing conditions allow. In this denition, good is undened. It embraces such properties as regularity, symmetry, simplicity and others. . . . For instance, slight irregularities in the shape of visual forms are usually not noticed. Kohler gives the following example: Faces of people usually appear to us symmetrical, in spite of the fact that they are rarely objectively symmetrical. We may notice this irregularity in another persons face when we look at his mirror image; but ordinarily we dont see it. Kohler has called attention to the fact that a tendency towards simplicity can also be observed in physical systems, as, for instance, Ernst Mach has pointed out. Kohler gives many examples in his book on Physical Gestalten [1924]. . . . Let us recapitulate: Wertheimer observed the tendency toward good form with percepts; Kohler then related this observation to a similar tendency found in physical systems. Of course the same tendency is then assumed to rule the process in the physiological brain eld. Since isomorphism is assumed, all this ts very well together. The thought model is one of a complex process with many part events which interact in such a way that a certain endstate is reached, an end-state which is in some way distinguished, and which has characteristics the other possible states do not have; as long as this end-state is not reached something will happen. On the other hand, when it is reached, the process attains an equilibrium and nothing more will happen.
xviii modeling complex systems Furthermore, the end-state will come about regardless of what the beginning state of the system is: thus one can talk about a tendency, which implies direction, a reaching of the same endstate by different routes. (Heider, 1960, pp. 145146) Heider follows his discussion of Khlers ideas with a perspective from Kurt Koffka: It is not surprising that Gestalt psychologists have applied this same thought model to behavior. For instance, Koffka does so in his book on The Growth of the Mind, which rst appeared in 1921. I should mention that Koffka uses the term closure for the distinguished end-state, a closed gure being a better gure than an open one. Tendency to closure is therefore only another name for tendency toward simplicity, or goodness of gure. This is what Koffka says (Koffka, 1925, p. 103): . . . The characteristics of closure . . . belong not merely to the phenomena themselves, but likewise to the behavior taken as a whole, including all reactions made to the environment. Instinctive activity then becomes an objective mode of behavior analogous to such phenomena as rhythm, melody, and gure. . . . He [Khler] calls the state toward which the processes in the organism are directed a standard state. It has to be distinguished from the state of equilibrium (Kohler, 1938, p. 325), and he describes it as follows (Kohler, 1938, p. 303): The essential characteristic of regulation is an invariance of direction. Whatever initial conguration may obtain in those systems when we begin to observe themif we observe long enough their inner displacements or transformations will always be found to bring them nearer to a standard status. The word standard points here to the fact that the nal status is independent of the initial conguration. . . . Essentially . . . the thought model of a system tending towards a standard state is applied to directed action, and this model had its origin in the principle of good gure. However, . . . when we try to nd out how it is carried out, we see that
xix Introduction two steps are necessary for the transition from the phenomena of the visual eld to action. First, we have to take into account not merely perceptual appearances but a space in which behavior occurs; and secondly, we have to consider the objective environment, and the way the organism effects changes in it. . . . Thus we have to substitute for the visual eld what Koffka called the behavioral eld and Lewin the life space. This behavioral eld is conceived of as having similarities with the visual eld. It also is a system containing a great number of part processes which interact, it exhibits forces and tensions, and tends to arrange itself in such a way that a distinguished end-state is reached. This distinguished end-state, in some way comparable to the simple gure, is the state of the person who has reached the goal in his life space. Now, this life space or behavioral eld is a concept which involves many difculties and unsuspected depths and snares. . . . I can only say that in a rst approximation, which, however, is not entirely correct, one can conceive of it as representing the environment of the person as the person himself experiences itand it is in some way related to the brain eld, to physical processes going on in the brain. The Gestalt psychologist would characterize this relation as one of isomorphism, i.e., of structural similarity. This is the rst step we have to accept when we apply the principle of good form to activity: namely, the step from perceptual to behavioral eld. The second step requires a more extensive consideration. So far we have only considered processes which are inside the organism in some way, which are encapsulated, as Brunswik says. How is it possible that they produce effects outside the organism, in his physical environment? We assume that this behavioral eld changes in the direction of a distinguished state, maybe a state of minimal tension, i.e., the state of the person being at the goal. But . . . we have to understand how the tension in the behavioral eld makes the person reach the goal in reality. (Heider, 1960, pp. 147149) He later continues: In action, not only a part of the organism, but the whole organism is involved. . . . The idea of the feedback or circular process
xx modeling complex systems can be applied also in this case: as long as the person has not yet reached the goal, there is a tension in the behavioral eld; this tension is communicated to the executive system, which changes the relation between organism and objective environment in such a way that the goal is reached; via perception this objective state is communicated to the behavioral eld; and thus the tension in this eld is removed. I have used the term feedback to characterize this process. However, one has to keep in mind that this circular process is not a simple feedback process. What distinguishes the circular process of Gestalt theory from simple feedback is the interpolation of the behavioral eld with its tendency towards a distinguished state [emphasis added]. (p. 150) Heider goes on to describe two models advanced by Kurt Lewinthe person model and the environment model, together constituting what Heider terms Lewins spatialized psychology. He notes the move from the perceptual sphere to the behavioral realm in Koffkas work and what Lewin terms the life space. Heiders discussion of Lewins concepts is thought provoking, and the reader is encouraged to review those concepts in the source material. Heider then offers a relatively brief summary of his own recent theorizing, describing what he calls his balance theory, which he feels answers some questions left inadequately treated by the classic Gestalt theories he has summarized: This theory of balance deals mainly with congurations consisting of a number of entities between which exist certain relations. The entities can be personsthe own person or other personsand other entities, as for instance, things, situations, or groups. The relations considered are mainly of two sorts: on the one hand attitudes of liking or disliking, and on the other hand unit relations of belonging. The main idea is that certain of these congurations are preferred, and that, if circumstances allow, they will be realized by the person either in such a mental reorganization as wishful thinking, or in an actual change through action. . . . In recent times a number of theories have been proposed which are similar to the one just outlined. I remind you, for instance, of Newcombs (1953) discussion of processes of com-
xxi Introduction munication [see below]. . . . These conceptions, symmetry, consonance, balance, and simplicity, are, of course, implied in that idea with which Gestalt theory started and which always was central to it, namely, the idea of a good gure. We therefore have returned to the model we considered rst. This model implies a number of different entities with certain properties and standing in certain relations, which make up a constellation of factors tending toward a standard state. The properties of these congurations which determine their meaning and their fate are whole-qualities. Consonance or simplicity of the structure cannot be derived from the properties of the parts . . . If we study the p-o-x system1 which is composed of the own person (p), another person (o), and an impersonal entity (x), then we nd that the state of balance depends on the attitudes of p toward o or x. That is, the attitudes toward the parts of the conguration, and the relation of these attitudes to each other enter as signicant factors, and determine the attitude toward the whole conguration. . . . Thus, we are able to specify more exactly the conditions of goal selection, at least in some cases. The goal is not taken to be an unanalyzed entity which in some way acquired valence, but is derived from the properties of the structure [emphasis added]. . . . [The] difference between Lewinian theory and balance theory [is] in regard to the role structure plays. In Lewins environment model . . . structure is not intimately connected with the conditions of tendencies, nor with their effects. Structure helps us to derive the direction toward means from direction toward goal; but it does not help us to derive the direction to the goal. . . . . . . Thus we see that in these [Lewins] models the dynamic factors are not very closely linked with structure. Neither do the properties of the structure imply forces, nor do the forces affect the structure in a speciable way. In the balance model the dynamic factors are intimately connected with the structure. The dynamic factors arise out of denable structural characteristics and the forces toward the standard state tend to change the structure in denite directions.
xxii modeling complex systems . . . [Unlike Lewins models], in the balance model structure in a state of equilibrium is denably different from one in a state of disequilibrium, and all the parts of the structure are relevant to this difference, not only the relation between two parts, person and goal. (Heider, 1960, pp. 167170)
xxiii Introduction ity, set, attitude, and trait, along with many other concepts in psychology, are reducible to names of intervening variables or theoretical constructs, each related to different sets of operations, but all, ultimately, determiners of choice behavior. It will be the argument of this paper that the concept of psychological complexity can be used to account for the termination of psychological events, and the choice of the next event over a wide range of traditional concepts of determiners of choice. Thus, the concept of psychological complexity can be useful in answering the rst two of the three basic questions. Psychological complexity can also be used to account for many of the phenomena associated with the trace of a past event, but this third basic question or problem is beyond the scope of the present paper. (pp. 5152) Under the subheading Psychological Complexity Theory, Walker asserts: The major distinction that must be made is between stimulus complexity on the one hand and psychological complexity on the other. The rst is a characteristic of the external stimulus, more or less independent of the individual organism. Psychological complexity is a characteristic of the event itself, the organisms response. Psychological complexity and neural process complexity will be assumed to be completely isomorphic. (Walker, 1964, pp. 5253) In the context of a basic denition of terms, he turns to an elaboration of his concept of psychological complexity: Psychological complexity is a characteristic of the event itself and is thus a characteristic of the interaction of the organism with the distal stimulus when the event in question is initiated by a stimulus. Thus it is possible for two organisms to react with equal psychological complexity to stimuli with very different distal stimulus complexity values. The same organism may also react with different degrees of psychological complexity at different times to the same stimulus. (pp. 5455) Walker offers a brief description of the (assumed) underlying nervous system basis of the experience of optimal complexity, under the heading Neural Net or Neural Process Complexity:
xxiv modeling complex systems Underlying any psychological event, is, of course, a pattern of neural events. Such events are spatially three dimensional and occur over a nite period of time, a fourth dimension. It is assumed that variation can occur in the relative complexities of two neural processes. We shall refer to the relative complexity of a four-dimensional neural process as the relative complexity of the relevant neural net. Furthermore, we shall assume a complete isomorphism between neural net complexity and psychological complexity (Walker, 1964, p. 55). He offers as relative characteristics of simple versus complex neural net in terms of the four dimensions noted above: simple nets consist of processes that are relatively small, short, focal, and central (i.e., origin within the central nervous system); complex nets are relatively large, long, diffuse, and peripheral. Having offered a neurologically oriented substrate, Walker turns to the concept of psychological complexity itself. Under the heading Optimal Complexity, he asserts: The key concept of the theory I am attempting to fabricate is the concept of optimal complexity. The simplest and most straight forward psychological denition of optimal complexity is the following: Optimal complexity is that degree of psychological complexity the organism will seek to maintain. If a psychological event is more complex than the optimum, the organism will behave in such a manner as to reduce the complexity of the event. If a psychological event is less complex than the optimum, then the organism will behave in such a manner as to increase the complexity of the event. Optimal complexity can be bracketed by other values of psychological complexity. In perception, an input level far above optimum produces mental dazzle. A lower limit is a level of complexity that is below the threshold of consciousness. In motor activity, psychological complexity far above optimum results in discoordinated tetany, and there is a lower value which constitutes the threshold for action. The optimum is a normal percept or a smoothly coordinated movement. . . . The sequence is inevitable and the fall in neural net complexity is an automatic result of observable and fairly well understood neurophysiological characteristics. Since neural net
xxv Introduction complexity and psychological complexity are assumed to be isomorphic, psychological complexity may be said to rise and fall as well. With a sufciently complex stimulus, psychological complexity will rise to and exceed the threshold of consciousness, will rise to and exceed optimal complexity, will fall below optimal complexity, will then drop below the threshold of consciousness automatically, and will usually be followed by another psychological event. [The reader] will recall that the rst of the three basic questions of behavior theory was: What is the mechanism which terminates an event? The answer is: Whether the stimulus for an event is continued or not, a psychological event undergoes a sharp and automatic drop in complexity during a period of approximately one-half second after its initiation. (Walker, 1964, pp. 5658) With respect to repeated activation of an event, he asserts: Repeated activation of a neural net will result in a progressive decrease in the psychological complexity of the event involved. (p. 59) With respect to the second major question (What are the determinants of the choice of the next event?), he continues: The principle of optimal complexity incorporates the dynamism that the organism will seek such a level. The termination of one event occasioned by the automatic reduction in its psychological complexity below the optimum level, literally forces choice of that event among available next events which will be nearest optimum. Therefore: Among available alternatives, an organism will choose as a next event that activity which is nearest optimal psychological complexity. It is assumed as a working hypothesis that many of the major determinants of choice behavior such as reinforcement, habit, motivation, curiosity and other collative2 variables, subjective probability and utility, and others ultimately can be reduced to a single conceptpsychological complexity. (p. 60)
xxvi modeling complex systems Walker offers a brief recapitulation: The theory can be stated in an abbreviated form. During the course of a psychological event that has a duration of approximately 500 [msec], the psychological complexity of the event rises abruptly and falls more slowly. The automatic reduction in the psychological complexity of an event insures that it will drop promptly below the optimum to be replaced by that one of the available alternative events which is nearest optimum. The psychological complexity of alternative behaviors or events will be a function of four variables. They are: (1) the stimulus complexity of the initiation stimulus; (2) the time since this particular event has occurred previously; (3) the number of times that event had occurred before; (4) the arousal properties of the stimulus or event. (Walker, 1964, p. 60) Walker addresses the issue of what behavior(s) might be expected when no near-optimum event is present. In the case of all available alternatives below optimum, an individual might commonly respond in one or more ways (slightly adapting Walkers text): 1. Search the environment or his own repertory for more complex events; 2. Find a more complex stimulus in the environment to which he had not attended previously, or he might fall to daydreaming; 3. React by locomoting, getting up and moving about; 4. Seek arousing stimuli; 5. Seek to differentiate previously unexplored potential complexity in his environment or in old thought sequences and problems. All of these devices would serve to increase the complexity level of the sequences of events which are occurring. All would serve to move the sequence nearer an optimal level of complexity. (Walker, 1964, p. 61) At the other extreme: Situations in which the psychological complexity levels are above optimum are usually situations in which the sensory inputs into the nervous system are providing more information than the organism can process. This may result when the exter-
xxvii Introduction nal environment is too complex, when the problem one is working [on] is beyond immediate solution, or when the motivational or emotional system is in a highly aroused state. (pp. 6162) In such circumstances, he notes common reactions: 1. The organism may shift attention or narrow attention to a limited portion of the stimulus input; 2. If the overload is of external origin, the organism may locomote to a less complex circumstance; 3. If either are difcult or impossible, the organism may attend repeatedly to the same stimulus in an effort, usually successful, to produce a reduction in the psychological complexity of the situation through repeated activation of the relevant event; 4. An associated result of repeated activation is to organize a very complex stimulus into a smaller number of chunks. (p. 62) Walkers contribution concludes with a survey and critique of relevant research and theoretical distinctions to provide support for and elucidation of the conceptual framework adduced therein. He buttresses his concept by applying it to existing experimental data that are not easily explainable by any other existing theory. For example, he applies his theory of optimal psychological complexity to the often-observed (but less frequently reported) decremental variations in conditioned responses following learning experiments that have been taken by many to be unexplainable anomalies. After noting a few such anomalous observations: For the sake of the argument I am certain to get, let me take the position that the appropriate learning curve shape in running studies, conditioning studies, and selective learning studies, is one that rises and falls to zero or to a steady level below the maximum performance. The curve that rises to a steady maximum and remains there indenitely is likely to be rare. The reason that we see few learning curves of the postulated type is that most experimenters know in advance what a learning curve is supposed to look like. As a result of this knowledge, they stop training when the asymptote is reached, or, if they obtain a curve which does not t their conception of what one should look like, they nd a great many other ways to respond
xxviii modeling complex systems . . . other than to publish their sin against respectability. They throw away their data. They restructure the apparatus. They change the parameters of the study. They change the design. This process is known as the establishment of experimental(er) control. Sooner or later they manage a situation in which they obtain the right answer. I can attest that this process is carried out in good faith and under the assumption that in so doing, one is behaving like a sound, rigorous, and careful experimentalist. I can attest to this because I am one of the sinners. Thus psychological complexity theory handles the usual learning curve, extinction, and the drop in performance that often occurs under continued reinforcement. It predicts that most experimental situations will produce a drop in performance if training is continued. (Walker, 1964, pp. 8586) In later comments concerning Walkers presentation, his fellow presenter, Frank Logan, in addition to suggesting caveats to the formers views, concludes with an important observation: There are several features of Walkers approach with which I am in strong agreement. A language that avoids the articial separation of stimuli and responses more nearly captures the unied, interdependent inseparability of psychological events. It is also becoming increasingly recognized that the fundamental behavioral operation of an organism is selection or choice. . . . Walkers attempt to develop a system that can deal with behavior dynamically, i.e., continuously over time, is perhaps the critical feature necessary to achieve a general integration. And, by whatever means, visualizing such disparate concepts as habit, motivation, and decision-making in terms of a single construct certainly is one we should applaud. (Logan, 1964, p. 98)
xxix Introduction or cognition). In fact, Leeper highlights the ultimately inseparability of processes of perception, motivation, and emotion: Still earlier, David Krech (1949, 1950a, 1950b, 1951), in his usual impassioned style, had reasoned that it is unrealistic to conceive of psychological phenomena in terms of separate processes of perception, motivation and learning. Instead, he urged, we ought merely to conceive of Dynamic Systems. These, he said, are so denitely organic unities that no single aspect of such a system can be changed without changing the other aspects as wellwe have been dealing in myths in believing that we could vary some one of these aspects while keeping the other aspects constant. Though proposing a less drastic statement on this point, E. C. Tolman (1932, 1948) had been suggesting some perceptual factors in motivation in his view that motivation is partly a matter of reward expectations and punishment expectations. Kurt Lewin similarly had been discussing many problems of motivation in terms of factors in the organisms psychological environment. . . . In my own previous writing, my original paper on a motivational theory of emotion (1948) was extended to some extent into the perceptual-motivational theory which has been elaborated in the present paper. . . . One odd fact about these various earlier discussions of a perceptual or conceptual interpretation of motivation is that their authors have made practically no references to the related ideas of the other papers. This is the more surprising in view of the fact that most of this group are more or less closely related to one another both personally and as regards their general theoretical outlooks and interests. It seems, therefore, as though each of these persons had to grope to the concept on his own, even though possibly helped in ways that he did not recognize by his predecessors or colleagues. I make this suggestion with somewhat more condence because I remember that, in my own case, when I rst read Krechs papers on Dynamic Systems, they did not make much sense for me. . . . And, peculiarly, it took me a long time to recognize that Lewins ideas might be thought of as a perceptual theory of emotion. . . . Maybe this sort of thing will continue to be the case. If a perceptual theory of motivation is to become more common,
xxx modeling complex systems perhaps each psychologist will have to gure it out for himself. (Leeper, 1965, pp. 111112) Leeper summarizes some major themes: The suggestion that comes from a number of sources, therefore, is, rst of all, that emotions are motives, and then, second, that emotional processes, along with all other motives, are perceptual or representational processes. The suggestion that comes is that emotions and other motives do not exist or operate in any less complex sense than this. . . . Even though perceptual habits are hard to change in some cases . . . it seems that all perceptual habits can be modied by learning and that sometimes such modications can occur suddenly and dramatically. If emotional habits are perceptual habits, these same possibilities should exist for them. (pp. 113, 115)
xxxi Introduction Newcomb goes on to dene terms and delimit his focus to communicative behavior between individuals: The properties of objects may be studied either objectively or phenomenallypreferably in both waysbut in any case, they are studied not as things-in-themselves, but as related to persons. Thus the characteristic way in which social psychologists study motivation is in terms of person-object relationships (the term object includes other persons, of course). Since, as we have all learned in recent years, motivational phenomena are intimately interlinked with perceptual phenomena, it is often necessary to distinguish two aspects of person-object relationships, which may be labeled the cathetic and the cognitive. Often, however, one does not need to make this distinction, while still bearing in mind that both aspects are involved, and in such instances the term orientation is a useful one. The term is similar to the concept of attitude, except that it connotes existing directedness and not merely a predisposition or a readiness. Orientations are known, of course, only as they are inferred from observable behavior. Insofar as such behavior involves reciprocal stimulation and response (or anticipations thereof) it is traditionally referred to as interaction. But one cannot observe interaction-in-general; one must observe discriminable units of behavior. I propose, therefore, to use as such an interaction unit the communicative act, dened as any observable behavior by which information, consisting of discriminative stimuli, is transmitted from a human source to a human recipient. For present purposes, it is assumed that the discriminative stimuli have an object as referent. Thus in the simplest possible communicative act, one person (A) transmits information to another person (B) about something (X). Human social behavior is thus to be studied in terms of the conditions and consequences of varying communicative acts. And problems of motivation in social behavior are to be studied in terms of orientations toward the two kinds of objects necessarily involved in communicative actsi.e., persons as recipients of transmitted information and objects (including persons) as referents of transmitted information. The relationship between orientations and communicative acts, as we shall see, is a
xxxii modeling complex systems circular one, so that it will be necessary to consider each of them, in turn, as varying with the other [emphasis added]. (Newcomb, 1953, pp. 140141) Following a second section summarizing relevant ndings concerning group membership, orientations, and communication, Newcomb moves to a section titled Communicative Behavior as Varying with Orientation toward Persons and toward Objects. In this section he explicates the systemic relations between communicative acts and the orientations of individuals in a communication setting: I can hardly imagine anything that would surprise you less than to hear that communicative acts are learned in ways that seem to have something to do with rewards and punishments. I shall stop, however, only to indicate in the most general kind of way what seems to be the nature of the learning conditions of communicative behavior. These conditions have to do with what I have already referred to as the individuals necessity for co-orientationi.e., relating himself simultaneously both to objects and to persons as actually or potentially related to those objects. . . . We may start with the assumption that orientations both toward persons as potential co-communicators and toward other objects have adaptive value; not to be oriented to them would mean to have no cognitive content regarding them and to have no hypotheses (in the Postman-Bruner sense) as to their potentialities for reward or punishment. The further assumption that co-orientation has adaptive value stems from what I believe to be the fact that neither kind of orientation occurs singly and independently of the other, in connection with communicative acts. First, the orientation of any communicator, A, toward B, a potential recipient of his communication, rarely, if ever, occurs in an objectless vacuum. . . . Secondly, and conversely, the orientation of the communicator, A, toward almost any conceivable X rarely, if ever, occurs in the total absence of an orientation toward B, the potential recipient of his communication. (Autistic verbalization, of the kind Piaget reports in young children, would, of course, represent an exception . . .). The very fact that B is a potential recipient requires some kind of orientation toward him. . . . Al-
xxxiii Introduction most invariably, moreover, there is included in this orientation toward B some assumptionhowever accurate or inaccurate about Bs orientation toward the object of communication. From this elaboration of what is perhaps only too obvious, I want to deduce a single pointthat there is a necessary interdependence between co-orientation (which itself involves an interdependence of orientations) and communicative acts. . . . Since, according to these assumptions, there are relationships of interdependence among several distinguishable orientations, it is convenient to regard them as together constituting a system. For some purposes the system is best treated as an objective onei.e., a model employed by the observer. The elements in this system are, minimally, A, B and X (a source, a recipient and an object of communication); the interdependent orientations among them are As toward B and toward X, and Bs toward A and toward X. . . . The implications of this model are: (1) that while at any given moment the system may be conceived of as being at rest, it is characterized not by the absence but by the balance of forces; and (2) that a change in any part of the system may lead to changes in any of the others. I shall make one further set of assumptions about the system. . . . These assumptions are that (under the stated conditions) communication tends to result in increased similarity, or congruence, of As and Bs orientations toward X, and that, as a result of learning, communicative acts are instigated by the anticipation of increased similarity or congruence (or, alternatively, by the threat of decreased similarity or congruence). (Newcomb, 1953, pp. 147149) After positing adaptive advantages of his concept of congruence, he goes on to articulate an important aspect of his A-B-X system, which he considers a strain toward congruence (p. 149). The systemic perspective of Newcombs contribution and his initial observations of the relation between explanations/models at the level of subareas of psychology concludes with a view to the future: I should like to suggest (with a good deal of tentativeness) that something along the lines of the framework of co-orientation which I have roughly sketched out may nd a place in
xxxiv modeling complex systems general motivation theory. Many, among the higher forms of animal life, at least, are capable of plural orientation, and the actual direction of behavior at any given moment often cannot be accounted for in terms of any single object-orientation, others being held experimentally or hypothetically constant. . . . I suspect that the study of social behavior can provide evidence, in ways other behavior cannot, of how behavior directedness varies with multiple orientations. If so, an adequate theory of motivated social behavior will have contributed something to a general theory. Last, and far from least, an adequate general theory would take fuller account than it does today of self-orientation. . . . Here, as in the case of other concepts of peculiar relevance to social motivation, it is my belief that more extrapolations from a general theory will not sufce. Theorists from McDougall and Freud to Murphy and Rogers have properly accorded to the self a central place; though not always, in my judgment, have all of them seen that place in its full social context. Not only are self-orientations part and parcel of other-orientations, I would insist; they are inextricable from the eternal triangle of self, other persons, and the common environment. A general theory of motivation, when it is mature enough to include these interdependent orientations, will have borrowed from a theory of motivation in social behavior, as well as helping to establish it. (Newcomb, 1953, p. 159)
xxxv Introduction dividual and environment into account. At a practical level, this resonates with major themes in rehabilitation, for example, the World Health Organizations recent emphasis on the construct participation as the ultimate aim of rehabilitation efforts (see World Health Organization, 2001). This construct is of importance because it underscores the importance of including assessment and modication of an environment, in addition to clinical treatment, as a vital part of the rehabilitation process. Barker outlines features of his concept of psychological ecology, including the central concept of behavior settings, which provides an important window on our understanding of a range of psychological phenomena as a system and is, at times, a very useful unit of analysis for psychology. The relevance of Barkers concepts for the issues addressed in the present volume is that, like Newcomb (1953), Barker describes a framework that explicitly relates systems concepts to adaptive processes at the social/community level. Barker begins by incorporating from the work of Egon Brunswik an emphasis on the critical importance of including in accounts of perception and behavior the environment in which an individual acts and perceives. In Barkers words: Brunswik (1955) described psychological schools and theories in terms of their positions upon a macro-unit he considered to be the true vein of psychological ore. This vein extends from the environment to the environment; namely, from distal objects in the ecological environment, through proximal stimuli at the receptor surfaces of a person, through the persons peripheral receptor mechanisms, through his central processes, and through his peripheral effector systems, to his proximal reactions, or means behavior; and it nally terminates in the focus of the total unit: the persons achievement with respect to the nonpsychological world of things. The three major sectors of this unit are . . . (1) the ecological sector of objects and physical stimuli (preperceptual), (2) the organism or intrapersonal sector, and (3) the behavioral sector which occurs, again, in the ecological environment. (Barker, 1960, p. 1) Barker, along with Brunswik, regards the entire span of the E-E (environment-environment) unit as the fundamental unit of analysis with respect to psychology; it is the basic psychological entity.
xxxvi modeling complex systems He takes issue with some of Brunswiks conclusions, advancing the hope that taking the entire E-E span into account can lead to more than a probabilistic framework for psychological explanations: I hold the hope that a more detailed, conceptual, and explanatory account of the whole course of events can be achieved, particularly at the junction point between the ecological and the intrapersonal sectors of the unit, and especially with respect to motivation. This, in fact, is the theme of my paper. (Barker, 1960, p. 3) The E-E unit, then, is to be taken as the ultimate unit of analysis; one obvious way to understand this unit is as a multisectored system. As can be seen in the material reproduced below, one theme that recurs in other examples of systems approaches is that of the close relation between the processes of change, perception, and motivation. A second major theme is the inuence of changes or properties in one part of the system on the qualitative status of other parts of the system, a dening aspect of all exemplars of models informed by general systems theory. Finally, the interaction of sectors of the E-E unit is considered with respect to emergent social aspects in Barkers system: For a psychology dened in terms of E-E units, the usual considerations of motivation are not adequate. These considerations almost always make personal motives the whole story of the energetics of behavior, and place them within the organism. But a unit is a unit; it is indivisible. When it is a psychological unit, the environment, the organism, and the behavior are all involved, and energetics must occur in all of the parts. Either the E-E unit is false, or motivation theory is too limited. (Barker, 1960, p. 4) Barker goes on to lay the groundwork for discussion of his theory of behavior settings by introducing some concepts that provide the context for his central theses. In a later section he provides further elucidation of the relation between the entity and the environment elements of his model: Ecology is concerned with relations between entity and environment. But before this statement has any useful meaning,
xxxvii Introduction entity and environment must be dened. . . . Where does each entity end and its environment begin? . . . . . . To clarify this problem, it is necessary to revert to the levels of phenomena in science. . . . I have emphasized that the essential distinction between levels is this: The laws, the explanations, which have been devised to account for occurrences on one level are inadequate to explain occurrences on a different level, yet the levels are coupled systems. Another distinction that is crucial for the denition of environment is that between inside and outside. Every entity has a discriminable boundary; what is within the boundary constitutes the entitys inside, and what is without constitutes its outside. . . . The environment of an entity is made up of those parts of the outside regions with which the entity is coupled by laws on a different level from those which govern the entity itself. . . . Here, for ecological problems, is the basis for delimiting an entity from its environment. The test is this: As we move from any discriminable thing to more remote, surrounding parts, a point is reached at which the governing laws, so far as we know them, become incommensurate, yet the linkage remains. This point marks the boundary of the entity and the beginning of the environment. (Barker, 1960, pp. 78) Barker asserts the desirability of taking the entire E-E continuum as the crucial unit of analysis, rather than abstracting only elements of it for psychological examination. He offers the study of psychological principles of learning as an example: The eld of learning is interesting in this connection. Learning is usually interpreted as the process, par excellence, by which the environment inuences the organism and its behavior. This is the predominant way, almost the only way, a culture is presumed to shape the personality and behavior of the individuals born into it. The facts of learning demonstrate, however, as almost all learning theory recognizes, that even here the organism is the locus of driving forces without which learning does not occur. Indeed, within the context of learning it is, paradoxically, the behaving organism that endows the environment with behavior-controlling properties; the guiding and coercing powers of the environment have been shown to
xxxviii modeling complex systems depend upon what activities the organism has previously had with it, and these depend more upon the organism than upon the environment. Indeed, learning studies have demonstrated that almost every discriminable part of the ecological environment can be coupled with almost every kind of behavior. This is important information; it denes the range of an organisms power to transform its connections with the ecological environment, and it implies that parts of the environment are almost equipotential. . . . It will be clear now where ecology enters the environment-environment unit, which Brunswik took as the realm of psychology. Psychological ecology deals with the relations between the nonpsychological sectors of this unit, governed by the laws of geometry, chemistry, economics, etc., and the intrapersonal and the behavior sectors, governed by psychological laws. (Barker, 1960, pp. 1112) He then attempts to formulate an account of how these incommensurable system elements might be related (or, at least, an approach to a satisfactory understanding) by exploring earlier ideas of Fritz Heiders. Barker makes a gradual transition to his concept of behavior settings. Because these ideas are readily available to the interested reader, a detailed presentation will not be offered here. However, one would highlight a particularly important element of his concepts concerning behavior settings: Barker highlights how behavior settings are regions in a community that offer certain opportunities and, along with these, require certain responsibilities. Furthermore, there is a relation between the peopling of these settings and both the number of responsibilities and the adequacy of performance that can be expected to occur. This relation between the demands of a given behavior setting and the impact on the behavior and life of the individuals populating these settings seems very compatible with more recent concepts. Implications for the analysis of complex settings and behavior are evident. Barker continues, in a section entitled Theory of Behavior Settings: Field studies in which I and my associates have been engaged, of the behavior of children in their natural habitats,
xxxix Introduction have brought us to the hypothesis that under certain precisely dened and frequently occurring conditions, people stand in the relationship of media to behavior settings; and that under certain other less common conditions, people stand in the relationship of things to behavior settings, imposing certain absolute constraints on them. This hypothesis brings some order into data upon American-English differences in the behavior of children and adults, into data upon differences in the behavior of individuals in settings of different sizes, and into data concerning the behavioral consequences of physical disability. The wide ramications of these simple ideas suggest that they may have a basic signicance for psychology, and particularly for the psychology of motivation. . . . It is rst necessary to describe behavior settings. When a mother writes, There is a baseball game in progress on the playground across the street, she does not refer to any individuals behavior, but to the behavior of children en masse. The same is true of a newspaper item which reports, The annual fete held in the St. Ambrose Church garden was a great success. These are behavior settings. They are highly visible behavior phenomena; laymen mention them in conversation and in writing as frequently as they do individual persons. . . . Here are . . . [some] behavior settings: Streets and sidewalks Kanes Grocery Cliffords Drug Store Gwyn Caf Pearl Caf Midwest State Bank Of special relevance in the present connection, however, are the following characteristics of settings: 1. Behavior settings involve ongoing patterns of extraindividual behavior whose identity and functioning are independent of the participation of particular persons. 2. A behavior setting has a circumjacent soma of physical objects: of walls, doors, fences, chairs, dishes, typewriters, ad innitum, arranged in a characteristic spatial pattern, at a particular temporal and physical locus.
xl modeling complex systems 3. Behavior settings are homeostatic systems; they normally persist, often for years, at a relatively stable, characteristic level. . . . A behavior setting is a behavior entity, but its laws of operation are not the laws of individual psychology . . . Most of what we know about behavior settings is simple description, with any conceptualizations being not far removed from the surface appearance of settings. However, this is enough to make a beginning in tracing the connections along the Brunswikian unit which has its origins in this part of the ecological environment. For our purposes, the self-regulatory characteristic of behavior settings is crucial and must be considered further; it is this, indeed, which gives behavior settings, under certain conditions, the position of things which impose their own patterns on the people within them, who have the position of media. Behavior settings exhibit a stability-within-change, a persisting functional level which is due to a balance of many inuences. Some of these issue from the larger community, some are intrinsic to the setting itself, and some originate within the individuals who populate the setting . . . . . . Forces operate in every setting. These multiple balanced forces assure that the level of a setting is more stable than most of its parts or conditions singly. One frequently occurring means of balancing the forces and maintaining the homeostatic level of a behavior setting is compensating for a deciency in the number or docility of the parts of the medium by an increase in the amount of energy applied to each of them, and vice versa. When the media of a setting, the machinery, the tools, or the workmen, for example, are in short supply, those available have to work longer and/or harder. (Barker, 1960, pp. 1521) In the last section of his contribution, Barker discusses People: Media of Behavior Settings: Six features of the relationship between people and behavior settings must now be mentioned. 1. People are part of the inside manifold of behavior settings.
xli Introduction 2. Of all the attributes of settings, people are the sine qua non. . . . 3. Each quasi-stationary level of a setting has its optimal population requirements. . . . 4. Of all the equipment and paraphernalia of a setting, people are among the most immediately malleable and adjustable. 5. Different behavior settings on the same level of functioning, and therefore with the same optimal population requirements, actually differ greatly in population. . . . 6. These ve features of the relation between people and behavior settings emphasize the position of people as the media of behavior settings. This is true. However, there is one important exception. When the number of people in a setting, its population, falls below the minimal number required by its homeostatic level, the setting will be modied. (Barker, 1960, pp. 2122) He concludes: Behavior settings with less than optimal people for their homeostatic levels are self-disciplining settings. The opportunities within them are matched by the obligations they contain. . . . We sometimes call them self-discipline. In reality they are controls built into the structure and the dynamics of the setting, into the ecological environment. . . . I would like to close with two remarks: (1) Brunswiks environment-environment unit appears to be subject to more than empirical probabilistic laws, and (2) the ecological environment appears to be, especially, the seat of motivating inuences. (pp. 4849) Barkers contributionas is his body of scholarly work in generalis novel and interesting and would seem to have continuing applications today. In particular, the growing acknowledgment that community reintegration and quality of life are vitally important ends of rehabilitation efforts and that rehabilitation cannot really be considered a successful endeavor unless an individual is supported to the point of maximal participation in the life of the community, with the greatest degree of independence possible, leads inevitably to the recognition that there must be a satisfactory awareness of the
xlii modeling complex systems environmentthe behavior settings and social/interpersonal environment to which an individual will be returningbefore an optimal rehabilitation treatment plan can be developed and delivered. Barkers (and his students) techniques for identifying and cataloging community venues can serve as a guide in expanding rehabilitation practice to include such analyses. In this regard, in addition to Barkers work and the models of Brunswik and Heider (already referenced), additional useful resources include Gibson (1979) and Wicker (1984).
Notes
The editors would like to acknowledge and express a special thanks to Mr. Joe Brown for his superlative work as copy editor of this volume of the Nebraska Symposium on Motivation. Mr. Browns consistently resourceful and creative suggestions concerning the substance, organization, and presentation of chapter materials contributed immeasurably to the clarity and coherence of the nal product. As will be evident, the conceptual breadth of material and diversity of perspectives reected in this volume are considerable. Mr. Browns timely and precise questions and observations, clear and patient counsel, and unagging good humor throughout were notably catalytic in successfully bringing the 52nd edition of the symposium together in its present form. All volume editors should be so fortunate! 1. Compare Theodore Newcombs A-B-X model, discussed below. 2. Variables such as curiosity, novelty, and stimulus change, stimulus aspects sometimes thought to provoke increased engagement, interest, and/or increase arousal.
References
Barker, R. (1960). Ecology and motivation. In M. R. Jones (Ed.), Nebraska symposium on motivation (Vol. 8, pp. 149). Lincoln: University of Nebraska Press. Brunswik, E. (1955). The conceptual framework of psychology. In International encyclopedia of unied science (Vol. 1, Pt. 2, pp. 656750). Chicago: University of Chicago Press. Burke, C. J. (1966). Linear models for Pavlovian conditioning. In D. Levine (Ed.), Nebraska symposium on motivation (Vol. 14, pp. 4966). Lincoln: University of Nebraska Press. Gibson, J. J. (1979). The ecological approach to visual perception. Boston: Houghton Mifin.
xliii Introduction
Heider, F. (1959). On perception, event structure, and the psychological environment. Psychological Issues, 1(3), 1123. . (1960). The Gestalt theory of motivation. In M. R. Jones (Ed.), Nebraska symposium on motivation (Vol. 8, pp. 145172). Lincoln: University of Nebraska Press. Klein, G. S. (1959). A note to the reader. Psychological Issues, 1(3), vvii. Koch, S. (1956). Behavior as intrinsically regulated: Work notes towards a pre-theory of phenomena called motivational. In M. R. Jones (Ed.), Nebraska symposium on motivation (Vol. 4, pp. 4287). Lincoln: University of Nebraska Press. Koffka, K. (1925). The growth of the mind. New York: Harcourt, Brace. (Original work published 1921) Khler, W. (1924). Die physischen Gestalten in Ruhe und im stationren Zustand: Eine naturphilosophische Untersuchung von Wolfgang Khler. Erlangen: Philosophische Akademie. Khler, W. (1938). The place of value in a world of facts. New York: Liveright. Krech, D. (1949). Notes toward a psychological theory. Journal of Personality, 18, 6687. Krech, D. (1950a). Dynamic systems as open neurological systems. Psychological Review, 57, 345361. Krech, D. (1950b). Dynamic systems, psychological elds, and hypothetical constructs. Psychological Review, 57, 283290. Krech, D. (1951). Cognition and motivation in psychological theory. In W. Dennis et al. (Eds.), Current trends in psychological theory (pp. 111139). Pittsburgh: University of Pittsburgh Press. Leeper, R. W. (1948). A motivational theory of motivation to replace emotion as disorganized response. Psychological Review, 55, 521. Leeper, R. W. (1965). Some needed developments in the motivational theory of emotions. In D. Levine (Ed.), Nebraska symposium on motivation (Vol. 13, pp. 25122). Lincoln: University of Nebraska Press. Lewin, K. (1938). The conceptual representation and the measurement of psychological forces (Contributions to Psychological Theory, Vol. 1, No. 4, Serial No. 4). Durham nc: Duke University Press. Logan, F. (1964). Comments on Edward L. Walkers paper. In D. Levine (Ed.), Nebraska symposium on motivation (Vol. 12, pp. 9698). Lincoln: University of Nebraska Press. Newcomb, T. (1953). Motivation in social behavior. In M. Jones (Ed.), Current theory and research in motivation (Nebraska Symposium on Motivation, Vol. 1, pp. 139161). Lincoln: University of Nebraska Press. Nissen, H. W. (1954). The nature of the drive as innate determinant of behavioral organization. In M. R. Jones (Ed.), Nebraska symposium on motivation (Vol. 2, pp. 281321). Lincoln: University of Nebraska Press. Simon, H. A. (1994). The bottleneck of attention: Connecting thought with motivation. In W. D. Spaulding (Ed.), Integrative view of motivation, cognition, and emotion (Nebraska Symposium on Motivation, Vol. 41, 121). Lincoln: University of Nebraska Press.
In every special doctrine of nature only so much science proper can be found as there is [applied] mathematics in it. So wrote Immanuel Kant about the status of natural philosophy, as he saw it in 1786 (Kant, 1970, p. 470). This declamation seemingly reected a certain frustration not unlike that possibly experienced almost two centuries earlier by Francis Bacon, who decried the distemper of learning, which occurs when men study words and not matter (Bacon, 1937, 200 (bk. 1)). These quotes are extracted from commentaries on activity done in the name of science by their authors contemporaries activity possibly characterized by considerable redundancy of effort to resolve the days purportedly key issues, contentious debate devoid of clear resolution, a lack of precise place keeping regarding the status of important segments of the body of knowledge, and an absence of salient direction for potentially protable future thrusts of investigation. More contemporary, and emanating from within our discipline, Paul Meehl (1978) bemoaned the slow progress of soft psychology (which presumably included psychological clinical science). He indicted the dominant research paradigm (Kuhn, 1962), of which a prominent modus operandi was, and for that matter remains, Fisherian- and Pearsonian-based statistical methods. Theoretical predic-
2 modeling complex systems tions were derided as somewhat anemic, typically embracing simply the presence of nonzero relations among ad hoc measured variables; as nature is said to abhor an absence of association, so-called support rested on statistical power, whereby the ever-popular anova and related tables of results were, in effect, summary statements of such power (cf. Cohen, 1988, pp. 16, 17; Steiger & Fouladi, 1997). Meehl noted that theories, at least in elds of soft psychology, typically fell by the wayside, not because they were toppled by uncorroborated risky predictions, but because the discipline became inured to them and their proponents grew weary of marketing them. It might, nevertheless, be added that their proponents need not worry because it is well-known that old theoretical ideas frequently reemerge in different guises and with modernized labels (Staddon, 1991). (Most readers will have their favorite accounts of such instances from their particular elds of expertise.) I would contend that the foregoing lamentation retains considerable currency when it comes to contemporary psychological clinical science. This observation provides much in the way of motivation for the developments presented here. It is maintained that the implementation of formal theory in psychological clinical science can go far to redress current quagmires and accelerate progress. Emphasis of course is on the eld with which I am most familiar, clinical cognitive science. However, the merits of formal theory disclosed in that domain of investigation are, arguably, general, providing additional incentive for the particular sphere of application. To begin, a clear delineation of what does and what does not constitute formal theory is presented. Benets and selected (practical) liabilities are described. Following on that is an exemplary case in point from the domain of clinical cognitive science. Novel developments include proposals for monitoring cognitive aspects of an individuals response to treatment over time and, likewise, dynamic assessment of treatment-program efcacy with respect to cognitive functioning. Some surprising noteworthy spin-offs from excursions into formal modeling, spin-offs that impinge directly on prominent issues in this area of study, are expounded on. One such by-product emanates from a personal labor of love, entailing the appropriation of stochastic modeling to the study of cognitive efciency as it relates to schizophrenia, to psychological stress, and to their combination. It constitutes a new form of construct validity, using
3 Formal Clinical Cognitive Science mathematical heuristics. Another bears on an issue that I have engaged more or less since its appearance in the literature, namely, the so-called psychometric-artifact problem in the study of differential cognitive decit. The problem is translated into simple but comprehensive measure-theoretic terms, revealing the shortcomings of typical efforts at redress. From there formal theory is shown to recast the entire issue in its own terms and in so doing render it as more or less obsolete.
4 modeling complex systems sures of dependent variables, guided, if not prescribed as such, by the nonformal theoretical apparatus (see, e.g., McFall & Townsend, 1998; Meehl, 1978), provide the data array submitted to statistical analysis, expedited by associated software. Statistical treatment is part and parcel of a formal subsystem because its constituent operations rest on closed-form formulas that, by and large, emerge from theorem-proof continuity (nowadays increasingly accompanied by numerical simulation). Note that so-called qualitative research methodology replaces this formal data-analytic arm of the logical-deductive system with nonformal, verbal operations. In this way, the entire logical-deductive enterprise is rendered informal. Conversely, formal theoretical approaches replace nonformal reasoning in the above mixed deductive design with a formal logical-deductive subsystem, thus creating an entirely formal logicaldeductive structure. Whereas the typical statistical apparatus (often elegant in its own right and with a virtual industry of software support) is substantively genericits application transcending theoretical-content domainsformal theoretical models are content specic. For example, Doob (1953) has described one such type of theoretical formulation, a stochastic model, as a mathematical abstraction of an empirical process, whose development is governed by probabilistic laws (p. v). Whereas statistical computations constitute mathematical methods for assessing arrays of values rendered by empirical investigations, formal theoretical models are concerned with processes that make the data what they are (although, as will be demonstrated below, formal theories are far from mute about ways of analyzing the data whose genesis they address). In other words, in the explanatory chain of events, such models weigh in with quantitative constraints ab initio. The nature of authentically formal theory can be thrown into sharper relief by considering procedures that may appear to qualify as the former, pending closer scrutiny. Integrative and evaluative reviews, for example, potentially useful in their own right to be sure, are a case in point. Apart from a rare application of synthesizing quantitative models of possible agents of surveyed outcomes, verbal schemata are, as a rule, restatements of what is already known (e.g., McFall, Townsend, & Viken, 1995). Formal theory can also be distinguished from prominent higherlevel statistical methods, including structural-equation causal mod-
5 Formal Clinical Cognitive Science eling (sem; e.g., Kline, 1988; Tomarken & Baker, 2003; cf. MacCallum & Ashby, 1986), taxometric procedures (Meehl & Golden, 1982; Ruscio & Ruscio, 2004; Waller & Meehl, 1998), and hierarchical linear modeling (Bryk & Raudenbush, 1992; see Piasecki, Jorenby, Smith, Fiore, & Baker, 2003). With respect to sem, the subject of prediction comprises its parameters,1 for all intents and purposes meaning the path coefcients of conjectured relations among the posited latent variables. (It is surmised that criterion values for contemporary indexes of overall statistical t, a sort of overarching prediction, have been met.) Note that, ordinarily, the hypothesized structure of interlatent-variable linkage and the relative magnitudes and/or signs of its associated path coefcients themselves are posited nonformally. In a similar vein, conrmatory factor analysis forecasts factor loadings and coefcients expressing interfactor relations, including those involving higher-order factors. Concern is with an empirical covariance structure, but one whose features are dened according to antedating empirical observations, or nonformal argumentation, rather than derivations spawned by specic axioms, denitions, and assumptions. The chief subject of prediction for taxometric methods, in turn, is certain congurations of data summaries (calculated via methods prominently known as Maxcov, Mambac, Maxslope, Maxeig, and L-mode) indicative of dichotomous versus continuous latent distributions of target variables (e.g., specic symptoms or syndromes). Selected parameters, such as base rates of a taxon (a discrete group with the symptom, genetic endowment, or other target characteristic), can be estimated. Once more, however, the data conguration corresponding to the conjectured distribution is not in and of itself the product of formal-system reasoning. By these characteristics, taxometric methods and sem fall into the category of statistical analysis and, in that way, constitute the formal subsystem of a mixed deductive system (elaborated on in Neufeld, 1989). Sufce it to say that similar observations attend typical applications of hierarchical linear modeling. Formal theoretical systems as well entail parameter estimation and testing for goodness of t between predictions and observations. Their parameters, however, are substantively signicant, as imbued by their explanatory roles in the theoretical model in which they participate (Braithwaite, 1968; exemplied below). Moreover, their
6 modeling complex systems variation generates selective changes in empirical data sui generis to the localized problem (e.g., Link, 1982; Townsend & Wenger, 2004a, 2004b). This treatment contrasts an essentially generic interpretation of parameter values, or other model properties, across even vastly divergent content areas and a focus on highly similar aspects of data (e.g., empirical covariance structures). I will return momentarily to the nonformal system widely known as qualitative research methodology because of its growing popularity, especially in the social sciences. Qualitative research increasingly has been touted as a legitimate, even superior, alternative to systems of science implanted with formal methods. Nonformal strategies give the appearance of exibility in dealing with concepts and measures whose implementation may strain formal systems. The former, however, accommodate such challenges handily, not because they somehow transcend the explanatory merits of formal systems, but because they simply sidestep the constraints on precision of expression demanded by formal systems or subsystems (Staddon, 1984). As such, despite their purported checks and balances (e.g., Miles & Huberman, 1994), they are vulnerable to the onslaught of frailties in reasoning and inefciencies accompanying dependence on essentially verbal contentions, defects that formal procedures are designed to confront (Hintzman, 1991; Kline, 1985). Indeed, the history of science makes it clear that systematic retreat from decidedly formal inferential methods inevitably detracts from progress (e.g., Braithwaite, 1968; Kline, 1985).
7 Formal Clinical Cognitive Science measurement. The actual physical mechanism of gravitational pull may be enigmatic, but the mathematical denition is not. Moreover, to be taken seriously, any proposed physical mechanism would have to comply with this mathematical statement. The function of each term, and of the law itself, in turn, is decipherable from the nature of the formal assertion and from the way in which the law serves in broader theoretical contexts. Similarly, something of the nature and function of Einsteins limiting constant, the speed of light, can be appreciated contextually by considering its operation in a formal deductive system. Such exposition, for example, discloses that any meaningfulness of a larger value rests on a solution to the imaginary number ( 1). In psychological clinical science, elucidation of obtained data congurations can be bolstered by stipulating relations between observed and inferred variables in terms analogous to the above. As opposed to ad hoc interpretation, statistical interactions, for one, can be understood from formal depiction of the psychological processes under study and associated scales of measurement (e.g., Busemeyer, 1980). An illustrative instance entails the stochastic dynamic modeling (see, e.g., McFall et al., 1995) of the effects of stress (ambient noise) on cognitive efciency (Weinstein, 1974). One such model expresses the expected (mean) latency for proofreading lines of prose (e.g., Glass & Singer, 1972). The expression is mr/(k 1) (Neufeld, 1996, 2002). In this formula, r conveys stress-related impairment of performance speed (r goes up with increasing stress, elevating the mean, reecting diminished speed), k is an increasing function of task-relevant competence (associated here with practice), and m indicates task load, quantied as the mean number of processed stimulus elements (chunks of letters scanned per line of prose). The expected latency for a given level of stress, and amount of practice, for the summary task load m can, thus, be predicted as a function of substantively signicant parameters. The expression also discloses the fabric of interplay between stress and practice, making for their interaction: the effects of stress can be mitigated by practice, practice becomes more important as stress goes up, and the effects of this interplay on performance latency are scaled according to the prevailing task load m. In fact, the form of the stress-practice interaction is one of superadditivity of the interactions second-order difference: [(mean latency under higher stress and lower practice) (mean latency under
8 modeling complex systems lower stress and lower practice)] [(mean latency under higher stress higher practice) (mean latency under lower stress and higher practice)] > 0 (Townsend, 1984). Note that each of the above parameters is also entrenched in its own quantitative infrastructure, endowing it, accordingly, with further analytic and substantive meaning (Evans, Hastings, & Peacock, 2000; elaborated on below). Beyond their revelatory function, closed-form, analytic expressions can also serve the practical cause of computational economy. It is enticing to depend on numerical simulations as surrogates for analytic derivations, given contemporary computational technology. Analytic derivations, however, not only enhance understanding, according to their detailed structures, but can also reduce computational steps, ease computer memory demands, and increase speed when simulations are inevitable. Partly for this reason, theorists in longer-established sciences will opt to take analytic solutions to the limit, before invoking numerical algorithms.2
9 Formal Clinical Cognitive Science Formal theories provide specic predictions that can compete directly in terms of differential conformity to observed values. Predictions set in quantitative terms, in turn, expedite quantitative expressions of their relative superiority (e.g., tests for goodness of t and comparative goodness of t). In addition, however, models can prescribe their own qualitative signatures of validity. The latter entail the contour of data across gradations of independent variables and their combinations (e.g., superadditivity of means, as discussed above; Neufeld & Williamson, 1996). Such patterns can, moreover, be mutually exclusive, in that, even with very weak assumptions, competing models cannot produce each others congurations (e.g., Townsend & Wenger, 2004a; cf. empirical equivalence, e.g., Townsend, 1990). If formal models mandate their own diagnostics, they necessarily ease the problem of empirical measurement. In fact, formal deductive systems stand the selection of measures on its head. As with longer-established sciences (e.g., physics examples, as discussed above), theory is potentially strong enough to mandate certain empirical ramications. Displaced, therefore, is the often ad hoc process of selecting off-the-shelf measures, conjecturally connected to dependent variables (McFall et al., 1995; Meehl, 1978). Owing to the precision of the formal deductive system, theory and measurement become intertwined. Assessment of the generated measures, for the usual statistical properties of fallible estimates (e.g., bias, maximumlikelihood status), to be sure remains applicable (e.g., Riefer & Batchelder, 1988; Townsend & Wenger, 2004a). Nevertheless, those properties apply to indexes substantively tied to the theory at hand. Individual differences in model expression can be accommodated within the context of model diagnostics, especially betting clinical science and assessment. Included is variation in the structure of cognitive transactions (e.g., serial processing, variations on parallel processing, and selected combinations of these structures; Townsend & Fic, 2004). Variation in model properties, notably parameter values, across individuals can also be protably accommodated through Bayesian techniques (Batchelder, 1998; Chechile, 1998; Neufeld et al., 2002). Not least among the endowments of formal theory is its inherent challenge to preferred interpretations of addressed phenomena. Explicitness of expression helps make salient current boundaries of
10 modeling complex systems knowledgeaws and limitations in extant formulations. In particular, empirical equivalence, on current measures, of nonetheless opposing positions is disclosed. However, so too are alternatives that can break the encountered experimental-mimicking logjams (Casti, 1989; Townsend, 1990).
11 Formal Clinical Cognitive Science Table 1. Schema of Theoretical Systems, with Selected Examples
Static Deterministic Verbal theory; Nondynamic ow diagrams Some decision-and-choice models (e.g., selected subjective-utility models) Dynamic Deterministic nonlinear dynamic systems (chaos-theoretic systems) Dynamic extensions of decision-and-choice models; classic stochastic models; nonlinear dynamic-systems models, with a stochastic element
Stochastic
system is represented by differential equations, each indicating the momentary change in a variable. The changes themselves are dened in these equations essentially by the variables own extant state and those of the other variables making up the network (Neufeld, 1999b). Stochastic dynamic theories are exemplied by time-related extensions of decision-and-choice models. The elimination-by-aspects model (discussed above), for example, has been reformulated by Marley and his colleagues (e.g., Marley, 1989; Marley & Colonius, 1992) so as to add temporal predictions about selection responses to the original models predicted frequencies of selection from among the competing items. Busemeyer and Townsend (1993; also Roe, Busemeyer, & Townsend, 2001) have presented an empirically compelling model of decision and choice, providing for speed of choice, type of choice, and their interconnections. Variability of expressed preferences and their latencies are, once again, accommodated, in that the crux of prediction involves frequency distributions. Stochastic dynamic models represent possibly the most widely used version of formal theory in psychology. Generally speaking, they depict probability distributions of events as a function of time (see, e.g., Johnson, Kotz, & Balakrishnan, 1994, 1995; Luce, 1986; Townsend & Ashby, 1983). Part of the model fabric, then, is variability of event latencies across occasions of observation (e.g., experimental trials). The provision for variability is graphically illustrated by considering the model-prescribed variance of time taken to complete a relatively simple cognitive process.
12 modeling complex systems The process comprises k stages, each stage having a specied frequency distribution, as a function of time. Each stage is executed at the rate of v stages per unit of time (whereby the mean stagewise completion is 1/v time units). Such a process may correspond to the encoding of k stimulus features (e.g., curves, lines, and intersections of an alphanumeric item) into a cognitive format facilitating further processing, such as selected manipulations involving the item in short-term memory (e.g., Sternberg, 1975). By the present model, known as the Erlang distribution (Evans et al., 2000), the mean time across trials for process completion is k /v. The across-trial variance of completion, in turn, is k /v2. Note that this statement of dispersion is squarely embedded in the substantively signicant parameters of the theoretical model of process completion. Further provision for stochastic properties is available by allowing for individual differences in the values of the parameters k and v. This extension (a mixture model) acknowledges interparticipant variation in the above variance and related model expressions (e.g., Neufeld et al., 2002). Latency variance for a given value of k and v maps onto the familiar within-participant variance in (m)anova, and variation across participants in k and v maps onto between-participant variance (see, e.g., Parzen, 1962, p. 200; also A Stochastic-Modeling Translation of the Statistical-Property Issue below). Other contemporary exemplars of dynamic stochastic models include selected extensions of multidimensional scaling (see, e.g., Schiffman, Reynolds, & Young, 1981). These extensions quantitatively relate categorizations, or ratings of multidimensional stimuli, to their cognitive-process underpinnings, where the variable of time plays a central role (e.g., Carter & Neufeld, 1999, in press; Nosofsky, 1992; Takane & Sergent, 1983). Before leaving this subsection, it should be noted that nonlinear systems theory can, in principle, be augmented to incorporate a stochastic component. A specically random element can be added to the otherwise deterministic differential equations describing intervariable inuences across time (e.g., Brown & Holmes, 2001; Huber, Braun, & Krieg, 1999). This interlaced stochastic thread perturbs the momentary change in each variable and the consequent inuence of those changes on the coupled network of variables. Interestingly, the temporal trajectories of such a stochasticized system can deviate dramatically from those of its strictly deterministic counterpart.
13 Formal Clinical Cognitive Science Sufce it to say that, overall, unaccounted-for variance in observed values may never be eliminated (cf. Gilden, 2001). However, such model-exogenous variance stands to be substantially diminished with increasing comprehensiveness of theory construction.
aesthetic appeal
P. A. M. Dirac, a Nobel laureate in physics, once declared: It is more important for our equations to be beautiful than to have them t the experiment (quoted in Freedman, 1993). Given the historical success of beautiful equations, their apparent failure was held to be a temporary aberration. Perhaps, in their presented form, the equations were an inadequate approximation of empirical events. Or, possibly, the experimental observations themselves were awed. Either of these shortcomings could be redressed. Beautiful equations were intrinsically self-vindicating, and the rest was detail. At the very least, if equations were to stand up to empirical challenge, they must spring forth from among the ranks of the beautiful, whereas equations that are not beautiful are inevitably doomed to experimental defeat (cf. Smolin, 2006). Formal theory thus potentiates the scientic merit of equational elegance. There is no inherent reason why psychological clinical science should be barred from this asset, which is, arguably, a proprietorship of formal developments. What, then, makes an equation aesthetically appealing? One criterion is succinctness. The simplicity of beautiful formal expressions belies their profundity, which becomes apparent when their copious and far-reaching implications are unveiled (Polkinghorne, 2003). One example, emanating from the eld of nonlinear dynamics, is Mandelbrots formal representation of fractals. Although the representation is parsimonious, a feature of its constituent variables is that they take on a multitude of intriguing trajectories. The proliferation of exotic patterns has been widely disseminated, thanks to modern computer-graphics technology (e.g., Gleick, 1988; Townsend, 1994). A more modest example is available from the Erlang distribution of cognitive-process completion (described above). Here, the probability of completing the stated process by a given time interval, t, is
(1)
Note that one can safely bypass the denition of each term in the equation and its equivalents (below) and still appreciate the pithiness of this expression. The somewhat involved operations, summarily contained above, include
1
00
vt
x k1 e x dx/(k 1)! =
00
j=k
j (vt)/j! evt.
(2)
These formulas, in turn, can be used to create visual aids that facilitate insight into the interplay of model parameters (for a treatise on theory-related intuitions afforded by visual imagery, see Clark & Paivio, 1989). Three-dimensional response surfaces convey variation in the present probability across alternate combinations of k , v, and the time window, t. Two examples are provided in Figure 1, for values of k ranging from 1 to 10 and values of v ranging from 0.01 to 3.0. The lower response surface is constructed for t = 4 and the upper one for t = 7. This gure presents but two instances of a virtually innite number of possible graphic portrayals of the formally stated process. This is why, if a picture is worth a thousand words, a formula is worth a thousand pictures.
liberating qualities
Formal logical-deductive systems are constrained by observationally and quantitatively principled initial assumptions, denitions, and axioms.4 Building on this foundation, rigorous derivations render deductions about events of principal interest. In this way, precisely framed empirical implications can, in effect, be evaluated for their potential importance, in advance of actual empirical testing. Fruits of these labors, then, include, not only the resulting insights, but also a potential increase in the overall economy of inquiry. Important questions can be posed before launching out empirically. Will the forecasted yield of information justify the associated investment? Is that which may be found actually worth nding? Does the theory posit results that, if corroborated, could signicantly affect explanation, measurement, and, indeed, future predictions? The point to be made is that the theorist is emancipated to safely
Figure 1. Probability of completing k stages in four time units (lower surface) and seven time units (upper surface) with each stage transacted at rate v (see Equation [2]; k ranges from 1 to 10, and v ranges from .01 to 3.0.
wade into the staked out substantive territory, with the above and related issues in hand, because her deliberations are constrained mathematically. These constraints represent an antidote to the hazards of armchair philosophy and psychologys ensuing reaction against exploration free of constant empirical vigilance (cf. instrumentalist vs. realist distinctions in scientic theorizing; e.g., Casti, 1989). As for the empirical enterprise, formal theory can increase the latitude of investigation in clinical science. It can do so by incorporating abnormalities associated with psychopathology into hypothesis formation and testing. Such possibilities are exemplied in the cognitive-science application presented below. In the enterprise of clinical science, certain hypotheses regarding pathognomonic disturbances may defy often-used methodological options, and formal theory can supply essential alternatives. To elaborate, one tack to take when testing hypotheses about cognitive debilities, for example, involves attempting to duplicate performance deviations in normals using experimental manipulations that are designed to mimic their etiology in patients. Formal theory can complement this strategy and, moreover, pick up where it leaves off. The approach is to modify quantitative mod-
16 modeling complex systems els of performance in ways implied by the hypothesized agent of deviation and then to evaluate the accuracy of the resulting predictions (elaborated on below). Such methodological endowments come into play especially where tendered sources of disturbance are beyond the reach of direct manipulations. Selected agents of dysfunction, although model hospitable, may be fundamentally alien to normal functioning or demand unethical extremes of experimental induction.
17 Formal Clinical Cognitive Science On balance, it can be said that such hurdles to formal theorizing may be daunting at rst blush. Current practices, however, need not be abandoned while cultivating the formal alternative as an increasingly friendly ally. Eliminating the present hurdles can reap the substantial dividends on outlay in time and labor that have been enumerated above.
philosophical orientation
Associations among cognitive science, clinical cognitive science, and applications to individuals for clinical assessment or other purposes
Figure 2. Transitions from basic cognitive science among normals to cognitive assessment among patients.
are depicted in Figure 2. Its arrows denote a smooth, if not seamless, transition among these investigative domains. In the rst instance, a model of cognitive performance addressed to the targeted territory of mentation is chosen. Deliberations include viability of the experimental paradigm and relevance of its response parameters. Addressed cognitive faculties, for example, may be those identied with visual, memory, and collateral processes (e.g., stimulus encoding and response mechanisms); the paradigm may take the form of a conventional visual- or memory-search task (Sternberg, 1975; Townsend & Ashby, 1983); and response parameters may include performance latency and/or accuracy.6 Several candidate models may present themselves. Selection from among contenders is based on congruence between the models features (e.g., its parameters or other properties) and concepts indigenous to the substantive research question (e.g., cognitive capacity, processing architecture [serial, parallel, or hybrid], efciency of capacity allocation, and so on). Extant models may require elaboration or modication, betting the currently driving issues. It may even be necessary to develop a model ab initio (see Barriers: Attitudes, Aptitudes, and Investigator Role above), depending on points of contact between the available armamentarium and the clinical-science queries that are pressing. The formal account of cognitive processing is then adjusted so as to accommodate performance deviations of disordered individuals. Alterations can be directed toward the architectural structure of the modeled processing system and/or toward model parameters.
19 Formal Clinical Cognitive Science Architectural structure refers to the arrangement of constituent processes involved in task transaction. To illustrate, in ascertaining and registering the existence of a presented alphanumeric item in a memorized set of items, component processes composed of encoding, memory-search, and response operations may be conducted serially, in parallel, or in some combination of the two. Structure refers as well to the arrangement of constituent stages (subprocesses) of the said processes (e.g., in the case of encoding subprocesses, parallel vs. serial input of the physical item features for template matching to those of the memorized item set; Townsend, 1984; Townsend & Ashby, 1983). Parameters refers to variables characterizing the processes, such as their size (operationalized as the number of subprocesses making up the process) and capacity (operationalized as the speed with which these subprocesses are transacted)the parameters denoted above as k and v, respectively. For the sake of parsimony, the normal model is minimally perturbed, pending realization of the target conguration. Necessitated adjustments purportedly indicate disorder-affected functions, and those remaining intact signify functions that are spared. In general, processing architecture tends be preserved; however, the same cannot be said for the values of model parameters (Neufeld et al., 2002; Neufeld, Vollick, et al., in press; Townsend, Fic, & Neufeld, in press). It is apparent from the foregoing that the models most closely aligned with the present purposes are those that are stochastic and dynamic (see Absorption of Theory-Exogenous Variance above). By certain extensions of their stochastic element, beyond that attending response parameters per se, these models become poised to accommodate unique aspects of individual performance. Specically, mixture models are fashioned to incorporate unequal values of performance-model parameters across units of observationin this case, members of diagnostic groups (Batchelder, 1998; Berger, 1985; cf. Busemeyer & Stout, 2002). The distribution of individual parameter values becomes, in effect, a prior distribution, in Bayesian statistical terms. Through the implementation of Bayess theorem (discussed below), performance-model parameter estimates, as well as performance-latency distributions, stand to be individualized with considerable precision. Analogous to the procedure followed in a medical-laboratory scenario, where a modest blood specimen is
20 modeling complex systems referred to the broader body of hematological knowledge, a modest cognitive-performance sample is now embellished with information conferred by the prior distribution of parameter values emanating from the participants group. Note that, with the introduction of Bayesian-prior distributions of parameter values, deviation-accommodating adjustments no longer take place directly at the level of the task-performance process model. In the case of mixture models, there is, instead, movement of the deviant groups Bayesian-prior distribution of parameter values. In this way, effects of disorder are registered as features of the prior, and, inasmuch as what has been altered is a distribution, provision for individual differences in the expression of these effects is retained. The arrows in Figure 2 are unidirectional, conveying the thrust of the current argument. There is, nevertheless, reciprocity among the above investigative domains. Successful capturing of deviant performance among clinical samples speaks to model robustness; those that readily accommodate performance extremes are preferred to those that are strained or fail in this regard (Neufeld, 1982; Neufeld & Mothersill, 1980). This widening of application represents a unique opportunity for generalization testing (Busemeyer & Wang, 2000) and, thus, potentially reciprocates benets back to basic cognitive science. Nor are the layers of investigation mutually exclusive (see the leftmost column of Figure 2). Successful titration of stochastic models can lead to parallel manipulations at the connectionist level of analysis, whose own success or failure adjudges theory robustness (Carter & Neufeld, in press; Roe et al., 2001). Neurophysiological levels of investigation also stand to benet. As described above, performance-model parameters (e.g., k , v) can be customized to individuals by amalgamating each ones performance sample with the parameter-distribution prior, ascertained for the associated diagnostic group. With the individual estimates in hand, parametrically homogeneous subgroups can be formed. The resultant specication of cognitive-process-latency distributions (e.g., that of cognitively encoding a presented item into a task-facilitative format) can, in principle, be used to establish the likely epoch of occurrence of the process for the constructed group (e.g., Neufeld et al., 2002, elaborated on below). When it comes to measuring evoked neurophysiological responses (as in fmri and erp), estimated time
21 Formal Clinical Cognitive Science windows of interest may, thus, be prescribed, complementing neuroanatomical regions of interest, in calibrating the space-time coordinates of measurement.
group and individual differences in the encoding process: bayess theorem in the current context
Recall that Bayess theorem states: Pr(A|B) = [Pr(A)Pr(B|A)]/Pr(B), (3)
where A and B are events; Pr(A) is the prior probability of A, or the probability of A preceding consideration of B; Pr(B|A) is the conditional probability of B, given A; and Pr(B) is the unconditional probability of B. As in the present context, A is often a hypothesized event, and B refers to experimental observation. This statement appears simple enough, until its wide-reaching ramications are unraveled. The parsimony of expression that, nevertheless, belies its profundity obviously qualies this theorem as one of statisticss aesthetically exquisite equations. Perhaps a more familiar general expression of Bayess theorem is Pr(Ai |B) = [Pr(Ai )Pr(B|Ai )]/[ iPr(Ai)Pr(B|Ai)], (4)
where Ai is one of a set of hypothesized events, Ai. In its present implementation, Ai becomes a parameter value of a test-performance process model, and B becomes a participants performance sample comprising N empirical latencies {t1, t2, . . . , tN},7 or {*} for short. The latencies refer to stimulus encoding because cognitive translation of presented items into a format that facilitates collateral functions (e.g., search for a matching memory-held item) tends to be elongated in schizophrenia (Neufeld et al., 2002). The following formula adapts these equations to the present setting of application and, in so doing, provides a comprehensive backdrop for explicating the potential assets enumerated at the beginning of this section:
G
w( |{*}) =
g=1
(5)
22 modeling complex systems In this equation, [w( |{*}) is the probability, or probability density (in the case of discrete and continuous parameters, respectively), of the parameter value , given the set of empirical latencies {*}; Pr(g|{*}) is the probability that the participant belongs to diagnostic group g, given {*}; and w( |{*}, g) is the probability (density) of , given {*} as well as the prior distribution of for group g. As stated above, it is through variation in the prior distributions of that group-related performance deviations become registered in Equation (5). It is the presence of cognition-relevant parameters in this equation, moreover, that takes Equations (3) and (4) beyond a theoretically barren actuarial status to substantive signicance. The precise role played by the group-specic priors in Equation (5) is pinpointed in its expansion (presented in Appendix A). Note that the equation is partitioned by elongated braces { } and elongated square brackets [ ] according to the respective terms bearing on the main developments (outlined below).
23 Formal Clinical Cognitive Science ods emanating from Paivios dual-coding theory (Paivio, 1971, 1986). Individuals with schizophrenia and controls indicated as quickly and accurately as possible whether or not the real-life sizeoverall volumeof a presented item (the probe item was either an object or an animal) was similar to that of a member of a previously memorized set of items (the memory set; xed-set procedure). Memory-set size, ranging from 1 through 4, varied randomly across trials. Presence versus absence of the probes size properties in a member of the memory set (positive vs. negative trials) was similarly scrambled. Half the participants in each group were presented with similarly sized drawings of probe items, whereas the other half were presented with names. Encoding demands were, accordingly, deemed to be greater in the latter instance because the initial verbal presentation of the probe would require referral to the imagery system for access to the demanded spatial-size properties (Paivio, 1986; for an examination of dual-coding and related predictions among schizophrenia participants, see George & Neufeld, 1984). Scanning memory-held items for the presence of the probes size properties was considered to be comparatively taxing. It required comparison of the probe and memory-held items overall volume, as set against subjective criteria of similarity in this property (Hockley & Murdock, 1987; Wright, 1977). Size attributes of items in the memory set were spaced such that the means of their normative-size ratings (Paivio, 1975) were at least 2 standard deviations of their Thurstonian discriminal-difference size dispersions apart (practically, at least 2 standard deviations of the distribution of difference scores between their normative-size ratings; Highgate-Maynard & Neufeld, 1986). Responses were designated as correct or incorrect on the basis of whether or not they conformed to the following criteria for positive and negative trials. Negative trials, whose correct response was a no button press, meant that the probe items size did not resemble that of a memory-set item. Here, the probe items mean real-life size rating was at least 1 standard deviation of the Thurstonian discriminal-difference dispersion from the normative mean rating of each item in the memory set. Positive trials, whose correct response was a yes button press, were those where the mean of the probe item, and that of a memory-set member, were, for all intents and purposes, identical. Practice trials, ensuring familiarity with task requirements and the nature of correct responding, were similar in number for each group.
24 modeling complex systems Further specics, including those surrounding provision for potentially confounding clinical and demographic variables and ascertainment of the viability of the paradigm for each diagnostic group of participants (e.g., applicability of normed item properties to each one), are detailed in Highgate-Maynard and Neufeld (1986) and summarized in Neufeld et al. (2002). Considering the current emphasis on latency, namely, estimated encoding durations for correct responses, it should be emphasized that error rates were comparable across groups and did not contribute to latency in any confounding way. Finally, trimming from the examined latency data those times estimated for processes residual to the one of current focusnamely, encodingis detailed in Neufeld et al. (2002; see also Neufeld, Vollick, & Highgate, 1993). Specic processes thought to be tapped by this task, then, include, rst, the cognitive translation of the probe item into a format abetting comparisons with the memory set or the memory sets relevant properties; second, the comparisons proper; and third, response processes. Results from Highgate-Maynard and Neufelds (1986) study represent a case in point of delayed completion of the encoding process among schizophrenia participants, amid integrity of remaining processes. Individuals with schizophrenia taking part in this study included paranoid and nonparanoid subgroups. Comparison groups involved nonschizophrenic psychiatric controls as well as nonpatient controls. The present developments are expedited by focusing on the paranoid-schizophrenia participants and the nonpatient controls. The clinical signicance of this delineation has been described in Nicholson and Neufeld (1993); paranoid subgroups have also tended to be more elongated in probe-item encoding than nonparanoids. Convergent support for this form of decit, evinced on the present and related paradigms, has been presented in other sources (Neufeld, 1991; Neufeld et al., 2002; Neufeld, Vollick, et al., in press; Neufeld & Williamson, 1996). Convergent support for the modeled composition of the decit (elaborated on below) arises from investigations of divergent paradigms and forms of analysis. These include studies analyzing expressly the architectures and parameters of item encoding (Vollick, 1994; Vollick & Neufeld, 2004), studies combining memory-trace theory and multidimensional scaling in assessing deviant similarity judgments (Carter & Neufeld, 1999), and neuroconnectionist assays of temporal and substantive aspects of item categorization (Carter & Neufeld, in press). Finally, potential signicance
25 Formal Clinical Cognitive Science as a source of symptomatology, notably thought-form disorder (delusions and thematic hallucinations), has been discussed in other sources (Neufeld et al., 1993; Neufeld & Williamson, 1996).
process-latency model
The process-latency model has already been alluded to in the earlier discussion of the nature of stochastic dynamic models (see, e.g., Absorption of Theory-Exogenous Variance). It consists of the Erlang distribution, which is invoked to portray the stimulus-encoding process. The encoding process is highlighted because, of the several processes putatively involved in task transaction in toto, it is one that is diagnostic of group-performance differences. The rationale
for the adoption of the Erlang and other distributions used here and the qualifying considerations surrounding the present applications have been detailed in Neufeld et al. (2002). The model postulates k stages of encoding (e.g., constituent operations of covertly accessing size properties of the probe item). The stages proceed at the rate v per unit of time (seconds, in this instantiation). In the deterministic case, the process would invariably terminate after an interval of k /v. Ambient trial-to-trial inuences naturally inltrate processing time and perturb this pristine expression. These inuences are summarized according to the probability density of process completion at time t: f(t) = (vt)k
1
/(k
1)! ve vt,
(6)
27 Formal Clinical Cognitive Science The values of parameters k and v are not invariant, but posited to vary across group members. Thus, the present mixture model provides for individual differences according to the groups Bayesianprior distribution of parameter values that randomly enter into the above process-latency model. Figure 3 displays the f(t) for four combinations of parameter values, for four hypothetical individuals.
spared; its distribution parameters remain unaltered. The estimates of r and k, obtained essentially through tting theoretical to empirical moments (e.g., Townsend, 1984), were 0.03735 and 2.5044, respectively. Note that, in order to circumvent certain interpretative hurdles apropos the present purposes, incorrect responses were excluded, as were positive trials (Neufeld & Gardner, 1990; Neufeld et al., 1993). The common distribution of v is presented in Figure 4. Unlike the distribution of v, that of k varies with diagnostic group and encoding-load condition. Specically, this distribution is shifted upward in a similar fashion, with elevation in the encoding load, and with paranoid-schizophrenia diagnostic status. In this sense, these factors represent exogenous and endogenous sources of encoding-load increase, respectively. The distribution of k is considered to be Poisson (Evans et al., 2000). The single parameter of this distribution is designated m. Its meanthe modeled average value of k is moved to the right by a constant amount for each source, as is its variance. This perturbation of the mixing distribution for k therefore signies an additive
29 Formal Clinical Cognitive Science increase in encoding subprocesses attributable to each factor of the diagnostic groups by encoding-load layout. The Poisson intensity parameter, m, is taken to indicate tendency toward encoding-subprocess recruitment on engagement of the probe item. The mean and variance of the Poisson distribution, then, are both m. The distributions base value for the nonpatient controls, under the low-encoding-load condition, is denoted m . The base value is increased by the amount h for each group under the higher-encoding-load, verbal-probe format and by the amount g with paranoidschizophrenia diagnostic status under both higher and lower encoding demands. Reference to the binomial distributions counterpart to m allows a more concrete understanding of this parameter. This counterpart, denoted np, approximates m as follows. The quantity n indicates the pool of encoding subprocesses potentially transacting the encoding process for any participant. The term p, in turn, portrays the probability of a latent subprocess being marshaled to the encoding operation. The value of n is m + h + g + x, where x is merely some very large number. That of p approximates m/n, where the base value of m, for controls under low-encoding demands, is m , to which is added h and/or g, depending on whether exogenous (encoding condition) and/or endogenous (paranoid schizophrenia) related increases in propensity toward greater subprocesses are present. The term n is tantamount to a dormant subprocess pool, and p may be viewed as a disinhibition term. Clearly, m increases with each term. The present interpretation designates p as the source of variation in m (Neufeld et al., 2002). In keeping with this stance, as n approaches the limiting value of innity and np approaches the limiting value of m, the Poisson and binomial distributions converge. It is the value m = np that governs the probability of a specic value of k . This probability is Pr(k ; m) = mk /k !e
m
(8)
The estimated value of m was .00001, that of h was 19.7390, and that of g was 70.0. The value of m suggests that the controls required hardly any processing to encode the rudimentary limned drawings of the probe items. Plots (smoothed) of the Pr(k ; m) for the present values of m are plotted in Figure 5. The general structure of the mixture-distribution model is sum-
marized in Figure 6. The gamma and Poisson mixing distributions of v and k , respectively, correspond to those displayed in Figures 4 and 5 above. The Erlang distribution of t, on which v and k bear in the individual case (exemplied in Figure 3 above), is the mixtures base distribution. With this structure in hand, estimation of the base distributions parameters for an individual can, in principle, be substantially sharpened; all that is needed is a modest sample of the participants own process latencies {*}.
one of the 10 participants in each of the four factorial combinations (2 diagnostic groups 2 encoding-load conditions). Each sample size was four, one for each memory-set size, under negative trials. Adjusting these latencies to highlight the encoding process has been described in Neufeld et al. (2002). The sets of trimmed values are presented near the top of Table 2. A graphic example of tailoring a parameter distribution to the individual is exemplied for a tentative case. Specically, consider the performance sample supplied by the control participant under the low-encoding condition {1.037, . . .}. Although unlikely according to Table 3, this sample still could have been produced by a member of the paranoid-schizophrenia group, under the low encoding load. If so, the customized distribution of k would have moved markedly to the left, relative to the prior distribution aligned with the paranoid schizophrenia/low encoding load combination. This relocation reects the joint probability of comparatively lower latencies making up this encoding-process performance sample. The Bayesian posterior distribution of k (smoothed) is presented in Figure 7, alongside the prior distribution corresponding to m = 70.00001. Note as well the increased precision of estimation, indicated by the diminished width of the posterior distribution. In fact, the standard deviation of this posterior distribution of k is 1.8110, whereas that of the prior is 8.36. Similar observations attend the individualized estimation of v.
Table 2. Unconditional Joint Densities of Encoding-Latency Performance Samples under Alternate Bayesian Priors
Representative Individual-Participant-Latency Sets Control High Encoding Load {1.137, 3.204, 1.371, 2.207} .54107(10-11) .70547(10-9) .00002 .80080(10-12) .29397(10-40) .23614(10-32) .75783(10-13) .65324(10-12) Low Encoding Load {1.037, 0.554, 0.191, 6.390}
Paranoid Schizophrenia Low Encoding Load {2.177, 3.184, 2.411, 2.498} .00684 .00737 .00083 .1486(10-12)
Classication
.17176(10-12) .999776(10-11)
.32930(10-7) Pr{*}
Parameters of prior distributions: r = .03735, k = 2.5044; m = 89.73901. Parameters of prior distributions: r = .03735, k = 2.5044; m = 70.00001. c Parameters of prior distributions: r = .03735, k = 2.5044; m = 19.73901. d Parameters of prior distributions: r = .03735, k = 2.5044; m = .00001.
Table 3. Posterior Probability of Group Membership, Given Representative Individual Latency Sets
Representative Individual-Participant-Latency Sets Control High Encoding Load {1.137, 3.204, 1.371, 2.207} .27083(10-7) .00007 .99993 .200186(10-12) Low Encoding Load {1.037, .554, .191, 6.390} .730536(10-29) .23473(10-20) .188326 .81167
Paranoid Schizophrenia Low Encoding Load {2.177, 3.184, 2.411, 2.498} .022 .9511 .027 .23972(10-11)
Classicationa
.52528(10-6) .00012
.99987 Pr{g} 0
Base rates for group membership = .05 for paranoid schizophrenia/high encoding load, .20 for paranoid schizophrenia/low encoding load, .5 for nonpatient controls/high encoding load, and .25 for nonpatient controls/low encoding load. b Parameters of prior distributions: r = .03735, k = 2.5044; m = 89.73901. c Parameters of prior distributions: r = .03735, k = 2.5044; m = 70.00001. d Parameters of prior distributions: r = .03735, k = 2.5044; m = 19.73901. e Parameters of prior distributions: r = .03735, k = 2.5044; m = .00001.
Figure 7. Posterior distribution of k , Pr(k |{1.037, .554, .191, 6.39}; m = 70.00001), and prior distribution of k , Pr(k ; m = 70.00001).
sons proximity to variously symptomatic groups. This scenario may apply to treatment evaluation over time: Is the individual being edged closer to normal functioning, as assessed by methods rooted in contemporary cognitive science? Estimates are available in terms of the posterior probability of a representative of the target groupa group with less or no symptom severitygenerating the sample, in contrast to a member of a symptomatic or more severely symptomatic group. In effect, what is being asked is: In Bayesian terms, what are the compatibilities of the obtained process-latency specimen with each groups pair of prior parameter distributions? The expression of Equation (5) above enclosed in elongated braces { } is now brought forth. Allowing that the performance is sampled at selected points in time, it is assumed that the representativeness of each symptom group along the way is known. In other words, transitions in prevalence of the candidate groups are, ostensibly, monitored so that base rates for a testees group membership, Pr(g), g = 1, 2, . . . , G, are available. It is assumed as well that the mixing distributions for these
groups, established at the outset, remain valid or, if they do not, that new priors have been formulated for the ensuing points of performance sampling. In the present application, the assumption of initial-prior stability is invoked. Finally, the present computations imply that the G classications are mutually exclusive and exhaustive. The classicatory system is, therefore, regarded as comprehensive. This feature underwrites the intended description of the presenting individual, which, in fact, is a nite, discrete posterior-probability mixture. The adopted approach entails an odds-ratio evaluation strategyan odds-ratio evaluation strategy for assessing the appropriateness for the current performance sample {*} of the G respective priors. It is deemed to facilitate the ow of the present developments; other Bayesian strategies, with unique statistical merits, are, nevertheless, available (Karabatsos, 2006). In contrast to the scenario painted above, it is possible that diagnostic-group base rates are, indeed, not available for the designated time points. Instead of monitoring an individuals progress by proling his or her likelihoods of belonging to symptom groups of known base rates, the primary issue may entail the base rates themselves.
Here, the focal problem may be one of evaluating the collective effectiveness of treatment by ascertaining movement in size of symptom-group membership. A question of this nature can, once more, be broached through the elongated-braced expression { } of Equation (5) above. This second case is taken up in its turn below. To expedite the exposition, the present classications formed by factorially combining diagnostic status and encoding-load condition are appropriated. These combinations again provide the Bayesian mixing-distribution priors of k and v as well as the representative performance samples listed in Table 2 above. We conveniently allow for a progressive decrease in symptomatology with movement across groups formed from combinations of paranoid-schizophrenia status/high encoding demands, to paranoidschizophrenia status/low encoding demands, through nonpatient status/high encoding demands, to nally nonpatient status/low encoding demands. Apropos the encoding manipulation as a surrogate for symptom
severity, to be sure, responding to increased encoding demands with an increased number of encoding stages is hardly indicative of disorder. It nevertheless conveys an escalation in recruited subprocesses, as a stand-in for the elevated number that may well go along with increased symptomatology (Nicholson & Neufeld, 1993). In all events, this pretense for expedience is inconsequential to the ensuing line of argument and in no way detracts from its analytic points. We posit a scenario where encoding-latency samples have been obtained from four participants, as arrayed in the rst row of data in Table 2 above. The values have been obtained, say, at a time of interest specied by the treatment regimen. We wish to estimate probabilities of symptom-group membership for each participant, given the obtained performance data. These groupwise posterior probabilities are endowed by the prior parameter-mixing distributions aligned with the candidate groups. To proceed, it is necessary to know the values of the uncondi-
39 Formal Clinical Cognitive Science tional joint probability density of each latency sample with respect to each group ucd({*}|g). The ucd({*}|g) is the joint density of the obtained sample of latencies, given group g, all values of v and k considered. The numbers assumed by ucd({*}|g) will change with g because the parameter-mixing distributions are group specic. The values of ujd({*}|g) for the current performance samples are presented in Table 2 above. Base rates assigned to the respective groups Pr(g), proceeding from the left to the right of Table 2 (from paranoid schizophrenia/high encoding load to nonpatient controls/low encoding load), are 0.05, 0.20, 0.5, and 0.25. The posterior probabilities of group membership are entered in Table 3 above. Each individual is proled in terms of distance from the respective groups, enumerated as the Bayesian posterior probability of group membership Pr(g|{*}). Not surprisingly, the highest values tend to be aligned with the groups from which the samples originated. The sole exception occurs for the sample from the paranoid schizophrenia/high encoding load combination. This result is attributable to both ucd({*}|g) and Pr(g) being highest in connection with the nonpatient controls/high encoding load amalgam. The point of the exercise, however, mandates that the clinical scientist/practitioner is blind as to the actual origins of the latency samples. Table entries can, then, provide pivotal information regarding participant status, as conveyed by the Bayesian expansion of the present mixture model. Allowing that all individuals were members of the same group at the outset, clearly the far-right column presents the most encouraging prole and the second-from-the-left column the least so. As intimated above, such multivariate proles can be charted across successive times of interest. The resulting individual congurations of treatment (non)response, at least as quantied by focal cognitive functioning, may, in turn, be statistically mediated to psychometric and other predictors.
40 modeling complex systems or her specimen of process latencies. In the present case, we entertain a sample of individuals who have been allocated to one or another group on the basis of presenting symptomatology at the selected time of interest. However, the present case differs from the rst case in that the base rates are now unknown, and it is this information that we seek. These densities of enclaves with differing symptomatology may be useful in assessing treatment efcacy across the designated times of interest. It may be argued that the desired course of action is simply to directly examine the array of diagnostic assignments that have been made at the selected times. It may be prohibitive, however, to diagnostically interview (or assess by some other means, such as psychometric) a large enough sample to ensure reasonable delity of estimates (see, e.g., Batchelder, 1998; Luke & Homan, 1998). If so, it seems desirable to bring to bear our available bank of information on cognitive functioning. In importing this information, we may settle, not only for a modest sample of individuals whose classications are updated, but also for a modest performance sample from each. The required sampling may, feasibly, be tacked onto sessions hosting symptom assessment at the successive times of interest. The strategy for estimation is straightforward enough. It consists of constructing a likelihood function for the updated participant assignments and then maximizing the function with respect to Pr(g), g = 1, 2, . . . , G. The likelihood function is made up of the joint probability of the updated participant categorizations. This joint probability comes down to the product of the posterior probabilities of those categorizations, given the latency samples {*} linked to the categorized individuals. The procedure constitutes a somewhat unusual but valid combination of maximum-likelihood estimation of a meta-Bayesian parameter (see Appendix A). The likelihood function and related computational details of maximum-likelihood estimation of Pr(g), g = 1, 2, . . . , G, are presented in Appendix B. The instantiation of this application once again makes use of entries in Table 2 above. For reasons of mathematical tractability, and in the interests of illustrative neatness, only three of the four data sets listed in this table are considered. The set corresponding to the nonpatient controls/high encoding load is excluded, deleting the third row and third column of ujd({*}|g) values in the table. However, in
41 Formal Clinical Cognitive Science practice, such a restriction is not necessary, as described momentarily. The maximum-likelihood estimates of Pr(g) were obtained by solving for each value using closed-form solutions to two simultaneous equations (two, rather than three, as the base-rate probabilities sum to 1.0; see Appendix B). It was found that increasing G beyond 3 became infeasible, even with the assistance of a computer-algebra program. In larger-scale applications, however, where G > 3 and/ or the number of diagnostically updated participants is somewhat larger, one may resort to a numerical-search algorithm (e.g., matlab Optimization Toolbox), rather than pristine closed-form solutions, to estimate the respective values of Pr(g). For the present purposes, the three sets of performance samples in the provisionally reduced Table 2 now represent those putatively obtained at the period of updated individual classication. It is, moreover, sufcient to allow each individual to have been assigned to the group adjacent to the displayed performance sample. In other words, three individuals have been classied, one to each group. Furthermore, the classied individuals have putatively supplied the very process-latency samples atop the columns of their respective groups. With these arrangements in place, base-rate estimates of the three groups, proceeding from left to right, become 0.6748, 0.0852, and 0.2399. The most symptomatic category thus remains in the majority at this time of evaluation, and those who are moderately symptomatic are sparsest. Nevertheless, the least symptomatic category is reasonably well represented, with an estimated relative frequency of nearly 0.25. It is tempting to insert these base rates into the expression enclosed by { } in Equation (5) above and, in turn, to evaluate the posterior probabilities of membership in the respective groups, given the associated performance samples {*}as was done in the rst case presented above. Note, however, that this exercise is nothing more than instructive, regarding the involvement of ujd({*}|g) and Pr(g) in arriving at the values of Pr(g|{*}). The latter, again proceeding from left to right in Table 2, are 0.119786, 0.119736, and 1.00. The rst amount resembles the second because the comparatively low ujd({*}|g) connected with the paranoid schizophrenia/high encoding load group for the individual so assigned is offset by that groups
42 modeling complex systems having the highest value for Pr(g). The value of 1.00 for the nonpatient controls/low encoding load assignment evinces the fact that the ujd({*}|g) associated with this specic group exceeds the next highest value by a factor of .2766325061(1021).
qualifying assumptions
The above estimates of Pr(g) assume that the parameter-mixing distributions initially prescribed for the symptom-based classications extend to the time of estimation. This assumption is intrinsically bound up with stability of conditions under which the current sample of task performance is obtained. The assumption is also yoked to continuation of the initial classicatory criteria leading to participant allocation. These requirements come down to retention of their earlier values by parameters of the mixing distributions r, k, and m, continuing tenability of the distributions themselves, and transference of the previous meaning of g in the presently estimated Pr(g). Dislodgement of the established priors from their host symptom groups may occur because of a shift in the nature of task performance, owing, for instance, to practice effects. Potential redress may require the substitution of different task items, analogous to using parallel forms of psychometric tests. To guard steadfastness of the G categories, symptom criteria for designations need, by and large, to be consistent. It is additionally assumed that differential dropout across the G groups and other sources of distortion in their representativeness are prevented. Finally, the computations once again imply that the G classications are mutually exclusive and exhaustive. In other words, transitions in symptomatology eventuate in one or the other of the G classes addressed at the outset. It is evident from these developments that tenability of the proposed analytic strategies makes for new and potentially useful avenues for formulating clinically signicant information. Available are inferences about the status of specic patients, as conveyed by cognitive-functioning aspects of symptomatology. Available as well are broader inferences about treatment efcacy. Furthermore, assumptional constraints are relatively salient. Finally, the developments remain entrenched in a formal cognitive-process model.
Figure 12. Cumulative probability distributions corresponding to probability densities of Figure 12 above.
tered circular reference to the very neurocircuitry whose functional signicance is under study. The trajectories in these gures have been constructed for an individual performance sample/prior distribution combination. They could, however, be extended to accommodate subsets of individuals whose sets of performance samples are similar and to whom the same priors apply. The resulting amalgams should render subsets with relatively homogeneous parameter proles (see Figure 7 above), along with associated latency trajectories. Aggregation across participants is often necessary to attenuate noise of mri signals or for selected statistical treatment. Amalgamation of parametrically homogeneous participants guards against an amalgam of systematic individual differences in the expression of cognitive functions under study.
multiple processes
Typically, even the simplest cognitive task recruits multiple stochastic processes (Smith, 1995; Townsend, 1984). Separability of the respective processes then comes into play. The processes may occur
45 Formal Clinical Cognitive Science in parallel, meaning that they commence simultaneously. Their nishing times are, nevertheless, staggered, according to their respective stochastic distributions (Townsend & Nozawa, 1995; Wenger & Townsend, 2000). Let the probability-distribution function for one of a pair of processes be F1(t) and that for the second be F2(t). It is stated here without elaboration that the greatest likelihood of the second process surviving, with the rst having elapsed, is dened by max([1 F2(t)] [1 F1(t)]) = max(F1(t) F2(t)). Measurement epochs addressed to the second process would be synchronized accordingly. The processes may, instead, proceed serially. Let the probabilitydensity function of the initial one be f1(t). Then the greatest likelihood that it has been accomplished, while the second has not yet elapsed, occurs at the modal value of f1(t), or max(f1(t)).
46 modeling complex systems type of coping basically takes the form of situating oneself in a multifaceted stressing context so as to engage the situations most innocuous option. Dissection of decisional control, through the application of mathematical combinatorics and related methods to the essential structure of this form of coping, discloses it to be a prototypical expression of selection and choice (Morrison et al., 1988; Neufeld, 1999b). Unveiled, in turn, are its demands on cognition, including option-cue encoding, visual and memory search, and response processes. Unveiled alongside is the undermining potential of debilities in the required cognitive functions, including those comprising retarded completion of encoding processes (Neufeld, 1991, 1999a; Nicholson & Neufeld, 1992). A factor potentially exacerbating the above risk is the toll taken on encoding speed by stress itself (Neufeld & McCarty, 1994; Neufeld, Townsend, & Jett, in press). Apropos the formal depiction of encoding, this effect translates into a general reduction in the parameter v of the base distribution of latencies (Figure 3 above). Recall that a stress-induced movement of the distribution of v to the left is identied with an increase in its mixing-distribution parameter r (Figure 4 above; Neufeld, 1994; Neufeld & Carter, 2000).
47 Formal Clinical Cognitive Science construct validity is that the measure act in accordance with theory (Wiggins, 1973, p. 406). In the present treatment, measures consist of model parameters. Construct validity for their imputed meaning, in turn, emanates from their mathematical properties, which are part and parcel of the stochastic dynamic models in which they participate. The light thrown on the nature of parameters by the mathematical properties they possess arguably embodies a new form of construct validity (cf. Cronbach & Meehl, 1955). The parameters of the appropriated process-latency model included v and k . Their interpretations stemmed from the assumed structure of the encoding process, as translated into a corresponding distribution of latencies. Distributions of the above parameters, in turn, have their own parameters, r and k, in the case of a gammadistributed v, and m, in the case of a Poisson-distributed k . Because they pertain to the mixing distributions for v and k , the parameters r, k, and m are a step removed from the empirical touchstones of an individuals latency data. Accordingly, called for is support for the constructs that they purportedly express, arguably from a commensurate level of analysis. Recall that the parameter m was aligned with encoding load and conveyed exogenous (task demands) and endogenous (paranoid schizophrenia) sources of increase. Its binomial-based dissection cast a certain light on the potential composition of this parameter (cf. construct representation, Embretson, 1983). The parameter r was identied with stress effects on processing speed. Its increase moves the stage-dispatching capacity parameter v to the left. Finally, the parameter k ostensibly corresponds to process-transaction competence, and its increase shifts the distribution of v to the right. The parameter k is endowed with additional mathematical properties that effectively illustrate the type of construct validity imbued by analytic derivations. The behavioral signicance of these properties hinges on the concept of maintenance of cognitive performance, over and against its breakdown, as follows. Within the present analytic context, integrity, versus collapsing, of performance requires that the value of k exceed a certain threshold, paralleling a critical level of performer skill. In contrast, increasing r and m prolongs performance times, but, in this case, as opposed to that of k, actual failure of the processing system is not precipitated when a nite range of values is violated.
48 modeling complex systems The current formal expression of performance breakdown takes the form of certain latency-distribution moments acquiring innite values: more succinctly, E(Tn) = , where n is the order of the computed moment. Behaviorally, the result signies that the population of trials occurring under the prevailing value of k includes a critical mass of extremely long, or essentially outstanding, completions (in effect, noncompletions). Consequently, the integral
00
f(t)t n dt = E(T )
does not converge (for further technical exposition, see the rst section of Appendix C). The rst four moments of a stochastic distribution (n = 1, 2, 3, 4) are somewhat familiar, comprising or being involved in its mean, variance, skewness, and kurtosis. In model evaluation, empirical moments beyond the second are seldom sampled because of instability. The present theoretical deliberations, however, operate in the realm of population properties, bypassing issues of sampling and measurement weaknesses. The order of the moment n can, therefore, be driven to whatever amount is prescribed by the governing theoretical analysis, which, in this case, concerns consequences of variation in the processing systems constituent parameters. At the same time, such theoretical explorations retain a bona de bearing on the interpretation of the parameter values, as fallibly estimated from empirical data. Observe that this approach to discerning the nature of a theoretical systems parameters has much precedent in elds ranging from mathematical ecology, to economics, to astrophysics. The above integral indicates how values of t and the order of the moment n act in concert to affect E(Tn). As n increases, its exponentiation of t correspondingly magnies the consequences of elongated latencies for the computed moment, including its transition to innity (Appendix C). Apropos the model parameter k, strengthening process-performance skill should inuence the distribution of t downward, resulting in E(Tn) being bounded to an even greater extent by nite values. That is to say, it should enable E(Tn) to withstand heightened values of n before transmogrifying to innity. In effect, k acts to diminish densities f(t) for higher latencies, including extremes, through its shifting the distribution of the process-completion rates, v, upward. Indeed, the threshold value of k
49 Formal Clinical Cognitive Science issuing in bounded moments is dened precisely in terms of n (see both below and Appendix C). Insufcient k implies that appreciable densities of v, f(v), are perilously close to v = 0, generating the critical magnitudes of f(t) at acutely high ranges. The above support for the construct validity of k as a capacityaffecting competence parameter is robust, as assayed from multiple analytic standpoints. It transcends process-latency base distributions with diverse characteristics; its interpretation is enhanced by pertinent Bayesian extensions; its effects are modied in expected ways by formally dened efciency of capacity application by the operative base distribution; its effects align with those of related parameters from similar mixing distributions; and it is evinced in familiar base-distribution statistical properties. Finally, its effects are distinguishable from those of parameters to which a different meaning is ascribed, notably, r, the avowed index of stress effects on processing capacity. These variations on analytic support are dealt with in succession. Their largely verbal descriptions are presented below, with technical justications allocated to the several sections of Appendix C.
Table 4. Process-Latency Base Distributions and Their Moments with a Gamma-Distributed Capacity (Intensity) Parameter, v
nth Moment k>n
t= 00
Erlang
]
For the exponential distribution, k = 1.0
Compound Poisson
( 1) d MT ( )/d |
where MT ( ) =
0
-e-M
f(v)[(em(M i (
T -M
1)
)-e |v]dv,
MT ( ) =
i
v , (v + )
f(v) =
and m is the parameter of the Poisson distribution of k . For example, mr ; (k 1) r2m(2(k 1) + m) (k 1)2(k 2)
E(T n|k ) = 00, hence E(T n) = 00, also by results for the general gamma distribution below
E(T) =
Var (T) =
continued
General gamma
k i=1
Stagewise rate transition Ck cin!r n (k n) , (k) where ci is a scalar of v for the ith stage, ci > 0; the rate thus in effect for stage i is civ; and
k
00 , as
(k n) = (0)
Cik =
j=1 j=i
( )
k
ci
cj ( 1) i+1
n+1
( 1) i+1
n
k i i=1 t= 0 k
k i
t= 00
(k () r n! (k)i n) ( 1) , 00 = 00
i+1
Weibull
00 , as (k n) = (0)
Expression of moment infinity as (0) or ln(00) merely rests on integration with respect to t first or v first, respectively.
52 modeling complex systems The above relation between k and n occurs for the Erlang distribution (Figure 3 above), with a xed-stage parameter k . A specic case is the single-stage exponential distribution, where k = 1.0. This relation includes as well the extension to k as a Poisson-distributed random parameter, otherwise known as the compound Poisson distribution (e.g., Feller, 1966; Ross, 1996). Together with the mixture on v, the latter distribution makes up the dual-parameter mixture model depicted in Figure 6 above. In the present context of base distributions whose parameter v is gamma distributed, the compound Poisson distribution is considered simply to be a base distribution whose Erlang stage parameter, k , is Poisson premixed. The rate parameter, v, is constant for each stage in both the Erlang and the compound Poisson distributions. The above relation between k and n, however, transcends base distributions with a static rate parameter. It extends to those with rate transitions across stages, known as the general gamma distribution (McGill & Gibbon, 1965). A notable version of the general gamma is a pure death system with a linear death rate, which is a parametric version of an independent parallel-processing model with unlimited capacity (see Townsend, 1990; Wenger & Townsend, 2000). This model has gured prominently in selected theorizing about the architecture of human information-processing mechanisms (see, e.g., Townsend & Ashby, 1983, pp. 8095). Another well-known and frequently used distribution to which the present relation applies is one with a single stage of process completion but with a rate that changes continuously with time (Weibull distribution). The selection of base distributions, and their distinguishing properties, to which the relation between k and n applies is summarized in Table 4 (further details are briey presented in the second section of Appendix C, additional elaboration being available in Neufeld, 1994). Also listed are the moments for k > n and k = n. Apropos the latter, note that, if the nth-order moment is unavailable (innite), as in the case of n = k, so are moments of order exceeding n (for k as an integer, see, e.g., Harris, 1966, p. 102; where k is not an integer and k < n, the function (k n) appearing in the moments expression takes on a succession of negative, innite, and positive values as k n decreases from 0, whereby moments about the origin become meaningless. At the heart of the matter is the attestation to k as a process-
53 Formal Clinical Cognitive Science transacting competence parameter, according to invariance of its innity-avoiding values across base distributions of heterogeneous makeup. Robustness of this nature bears a certain resemblance to empirical convergent validity, which entails agreement among divergent methods of measuring the same trait (Haynes, 2001).
bayesian extension
Additional light can be thrown on the proffered interpretation of k by dissecting its properties within a Bayesian context. Insight is afforded according to the give-and-take between k and both empirical performance samples and collateral parameters. A poignant exemplar involves the Erlangk ,v distribution whose rate parameter, v, is gamma mixed. The current Bayesian extension once more incorporates a set of process latencies {t1, t2, . . . , tN} = {*}, and the nature of latency-distribution moments again occupies center stage. The xed-parameter status of k streamlines the present exploration without curtailing the generality of available inferences. Findings tenably apply to classes of individuals sufciently homogeneous regarding k that this simplication is a reasonable approximation of the quantitative goings-on. The required moment expression, so conditionalized on k , can be immediately extracted from derivation (A.10) of Neufeld et al. (2002). It appears as follows:
E(T |{*}, k ) =
n
(9)
This expression affords a certain perspective on the posited interpretation of k from the standpoint of rudimentary classic psychometric theory surrounding performance-sample size or test length. The excursion into the psychometric analogy is followed once more by observations on the viability of the proposed interpretation of k. Isolating rst on the term (Nk + k n), an innite value of n E(T |{*}; k ) is avoided if Nk + k exceeds n. In this case, it is possible to have k = n without E(Tn|{*}; k) = if and only if Nk > n k, allowing, of course, that the present ti < , i = 1, 2, . . . , N. A value of k = n signies that a nite E(Tn|{*}; k ) does not characterize all members of the class of individuals to whom the current value of k applies.
54 modeling complex systems The specic member at hand is, nevertheless, characterized by a latency moment bounded by nite values. The psychometric analogy seemingly throws light on this conclusion as follows. Within this analogy, the stage parameter, k , corresponds to the number of parallel items composing a psychometric test. Second, N corresponds to the number of independent test administrations. The collective item sample is, therefore, Nk . With a sufciently large empirical sample of nite latencies, combined with the extant value of k, a nite value of E(Tn|{*}; k ) is assured for the testee at hand. The argument can be carried further as follows. Consider that the Bayesian posterior moment E(Tn|{*}; k ) can, in principle, be partitioned according to a probability mixture of two partial expectations, (E(Tn|{*}; k ) = ) and (1 )(E(Tn|{*}; k ) < ), where is the mixing parameter, 0 1.0. A sufciently large performance sample of nite process-completion latencies, in concert with k, in effect renders equal to 0. Conversely, a value of Nk such that Nk + k = n fails to nullify the viability of the partial expectation E(Tn|{*}; k) = , 0 < 1.0, and an innite moment ensues for the present individual. Bringing the above formulation to bear on convergent support for k as a process-transaction competence parameter, the following inference from Equation (9) is apparent. A performance sample size that is inadequate to ensure stability of nite latency values for the person at hand can be offset by a sufciently large value of k for the class of individuals to which the present performer belongs, specically, k > n Nk . Inspection of Equation (9) reveals further interplay between k and N and between k and k , both informative as to the nature of k. Allowing Nk + k > n, consider the following ratio, extracted from Equation (9):
1 (Nk + k n) . = (Nk + k) (Nk + k 1)(Nk + k 2) . . . (Nk + k n)
E(Tn|{*}; k ) evidently decreases as N and k increase, other things remaining constant. This result stems from the participants performance-based maximum-likelihood estimate of v, mle(v) being
Nk .
N
ti
i=1 N
ti ,
55 Formal Clinical Cognitive Science stant. Likewise, the estimated process-completion capacity of the current participant is raised in a parallel fashion with elevation in k (Equation (A.6) of Neufeld et al. 2002), as should be the case, considering the imputed meaning of k as a capacity-enhancing competence parameter. Finally, Equation (9) discloses that the interplay between k and k is comparatively complex. Computer computations indicate that changes in E(Tn|{*}; k ) with elevation in k can be nonmonotone, decreasing and then increasing. The decrease is understandable in light of the inuence of k on mle(v), considered above. Note, however, that, as k goes up, so does the process magnitude, with respect to the number of stages involved. Evidently, effects of the latter ultimately surpass those on mle(v), eventuating in a movement upward of the Bayesian posterior moment. Apropos the current thrust regarding k, such eventual elevation can, interestingly, be replaced with continuing decline if k is made increasingly large. This result again should be observed if k represents a competence parameter, whose inuence can override that of a progressive increase in task load k on E(Tn|{*}; k ).
f (t|v)dt = 1.0.
Consequently,
E(T n ) = =
00 00
rn/
On inspecting the structure of this moment, it is apparent that the burden placed on k for retention of nite values is now alleviated, or, indeed, intensied (for > 1.0 and < 1.0, respectively), by a factor of 1/ . Observe in passing that stress effects on moment magnitude are also modied according to in the term rn/ . Similar to the escape of E(Tn) from nite values if k = n/ , E(Tn) breaches nite bounds if = n/k. In this way, performer competence as conveyed by k and capacity-application efciency as conveyed by reciprocate with respect to moment retention of nite values. A complementary angle on the relation of k to capacity utilization, , is afforded by a rigorous and robust index of process-completion potential, as articulated in contemporary cognitive science (Townsend & Ashby, 1983; Wenger & Townsend, 2000 with respect to the clinical-science implementations, see Neufeld, Townsend, & Jett, in press). This index is aligned with E(Tn); it consists of 1n(F (t)), where F (t) is the process-latency distributions survivor function, or the complement of the cumulative probability-distribution function, 1 F(t) (see the rst section of Appendix C; also Figure 12 above and related text). The current interpretation of k implies effects on 1n(F (t)) that should closely resemble or mimic those of . The parameters k and should also play off one another in maintaining specic levels of 1n(F (t)). These associations are apparent in Figure 13, which plots the above index of process-completion potential as a function of k and for selected values of r, k , and t (formulas underlying Figure 13 are presented in the third section of Appendix C). The parameter k, then, behaves as it should as a capacity-supply competence index, as set against as a capacity-application efciency index. It does so with respect to both E(Tn) and the allied expression of process-completion potential ln (F (t)).
Figure 13. Values of process-completion potential 1n(F (t)) for the nonhomogeneous Poisson distribution, with rate parameter vt 1 and v gammar,k distributed, for k varying from 1 to 3 and varying from 1 to 5, r = .03735, k = 18, t = 2.25.
58 modeling complex systems Table 5. Moments of Weibull and Erlang Process-Latency Base Distributions, with Capacity Parameters Either WeibullV, or Gammar,k Distributed.
Base Distribution Mixing Distribution Weibull; v Is Randomly Mixed Vn (n/ + 1) (1 n/B) Vn/ Weibull; v Is Randomly Mixed Erlang; v Is Randomly Mixed
Weibull
Gamma
rn (n/ + 1) (k n) (k)
rn/
A notable example whose distributed variate, like gammas, ranges from 0 to , with intensity parameter v and shape parameter , is the Weibull distribution, described above. As the Weibull now serves as a mixing distribution, these parameters are designated V and B, respectively. The second section of Appendix C shows that, whether they are mixing or base distributions, the gamma and Weibull are nested in the generalized gamma distribution (additionally allowing k of the generalized gamma to assume the status of k as a continuous variable). An exhaustive analysis of the two distributions effects seems unnecessary. Sufcient to make the point are denitions of moment niteness in relation to k and B. Selected process-latency base distributions now include the Erlang and the Weibull itself. In the latter case, the intensity (capacity) parameter v alone and, then, v exponentiated by , v , constitute the gamma- or Weibull-mixed random variate. The respective moments are presented in Table 5. Isolating on the far-right-hand gamma function, (x), in each case, maintaining moments within nite bounds evidently requires identical values of k and B from gamma and Weibull mixing distributions, respectively. Convergently, then, these equivalent thresholds of k and B add to the constellation of analytic properties endorsing the construct validity of k. Finally, similar observations attend the discrete-variable analogue of the mixture involving the Erlang base whose parameter v is gammar,k distributed. In contrast to the Erlang base distribution, whose functions are those of continuous time t, the Pascal distribu-
59 Formal Clinical Cognitive Science tion denes the probability of completing a process comprising k stages on the jth trial, j = k , k + 1, . . . . Its parameters are k and p, where p denotes the probability of successful stage completion on a trial (Patil & Joshi, 1968). The parameter p is, thus, the discrete-variable analogue of stage-completion capacity and, in that way, maps onto v of the Erlang distribution. The Pascal-gamma mixture denes p as e u, where u is gamma distributed, with parameters r and k. Clearly, the directions in which r and k affect the distribution of the base processes capacity parameter are now reversed, suggesting that moment niteness may depend on critical values of r. Accordingly, the mean and variance of process-completion trials, for example, escape nite bounds if r = 1, and the variance does so if r = 2.
60 modeling complex systems In like fashion, v for an individual whose variance in process latencies is equal to that of the expected valuethe average varianceis ((k 1)(k 2))<1/2>/r. The result follows because the latency variance, given v, is k /v2 and the expected or mean variance is k r2/ ((k 1)(k 2)). (Note that the theoretically prescribed value of v for the expected standard deviation and that for the expected variance are not identical, as E((SD(T|v))2) = E(Var(T|v)) (E(SD(T|v)))2.) Here, nite values of the summary statistic require that k > 2.
in summary
The method of illustration was used to motivate and exposit engagement with analytic aspects of construct validity as it related to interpretation of model parameters. Focus centered on support for
61 Formal Clinical Cognitive Science espoused meaning of a parameter bearing on the capacity-resource pool of a cognitive process (e.g., stimulus encoding). The capacityresource pool was operationalized as a stochastic distribution of a process-model parameter conveying speed of process transaction. Construct validity of the imputed substantive signicance of the capacity-affecting parameter, k, in turn, took the form of its mathematically derived analytic properties. The stochastic distribution of capacity values was gamma, with parameters r and k. Increasing values of these parameters shifted the capacity distribution downward and upward, respectively. The shape parameter, k, was tendered as an indicator of process-completion skill or competence. Evidence constellated for the posited interpretation of k included substantively meaningful properties transcending specic processlatency (base) distributions. Contrasting structures of base-distribution models were employed, tenably standing for diverse types of cognitive-task operations. Bayesian extensions lent additional support to educed inferences. Consequences of varying k interfaced with the status of other processing-system parameters in ways to be expected with the interpretation accorded k. Convergent support was also endowed by alternate-distribution parameters, whose effects on task performance should resemble those of k, given that they occupy a similar mathematically prescribed role. Next, performance of hypothetical participants, whose distributions of process latencies typied those of their embedding class, were inspected for workings of k. Performance properties (summary-statistic functions of k) of these idealized participants again conformed to the claimed substantive signicance of this parameter. Finally, the compatibility of analytic results with empirical support for construct validity was described.
Stochastic Modeling and the Psychometric-Artifact Issue in the Study of Differential Cognitive Decit
Stochastic dynamic modeling of cognitive performance can speak to certain long-standing measure-theoretic predicaments accompanying the study of differential cognitive decit in schizophrenia and
62 modeling complex systems other disorders (Carter & Neufeld, 1999; Neufeld et al., 2002). At the forefront is the psychometric-artifact problem, which states that the apparent separation between pathological and control groups on a given measure conates differences on the addressed cognitive faculty with the measures psychometric properties. A measure can, therefore, spuriously indicate greater decit on the faculty it putatively taps because it possesses greater psychometric precision than measures directed to other faculties (Chapman & Chapman, 1973, 2001). Advocated redress of the problem has entailed prematching measures psychometric properties using a standardization group composed of nonpatients who vary widely on the faculties being examined. Matching should, purportedly, incorporate variance in the measures observed scores and in their respective reliabilities. In addition, simplicity of factorial structure is recommended (Strauss, 2001; cf. Neufeld, 1984; Neufeld & Broga, 1981). The issue and its endorsed solution continue to elicit considerable adherence (e.g., Brenner, Wilt, Lysaker, Koyfman, & ODonnell, 2003). The proposed solution of psychometric-property matching has also received substantial challenge, and alternatives have been articulated (Carbotte, 1978; Knight & Silverstein, 2001; Neufeld, 1984, 1989; Neufeld & Broga, 1977, 1981; Neufeld et al., 2002). The issue is dissected here by translating it into simple but arguably sufcient measure-theoretic terms and then evaluating effects of typical psychometric-adjustment practices. Boiled down to essentials, the argument has, for some three decades, been more or less as follows. More psychometrically precise measures will have more statistical power for detecting control/ pathological-group separation and, thus, be differentially favored over psychometrically inferior measures in indicating statistically signicant decit or disproportionate decit (expressed as an interaction with corresponding groupwise simple main effects). Comparative effect-size estimates should, of course, conform to the above intermeasure differences in statistical properties, other values (notably, sample size) being equal. Note that, for two groups, statistical power for detecting a difference on a selected measure is an increasing function of the squared noncentrality parameter n(1/4d2), where d refers to Cohens (1988) d, or the standardized population-group difference in means ( 1 )/ 2 . In this expression, is the common within-group standard devia-
63 Formal Clinical Cognitive Science tion, and n is the common group sample size. The expression 1/4( 1 )2 conveys between-group effects and is, tenably, composed of 2 some amalgam, rst, of those that are unique to group membership and constant among its members (corresponding to taxonic variables in Meehl & Golden, 1982), denoted here as , and, second, of those that represent between-group inequalities with respect to latent variables that also generate true-score variance across participants, within groups . The net contribution to the numerator of 1/4d2 from the latter source is denoted c , where the scalar c increases as sources of between-group separation on the measure overlap with those of within-group, interparticipant variation, 0 c 1.0. The expression 1/4( 1 )2 can, thus, be replaced with c + 2 . Finally, duly replacing 2 with + e, where the term e represents measurement-error variance, makes 1/4d2 itself become (c + )/( + e). (10)
Although not essential, for mathematical tractability and the purpose of illustration, it is assumed that none of c, , e, and are functions of one another. Turning to the issue of psychometric adjustment using standardization-group data, it is apparent from Equation (10) that the nub of between-group inequality, c and , is bypassed; within-standardization-group measure reliability is equal to /( + e), and observedscore variance is equal to + e. Interestingly, by Equation (10), where all between-group variance is conveyed by , > 0, and c = 0, 1/4d2 can, evidently, be increased with reduction in . That is to say, the measure becomes less group discriminating as its standardization-group psychometric precision goes up. In fact, the effects of reducing are unveiled by differentiating (10) with respect to , the result being (ce )/( + e)2. Clearly, increasing and, along the way, increasing measure reliability, /( + e)has adverse effects on power if ce < and favorable effects if ce > . Thus, the consequences for d2 of varying are complex, and, in fact, unknowable from standardization-group variance, because the latter is mute about the structure of group differences composed of c and . Of course, decreasing e invariably increases reliability as well as d2, as would be expected, and as is obvious from Equation (10). But c and remain unknown, leaving psychometric matching to grope blindly when it comes to amounts by which e
64 modeling complex systems should be altered (say, through increased accuracy of instrumentation or preaggregating measurements; see, e.g., Neufeld & Gardner, 1990). The upshot is that equalizing a pair of measures on and e has no problem whatsoever in leaving them grossly unequal in their group-discriminating power.8 Empirical examples instantiating the theoretical dissociation between control/pathological-group discriminability and classic psychometric properties have been amply documented elsewhere (Knight & Silverstein, 2001; Neufeld, 1984, 1989; Neufeld & Broga, 1981). These instantiations need not be recapitulated here. More productively, it is shown that Equation (10) can be superposed over and against a stochastic-modeling formulation. The conversion metamorphoses the equation and arguably leaves behind as vestigial the statistical-property issue.
1,2
,r,k
Here, m1,2 signies that the expected variance is taken across both control and pathological groups, which, for simplicity and in keeping with the design depicted in Figure 6 above, differ only with respect to m. The actual amount prescribed by the mixture model is
(r2m1,2)/((k 1)(k 2)).
(11)
The model counterpart of is the variance in the conditional mean or expected latency, given k and v:
Var(E(T|k v)m
1,2 ,r,k
)=
(12)
Both the expected conditional variance and the variance in the conditional expectations (discussed above) provide for within-group random variation in v and k among participants. Model analogues of c and cast k as Poissonm mixed and xed parameter, respectively. In the rst case, k is randomly distributed across participants within groups. The distributions themselves are separated according to the values of m1,2. For example, m1 may be m of Figure 6, and m2 may be m + g. The differences in m1 and m2 are, thus, afliated with c of Equation (10) above. The scaled value of , c , therefore has a model representation as the following function of the between-group difference in means:
1/4 (k 1) (k 1) .
Equations (11)(13a)/(13b) together embody twelve model predictions of empirical terms. The data of each diagnostic group, under each experimental condition (encoding load, of Figure 6), can be assembled into the expected conditional variance, variance in conditional expectations, and group mean. Parameter estimates, obtained, for example, through moment-tting procedures, are ve in number according to the prevailing model design. It is apparent that each parameter is called on simultaneously to participate in empirical values whose sources range from within-participant variation to intergroup separation. Among other considerations, severity of model
( (
(m1r)
(m2r)
(k r)
(k r)
) )
(13a)
(13b)
66 modeling complex systems testing increases with the ratio of empirical observations to parameter-set size and the heterogeneity of conditions to which the model applies. Inferential validity regarding proposed dysfunction, then, ultimately rests on the rigor of the model diagnostics that are exerted. Predictions desirably include qualitative properties of empirical congurations. For example, the present model prescribes additive effects on means of diagnostic groups and experimental conditions, which is tantamount to a predicted second-order difference (see above) being equal to 0 (e.g., Neufeld et al., 1993; Neufeld & Williamson, 1996). Augmenting signicance tests for goodness of t involving predicted distribution moments (crafted from Equations (11)(13)) can also be constructed (Neufeld, Vollick, et al., in press). Further tests are available from distribution properties, which, in the case of dynamic stochastic modeling, are functions of time t (Garca Prez, 1994, 2000; Van Zandt, 2000). These properties ow from the very design spawning the above distributional moments (e.g., density functions, illustrated in Figures 10 and 11 above). Moreover, Bayesian extensions stand to further proliferate model predictions. Posterior density and distribution functions, given an individual sample of latencies (e.g., Figures 89 and 12 above), once more are engineered by the stochastically modeled system giving rise to Bayesian-posterior distributional moments themselves (illustrated in Neufeld et al., 2002).
Concluding Comments
Placing clinical cognitive science on a decidedly formal platform can unearth productive routes of study and reveal otherwise hidden substantive nuances. Brokering the clinical ramications of contemporary cognitive science makes for a potentially seamless transition between these two domains of investigation. Overarching model designs can be constructed so as to mediate group-level ndings to individual differences in their expression among constituent members. Developments self-disclose implications for evaluation of the cognitive side of treatment-program efcacy, and progress of an individuals own response to treatment, with respect to cognitive functioning. Also disclosed are assumptive underpinnings of the ad-
67 Formal Clinical Cognitive Science umbrated assessment technology, along with lingering ambiguities awaiting resolution. Formal clinical cognitive science uniquely convenes a quantitative infrastructure for assessing construct validity of meaning assigned to its parameters and, potentially, other properties. It also engages measurement and statistical challenges to inferences about cognitive dysfunction. The current offering at minimum represents a beachhead from which, it is to be hoped, a stronghold of formal clinical cognitive science and assessment methodology may be consolidated.
({ 1/[
G g=1
In this expression, Pr(g) is the current base rate of group g, corresponding to the relative frequency of members of g in the participant sample at large; ujd({*}|g) is the unconditional joint density of the obtained latency data set, given group g, meaning the combined probability density of {t1, t2, . . . , tN}, given group g, all values of considered; w( |g) is the probability (density) of , given group g; and cjd({*}| ) is the conditional joint density of the obtained latency data set, given the parameter value . Note that the above expansion reveals two sets of Bayesian priors. One comprises the groupwise mixing distribution of . The other comprises the discrete, nite mixing probabilities of the groupwise priors being operative, corresponding to the base rates Pr(g). The latter is a meta-prior of this compound Bayesian formulation. Note as well that the densities of , including that of any specic value , are determined by the group at hands prior distribution of . Where becomes a set of parameters (e.g., v and k ; see the text), each with its own independent group-related mixing distribution, one or another member of the set may be considered separately. Separate consideration occurs, for example, when constructing a posterior parameter-mixing distribution, following the acquisition of a
68 modeling complex systems performance sample {*}, as exemplied in Figure 7 above. Computations entail aggregating over values of the set-aside parameter(s) according to its (their) group-prescribed prior probabilities (densities) while addressing specic values of the parameter that has been singled out (marginals), also according to its group-specied prior probabilities (densities). Computational details undergirding the present developments have been presented in Neufeld et al. (2002).
Pr(g|{*}) j =
J j=1
Pr(g) ujd({*}|g) j /
G g=1
Pr(g) ujd({*}|g) j .
(B.1)
This product can be show to be a reduction of the joint multinomial probability of the J observed classications under the model Pr(g|{*}), each probability being an expression of the multinomial likelihood of classication j (see, e.g., Garca Prez, 1994; Riefer & Batchelder, 1988). Maximum-likelihood estimates of the Pr(g), g = 1, 2, . . . , G 1 G1 (recalling that Pr(G 1) = 1 Pr(g) ), are obtained by differentiatg=1 ing (B.1) with respect to each term in question, setting each derivative to 0, and simultaneously solving the G 1 equations in the G 1 unknowns. A numerical-search algorithm will be required to replace the analytic solutions as computations become unwieldy with increasing J and/or G. Note that Equation (B.1) beguilingly invites a 2 test for empirical goodness of t of the prevailing model of Pr(g|{*}). Specically, 2 ln(B.1), which here can be shown to equal 2 ln(B.1)/1.0, is tantamount to 2 ln(multinomial-likelihood ratio), an expression that stands to be distributed as 2 with its associated degrees of freedom (cf. Carter & Neufeld, 1999). In fact, this expression is dubiously distributed as 2df=J (G 1) in the present case because the current data array suffers from the problem of extreme sparseness of cell-wise observations (see Delucchi, 1993; Tollenaar & Mooijaart, 2003).
f (t) tn dt = 00 .
n
Equivalently,
( 1) nE(T n) = dm/d
00
MT ( )|
=0
= 00 .
f (t) e tdt,
encounters a singularity at
(1 F(t))tn1dt = n
00
F (t)tn1dt,
where F(t) is the cumulative probability-distribution function (see Figure 12 above and related text) and 1 F(t) = F (t) is the survivor function. This version of E(Tn) relates distribution moments, and their availability as nite values, to process-completion potential. The latter is dened as f(t )/F(t )dt , (C.1) 0
t
where f(t )/F (t ) is the hazard function H(t ), indicating the conditional rate of process completion at time t , given noncompletion by t . Because F (t) = exp( 0 (H(t )dt ), (C.1) is readily seen to be simply 1n(F (t)) (see the discourses of Townsend & Ashby, 1983; Wenger & Townsend, 2000). This expression of process-completion potential, and its association with E(Tn), is called on in developments presented in the text under Interplay with Capacity Utilization and the third section of Appendix C below.
t
70 modeling complex systems pacity (distributional intensity or scale) parameters are randomly distributed according to a gamma distribution whose own intensity or scale parameter is denoted r and whose shape parameter is denoted k. Moments are computed for the case where k > n, n being the order of the moment, and where k = n. Three of the base distributions, the exponential, Erlang, and Weibull, are nested in the generalized gamma distribution, whose probability-density function is (vt va)
k 1
va) ).
(C.2)
The probability-density function for the Erlang distribution is obtained from (C.2) by setting equal to 1.0 and a equal to 0. The density function for the Weibull distribution results when k = 1.0 and a = 0. The exponential distribution, in turn, is nested in both the Erlang and the Weibull distributions, as its density function is derived from (C.2) when k = = 1.0 and a = 0. Loosely speaking, each base distribution has a conjugate prior, in that both prior and base entail exponentials, making for mathematical tractability (see, e.g., Geddes & Scott, 1989). The probability-density function of mixture-model latencies f(t) for the Erlang base distribution, with its capacity (also rate) parameter v being distributed as gammar,k, is
rktk 1 (k + k ) . (k)(k 1)! (r + t) k + k
Then
f(t)tndt = E(T n)
The next base distribution is the compound Poisson (e.g., Feller, 1966; Ross, 1996), of which the special case at hand is composed of an Erlang distribution whose stage (distributional-shape) parameter k is Poisson distributed with parameter m (see Figures 5 and 6, as well as Equation (8), above and the related text). The Erlang rate parameter of this base distribution, in turn, is gamma distributed, with parameters r and k. A concise method of obtaining the moments of this second-order mixture capitalizes on its moment-generating function
< 00 ,
if k > n,
t= 00
= 00
t= 0
if k = n,
71 Formal Clinical Cognitive Science MT( ); the nth-order derivative of the latter, evaluated at = 0 and scaled by ( 1)n, yields the nth-order moment (for an especially lucid and still current exposition of the moment-generating function, see Kenny & Keeping, 1951). Proceeding accordingly,
MT ( ) =
00
00
f (v)
k = 00 k =0
00
f (v)
k = 00 k =0
00
where MT ( ) is the moment-generating function for each of the i k -stage intercompletion times, or v/(v + ) for the Erlang distribution with rate parameter v. As the k intercompletion times are identically and independently distributed for a given value of v, MT ( )|(k v = M k ( )|v . Note that T
1
MT ( )|(k = 0 ) = 0.
Thus,
MT ( ) =
n
00
)1)
dv e m
E(T ) = ( 1) d MT ( )/d | = 0 . and The nth-order moment of this second-order compound Poisson distribution will be innite if k = n, as follows: E(T n) = =
k = 00
m k e m E(T n|k )| k =0 k !
k =n
k = 00
m k e m. 00 e m. 00| k =0 k !
k =n
= 00| .
k =n
By similar reasoning, E(Tn) = |k=n also for the general gamma base distribution (discussed below). Turning to the general gamma base distribution, unlike that of the Erlang, the rate parameter is stage specic. In the present implementation, the rate for stage i is civ, ci > 0, i = 1, 2, . . . , k . For this distribution,
00
=0
Cikci n! rn
(k n). (k)
(i 1)), the density function f(t) For the special case, where ci = (k can be computed, from developments supplied by Morrison (1979), as i
k
( 1)
+1
i=1
(r + it) k + 1
(ki ) r ik .
k
f(t) tn dt
k i=1
( 1) i + 1 k rn n! i
()
(k n) (k)i n
< 00 ,
if k > n
k i=1
( 1) i + 1 k i
()
t= 00
= 00
t= 0
if k = n.
The probability-density function for the Weibull distribution is v(vt) whereby E(Tn|v) = (n/ + 1)/vn. E(Tn), in turn, is readily available as
00
exp( (vt) ),
13
1
For the nonhomogeneous Poisson process, with dynamic rate vt having been gammar,k distributed,
,v
and
k 1 F(t) =
j= 0
The hazard function H(T) = f(t)/F (t) enters into the index of process-completion potential as follows: 0 H(t )dt . Happily, this integral is available as 1n(F (t)), especially considering the possibility of complex f(t) and F (t), such as those above (see the rst section of Appendix C).
Notes
Sources of support for this research include operating grants from the Social Sciences and Humanities Research Council of Canada and the Workplace Safety and Insurance Board and operating grants and a New Emerging Teams grant from the Canadian Institutes of Health Research. Appreciation is expressed to the University College of the Fraser Valley, the Department of Psychology, University of British Columbia, Kenneth D. Craig, sabbaticant host, and to Ruth and Douglas Barr, private citizens, for providing facilities during the writing of this manuscript. 1. A parameter is a variable in a mathematical expression whose values can change without altering the structure of the expression in which it is embedded (cf. Borowski & Borwein, 1989, p. 435). 2. Consider, as a representative example, numerical vs. symbolic integration, a special case being numerical solutions to differential equations composing nonlinear dynamic (chaos-theoretic) systems, as contrasted with desired (but most often intractable) explicit solutions of the initialvalue problem (Koak, 1989). 3. Formal statements foster place keeping regarding the current state of progress in the problem area. They make clear what has been formulated, whether new formal proposals differ, and, if so, in what respects. Blind alleys are agged and promising leads discerned. The place of precedent is usually more fuzzy in the case of verbal statements. Even if they seem to rediscover preexisting ideas, subsequent proposals almost always escape indictment for ignoring antecedents and, if taken to task, usually appeal to allegedly important nuances of difference. 4. These explicit starting points can always be challenged, precisely because they are explicit. Informal theories escape because they are not forced to lay bare their assumptive moorings before commencing (cf. Bandura, 1984; Staddon, 1984). 5. Neuroconnectionist modeling, entailing computer simulations of task-transacting activations among neuronal units or unit modules (Rum-
References
Aguirre, G. K., Zarahn, E., & DEsposito, M. (1998). The variability of human, bold hemodynamic responses. Neuroimage, 8, 360369. Averill, J. R. (1973). Personal control over aversive stimuli and its relationship to stress. Psychological Bulletin, 80, 286303. Bacon, F. (1937). The Advancement of Learning. In Richard Foster Jones (Ed.), Essays, Advancement of Learning, New Atlantis, and Other Pieces (pp. 169 235). New York: Odyssey. (Original work published 1605) Bandura, A. (1984). Representing personal determinants in causal structures. Psychological Review, 91, 508511. Batchelder, W. H. (1998). Multinomial processing tree models and psychological assessment. Psychological Assessment, 10, 331344. Batsell, R. R., Polking, J. C., Cramer, R. D., & Miller, C. M. (2003). Useful
86 modeling complex systems The perspective for psychopathology that we wish to put forward in this chapter is in itself interdisciplinaryof its two ingredients, one is dynamic systems theory, the other cognitive science. In this introductory section, we will outline our theoretical guidelines and suggest how the two main ingredients of our approachcognitive science and dynamic systems theorymay be combined. We will argue that cognitive acts can be viewed as coherent patterns emerging spontaneously in a complex cognitive system. Pattern formation via self-organization is the cornerstone of this approach to psychopathology. Motivational processes are, consequently, dened as those forces or gradients that drive the complex cognitive system. It is with respect to such motivational gradients that cognition is more or less optimal. Cognitive science is a eld to which neuroscientists, psychologists and psychiatrists, philosophers of mind, workers in articial intelligence (ai), and many others contribute. This eld has gained momentum in the last decades especially because of advances in brain research (e.g., the use of brain-imaging techniques and the profound implications of those techniques for psychiatry) and because of the perennial questions of consciousness that have been receiving increased attention recently (Carter, 2002). Cognitive science is, thus, a natural foundation from which psychopathological states of cognitive functioning may be regarded. Among the current approaches in cognitive science, the distinction between the computationalist perspective and the dynamic approach has been debated intensely. The computationalist perspective in cognitive science elaborates on the theory of information processing and continues the tradition of classic ai, which is based on the physical-symbol-systems premise (Newell & Simon, 1972). According to this premise, the mind is regarded as a physically implemented symbol system. Consequently, mental processes are identical to symbol processing, which can, in principle, be simulated using digital computers. The major part of todays cognitive psychology, as well as the cognitive-behavioral therapies currently in use in psychiatry and clinical psychology, is derived from computationalist theory. The dynamic approach to cognitive science (e.g., Clark, 1997; Pfeifer & Scheier, 1999; Tschacher & Dauwalder, 1999, 2003) has challenged this idea of cognition as computation. Bridging assumptions of dynamic systems approaches to cog-
87 Approach to Psychopathology nition have been identied: ofine reasoning is continuous with online motor-control strategies in the sense that abstract cognition may be decoupled from behavior in an actual environment but still work with the same principles (Mechsner, Kerzel, Knoblich, & Prinz, 2001). There is agreement in the dynamics community that cognitive patterns are not provided by programs, as in the computationalist perspective, but are emergent properties of the complex brain/mind system. Thus, the colloquial distinction between software and hardware makes sense only with respect to computers, not with respect to naturally occurring cognition. By these assumptions cognitive science and dynamic systems theory have joined forces to formulate a non-Cartesian theory of cognition. Emergent properties are modeled in the mathematical framework of self-organization theory (Haken, 1983, 1996; Kelso, 1995; Nicolis & Prigogine, 1977), the core part of the theory of complex systems. Self-organization theory (synergetics) deals with systems composed of several or many components. By means of their interaction, these components can produce new qualitative features on macroscopic scales; such features are often directly observable as patterns. In other words, the emergence of new qualities is studied. The main question is whether there are general principles that govern the behavior of complex systems when qualitative changes occur. These situations are probably of particular interest. And it has, in fact, been shown that a large number of systems are accessible to unifying mathematical and conceptual approaches. Synergetics starts from the observation that the behavior of many systems is strongly determined by environmental conditions. These conditions may be divided into constant conditions or constraints (e.g., those constants of a physical environment that conne behavioral systems) and further environmental conditions that energize or drive the systems (e.g., free energy in physical systems, task affordances in psychological systems). In the mathematical approach these latter environmental conditions are taken care of by control parameters. In many cases control parameters have the form of externally applied gradients. The general strategy of synergetics is as follows: It sets out from a state of a system that is already known under a certain control-parameter value. When one or several control parameters are changed, the system may become unstable. Then it may tend to leave its state and develop a new structure or behavior.
Figure 1. Bistable Gestalt array (Rubin, 1921). Either two (black) faces or a (white) vase is recognized (for details, see the text).
Synergetics shows that the behavior of the system close to instability points is described and determined by order parameters. According to synergetics, thein general feworder parameters determine the behavior of the many individual components. This implies an enormous information compression because it sufces to describe the order parameters instead of all the components. On the other hand, the individual components react on the order parameters and, in this way, even generate the order parameters. Thus, the relation between order parameters and components is based on circular causality. Quite often order parameters show very simple behavior, for instance, bi or multistability; that is, a system can acquire different states under the same external conditions. An example of visual perception is shown in Figure 1, Rubins (1921) well-known vase-face Gestalt, a bistable Gestalt array. In this gure two order parameters compete with each other: perception of the pattern face(s) or perception of the pattern vase. Both order parameters organize the components (the black-and-white details of the display) in a specic fashion. Perception of these patterns is mutually exclusiveit is impossible to perceive both at the same time. What are viewed as noses in the domain of one pattern become merely background as soon as a vase is perceived. Under neutral circumstances, transitions between the patterns occur frequently; such Gestalt ips are analogous to phase transitions in dynamic systems theory. Moreover, control parameters such as task affordances can
89 Approach to Psychopathology
1 "chaos" 2
"order"
"chaos"
"chaos"
"order"
4 "order"
Figure 2. Sequence of Gestalt-like written words (left column, cells 14) that give rise to the perception of either chaos or order. Right column, depiction of respective potential landscape (for details, see the text). After Tschacher (1997).
modulate the emergence of perceived patterns. If, for example, Figure 1 is presented in a series of similar displays together with the instruction to determine the gender of faces in the pictures, the vase pattern is hardly ever perceived. A further essential property of pattern formation via self-organization can be illustrated with the help of Figure 2. Words of a somewhat ambiguous quality are presented (see the left column of Figure 2); in cell 1 the word chaos is quite reliably identied. Here again, the percept chaos emerges as the order parameter, the qualitative pattern recognized instantaneously in the stimulus of cell 1. The stability of the percept chaos resulting from this stimulus is represented by the potential sink depicted in Figure 2 (see the right column next to cell 1). The state of the system, symbolized by the black circle, rests in the minimum of a potential landscape (the attractor of the system). The stimulus is, then, altered gradually in the sequence of cells 2, 3, and 4. Most viewers, however, still read chaos in cell 2 and cell 3 (the system remains in the respective attractor of the potential land-
90 modeling complex systems scape), a phase transition occurring only in cell 4. Here, the attracting capacity of the Gestalt chaos has given way to the Gestalt order, a phase transition has taken place, and the state of the system has been qualitatively changed. In this way the stability of emergent patterns can be explored: a Gestalt may, and often does, remain unchanged in spite of alterations of the actual stimulus. Self-organized patterns have attractor-like dynamic properties. Interestingly, when the stimulus sequence is reversed in experiments with several subjects, it is found that the timing of the phase transition may depend on sequence. When the state of the system is started from order (see cell 4; the circle dwells in the potential minimum on the right-hand side) and passes through cells 3 and 2, the phase transition occurs in cell 1 instead of cell 4. This phenomenon of delayed phase transitions depending on past sequences is called hysteresis. Hysteresis is usually viewed as a hallmark of nonlinearity in system dynamics. In the present context, hysteresis is an additional ngerprint of attractors governing the perceptual system. In summary, we have shown that Gestalt perception may be analogous to pattern-formation processes described in the framework of dynamic systems theory. We claim that these are pivotal processes found throughout cognition and action. As we cannot go into more detail here, it must sufce to point to some active elds of research where pattern formation is at the center of investigations. In the movement sciences, for instance, cognitive coordination and movement coordination are modeled along such lines (Kelso, 2003). In neuroscience, synchrony in transient aggregates of neurons in the cortex is seen as the hallmark of the self-organization of cell assemblies (Miltner, Braun, Arnold, Witte, & Taub, 1999; Singer & Gray, 1995; Varela, 1995). In clinical psychology and psychiatry, therapeutic strategies have been developed on the background of dynamic systems concepts, for example, in psychotherapy research (Grawe, 2004; Mahoney, 1991) and in theoretical psychiatry (Ciompi, 1997). Patterns of the mind emerge from the complex mind/brain system via self-organization processesthis general premise puts great emphasis on the connectivity among the components as well as on the driving of systems by environmental-control parameters. We will briey explore both topics in the rest of this section.
91 Approach to Psychopathology
connectivity
There are two general strategies employed to investigate neurocognitive systems with respect to their interaction and self-organization, top-down and bottom-up. The top-down strategy begins with observable (macroscopic) phenomena, such as the perception of Gestalts and the ngerprints of phase transitions introduced above. The bottom-up approach rst identies components of a system and then bases the pattern emerging from the system on a description of the components. In focusing on the connectivity of systems components, we would adopt a bottom-up approach. We have argued that the synergy (i.e., the working together) of many components leads to the emergence of macroscopic patterns. The comparable term coined by the Gestalt-psychological tradition is nonsummativity. The Gestaltist holistic position claimed that aggregates are not identical with the sum of their components (e.g., Khler, 1920). With this core tenet of Gestalt theory, psychology departed from associationist psychologywhereas associationism stated that properties of aggregates can be derived from the properties of their components, Gestalt psychology insisted on aggregates possessing qualities that are absent at the level of components. Therefore, the historical roots of contemporary emergentism and of complexity theory lie in Gestalt theory. The crucial point about the bottom-up approach is that we must clarify the nature of the components. What, however, are the components of (neuro)cognition? Dening an elementary cognitive act is notoriously a nontrivial task in phenomenological psychology, by which one can easily be lead into the shallow waters of introspectionism. Attempts to dene the components of behavior in a cognitivebehavioral fashion, for example, as conditioned stimulus-response associations, have remained rather hypothetical. In a biopsychological framework, however, the task appears much easierthe activity of a single neuronal cell in the brain is a good candidate for a component because it is well dened and, at least in principle, observable. Ensembles of neurons are formed in concordance with Hebbs law: Cells that re together wire together. The research program of connectionism is, thus, based on the connectivity of single cells. Connectionist models are good examples of emergent complex systems (e.g., Haken, 1987).
92 modeling complex systems Several issues in psychopathology have recently been approached with a focus on connectivity. Schizophrenia, for instance, can be seen as a misconnection disorder characterized by altered interactions among brain regions. The cognitive-dysmetria hypothesis (Andreasen, Paradiso, & OLeary, 1998) predicts that the circuitry between areas of the (especially frontal) cortex, the thalamus, and the cerebellum is unbalanced in schizophrenia spectrum disorder. A related approach views the ion channels of glutamatergic synapses as responsible for the loosening of associations that Bleuler (1911) identied as the core symptom of spectrum disorders. Here, the emphasis is again on changed neuronal connectivity, which may result in a host of cognitive impairments, for example, Gestalt perception changes (Phillips & Silverstein, 2003) and decreased binding capabilities (Garcia-Toro, Blanco, Gonzalez, & Salva, 2001).
control parameters
In this section, we introduced the synergetic theory of pattern formation. In addition to the connectivity of many components in a complex system, we pointed out that the driving of the system by control parameters constitutes a necessary condition for self-organization to occur. Such energetic driving of cognitive systems is traditionally termed motivation. In the previous two perception examples (Figures 1 and 2 above), task affordances and attention/vigilance determined what was being perceived. In the absence of motivation, or in situations with competitive motivational parameters, some order parameters may never be generated, even if they appear obvious to unbiased observers. This has been demonstrated, for example, in experiments on inattentional blindness (Simons & Chabris, 1999) where observers were found to be functionally and consciously blind to certain items in a visual scene when their attention was focused on an alternative task, even if this task was performed using the same visual stimulus material. Most psychopathological states, in addition to the obvious affective disorders, are associated with specic motivational and affective decits or excesses. The prototypical courses of schizophrenia, for example, often have uctuating levels of negative symptoms, such
93 Approach to Psychopathology as at affect and emotional withdrawal (see the next section). Thus, it is to be expected that the motivational driving of the cognitive system in such patients may undergo marked changes during the course of the illness. We conclude the theoretical introduction to this chapter by mentioning the interaction of (cognitive) pattern formation and (motivational) control parameters. So far we have concentrated on the control side of the system-environment coupling, that is, how control parameters affect the generation of patterns. Yet this reects only one side of the coin of interaction: it is self-evident that the patterns, in turn, exert an inuence back on the control parameters. As we have claimed above and discussed elsewhere (Tschacher & Haken, in press), motivation interacts with cognition in a manner analogous to the driving of complex systems by gradients of free energy (also called exergy in ecology; Schneider & Kay, 1994). In such systems, emergent patterns are consistently found to present the most efcient ways to reduce the gradients that caused their emergence in the rst place. We argued that, when applied to cognitive systems, this principle of efciency or optimality in reducing environmental gradients was a key to understanding intentionality in cognition. In this theoretical section we have sketched a formal systems model for psychopathology. We think that this may turn out to be an encompassing big picture, from a natural science perspective, of how cognition works. Admittedly, however, this big picture must be lled out with empirical ndings and scrutinized for its predictive value. In the remainder of this chapter, we will examine the evidence that has been accumulated and put forward ideas about how one might proceed in this line of research on psychopathology.
94 modeling complex systems approaches can be summarized under three levels of analysis. On a rst level of analysis, the study questions relate to how things change in time. The results on this level are descriptions of longitudinal developments, called trajectories in the language of dynamic systems theory. This type of analysis is mainly descriptive. A second level of analysis investigates how things interact in time. In other words, the interrelation of the systems variables is analyzed. Systematic interrelations result in stable patterns of a system, that is, attractors. Finally, a third level of analysis is related to the emergence of patternsnonlinearity, degrees of order, and pattern formation. It is important to note that these different levels of analysis are complementary rather than mutually exclusive. Both descriptive and more rened analytic approaches are necessary to illuminate self-organizing systems.
95 Approach to Psychopathology
Prototypical Dynamics and
Dynamics
Convergent Cyclic Chaotic
Attractors
Time
Time
Time
Attractors
Point Limit Cycle Chaotic
0-dimensional
1-dimensional
>2-dimensional
called points attractors. Another prototypical dynamics is characterized by oscillations: limit cycle attractors. Finally, unpredictable yet nonrandom uctuations are produced by certain nonlinear dynamics that have been discovered only recently (Rssler, 1976). In such a way psychopathological systems may be classied as exemplars from a zoo of prototypical attractors arising from various linear and nonlinear dynamics. Psychopathology in acute schizophrenia, for example, is often characterized by dramatic and repeated changes in symptoms. To date, however, research on the dimensions of psychopathology in schizophrenia has been mainly cross-sectional in nature (Peralta & Cuesta, 2001). Correspondingly, whereas the cross-sectional structure of schizophrenic symptoms has been studied extensively, little is known about the development of symptoms during acute episodes. In a study of symptom trajectories, 46 schizophrenia spectrum patients (18 females; mean age 24.7 years) were examined daily during an average treatment period of 104 days (Kupper & Tschacher, 2002). A novel time-series approach was used to identify initial phases of response and other descriptive features of the trajectories. Figure 4 shows the symptom processes of four exemplary pa-
tients. Each process consisted of three trajectories, termed psychoticity, excitement, and withdrawal, depicting the positive symptoms, anxiety and tension, and negative symptoms together with depression observed in patients. The time-series method used in this study allowed the identication of dynamic factors, consisting of trends, levels, and durations of these trajectories that were assessed in two periods of treatment, the initial improvement phase and the ensuing stabilization phase. These factors summarized the temporal structure of symptom evolution throughout the psychotic episodes. Five such dynamic factors were found: (1) overall level of positive symptoms, (2) duration of nonspecic response, (3) slope of response in all symptom domains, (4) enduring negative symptoms, and (5) duration of response regarding psychoticity. Compared to patients with an acute schizophrenia-like psychotic disorder, schizophrenia and schizoaffective disorder patients ranked higher on factor 4 (enduring negative symptoms). They tended toward a lower level of positive symptoms and showed a less prominent response to treatment. The examination of a subsample of 19 patients with relapse indicated a
97 Approach to Psychopathology prolonged duration of initial treatment response regarding psychoticity. These results supported the validity of this approach for the description of symptom trajectories. In this study the symptom change throughout psychotic episodes showed distinct temporal features. Compared to a merely cross-sectional view, this approach provided a dynamic analysis of the dimensions of schizophrenic psychopathology. Therefore, it was possible to separate the stable and the variable components of psychopathology in the patients. The trajectories of psychopathology found suggested that, whereas levels of symptoms are, in fact, often independent cross-sectionally, their change within a given patient was often strongly interrelated. Moreover, these results suggested that the different aspects of psychopathology might be related to some single, common process. In schizophrenia meaningful temporal patterns in psychopathology are not restricted to acute phases of the disorder. In a study of rehabilitation (Kupper & Hoffmann, 2000), courses of psychosocial functioning during intensive vocational rehabilitation of schizophrenia patients were examined. Although considerable research has been undertaken on psychosocial treatment and rehabilitation of patients with chronic schizophrenia, few studies have examined individual courses by means of repeated and frequent observations. A more dynamic view of rehabilitation was thought to disclose patterns of response useful for both understanding and treating symptoms and disabilities associated with chronic schizophrenia. In this exploratory study time series of 35 schizophrenia outpatients participating in a vocational rehabilitation program were examined by a novel quantitative approach by which dynamic patterns were identied. Time-series regression was applied on weekly behavioral ratings of psychosocial functioning, provided by the nurses observation scale (nosie). The mean, trend, and variability of each trajectory were calculated. Cluster analysis revealed ve groups of trajectories: (1) stable at a high level, (2) uctuating at a middle level, (3) middlelevel functioning, tending toward a slight descent, (4) steep descent of functioning, and (5) unstable at a low level of functioning. Examples from groups 1 and 4 are shown in Figure 5. The ve groups of patients with different trajectories varied at intake with respect to psychopathology, cognitive dysfunction, and measures of self-concept, locus of control, and coping. At program termination pronounced differences were found among the groups
in vocational reintegration. The different trajectories can, thus, be understood as typical pathways linking patient characteristics to rehabilitation outcome. This study exemplied that a broader use of dynamic designs can substantially clarify the variety of reactions of patients to psychosocial interventions. From a systems-theory point of view, these trajectories arose from patients characteristics interacting with environmental constraints that were induced by the introduction of intensive rehabilitation with increased levels of stress. The task affordances related to the rehabilitation environment acted as gradients (i.e., environmental-control parameters), leading to a potential destabilization of the system. This resulted in a transfer either to a higher level of functioning or to a deterioration in functioning. Similar patterns are known from other forms of psychological intervention. As an example, it is observed that psychotherapy can initially amplify the differences among clients. It should be noted that, generally, these few examples showed that the descriptive level of how things change in time provides information that is complementary to traditional research approaches. Cross-sectional analysis cannot explain this rich source of variation. On the grounds of dynamic systems theory, we may suppose that the diverse courses described here have resulted from attractors, the driving forces of longitudinal evolutions of psychopathology. The time-series methods used in the studies discussed above allowed the description and clustering of such evolutions but gave no indications as to the attractors and patterns underlying them. For this type of exploration additional methods must be applied, methods that
99 Approach to Psychopathology provide a closer look at the mechanisms of change leading to observable dynamics.
Figure 6. Individual time-series model of a patient (see the text). * p < .05; ** p < .01.
Multiple and stepwise regression analyses were then performed on the basis of the parameters of var models of all 84 patients. Two var parameters were found to be associated signicantly with favorable outcome in this sample: withdrawal reduction of psychoticity as well as excitement increase of withdrawal. These ndings were interpreted as indicating how patients cope with psychotic episodes. It may be assumed that this illuminated mechanisms that drove the symptom evolutions in this sample. These ndings must, however, be regarded as preliminary assumptions, as hypotheses about the underlying attractors of the symptom dynamics. As a further step in a consecutive unpublished study, we combined this analysis of schizophrenia courses with a comprehensive assessment of neuropsychological functioning and psychopathology (Spaulding et al., 1999). In the Schizophrenia Process Study, symptom courses of 114 schizophrenia spectrum patients were observed using daily ratings of psychopathology (mean duration of observation: 88 days). A 10-item scale for daily symptom assessment was applied (Todays Evaluation of Psychopathology, or tep). The rating scale was composed of three factors: positive symptoms, negative symptoms, and anxiety-depression. The courses were analyzed by time-series methods using var to model the day-to-day interrelations between symptom factors. Comprehensive neuropsychologi-
101 Approach to Psychopathology cal assessments were performed at the beginning and at the end of the observed courses. Associations between the day-to-day patterns of symptoms, neurocognition, and outcome were calculated. Preliminary results (Kupper, Tschacher, & Hoffmann, 2004) revealed specic relations between longitudinal symptom patterns and both outcome and neurocognition. The interrelations between positive and negative symptoms were associated with outcome. If positive symptoms entailed negative symptoms, outcome tended to be unfavorable. In cross-sectional studies different symptom domains as well as symptoms and neurocognition are often found to be independent. The results of this study, again, suggested that the various symptom domains as well as symptoms and neurocognition are, in fact, associated if longitudinal patterns of symptoms in individual patients are examined. In a study of crisis intervention (Tschacher & Jacobshagen, 2002), the remediation processes after psychosocial crises were modeled. We included patients with diagnoses such as adjustment disorder and an affective disorder who were often hospitalized with a status of suicidal behavior or ideation. The focus of this study was on cognitive orientation (outward events centered vs. inward centered). A sample of 38 inpatients who were assigned to treatment in a crisis-intervention unit was monitored in order to study the interrelations of process variables. The process data consisted of patients self-ratings of the variables mood, tension, and cognitive orientation, which were assessed three times a day throughout hospitalization (on average, over 22.6 days). Again, var models of the process data were computed to describe the prototypical dynamic patterns of the sample and explore process-outcome associations. Outcome of crisis-intervention treatment was evaluated by pre/postquestionnaires. On a rst level of analysis, the expected linear trends were found pointing to an improvement of mood, a reduction of tension, and an increase of outward cognitive orientation. Time-series modeling (level of analysis 2) showed that, on average, outward cognitive orientation entailed improved mood (see Figure 7). The time-series models predicted the treatment effect, notably the outcome domain reduction of social anxiety, yet were unrelated to the domain of symptom reduction. From the results it was concluded that crisis intervention should focus on having patients increasingly engage in outward cognitive orientation, that is, reduce the levels
.34*
TENSION
.97**
COGNITIVE ORIENTATION
.90***
Figure 7. Interrelations among system variables in crisis intervention. * p < .05; ** p < .01; *** p < .001.
of self-focused attention. This therapeutic approach should support stabilizing mood, reducing anxiety, and activating the resources of patients. Similarly, driving forces of change were extracted from psychotherapy courses (Tschacher, Baur, & Grawe, 2000). A sample of 91 courses of dyadic psychotherapy using different treatment modalities was analyzed in order to study session-by-session dynamics. The presented problems generally belonged to the spectrum of neurotic disorders, the largest diagnostic groups being adjustment disorders, anxiety and phobic disorders, eating disorders, and relationship problems. The process data consisted of therapists and patients session reports. Therapy outcome was evaluated by pre/ postquestionnaires and direct measures of change. After data reduction by principal-component analysis, linear time-series models of the resulting factors were computed to describe the prototypical dynamic patterns of the sample and of the modality subsamples (cognitive-behavioral, client-centered, schema-theoretical cognitive psychotherapy). It was found that the factor patients sense of selfefcacy/morale governed the observed dynamics of the sample, whereas the therapeutic-bond factors did have less of an effect on the dynamics (Figure 8). The dynamic patterns of client-centered therapies differed from other modalities. The dynamics-outcome ndings showed that direct measures of change were associated with a spe-
Figure 8. Interrelations among process factors in psychotherapy. * p < .05; ** p < .01.
cic process pattern in which the patients sense of self-efcacy was, in turn, supported by other process factors. This time-series approach was also utilized to uncover psychopathological changes and side effects related to different substances employed in psychopharmacology. Clinical observations and recent ndings suggested different acceptance of morphine and heroin by intravenous drug users in opiate-maintenance programs. Such programs are conducted nationwide in Switzerland. In a psychopharmacological study (Tschacher, Haemmig, & Jacobshagen, 2003), we postulated that this different acceptance may be caused by differences in the perceived effects of these drugs, especially how desired and adverse effects of both drugs interacted. Thus, we assumed different mechanisms of pharmacological action underlying the manifest effects. We measured the desired and adverse effects of high doses of injected morphine and heroin in patients included in the program. The interrelations between both types of effects were determined to test the hypothesis of a differential mechanism of action. Thirty-three patients (5 females, 28 males; mean duration of previous street heroin use 10.7 years, mean age 30.1 years) were randomly allocated double-blind to the substance groups. The average daily dose per participant in the heroin condition (n = 17) was 491 mg, in the morphine condition (n = 16) 597 mg. The observation period lasted 3 weeks; an average of 70 injec-
104 modeling complex systems tions was received. After each injection of either substance, various aspects of drug effects were recorded systematically. Ratings were summarized into the factors euphoria and adverse effects. Time-series models were computed for each participant on the basis of the factor scores, using the var format described above. A highly signicant difference between the substances was found in the interrelation between euphoria and adverse effects. Adverse effects of heroin preceded higher euphoria, whereas adverse effects of morphine preceded subsequent lower euphoria. Additionally, the nding of a generally higher level of adverse effects in morphine was replicated. These results pointed to different mechanisms of action of the two opioids when the perceived drug effects are evaluated in a eld setting. This may explain the better acceptance of heroin in opiate-assisted treatment of intravenous drug patients. In addition to the results concerning the studied sample, we proposed the method as a promising pharmacological tool for the comparison of substance groups other than opioids. In general, this second level of analysis allowed for the identication of interrelations among process variables of many different origins and in various time frames. The monitoring units may be xed (e.g., daily ratings), natural (e.g., at each injection of a substance), or a combination of both (e.g., at each psychotherapy session). The interrelations of systems variables can be studied under naturalistic conditions, giving this approach high heuristic value and ecological validity. We view such interrelations as core attributes of dynamic systems that give rise to the patterns and attractors described above. Some premises for the application of level of analysis 2 methods must be regarded, notably, that the approach assumes stationarity of the processes. For this reason we used several techniques in the aforementioned empirical studies to ensure stationarity, for example, detrending of courses prior to modeling. On a next level of analysis, transitions in order and pattern formation were analyzed. As described in the introductory section of this chapter, such transformations of systems leading to the formation of new patterns are a striking feature of self-organizing systems. The next section therefore presents studies of nonlinearity and pattern formation in psychopathology.
106 modeling complex systems items of session reports, t = session number). One minus this ratio then provides the measure of order. We computed this measure for the 20 rst and the 20 last consecutive sessions using the 33 items recorded in therapy-session reports. Results clearly supported the hypothesis of a signicant increase of order during the course of psychotherapy in a comparison of rst versus last sessions of each therapy course. This nding was corroborated using an extended data set of a later study (Tschacher et al., 2000). In a second step of this study we addressed a question that follows naturally from this nding: Is there a connection between order and therapy outcome? The general impression achieved by the outcome study was unequivocal: Order of the therapy system as well as increase of order were related to positive outcome of therapy. We monitored 41 outcome measures, 20 of which were correlated with Landsberg order in a statistically signicant way. These associations pointed to order and order formation as being functional for good treatment results. Improvement of the self-image of patients and a decrease in social anxiety were linked with both order and increase of order. Psychopathology of the psychotherapy patients was found to be connected to increase of order. Order was also computed to investigate the course of schizophrenia in acute treatment (Kupper, Tschacher, & Hoffmann, 1997) and to study rehabilitation processes in psychiatry. In these data the assessment of order was based on daily symptom ratings using the instruments introduced above and on weekly evaluations of progress in a psychiatric vocational rehabilitation program (Kupper, 1999). In the study of acute treatment we analyzed psychopathology courses of 21 patients (mean age 24 years, 10 of 21 female) during a psychotic episode. Daily ratings of psychopathology were performed. As in the study reported in Tschacher et al. (1998), order and order formation were calculated using the measure Landsberg order. The results of this study showed that order and order formation correlated negatively with the overall level of psychopathology and with the nal level of psychopathology at discharge of patients. Order generally increased throughout the treatment of psychotic episodes. In the vocational rehabilitation data (Kupper, 1999), order and order formation in courses of 30 participants (mean age 27 years, 10 of 30 female) with schizophrenic disorders were explored. The same methodological approach to order formation as in the previous study was
107 Approach to Psychopathology Table 1. Summary of Order and Order-Formation Results from Studies of the Treatment and Rehabilitation of Schizophrenia Patients
Vocational Rehabilitation Initial phase: no correlation Correlates of order (r approximately 0.40.7) Middle phase: positive outcome (level of integration) Acute Treatment Initial phase: more negative symptoms and depression (outcome) Middle phase: less confusion and tension (outcome) Final phase: no correlation Yes Initial-middle: less overall tension, fear, and ambivalence Initial-nal: less negative symptoms and depression
No
used. This program for vocational rehabilitation included vocational training, social skills training, family interventions, job placement, and support by job coaches. The assessments were weekly ratings of behavioral and emotional indicators of functioning, from both participants and supervisors perspectives. The primary outcome was level of vocational reintegration after the conclusion of the program. The results showed that order and order formation were again related to a positive outcome, here regarding vocational reintegration. No general increase of order was found. Overall, the majority of results in these data were consistent with what we labeled order effect (level of Landsberg order correlating to favorable outcome) and ordering effect (increase of Landsberg order correlating to favorable outcome; see Table 1). Moreover, meaningful differences between studies are evident from Table 1. Acute treatment seems to consist in a transition from a disordered acute state to a more stable state toward the end of treatment. Rehabilitation programs with their higher demands, however, appear to start from a stable entry point. The increased stress induced by these demands does not allow a general ordering effect but leads instead to largely varying developments among patients. Generally,
108 modeling complex systems it can be hypothesized that, in these studies order reected the stability of the mental system of the patients. Therefore, the tacit undertone of the word disorder (meaning mental illness) seems justied, even when we dene order in the strict mathematical sense. Fingerprints of Nonlinearity and Chaos One of the hypotheses that have been derived from dynamic systems theory is that seemingly random temporal behavior may be generated by relatively simple deterministic nonlinear systems. Chaos theory, a branch of systems theory, has pointed out how such systems may provide both stability (attractor-driven dynamics) and innovation (sensitivity to environmental changes). It was, therefore, suggested that nonlinear systems might provide explanations for the often unpredictable trajectories of psychopathological disorders, especially schizophrenia (Ciompi, 1997). The review of Paulus and Braff (2003) lists the recent dynamic approaches used to investigate the temporal architecture and the disease process underlying schizophrenia. In a study of daily ratings of psychopathology in 14 schizophrenia spectrum patients where observations were possible for extended periods of time (between 200 and 760 consecutive days), we modeled the attributes of the disease process (Tschacher, Scheier, & Hashimoto, 1997). To do this, we rst estimated the forecastability of the time series of each individual patient. Then we applied a series of bootstrap tests that implemented a hierarchy of potential process attributes, such as random, linearly autocorrelated stochastic process, and linear stochastic process with identical power spectrum. Finally, the actual forecastability of the observations was compared to the forecastabilities of the bootstrap time series. Whenever signicant differences occur, we may conclude that the dynamics of the observed time series was not sufciently explained by the process attributes, thereby pointing to nonlinearly generated dynamics of the observed data. The analysis suggested that 8 of 14 schizophrenia patients presented nonlinear dynamics. These investigations were compatible with the hypothesis that schizophrenia may be characterized by chaotic evolutions. Yet direct proof of deterministic chaos in relatively short empirical time series is unattainable, therefore, one cannot assess the chaoticity of the schizophrenia courses we studied. Visual inspection of forecasting accuracies showed, however, that the decay of predictability typical of deterministic chaos was present in the larger
109 Approach to Psychopathology part of the sample. Together with the rejection of alternative hypotheses (especially linearly correlated stochasticity), this allows that chaotic processes may underlie psychopathology in schizophrenia. We think that the chaos hypothesis of schizophrenia and other diseases (e.g., depression; Pezard et al., 1996) must be treated with considerable caution. Although there is some evidence from a variety of data sources, both behavioral and neurophysiological, the prerequisites for the generally very sophisticated dynamic methods are seldom realized in psychopathology data. Existence of chaos may well remain undetected because of the constant uctuations found in any environment outside a physicists laboratory. In addition to chaos representing a quite ephemeral quality of (some) nonlinear systems, it is probably not their most important property. We already introduced pattern formation as a functionally valuable and, at the same time, ubiquitous phenomenon and will concentrate on ngerprints of cognitive patterns and their stability in the following. Stability of Cognitive Patterns in Schizophrenia Perceptual and cognitive patterns are characteristically changed during psychotic experience. The tradition of Gestalt psychiatry (Conrad, 1958; Matussek, 1952) views the evolution of schizophrenia by a succession of stages: everyday cognitive patterns may be impaired prior to the outbreak of psychosis (derealization as Gestalt loss) and may then be resurrected in a bizarre way (delusions and hallucinations as malfunctional Gestalt re-formation). These phenomenological studies thus pointed to a variety of alterations in Gestalt perception and cognition indicating that both disorganization and delusional hyperorganization should be found. A number of empirical studies have, consequently, focused on the changes in (predominantly perceptual) organization in schizophrenia spectrum disorder (Uhlhaas & Silverstein, 2005). Cognitive psychology views Gestalt-like processes as essential during preattentional stages of parallel stimulus processing; in these early stages stimulus features are grouped and bound together. The large majority of these studies suggested that perceptual organization is dysfunctional in schizophrenia. These changes in perceptual organization may reect a more encompassing impairment in the coordination and binding of spatial and temporal information and of information originating from different sensory modalities (Phil-
110 modeling complex systems lips & Silverstein, 2003). Interestingly, Gestalt dysfunction does not necessarily go hand in hand with patients showing decreased task performance. Place and Gilmore (1980) and others found that schizophrenia patients did not present a generalized decit when compared to controls; in fact they were, for example, more accurate than controls in counting tasks where grouping of stimuli did not support performance. Research has shown that alterations of perceptual organization differ between subgroups of patients and, especially, correlate with symptomatology. A number of clinical symptoms were linked with Gestalt perception. In their review Uhlhaas and Silverstein (2005) cited evidence for a correlation of Gestalt dysfunctions with cognitive disorganization as well as with both positive and negative symptoms. In a recent study of Gestalt perception (Tschacher, Dubouloz, Meier, & Junghan, in press), we investigated several apparent motion paradigms. We included 32 schizophrenia patients (66% dayclinic patients, 34% inpatients; 81% males; mean age 27.3 years) and 32 control subjects matched for age, sex, and education. Clinical symptomatology of the patients was measured by the Positive and Negative Symptom Scale (panss). Although well in need of treatment, most patients were not acutely symptomatic, receiving average panss scores below 3 (low symptoms). For all computations we used the factorization of the panss according to Lindenmayer, Grochowski, and Hyman (1995), which consists of the ve factors positive, negative, excitement, cognitive (disorganization), and depression. Nevertheless, this schizophrenia group had reaction times markedly decreased compared to controls. The mean reaction time was increased by 1.7 standard deviations in patients, pointing to their generalized impairment of cognitive functioning. Circular Apparent Motion (cam) The frames displayed in Figure 9 were presented alternately for 500 ms each. Generally, Gestalts in the shape of movement illusions result. The impression is always either that of a wandering of the dots in a clockwise or counterclockwise circular direction, or that of a rapid oscillation of the two apparent motions. Thus, the stimulus presentation is inherently multistable (Kruse, Stadler, Pavlekovic, & Gheorghiu, 1992) because the same display can give rise to different illusions in the same person under
Figure 9. Display shown to test circular apparent motion (cam). The left- and the right-hand panels are shown alternately for 500 ms each, producing the motion perception.
identical circumstances. The perceived durations of respective Gestalts were suggested as a measure of cognitive stability (Kruse & Stadler, 1995). In the present sample, multiple regression showed that durations of Gestalt perceptions were signicantly linked to the panss factors. Especially the negative symptom factor was associated with prolonged durations. Yet there was no difference between the patient and control groups. Stroboscopic Apparent Motion (sam) Corresponding to the previous task, the frames displayed in Figure 10 were presented alternately for 500 ms each. Two qualitatively different Gestalt illusions are generated, vertical apparent motion (vsam) or horizontal apparent motion (hsam) of the dots. In the case of vsam, dot A is perceived as moving down to dot D and B as moving up to C. In the case of hsam, the perception is of A and B moving to C and D, respectively. The Gestalt impression depends on the horizontal distance d. The smaller is d, the more likely hsam is seen. With d gradually reduced from a maximum value to 0 and then increased again, ip events are enforced (initially, vsam transiting to hsam, then the reverse, hsam giving way to vsam). We monitored the timing of these ip events by having subjects press a button when a ip occurred. The asymmetry of the timing of the initial and the reversed ip events (i.e., the hysteresis effect) is a hallmark of phase transitions between stable attractors, as was seen above. Haken (1996) discussed the size of the hysteresis effect as a prominent measure of Gestalt stability.
Figure 10. Display shown to test stroboscopic apparent motion (sam). The left- and the right-hand panels are presented alternately for 500 ms each.
Again, no group effect was observed in our data. In other words, Gestalt stability measured by the hysteresis effect did not generally deviate in patients. The timing of the initial Gestalt ip (vsam to hsam) was, however, strongly linked with symptomatology. The ve panss factors explained 48% of the variance in this variable in a whole model test. The negative and the cognitive (disorganization) factors both delayed the initial Gestalt ips, whereas positive symptoms accelerated the occurrence of ip events. Motion-Induced Blindness (mib) The following phenomenon was described recently by Bonneh, Cooperman, and Sagi (2001): Subjects are instructed to observe the middle of a display such as that depicted in Figure 11. The display contains three stationary dots colored bright yellow and a grid of dimly illuminated blue crosses. The grid is rotated slowly. The phenomenon that most observers perceive is apparent blindnessone or all of the yellow dots subjectively disappear from the visual eld for a period of time, sometimes for several seconds. The subjects are asked to press a key whenever, and for the duration that, apparent blindness occurs. It was our premise that this task is, in fact, a Gestalt phenomenon in the sense of gureground rivalry. The blue grid may acquire the function of gure after some time has elapsed, with the result that the yellows dots are put in the background. This exchange of gure and ground may entail the subjective perception of subjects becoming blind to the yellow dots. The mib phenomenon is similar to inattentional blindness as investigated by Simons and Chabris (1999). The variable of interest in this paradigm was the number of mib
phenomena perceived by each participant during a 3-min period. There was a tendency in the patients group toward fewer mib counts (29.3, vs. 42.1 in the control group), yet this difference did not reach the signicance level. In multiple regression analysis of the patients group, however, we found a strong association of number of mib phenomena with symptomatology. The ve panss factors explained 49% of the variance in this variable. The positive and the excitement factors were both related to increased numbers of mib phenomena (Tschacher, Schuler, & Junghan, 2006). Intersensory Binding The nal task of the Gestalt test battery addressed temporal binding of different sensory modalities (here, auditory and visual) into a coherent perceived pattern. We used a program developed by Scheier, Lewkowicz, and Shimojo (2003). Two disks are presented moving from either sides of the screen toward the opposing sides, following the same trajectory (Figure 12). When the disks approach one another in the middle of the screen, they are perceived either as bouncing off each other or as streaming through one another at the moment they overlap visually. Similar
Figure 12. Screen display (schematic) for the intersensory binding task. Two disks wander from the sides of the screen toward one another; at the time of their collision and visual overlap, a click sound is given. Which causal relation is attributed to this scenario, bouncing or streaming?
displays have been investigated in studies of apparent causality by Michotte (1954). In addition to the optical stimuli, an acoustic-click stimulus is presented with varying temporal latency relative to the time of disk overlap. The observer is instructed to indicate which of the two possible events (bouncing or streaming) was perceived. We report here the ndings regarding the probability of bouncing versus streaming perceptions (Tschacher & Kupper, 2006). There was a tendency toward patients having fewer bouncing perceptions. In the patient group, a strong relation was found between symptoms and preferred causal attributions. The positive factor and the cognitive (disorganization) factor of the panss were strongly linked with increased probability of bouncing perceptions. As a whole, the Gestalt test battery showed few general group differences between patients and matched controls. This may be attributed in part to the low symptom burden in the patient group. Within the patient group, however, the panss ratings suggested that specic associations exist between psychopathology and Gestalt perception. Psychopathology has opposing effects on perception, effects that can explain the weak signicance of group comparisons. More specically, we found that positive symptoms were linked with Gestalt instability and faster production of Gestalt perceptions, negative symptoms and the cognitive disorganization factor with persistence of Gestalts. Hysteresis, however, was unaffected by psychopathology. Additionally, apparent causality depended on symptoms. Attri-
115 Approach to Psychopathology bution of causality (the bounce perception) was a marker of positive and disorganization symptoms.
Discussion
A common nding in research on mental disorders is that the level of psychopathology usually varies in each individual case in the course of time. Such uctuations have frequently been perceived as a problem for the validity and reliability of quantitative research, even more so as some psychopathological phenomena are inherently unstable or show unpredictable and erratic dynamics. With the theoretical and empirical approach presented here, such dynamics poses not so much a scientic problem as an opportunity for a novel approach to modeling. In this, a long-standing call for longitudinal studies is met. The dynamics-oriented approach views change as primary, especially in complex open systems. Structure and stability are, consequently, seen as the secondary results of dynamic processes that lead to the emergence of attractors. In the second section of this chapter, we have demonstrated the validity and feasibility of this approach. We have quantied and described the trajectories and processes ubiquitously found in various elds of practice; the resulting attractors and patterns have been detailed; some of the specic hypotheses of the dynamic approach have been tested in the data sets. The main results were these: Studies show a variety of dynamic patterns in courses of psychopathology concerning psychiatric rehabilitation and acute treatment of psychoses. For a wide range of settings and problems we applied methods by which the trajectories could be translated into condensed models, especially var models, that are capable of capturing the essence of the change processes. We proposed that these models reect the attractors underlying the observed dynamics. We concluded that these models were psychologically meaningful because the attractors were related to outcome and/or patient characteristics. The nal part of the second section addressed some of the core assumptions of dynamic systems theory applied to elds of psychology, particularly order formation, nonlinearity, and the stability of Gestalt-like attractors. On the basis of the empirical evidence we may claim that the search for ngerprints of nonlinear dynamics has been successful.
116 modeling complex systems The data again suggested that the phenomena of nonlinearity, such as order and stability of patterns, are associated with markers of psychopathology and clinical outcome. From these applications of the dynamic approach can be drawn a series of implications for psychiatry and psychology, especially as regards the elds of therapy, diagnosis, and the theory of cognitive science.
therapy
We expect that time-series methods can elucidate mechanisms of action in practice elds of psychiatry and psychology, such as rehabilitation and psychotherapy. The kind of longitudinal assessment that we proposed stands in contrast to the established cross-sectional approach; this has signicant clinical implications. The most important benet lies with the possible identication of change mechanisms in therapeutic settings. As time-series methods enable the modeling of single systems, the effective mechanisms can be assessed directly in the respective elds of practice and in natural clinical situations with uncompromised ecological validity. The aggregation of many single systems, as demonstrated repeatedly in the previous section, allows the distinguishing of favorable from unfavorable patterns. New methods of online control and the adaptation of interventions in the context of quality control may, thus, come within reach as soon as favorable patterns are reliably identied. It is, nevertheless, necessary to be aware of the limitations of dynamic methods, limitations that are brought about by the demands of time-series methodology. These demands are the large number of sequential measurements necessary for modeling; during the necessary observation period the underlying causal mechanisms must remain unaltered (the premise of stationarity). Another caveat must also be kept in mind: var models give an account of sequential interrelations among the process variables that may likely, but need not necessarily, reect causal mechanisms. Time-series models are so-called Granger-causal modelsas in all correlational models, the relation could be affected by unknown third variables that have not been monitored. Therefore, dynamic approaches should go hand in hand with classic experimental therapy research wherever possible. When these limitations
117 Approach to Psychopathology are taken into account, the dynamic approach has proved fruitful in process-outcome research, with a vast potential for further studies of therapeutic applications.
diagnosis
Categorical diagnosis and the dynamic view of psychopathology follow conicting concepts of disorder. For example, the concept of dynamic diseases (cf. Blair, Glass, an der Heiden, & Milton, 1995; Emrich & Hohenschutz, 1992) deviates from the common structural concept in that (some) diseases are viewed as emergent states and attractors. This is in line with what was found above. Yet the dynamics-oriented approach holds further implications for psychiatric diagnoses, implications that have as yet received little attention. Current dsm-style diagnostics is based on set theory; for example, to diagnose borderline personality disorder, ve of nine criteria must be fullled. The denition of concepts in terms of a list of criteria, however, is a computationalist notion that cognitive science, especially dynamic cognitive science, has largely abandoned. Categorization along the lines of prototype theory (Rosch & Lloyd, 1978) has clear advantages and, in addition, would be more compatible with the approach favored in this chapter. We propose that psychiatric diagnosis must be reformed at its very basis by considering what is known about psychological categorization.
theory
Our nal remarks, however, go beyond psychopathology: How can we address cognition in general on the basis of the dynamic view? Arguments for a novel view of cognition, one that treats cognition as an emergent property of the mind, have accumulated in the recent past (e.g., Tschacher & Dauwalder, 2003). As we said in the introductory section, this understanding of cognition deviates from the computationalist approach of cognitive psychology, where cognition is an abstract semantic structure running on neuronal hardware, much as a computer program would run on silicon chips. Dynamics-oriented theory is more encompassing in several respects. First, this theory is
118 modeling complex systems in accordance with transdisciplinary systems theories in the natural sciences that study and predict self-organization phenomena. This unitary approach has advantages with respect to the dissemination of both methods and phenomenological ndings across disciplines. Second, in the dynamic framework, information processing and cognition is never separable from emotion and motivation. The exploration of cognition by abstracting from its very driving forces is increasingly viewed as unfeasible in psychology and neuroscience. This integrated view of cognition in context (embodied cognition) is a principal attribute of our model. Third, dynamic systems theory deals with complexity. We have demonstrated that, in cognitive science, two separate sources of complexity must be addressed. On the one hand, cognitive systems themselves are always complex, as only complex systems can, via a self-organizational process, produce adaptive pattern formation. On the other hand, cognitive systems always deal with a complex world, a world that enforces and affords complexity reduction. Continuous simplication is, therefore, of the utmost survival value. The main demand on cognition is just thissimplifying the world by the extraction of signicant patterns on the basis of motivation.
References
Andreasen, N. C., Paradiso, S., & OLeary, D. S. (1998). Cognitive dysmetria as an integrative theory of schizophrenia: A dysfunction in corticalsubcortical-cerebellar circuitry? Schizophrenia Bulletin, 24, 203218. Banerjee, S., Sibbald, P. R., & Maze, J. (1990). Quantifying the dynamics of order and organization in biological systems. Journal of Theoretical Biology, 143, 91111. Blair, J., Glass, L., an der Heiden, U., & Milton, J. (Eds.). (1995). Dynamical disease. Woodbury ny: American Institute of Physics Press. Bleuler, E. (1911). Dementia Praecox oder die Gruppe der Schizophrenien. Leipzig: Deuticke. Bonneh, Y. S., Cooperman, A., & Sagi, D. (2001). Motion-induced blindness in normal observers. Nature, 411, 798801. Carter, R. (2002). Consciousness. London: Weidenfeld & Nicolson. Ciompi, L. (1997). Die emotionalen Grundlagen des Denkens: Entwurf einer fraktalen Affektlogik. Gttingen: Vandenhoeck & Ruprecht. Clark, A. (1997). Being there: Putting brain, body, and world together again. Cambridge ma: mit Press. Conrad, K. (1958). Die beginnende Schizophrenie. Stuttgart: Thieme.
Developing ComputerBased Learning Environments Based on Complex Performance Models Susanne P. Lajoie
McGill University
A universal theme in my research, regardless of the population or the subject material studied, is the importance of modeling performance. The word model can be dened in many ways. My use of the term is close to that of the online Merriam-Webster dictionary, which refers to a model as being something set or held before one for guidance or imitation. Interestingly enough, from a psychological perspective, modeling has been dened in a similar manner. Bandura (1977) dened it in terms of how learning can be facilitated by observing others perform the task in question. The benet of modeling would be that learners do not always have to learn from their own mistakes but can sometimes learn by observing others do a task efciently and then incorporating the ideas thereby obtained. The relevance to learning of such modeling is that learners practice deliberately rather than indiscriminately (Ericsson, 2006). Even though there is some truth to the notion that practice makes perfect, we also know that practicing the piano for 10 years does not make one a concert pianist. Expertise can be developed when practice is sufciently deliberate with feedback regarding performance. There are still three key questions regarding modeling. The rst one is what to model, the second who or what does the modeling,
124 modeling complex systems and the third how to model effectively. Some answers to these questions are presented, followed by examples from my own research that will help contextualize the arguments made in this chapter.
What to Model
In learning and performance contexts it is often the cognitive processes involved in problem solving that need to be modeled. Studies of expertise and prociency have helped in the identication of such processes in specic contexts and have also led to the discovery of commonalities in such components across expertise theories (Alexander, 2003; Chi, Glaser, & Farr, 1988; Ericsson, 2002; Glaser, Lesgold, & Lajoie, 1987; Lajoie, 2003). The groundwork for examining expert/ novice differences was summarized in Chi et al. (1988). Many of the principles of expertise discovered then, and the methodologies for studying expert/novice differences, are used today. The assumption is that identifying dimensions of expertise can lead to improvements in instruction that will, ultimately, result in helping learners become more competent. Instructional improvements are based on modeling these dimensions of expertise. For instance, self-monitoring and metacognition are key elements of expertise; however, we need to identify what experts monitor in a specic context before we can model these processes for novices. The domain knowledge, the structure of the knowledge, and the strategies that lead to effective problem solutions can all be modeled for the learner, but it is not simply observing things that leads to learning but, rather, using such knowledge and interacting with it with specic feedback. Studies of expertise have been criticized for being too coldly cognitive (Alexander, 2003). Although the mind, body, spirit connectionwhere affect and emotions are considered to go hand in hand with the mind (James, 1890)has long been recognized, we are still struggling as a community to see how these elements interact. Lepper (1988) reintroduced the connection when he called for the intersection of cognition and motivation theories, comparing cognitive psychology to the Tin Man in The Wizard of Oz (Baum, 1969), who wanted a heart, and motivation theory to the Scarecrow, who wanted a brain. It seems that the cognitive scientists are often considered heartless in the sense that they focus on the mind and how people
125 Developing Learning Environments learn. However, learning theories are increasingly more inclusive in that cognition, motivation, and the social context in which learning takes place are considered as interconnected (Anderson, Greeno, Reder, & Simon, 2000; Cordova & Lepper, 1996; Mayer, 1997). Mahoney (in this volume) describes lifelong learning from the perspective of embodied cognition, where higher learning and cognition are linked to emotion. Examining learning processes as they occur within situations or meaningful contexts goes hand in hand with personal interests when considering what is meaningful. Learning situations in the real world are often social in nature, and team members are part of the situation. Salas, Stagl, Burke, and Goodwin (in this volume) speak to this issue with reference to developing expert teams in natural settings. Their chapter addresses the notion of distributed expertise, which is important in real-world settings. The intersections of learning, motivation, and social situations are becoming elements of research design (Cobb, Confrey, diSessa, Lehrer, & Schauble, 2003). Alexander has edited a special issue on the topic of expertise in the journal Educational Researcher (Vol. 32, No. 8) that points to how researchers study expertise today, pointing toward a shift to a more inclusive theoretical model that parallels the discussion in the editors introduction to this volume of the conuence of social, cognitive, and motivational theories in interesting combinations that lead us to a better understanding of human performance (see esp. Alexander, 2003). New models must consider the interleaving of such principles in specic contexts.
126 modeling complex systems principles about learning and knowing that can assist learners along a learning trajectory within a specic eld. They report that effective learning environments are learner centered and, hence, that instruction must provide scaffolding for solving meaningful problems and supporting learning and understanding. Again, we cannot scaffold learning unless we understand the learning process; only then can we model these processes for novice performers. The same principles would apply in complex learning situations in the real world (see Lajoie, 2003). The implications for teaching and the design of instruction and assessment (Pellegrino, 2002) are clearly based on identifying what students know and what they need to know and having them reect on the processes. Studies of expertise have provided the foundation for developing models of what students need to know with respect to complex performance across domains. Such studies supported the design of methodologies used today to support our understanding of verbal data (see Chi, 1997; Ericsson & Simon, 1993), which are central to documenting problem-solving activity. Cognitive task analysis (cta) is still an appropriate means of discovering what experts know with regard to their problem-solving plans, actions, and mental models (Lesgold, Lajoie, Logan, & Eggan, 1990). In fact, cta is an excellent mechanism for documenting an effective problem space that reects a more inclusive analysis of how individuals who differ in their performance skills go about solving problems. Two examples from complex decision-making activities in the real world, avionics troubleshooting and clinical diagnostic reasoning, reveal that experts can differ in their solution strategies and steps but be guided by common underlying mental models to the correct solution. Analyses of avionics troubleshooters revealed no ideal solution paths and demonstrated that experts solve similar problems differently (Lesgold et al., 1990). However, similarities were found in the underlying mental models that guided their problem solving. Similarly, a cta of the clinical decision making of surgical nurses revealed that expert nurses reached similar decisions, albeit by different routes, demonstrating that there are many paths to solving problems (Lajoie, Azevedo, & Fleiszer, 1998). Differences were observed in hypothesis generation, planning of medical interventions, actions performed, results of evidence gathering, interpretation of the results, heuristics, and overall solution paths.
127 Developing Learning Environments Documenting such problem spaces gives clear guidance for how we might tutor individuals in their problem-solving decisionmaking strategies. Assessment must consider learning at a higher level of abstraction and a wider effective problem space in which assessment of learners takes place (Lajoie, 2003). Developing computer-based learning environments (cbles) and intelligent tutoring systems on the basis of cta can validate performance models by demonstrating that learning does occur in the predicted manner and that providing feedback based on such performance models can lead to greater prociency in trainees. Such an approach was employed in the avionics troubleshooting example mentioned above. The cta led to the development of a cble, Sherlock, that provided a practice environment allowing rst-term airmen to become procient troubleshooters (Lesgold, Lajoie, Bunzo, & Eggan, 1992). The tutor teaches troubleshooting procedures for dealing with problems associated with an f-15 manual avionics test station. A study was conducted evaluating Sherlocks effectiveness using 32 trainees from two separate air force bases (Nichols, Pokorny, Jones, Gott, & Alley, 1992). Pre and posttutor assessment was done using verbal troubleshooting techniques as well as a paper-andpencil test. Two groups of subjects per air force base were tested: (1) subjects receiving 20 hours of instruction on Sherlock and (2) a control group receiving on-the-job training over the same period of time. Statistical analyses indicated that there were no differences between the treatment and the control groups on the pretest (means = 56.9 and 53.4, respectively). However, on the verbal posttest as well as on the paper-and-pencil test, the treatment group (mean = 79.0) performed signicantly better than the control group (mean = 58.9) and equivalently to experienced technicians having several years of on-the-job experience (mean = 82.2). That is, the average gain score for the group using Sherlock was equivalent to almost 4 years of experience. All this is to say that identifying robust performance models in complex domains can lead to effective instruction for those less procient in problem-solving performance (Shute, Lajoie, & Gluck, 2000). Hence, knowing what to model can lead to better performance. Similarly, sicun, a computer tutor for nurses working in a surgical intensive-care unit (sicu), was designed on the basis of the cta described by Lajoie et al. (1998). The analysis led to the design of
128 modeling complex systems a system that encourages self-monitoring of decision-making processes. Nurses who worked with sicun were required to post their goals before conducting patient assessments and then specify outcomes of their assessments. For example, if their goal was to check the patients circulatory system for adequate blood supply to the heart, they might check pulse rate as well as skin (for swelling, edema, coloration, capillary rell, and temperature). Prior to their moving on to a new goal or body system, the tutor would prompt them about the results of their assessment. Hence, plans, goals, actions, and outcomes were all built into the system, with decision trees designed to encourage self-monitoring and comparison to expert problem solvers at various phases of problem solving. A full-scale evaluation of this tutor was not conducted, but the subjects tested became more planful in their patient assessments. The discussion presented above has described how decisions can be made about what to model to learners. The issue of deciding who or what does the modeling is another matter entirely, one that can lead to choices about how to successfully model.
129 Developing Learning Environments in the context of contained, well-dened laboratory tasks that had solutions. Well-dened tasks are easier to model than some of the ill-dened real-world tasks being modeled today. In moving toward more complex problem-solving tasks, even the most seasoned ai experts have moved away from the modeling camp, Elliott Soloway (1990), for example, suggesting that extracting expert models in more ill-dened problems takes too much time and involves less certainty that all the possible solutions patterns can be captured with computer models. Instead of living with incomplete computer models, Soloway moved away from computer models to human models involving teachers and peers serving as guides on the side. It is somewhat paradoxical that we do not question whether human coaches, teachers, and tutors have all the possible models in their heads, but, when we build computer tutors, we are not content without robust testable models. Musen (in this volume) describes how intelligent databases and expert systems can serve as vehicles for expertise but that domain ontologies that reect problem solving in a domain are becoming increasingly more important in assisting learners. Rather than debating whether human or computer coaches are better at modeling, it might be better to consider the two in tandem than to replace one with the other. Computers can serve as cognitive tools that can amplify, extend, and enhance learning (Jonassen, 1996; Jonassen & Reeves, 1996; Pea, 1985; Salomon, Perkins, & Globerson, 1991). They can do so by providing a mechanism for learners to generate and test hypotheses in complex problem-solving situations (de Jong & van Joolingen, 1998). Furthermore, such tools can be designed to provide information through multiple modalities, thus ensuring opportunities for multiple representations of knowledge (Kozma, Russell, Jones, Marx, & Davis, 1996; Mayer & Moreno, 2002). In addition, given the identication and use of complex student models, cbles can be designed to provide scaffolding that is in the form of diagnostic feedback embedded in the learning situation to be offered when assistance is needed (Lajoie, Faremo, & Wiseman, 2001; Lajoie & Lesgold, 1992; Mislevy, Steinberg, Breyer, Almond, & Johnson, 2002). If such sophisticated models are not available, computers can still be used as cognitive tools and used in conjunction with human coaches and peers to scaffold the process (Palincsar & Herrenkohl, 2002).
How to Model?
Just as theories of expertise have guided my decisions about what to model, other frameworks have led me in deciding modeling performance itself. The cognitive-apprenticeship (Collins, Brown, & Newman, 1989) and communities-of-learning (Brown, 1994; Brown, Ellery, & Campione, 1998; Palinscar & Herrenkohl, 2002) frameworks lead the way in terms of linking cognitive science and pedagogy. The cognitive-apprenticeship model (Brown, Collins, & Duguid, 1989; Collins, Brown, & Newman, 1989; Gott, 1989; Lave & Wenger, 1991; Williams, 1992) provides a template for connecting abstract and real-world knowledge by creating forms of pedagogy that are based on a model of learners, the task, and the situation in which learning occurs. When considering a traditional apprenticeship setting, such as apprenticing to become a chef, we think of novices learning from a head chef by observing and modeling physical skills. For example, an apprentice chef observes and learns from the head chef and usually starts with easier tasks, cleaning and preparing vegetables, rather than planning menus and meals for heads of state. The novice, in this example, is supported by the expert chef, who provides coaching in, for example, the correct cutting procedures, and whose active intervention gradually fades into the background when the novice demonstrates that he or she can perform the task independently. A cognitive apprenticeship is more difcult to construct since cognitive rather than physical processes are being modeled for the learner. The literature on expertise reveals that experts are not always able to articulate their problem-solving strategies or their thinking processes since those things are highly routinized or automated (Anderson, 1983). Hence, as described above, methods for externalizing these decision-making strategies must be used in order to build a model of complex performance prior to modeling for novice learners. There are four aspects of the pedagogical model, namely, content, methods, sequence, and sociology. I have described methods for documenting the content knowledge earlier in this chapter. The pedagogical methods employed by this framework involve modeling expertise and coaching learners with feedback in a manner that supports differences in knowledge acquisition. Coaches scaffold knowledge by providing the right amount of support to learners while they are performing a task. Not everyone needs the
131 Developing Learning Environments same amount of help or scaffolding. Once students demonstrate that they can do something on their own, the scaffolding is gradually removed since the goal is for students to become independent learners and problem solvers. The methods include opportunities for articulation, reection, and exploration. Learners articulate what they know either through performances or through communication of some sort. Reection can consist of self-assessments, and it can include comparisons of student processes with more procient models of performance. Exploration and active problem solving in meaningful contexts are key features of the model. The sequence of instruction takes the complexity of the content material into account, ensuring that students are challenged with meaningful tasks but not overwhelmed by information that they are not ready for. Hence, students participate in the global or overall task by performing smaller components of the overall task prior to engaging in the specic local skills from the start. In this way they are still participating in producing a nal product and developing a conceptual model of how the overall task is performed rather than working in isolation. The sociology of such real-world tasks usually occurs in small groups or teams of people working together. Once again, head chefs have a team of sous chefs who help in the overall preparation of a banquet. Similarly, with tasks requiring cognitive skills, the work can be distributed among a team of individuals working toward a common goal. Learning in this way must take the sociology of the group into consideration. Learning theories and pedagogical models alike are considering issues of shared cognition more carefully. We see similar considerations with training situations in the real world (Cannon-Bowers & Salas, 2001). In both in- and out-of-school learning situations, communities of learning (col; Brown, 1994; Brown et al., 1998) and communities of practice (cop; Barab & Duffy, 2000; Wenger, 1999), respectively, are being described, created, and empirically studied. The col model engages students in meaningful research activities, each student in the community playing a role as the community progresses toward the goal of learning. The cop model is employed in real-world situations, allowing students of such professions as medicine and law to work together to solve cases. The social sharing of information in this approach is essential to the nal outcome of the task. Similarities exist between these models and the cognitive-ap-
132 modeling complex systems prenticeship framework discussed above. Both approaches include in their descriptions reference to the sociology of the environment or the notion of community. Both suggest ways of making schooled learning more meaningful as a way of easing the transition between school and practice. Both use models of scaffolding that are based on the assumption that multiple zones of proximal development (Vygotsky, 1978) exist within the group and, hence, that feedback must adapt to those developmental zones. Both approaches use the models of expertise within their design of effective instruction. The col approach is unique in the types of participant structures that it provides as examples of modeling, turn taking, and role-playing to serve as new ways of sharing expertise and constructing learning within a community. So far I have described ways of discovering what to model, who or what can do the modeling, and how to model cognitive processes. The remainder of this chapter will contextualize these ideas with examples from my own research.
In determining what to model, we asked medical experts to solve the patient cases designed for BioWorld, in an attempt to get a robust expert model of problem solving for each case. By identifying what experts know, it is possible to develop assessment genres for examining transitions in learning in novice performers. Expert problem-solving traces were monitored and collected dynamically, as were verbal protocols. These data determined the unique proles of experts in terms of plans, strategies, and actions within BioWorld demonstrating different clusters of competencies that served to establish benchmarks of performance (Lajoie et al., 1995). Figure 1 presents an annotated screen shot of the BioWorld interface. The patient case is presented and evidence from this case can be collected and stored in the evidence table. Hypotheses can be selected and changed throughout the problem-solving activity, as can ones beliefs about said hypotheses. For instance, you might select cirrhosis as a hypothesis, but your condence that it is a correct hypothesis might be weak. Results have shown that learners condence levels increase as they acquire more knowledge (Lajoie, La-
vigne, et al., 2001). Interestingly enough, similar ndings are found in clinical psychology, where treatment outcomes are related to increases in clients self-efcacy (Tschacher & Kupper, in this volume). Condence and knowledge building are intertwined. BioWorld also presents an online library where factual information is provided, as is a patient chart, allowing learners to practice ordering diagnostic tests. Consultation is also provided online. As experts solve patients cases, they select relevant evidence from the patients medical histories, collect pertinent information in an online medical library, perform diagnostic tests pertinent to the hypothesis (see Figure 2), use systematic plans and actions throughout problem solving, and make nal arguments based on evidence (see Figure 3). Records of expert competency clusters are established and used to monitor novice students electronically in terms of the proportion of overlap of their actions with each predetermined cluster. Instructional feedback was based on the dynamic assessment of actions throughout problem solving. Students also learn to self-assess by comparing their reasoning with that of an expert (as shown in Figure 3). With practice, students became more systematic in their
reasoning and more strategic about evidence collection and hypothesis testing, indicating that instruction based on dynamic assessment can lead to transitions in learning (for evidence of these claims, see Lajoie, Lavigne, et al., 2001). Decisions regarding what to model led to design decisions regarding the types of cognitive tools that BioWorld provides to learners. In essence the evidence palette provides a mechanism for learners to monitor what they thought was important to the solution process. They can reect on their own thinking and also compare their solutions with an experts path at the completion of the problem. Rather than modeling a sequence of problem-solving processes, BioWorld highlights relevant information in the decision-making process. The consult button was designed to model expertise in the form of feedback. The success of BioWorld with high school students has led us to consider ways to reinvent this cble for medical students. In order to better understand developing expertise, Faremo (2004) examines how medical students and residents solve medical cases, compared
136 modeling complex systems to experts. Her analysis of the cognitive processes involved in diagnostic reasoning represents the knowledge levels and qualitative differences in argumentation and reasoning patterns in this medical cohort. Through verbal-protocol analyses she was able to capture a picture of emerging expertise as individuals reason through cases. Her coding schemes are based on the plans, evidence identication, and factual and procedural knowledge specic to a task. As an example, Figure 4 presents a representation of how a medical students problem-solving trace compared to that of a resident and expert in diagnosing a patient with celiacs. The patient case took the form of a narrative that presents a 27-year-old male with a 2-year history of intermittent diarrhea, weight loss, and anxiety. Celiacs is a chronic intestinal malabsorption disorder caused by intolerance to gluten. It presents in a variety of ways, and some direct clues (diarrhea, abdominal discomfort, and distention) may be present. Iron deciency is found in affected adults, and there can be (1) low albumin, calcium, potassium, and sodium and (2) elevated alkaline phosphotase and prothrombin time. The disorder is diagnosed by means of symptoms, lab studies, and X-rays (antigliadin antibodies are the gold standard test for detection). Figure 4 demonstrates the relation between the plans or goals experts had, the types of data they collected and reviewed, and the hypotheses they generated in the context of reasoning about this case. It also shows that the students identied some of the same information as relevant, but did not engage in the same systematic planning and evidence evaluation, as experts. This type of in-depth analysis of diagnostic reasoning across cohorts can eventually lead to robust models of problem solving that can extend the use of BioWorld to higher-education classrooms and complete the picture of emerging expertise in this domain.
Figure 4. Expert model with student overlay for the case of a celiac.
138 modeling complex systems the setting is that realistic patient cases are seen and handled. The weakness of the natural apprenticeship is that what you see is what you get. In other words, there is no standardization of curriculum or guarantee that all students will see the types of patient cases that will build on their existing knowledge. We were interested in documenting the types of diagnostic feedback that human tutors gave to students as well as describing the learning context. In doing so our attempt was to develop a cognitive-apprenticeship model for diagnostic feedback that might be built into a cble. What we discovered was that a col approach was common in these tutoring sessions, whereby small groups of individuals shared the problem and distributed their learning through shared representations, articulation, and reection, with a human tutor scaffolding as needed. Two lessons were learned from this setting. The rst lesson was that contextualized and diagnostic feedback could be studied in these settings and that elements of scaffolding and gradual removal, articulation and reection, were, in fact, all part of the tutorial process. Hence, an argument could be made that the pedagogical methods of the cognitive-apprenticeship model were used and could be documented. The second lesson was that this environment had to consider the sociology of the group and the roles of each member if it was to be an authentic tutorial.
139 Developing Learning Environments tional moments for their students. They can also run the case with students to see whether they need to revise it for better instruction and assessment purposes. Incorporating the social context into performance modeling is more difcult. Medical students learn from experienced practitioners and teachers (experts) and other students and residents (intermediates) during their training. Students attend lectures, participate in problem-based groups, observe others, and participate in discussions about hospital patients. We are currently looking at ways to build a more distributed model of learning for this population in the context of problem-based learning situations. Lu and Lajoie (2003) describe three types of cognitive tools that will be considered in this situation: visualization tools, to make knowledge representations more visible to the teams involved; knowledge-building tools, to help people add on to existing knowledge in specic ways; and management tools, to facilitate the structuring and tracking of information between small groups. On a clinical teaching ward, there are experts in many different content areas collaborating hierarchically to work on a given case. These experts also continuously interact within a hierarchy of novice to intermediate learners (medical students, interns, and residents). We are still researching ways of designing tools that will facilitate these exchanges in a somewhat controlled setting, allowing individuals to share knowledge collaboratively.
Conclusion
By asking and answering a set of questions, I have attempted to address the intersection between performance modeling and the design of cbles. These questions are: What do we chose to model? Who or what should we model? How do we model complex decision making? Theory-based arguments were provided in response to each question. The literature on expertise and competency was discussed in the context of the empirical ndings from this literature as well as the methodological tools that were offered from that literature. In essence, selecting what to model must be specic to the domain and problem tasks in question. Better-informed models of competency can provide transparent trajectories for learning in such domains.
140 modeling complex systems Decisions regarding who or what should do the modeling are based on guiding learning paradigms. For instance, social constructivists may prefer that modeling occur in groups, with more expert peers assisting the learning process. Cognitive constructivists may agree that models provided in a cble can extend learning effectively when individuals interactively engage in the problem-solving activity. Not all modeling is done through explicit coaching, whether the tutor be human or computer. The design of cbles with cognitive tools can support learning by constraining problem solving through support of the overall learning process. In answering the how-to-model question, the cognitive-apprenticeship, col, and cop approaches were described, approaches that support specic pedagogical methods in order to promote learning through effective models. Subsequent to these discussions, examples from science and medicine were presented to demonstrate how such approaches guide the design of instruction and assessment in complex learning situations. There are still unanswered questions. As modeling moves beyond the individual learner and includes a small group of learners, the complexity of assessment is increased. Although it is widely recognized that real-world learning often takes place in small groups, it is difcult to develop both individual and group models of performance when deciding appropriate levels of feedback to move the problem-solving process forward. New ways of supporting differences will need to be developed, and new ways of sharing and documenting the group process will need to parallel authentic practice.
Note
Research reported in this chapter was made possible through funding provided by the following granting agencies: the Canadian Social Sciences and Humanities Research Council; the Quebec Ministry of Industry, Commerce, Science and Technology; Valorisation Recherche Quebec; Ofce of Educational Research and Improvement. Many graduate students (former and current) have contributed to the work that is reported here. Special thanks to Arshad Ahmad, Roger Azevedo, Gloria Berdugo, Janet Blatter, Andrew Chiarella, Lucy Cumyn, Sonia Faremo, Genevieve Gauthier, Claudia Guerrera, Nancy Lavigne, Susan Lu, Carlos Nakamura, Thomas Patrick, and Jeffrey Wiseman.
References
Alexander, P. A. (2003). The development of expertise: The journey from acclimation to prociency. Educational Researcher, 32(8), 1014. Anderson, J. R. (1983). The architecture of cognition. Cambridge ma: Harvard University Press. Anderson, J. R., Greeno, J. G., Reder, L., & Simon, H. A. (2000). Perspectives on learning, thinking and activity. Educational Researcher, 29(4), 1113. Bandura, A. (1977). Social learning theory. Englewood Cliffs nj: PrenticeHall. Barab, S. A., & Duffy, T. M. (2000). From practice elds to communities of practice. In D. H. Jonassen & S. M. Land (Eds.), Theoretical foundations of learning environments (pp. 2555). Mahwah nj: Erlbaum. Baum, F. L. (1969). The wizard of Oz. Chicago: Childrens Press. (Original work published 1900) Brown, A. L. (1994). The advancement of learning. Educational Researcher, 2(8), 412. Brown, A. L., Ellery, S., & Campione, J. C. (1998). Creating zones of proximal development electronically. In J. G. Greeno & S. Goldman (Eds.), Thinking practices: A symposium in mathematics and science education (pp. 341367). Hillsdale nj: Erlbaum. Brown, J. S., Collins, A., & Duguid, P. (1989). Situated cognition and the culture of learning. Educational Researcher, 18(1), 3242. Cannon-Bowers, J. A., & Salas, E. (2001). Reections on shared cognition. Journal of Organizational Behavior, 22(2), 195202. Chi, M. T. H. (1997). Quantifying qualitative analyses of verbal data: A practical guide. Journal for the Learning Sciences, 6(3), 271316. Chi, M. T. H., Glaser, R., & Farr, M. J. (1988). Introduction to M. T. H. Chi, R. Glaser, & M. J. Farr (Eds.), The nature of expertise (pp. xvxxvi). Hillsdale nj: Erlbaum. Cobb, P., Confrey, J., diSessa, A., Lehrer, R., & Schauble, L. (2003). Design experiments in educational research. Educational Researcher, 32(1), 913. Collins, A., Brown, J. S., & Newman, S. E. (1989). Cognitive apprenticeship: Teaching the craft of reading, writing, and mathematics. In L. B. Resnick (Ed.), Knowing, learning, and instruction: Essays in honor of Robert Glaser (pp. 453494). Hillsdale nj: Erlbaum. Cordova, D. I., & Lepper, M. R. (1996). Intrinsic motivation and the process of learning: Benecial effects of contextualization, personalization, and choice. Journal of Educational Psychology, 88(4), 715730. de Jong, T., & van Joolingen, W. R. (1998). Scientic discovery learning with computer simulations of conceptual domains. Review of Educational Research, 68(2), 179201. Derry, S. J., & Lajoie, S. P. (1993). A middle camp for (un)intelligent computing. In S. P. Lajoie & S. J. Derry (Eds.), Computers as cognitive tools (pp.1 11). Hillsdale nj: Erlbaum. Ericsson, K. A. (2002). Attaining excellence through deliberate practice: In-
Technology for Building Intelligent Systems: From Psychology to Engineering Mark A. Musen
Stanford University
146 modeling complex systems Because computers have been mass-produced for only a few decades, it is not surprising that the eld of articial intelligence (ai) is relatively young. What is surprising is that, nearly as soon as computers became available for general research, investigators began to study how computers might be programmed to do intelligent things. The 1950s saw the birth of quite sophisticated programs for playing checkers, for proving mathematical theorems, and for processing natural language. By 1956 there was sufcient interest in the challenge of getting computers to demonstrate intelligent behavior that John McCarthy and several colleagues convened a two-month meeting of workers in computer science, psychology, philosophy, and economics at Dartmouth College to study a new eld that McCarthy dubbed articial intelligence. In the original proposal for the Dartmouth conference, McCarthy, Minsky, Rochester, & Shannon (1955) declared: The study is to proceed on the basis of the conjecture that every aspect of learning or any other feature of intelligence can in principle be so precisely described that a machine can be made to simulate it. An attempt will be made to nd how to make machines use language, form abstractions and concepts, solve kinds of problems now reserved for humans, and improve themselves. This grand vision for ai has been achieved, to date, only in science ction. Enormous progress has been made on components of this montage of capabilities, but no computer system can be viewed as generally articially intelligent. Some programs can analyze or generate natural language. Others can form abstractions from primitive data. Some can solve problems typically reserved for humans. For a host of reasons, no one has created ai in the composite sense. The computer programs that help us prepare our taxes and that offer driving directions are extremely useful, nevertheless, and represent a class of computer programs that have been enormously successful in our society. We refer to such programs as intelligent systems. This chapter traces some of the history of intelligent systems, outlining the obstacles that developers have overcome, and suggesting methods by which such systems can be built and maintained. The emphasis is on the representation of human knowledge in computers and on the primitive building blocks that make such representations possible. The story begins with a close alignment between developers of intelligent systems and scholars of cognitive psychology. It ends, however, with a perspective that construes the devel-
147 Building Intelligent Systems opment of intelligent systems more as software engineering than as applied psychology. By not attempting to replicate human cognition but instead attempting to take advantage of what is known about the construction of good software systems, developers now can build computer programs that are extremely useful, reliable, maintainable, andmany would sayintelligent. Our goals, however, must be more focused than those of the participants in the 1956 Dartmouth conference. We must recognize that the computer systems that will be most useful to society are not necessarily those that might replicate human cognition but rather those that allow us to store, communicate, and apply human knowledge in specic contexts when it is needed. The elucidation of the methods that are needed to enable system builders to create such software artifacts can be just as exciting as the elucidation of human cognition itself. I nd it particularly helpful to look at the evolution of thought regarding knowledge-based systems in a historical context. The broad picture of how developers of intelligent computer programs have viewed the role of psychology in their craft over the years tells a very interesting story. This is a book chapter, however, not a book, and, therefore, I have to be selective in what I say. The reader should be aware that the need to streamline the presentation requires that I simplify (and sometimes oversimplify) parts of the story.
148 modeling complex systems of difcult tasks in application areas that ranged from medicine to geology to computer-hardware conguration. The notion of rulebased programming became so popular that the terms rule-based system and expert system began to be used nearly interchangeably; if a computer behaved intelligently, there was an assumption that the program was using rules. (In this chapter, I avoid the term expert system, which is now becoming dated. I use the terms intelligent system and knowledge-based system as synonyms that refer to any computer system that encodes domain-specic knowledge for the purpose of driving problem solving, regardless of whether it uses rules.) Rule-based systems encode domain knowledge in an intuitive if-then formalism. In the mycin system, for example, there were hundreds of rules such as the following: Rule 124 if: 1) The site of the culture is throat, and 2) The identity of the organism is Streptococcus then: There is strongly suggestive evidence that the subtype of the organism is not Group D Although the rules were actually written in a dialect of the lisp programming language, the computer could translate them into understandable English. The mycin system used the rules to reason about descriptions of patients entered into the program (Figure 1). A user of mycin presumed that the patient to be described had either bacteremia (bacteria in the blood) or meningitis. At the time of the users interaction with mycin, the patient ostensibly was ill, but it was not clear what the source of the infection was. The program would analyze the case description, asking questions of the user to gather all the relevant data, and then determine (1) the organism or organisms most likely causing the patients infection and (2) the antibiotics that should be administered for optimal treatment (Buchanan & Shortliffe, 1984). A program known as an inference engine interpreted the rules in mycin and allowed one rule to invoke other rules in order to reach the programs nal conclusions. For example, when evaluating Rule 124 (given above), the mycin inference engine would examine the rst clause of the rule and ask: Is the site of the culture the throat?
Figure 1. A trace of a users dialog with the mycin system. The program posed questions to the user, allowing the entry of information regarding a patient to be treated for presumptive bacteremia or meningitis. Each question seeks information necessary to determine the truth value of one of the programs many production rules.
If the site from which the culture under consideration was taken was not the throat, then the rule would fail, and further evaluation of Rule 124 would not take place. If, however, the site of the culture was, indeed, the throat, then the mycin inference engine would consider the second clause of the rule and ask: Is the identity of the organism under consideration Streptococcus? If the organism was not streptococcus, then the rule would fail. If the organism was, indeed, streptococcus, however, then the rule would succeed, and mycin would conclude that the subtype of the streptococcus is not Group D. What if the identity of the organism under consideration were not known one way or the other? The mycin inference engine would examine the rule base to see whether there were any rules that had not yet red that would allow the system to conclude whether the identity of the organism was streptococcus. If there were such rules, mycin would suspend its evaluation of Rule 124 and go on to evaluate those other rules; once it was known whether the identify of the organism was streptococcus, the mycin inference engine would return to Rule 124 to determine whether the second clause was true
150 modeling complex systems or false. If the second clause evaluated true, then the system would conclude that the subtype is not Group D. mycins inference engine thus used the hundreds of rules in the knowledge base to evaluate the case at hand and to generate recommendations for the user. Starting with an overarching goal rule (basically: if you know what are the likely organisms causing the infection then recommend treatment), the inference engine would invoke rule after rule recursively to conclude everything it could about the case at hand so that the program could then determine the appropriate course of action. This style of computer programming is known as a production system. Each production rule in the system has a left-hand side (a condition) and a right-hand side (a conclusion). An inference engine such as the one in mycin starts with the goal rule and determines whether the condition predicating the rule is true. If so, then the rule produces its conclusion; if not, then the inference engine considers the truth value of other rules that can produce as their conclusion the condition of the rule currently under consideration. Examination of these rules may cause yet other rules to be invoked. The procedure stops when the truth value of the goal rule becomes known. The wonderful thing about the mycin inference engine was that it was a simple procedure. The inference engine contained no knowledge of infectious diseases or their treatment. It was designed only to invoke rules to follow chains of inference; all the domain-specic knowledge was encoded strictly in the rules. Thus, it was a relatively straightforward matter to apply mycin-style inference engines to an entirely different rule baseallowing the system to evaluate cases and make recommendations in new application areas far aeld from medicine. The idea of rule-based systems caught on like wildre. It was extremely appealing to view knowledge bases as collections of rules. The rules were viewed as modular chunks of knowledge that developers could drop into an expanding knowledge base as new capabilities were dened. Rules were viewed as independent of one another and not subject to interaction effects as a knowledge base grew. Most important, workers in ai at the time believed that mycin-style rules had verisimilitude with components of actual human problem solving.
Figure 2. A cryptarithmetic problem. Newell and Simon (1972) asked subjects to replace each letter consistently with a digit in order to create a mathematically correct formula.
152 modeling complex systems and internal stimuli. They saw in their data a common problem-solving procedure that looked as if subjects were performing rule-based reasoning. They intentionally were guarded when they claimed to confess to a strong premonition that the actual organization of human programs closely resembles the production system organization (p. 803), but there was an obvious attractiveness to this conjecture. Human problem solving suddenly could be understood in terms of production systems (Young, 2001)as the simple application of rule-based reasoning, using mechanisms that had been well studied by computer scientists and philosophers for decades. For the ai community, this work had a rather unguarded implication: human knowledge must surely be encoded as production rules, and the builders of intelligent systems can construct their electronic knowledge bases by nding out what those rules are. By the end of the 1970s, it was commonplace for developers of intelligent systems to view the problem of building electronic knowledge bases as one of eliciting rules from experts. More important, there was widespread belief that the rules in the knowledge base had correlates in representations in the head of the expert who served as the informant. Building an intelligent system was seen as a problem in the transfer of expertise, where the goal was to get the rules out of the experts brain and into the computer. Pioneers in the eld, such as Feigenbaum (1984), frequently used a metaphor in which system builders were said to mine the knowledge of human experts, capturing golden nuggets of expertise that could be added to a growing knowledge base, one rule at a time. The prevailing view was that the rules in the computer system were precisely the same rules that the experts used to solve real-world problems. By denition, every knowledge-based system became a cognitive model. Granted, there were some vocal naysayers at the time. Dreyfus (1981), for example, pointed to the qualitative transformation of problem-solving abilities that seems to take place as novices become experts and argued that the tacit nature of much professional knowledge makes it impossible for computers ever to reason like human experts. The ai community paid some attention to these concerns, but the apparent success of mycin and a host of other rule-based systems seemed too compelling to ignore. There was face validity to the claim that human expertise could be packaged within a com-
153 Building Intelligent Systems puter, and Newell and Simons suggestion that simple production systems could form the basis of human cognition seemed to be playing outat least in silico.
154 modeling complex systems such as Intellicorp, Teknowledge, and Symbolics were busily busing hundreds of conventioneersincluding bright-eyed graduate students who had no money at all to spendto exclusive locations in Beverly Hills, offering abundant free food and a chance to see the companies products in action. Scruffy-looking researchers who had agreed to give tutorials on basic aspects of rule-based systems were stunned, not only by the numbers of attendees cramming into their lecture halls, but also by the unprecedented per capita honoraria that they received, allowing at least one to brag about buying a new sports car as a result! The word on the street was that human intelligence could be mined and put into machines. By 1988, ai had become a $1 billion industry in the United States. Within a few years, however, the expert-systems industry would be largely gone. The expectation that knowledge-engineering groups would be springing up in every cranny of the corporate world was wildly optimistic. It proved much more difcult to encode and to maintain electronic knowledge bases than most ai practitioners had ever imagined. As we shall see in the next section, building intelligent systems is an enterprise that is extraordinarily more complicated than simply mining nuggets of knowledge. Rule-based systems, nevertheless, still remain essential commodities and are woven into the fabric of modern society. Such systems continue to review our applications for commercial credit, to determine the appropriateness of requested medical procedures, to audit our income tax returns, to interpret our electrocardiograms and other medical tests, to assist in industrial manufacturing, even to control the environment of the buildings in which we work. We have learned, however, that the knowledge needed to drive such systems does not exist as sets of production rules waiting to be plucked from the heads of human experts and dumped into computer programs.
155 Building Intelligent Systems generally in the form of production rulesfrom human experts to computer programs and transferring knowledge from computer programs to naive users. Knowledge was a commodity, a substance, and the goal was to get it from one place to another. The view of knowledge-based development as the transfer of expertise hindered the now-obvious recognition that building electronic knowledge bases is a creative and inventive activity. When developers interact with application specialists to build knowledgebased systems, the developers form mental models of how the experts solve problems; the experts, of course, have mental models of their own that attempt to capture their professional problem-solving behavior. In the course of building the system, both the developers and the experts continually revise their respective mental models. Although the developers and the experts may have very different mental models at the outset of their collaboration, the models eventually converge. This convergence is possible (1) because the development process forces all parties to commit their mental models to a xed, publicly examinable formtypically the emerging knowledge baseand (2) because the frequent consideration of examples and test cases forces the system builders to assess, corroborate, and revise their respective models (Regoczei & Plantinga, 1987). By the 1990s developers of knowledge-based systems began to realize that the nuggets of knowledge were not always there for the mining. Builders of intelligent systems began to accept that the application experts, whose professional acumen was to be encoded as a knowledge base, often have no preexisting mental model of how they do their work. Often, the experts cannot begin to verbalize how they actually go about solving problems. Dreyfus (1981) and others had been warning about these difculties all along, but suddenly the warnings were beginning to be take root. The problem of building knowledge-based systems was no longer seen as one of transferring expertise; much of the expertise often had to be invented in the rst place. At the core of the matter is the automatization of professional behavior. Human cognitive skills appear to be acquired in at least three generally distinct stages of learning (Fitts, 1964; Johnson, 1983). Although different authors have used different terms to describe the three phases, there is concurrence regarding the qualitative changes that occur in the way in which people seem to retrieve information during problem solving. Initially, there is the cognitive stage, during
156 modeling complex systems which an individual identies the actions that are appropriate in particular circumstances, either as a result of direct instruction or from observation of other people. In this stage, the learner often verbally rehearses information needed for execution of the skill. Next comes the associative phase of learning, in which the relations noted during the cognitive stage are practiced and verbal mediation begins to disappear. With repetition and feedback, the person begins to apply the actions accurately in a uent and efcient manner. Then, in the nal autonomous stage, the learner compiles the relations from repeated practice to the point where he or she can perform them without conscious awareness. Suddenly, the person performs the actions appropriately, prociently, and effortlesslywithout thinking. The knowledge has become tacit (Fodor, 1968). There is substantial evidence that, as humans become experienced in an application area and repeatedly apply their know-how to specic tasks, their knowledge becomes compiled and, thus, inaccessible to their consciousness. Experts lose awareness of what they know. The knowledge that experts acquired as novices may be retrievable in a declarative form, yet the skills that these professionals actually practice are procedural in nature (Anderson, 1987). The inability of experts to verbalize these compiled associations is well accepted (Lyons, 1986; Nisbett & Wilson, 1977). The consequence is that the special knowledge that we would most like to incorporate into our intelligent systems is often that knowledge about which experts are least able to talk. Johnson (1983) has identied this phenomenon as the paradox of expertise. The problem for builders of intelligent systems is that, when professionals are asked to report on their compiled expertise, they often volunteer plausible answers that may well be incorrect. In experimental situations, subjects have been shown to be frequently (1) unaware of the existence of a stimulus or cue inuencing a response, (2) unaware that a response has been affected by a stimulus, and (3) unaware that a cognitive response has even occurred. Instead, subjects give verbal reports of their cognition based on prior causal theories from their nontacit memory (Nisbett & Wilson, 1977). Furthermore, because Western culture mistakenly teaches us that accurate introspection somehow should be possible (Lyons, 1986), people freely explain and rationalize their compiled behaviors without recognizing that these explanations are frequently incorrect.
157 Building Intelligent Systems Developers of knowledge-based systems are, thus, not mining preexisting nuggets of knowledge. They are creating de novo theories of professional problem-solving behavior and representing those theories in terms of electronic knowledge bases. The developers serve the important function of detecting gaps in the articulated knowledge of their informants and helping them ll in those gaps by dening plausible sequences of actions that can achieve the necessary goals. The intelligent systems that result from this work may not achieve the same level of nuanced performance associated with the procedures actually used by domain experts. The underlying knowledge bases can, nevertheless, be observed, extended, and easily disseminated to other people in need of advice. It is simply incorrect to view an electronic knowledge base as an embodiment of actual professional knowledge; knowledge bases instead represent only models of surface-level behaviorsmodels that attempt to approximate, but that do not reproduce, the actual problem-solving steps used by humans (Clancey, 1989).
158 modeling complex systems a formal psychological theory (Kushmerick, 1996), but it had a jarring effect on the way people viewed the role of representations in cognitive science. The situated-action perspective cast aside Newell and Simons hypothesis that human problem solving is fundamentally grounded in the processing of symbols (Newell, 1980). Intelligent behavior was seen instead as the consequence of myriad specic, often-unknowable stimuli in the environment leading to myriad specic actions. There were no preexisting abstract representations in the brain to dene general plans of behavior; there were only particular behavioral responses to particular situations. Such responses might require the dynamic creation of a mental representation to guide the agents behavior, but there were no prefabricated representations for the abstract problem space (e.g., no production rules) to drive cognition. Although the emergence of the situated-action perspective caused a minor uproar in the cognitive-science community, its effects on the developers of knowledge-based systems were, ultimately, more muted (Menzies & Clancey, 1998). The primary consequence of the situated perspective was that virtually no developers of knowledge-based systems claimed that they were creating cognitive models anymore. The emphasis shifted almost completely away from psychology toward the construction of useful pieces of software. This move toward a software-engineering perspective was reinforced by other events in the ai community. Perhaps none was more important than the appearance of computer programs that could, nally, beat grand masters in games of chess. Workers at CarnegieMellon University built Deep Thought (reborn at ibm as Deep Blue) and achieved outstanding game-playing performance. They did so, not by building a computer system that mimicked the cognitive behaviors of human experts, but by capitalizing on things that only computers can do: blazingly fast computation, the execution of multiple computations in parallel on distributed processors, and the ability to use brute force to look ahead many moves to see the consequences of potential actions (Newborn, 2002). Deep Blue ultimately beat Gary Kasparov at his own game, but not because the program was cleverer than Kasparov or because it modeled the thought processes of a better chess player. Deep Blue demonstrated convincingly that knowledge-based systems would be successful when they could
159 Building Intelligent Systems take advantage of the unique abilities of the hardware and software of which they were made. The emphasis was placed squarely on better techniques for systems engineering, rather than on better cognitive modeling.
Figure 3. Stages in the development of intelligent systems. From a conceptual model of system performance, developers create a design model that makes a commitment to an abstract pattern for how the system might be created using software. The design model then drives the implementation, which results in a physical software system that users can run.
system. At its core, kads was a software-engineering methodology for building intelligent systems. For many development groups, kadsand its subsequent renement, commonkadstransformed the process of knowledge engineering from an informal enterprise of mining nuggets to a stepwise, methodical practice for building complex software systems that manifested intelligent behavior. An essential contribution of the kads project was the view that, like conventional software, intelligent systems need to be developed in stages (Figure 3). Developers should never simply sit down and try to build a large, complex system all at once; instead, they need to approach the design process in a principled manner. The kads project viewed system building as the construction of a set of models and argued that those models should be made explicit and examinable. First developers should build a conceptual model that denes what task the system is to perform in a functional manner: What are the inputs? What are the outputs? How does the nature of the inputs constrain the nature of the outputs? Can the task be decomposed into subtasks, both to simplify the computation and to make the system more cognitively tractable? In the case of mycin, for example, the conceptual model would identify patient data as the input and a set of antibiotics to prescribe as the output. The overall task of determining what presumptive therapy to administer might be decomposed
161 Building Intelligent Systems into two subtasks: (1) identify what classes of organisms most likely were causing the infection and (2) construct a small set of antibiotics that, together, were likely to be effective against the most likely organisms. The conceptual model is just that; it is conceptual. There is no discussion of how the computer might actually perform the tasks required by the nished system. Next comes the creation of a design model. The design model is just that; it is a design. It is not a working piece of computer code but, rather, an outline of how software components might be brought together to implement a working system. A design model for mycin as it was originally implemented might describe how production rules could be used to program the organism-identication task, how a constraint-satisfaction program could be used to determine the minimal set of antibiotics to prescribe, and so on. Just as developers of methodologies for software engineering were arguing that design must precede implementation, the authors of kads and commonkads insisted that builders of intelligent systems needed to articulate a coherent design before they actually began to encode the production rules and write the program components that would form the basis of the executable system (Schreiber et al., 2000). Construction of a design model is facilitated if developers have access to a set of primitives that can serve as building blocks for constructing a design. Those primitives could be the components of expert-system-building shells or abstract algorithms for performing tasks such as constraint satisfaction. I will have much more to say about possible building blocks for the design model in the next section. From the design model would come an implementation. The implementation would be the software artifact with which end users would interact and that actually would provide them with advice. Because any creative task is simplied when it can be decomposed into more tractable elements, ideally there are software components that can serve as building blocks for developing the implementation. When creating conventional software, object-oriented programming often plays an important role in helping developers decompose the implementation task into the construction of cognitively tractable software modules. The kads project initially argued for a waterfall model of software development (Sommerville, 2004), with information owing
162 modeling complex systems from the conceptual model to the design model to the implementation. A primary problem with the approach is that developers rarely get things right the rst time; they must revisit their conceptual models and design models as they gain experience with a particular implementation. Unfortunately, well-meaning developers can choose to tinker with the implementation and ignore the design model. In time, the design model may become increasingly irrelevant for explaining the programming choices made in the implementation. Similarly, the conceptual model can rapidly become out-of-date unless system maintainers are particularly conscientious about documenting all their changes. Thus, the behavior of the knowledgebased system can come to represent something different from the specication documented in the original conceptual model. System maintainers can guard against such drifts when they use building blocks that have well-understood meanings and behaviors to create their design models and implement their systems. If each of the building blocks is xed and invariant in its semantics, then developers can have condence in the consequences of replacing one building block with another. Any drift in behavior can be cleanly ascribed to a change in the components and can be documented fairly easily (Krueger, 1992). As system builders have thought about the development process shown in Figure 3 above, a number of questions surfaced: What are the optimal building blocks for creating design models and, ultimately, for implementing intelligent systems? How can a set of building blocks help developers construct their models in a way that ensures comprehensive and useful designs? How can a set of building blocks ensure that the software that implements an intelligent system is maximally reliable, maintainable, and traceable in its behavior? As the knowledge-engineering community evolved in the 1990s, the focus for research shifted radically to elucidating components both for design models and for implemented systems. The emphasis was still on nuggets, so to speak, but suddenly the community was much more concerned about software components and elements of design than it was about nuggets of expert knowledge. Psychology still was important to the builders of knowledge-based systems, but their attention was turning to the cognitive processes of the system developerswith the goal of creating modeling and implementation primitives that would help software engineers manage
163 Building Intelligent Systems the tremendous complexity inherent in the construction of intelligent systems.
164 modeling complex systems ate each class into subcategories. The study of ontology remained a relatively obscure branch of metaphysics until the last decade of the 20th century, when computer scientists co-opted the word, put an article in front of it, and began to speak of an ontology as a description of the concepts and relationships among concepts that existed in some world being modeled. Suddenly, the elucidation of ontologies was, not only the work of philosophers, but also that of software engineers trying to model the concepts and relationships needed to dene some reality for the purposes of building a system (Guarino, 1998). There are now conferences and workshops devoted exclusively to the subject of building ontologies. There are specialty journals and an explosion of books on the subject. The study of ontology has swelled because formal data structures that can be used to represent some aspect of the world and to provide an enumeration of essential categories have become more important commercially than any philosopher could ever have imagined. Yahoo! initially had enormous success because it provided easy entre to the Internet by allowing users to browse a detailed ontology of the kinds of Web pages that people had created. Soon, other commercial Web sites such as Amazon.com offered a detailed ontology of books and music that allowed Web surfers to examine an enormous inventory of online products. Ontology became a big business as the Internet revolution required computer programs to be able to represent detailed descriptions of hierarchically organized informationfrom Web sites, to commodities, to e-commerce transactions (Staab & Studer, 2004). This notion of ontology was, of course, hardly something that the Internet community could claim as its own. The very idea stemmed from centuries of work in metaphysics. In the modern era the formal classication of entitiesfrom Linnaeus and the speciation of living organisms, to Roget and the thesaurus of the English language, to Dewey and the classication of literaturehad had profound effects on human thought. Indeed, much of human cognition centers on the creation and application of categorizations and taxonomies (Bowker & Star, 1999). The rediscovery of the notion of ontology in the late 20th century, however, provided a principled way for computer scientists to begin to think about the world that their programs were attempting to model and allowed them to put their models into a well-established context (Guarino, 1998).
Figure 4. An ontology of oncology protocols. The ontology models the entities and relations needed to specify how medical care should be given over time to patients who have cancer. In the ontology, clinical-trial protocols are a kind of clinical guideline. Clinical trials have a number of attributes, including the authors who developed the protocol, the eligibility criteria that determine the set of patients to which the protocol applies, and the end points that determine when treatment should terminate.
When building knowledge-based systems, developers now concentrate on fashioning a formal model of the entities in the domain and the relationships among those entities. For example, Figure 4 shows an ontology of a medical application areastandard protocols for treating patients who have cancer that specify how chemotherapy and other treatments should be administered to patients, either as part of a clinical trial or in accordance with a clinical management guideline that attempts to dene a set of best practices. The ontology describes what may existwhat are the possible components of a protocol (e.g., eligibility criteria that determine whether a clinical trial actually applies to a given patient, algorithms that specify the sequence of treatments to offer, tests to perform that may predicate decisions in the clinical algorithm) and what are the relationships among those components (e.g., that
166 modeling complex systems steps in the clinical algorithm are related temporally). The ontology does not describe any particular protocol (i.e., a specic plan of care for a specic medical problem), instead laying out the classes of entities that exist in protocols in general. The protocol ontology can then form the basis for acquiring the content knowledge for specic protocolsproviding a structure that allows developers to build multiple knowledge bases such that each knowledge base encodes the description of a particular type of guideline or clinical trial (Musen, 1998). In our laboratory, we have developed a computer-based workbench that helps system builders use ontologies to create detailed knowledge bases. This workbench is known as Protg (Musen, Fergerson, Grosso, Noy, Crubzy, & Gennari, 2000; Gennari et al., 2003). The screen shot in Figure 4 shows how an ontology appears within the Protg tool. The tree browser on the left-hand side of the screen shows the entities in the ontology. The form on the right-hand side of the screen shows the attributes of whatever class is highlighted on the left. Ontologies are typically represented within computer systems as hierarchies of objects. At the top of the hierarchy may be an object with a very general description (e.g., thing). There will be subclasses of this object that represent more specialized concepts. The subclasses in turn will have subclasses. Each object in the hierarchy has attributessome of which are particular to that object, others of which may be inherited from objects higher up in the hierarchy. In Figure 4, for example, the concept guideline has three distinct subclasses: management guideline, consultation guideline, and trial protocol. In the gure, attributes of trial protocol include items such as authors and eligibility criteria. Some of the attributes of trial protocol are inherited from more general concepts, such as guideline. For any ontology, the Protg system can generate a user interface with which developers can enter the content knowledge structured by that ontology (Figure 5). For each generic entity in the ontology (e.g., trial protocol), Protg provides an interface with which users can enter instances of that entity (e.g., a particular clinical trial for treating breast cancer). The interface also allows developers to specify relationships among the instances (e.g., the sequence of steps that determine how chemotherapy and other inter-
Figure 5. A knowledge-entry tool for the specication of clinical trials. Protg generated the user interface for this tool directly from the ontology shown in Figure 4 above. The ontology denes the classes of entries that a user can make (in this case, the particular blanks to be lled in and the choices in the palette available for drawing diagrams); users ll in the blanks and draw diagrams to specify the content knowledge that constitutes particular knowledge bases. In this example, the Protg-generated tool allows developers to create knowledge bases that dene clinical trials. Here, the user is entering knowledge about a clinical trial for treating patients who have breast cancer. The particular protocol compares the effects of conventional chemotherapy and radiotherapy with those of high-dose chemotherapy followed by bonemarrow transplantation.
ventions should be administered over time). Tools such as Protg divide the process of creating an electronic knowledge base into two phases: (1) description of a general ontology of the application area and (2) instantiation of that ontology to dene particular knowledge bases. We and many other groups routinely use Protg (and tools like Protg) in this two-step process of knowledge-base denition.
Figure 6. Heuristic classication in mycin. Problem solving in the mycin system can be construed in terms of three major inference patterns: (1) feature abstraction, (2) heuristic match, and (3) solution renement. After Clancey (1985).
State group described recurrent problem-solving approaches that, unlike weak methods, could take advantage of domain knowledge to address tasks in more sophisticated ways than are possible with a single evaluation function. For example, Chandrasekaran described a method that he called establish-rene, which performed diagnosis by examining a hierarchy of possible diagnoses (the most general diagnoses were at the top of the hierarchy) and performed a top-down analysis of the hierarchy to determine the most specic diagnosis that applied to the situation at hand. Each time a diagnosis was established and then rened into a more specic diagnosis, additional domain knowledge was considered by the problem solver. Soon thereafter, Clancey had fully dissected the problem-solving activities of the mycin system. Clancey (1985) showed that the behavior of mycin was not the emergent result of hundreds of modular rules interacting seemingly at random. Rather, Clancey maintained, the developers of mycin had deliberatelyalthough perhaps unintentionallybuilt into mycin a coherent set of inference patterns that collectively he called heuristic classication (Figure 6). For example, the use of data about the case at hand to infer generalizations about the patient to be treated involved a collection of inferences known as feature abstraction. In Figure 6, for example, mycin is shown to use its domain knowledge to reach the conclusion that the patient under consideration may be a compromised host either because the
170 modeling complex systems white-blood-cell count is low or because the patient is an alcoholic. The system is, thus, using input data about the case to determine resulting abstractions about the case. Other inferences allow mycin to perform solution renement, using domain knowledge to distinguish among a group of organisms that may be causing the infection. In this case, mycin uses knowledge of microbiology to identify the most specic organism that may be causing the infection, adopting an approach much like establish-rene. Special inferences, known as heuristics, allow mycin to associate an element of one abstraction hierarchy (e.g., that the patient is a compromised host) with an element of another abstraction hierarchy (e.g., that the organism causing the infection may be of the class Gram-negative rod). In the case of mycin, heuristics links the abstractions related to the condition of the patient to abstractions related to the possible causes of the infection. Clancey (1985) demonstrated a new coherency in the way in which mycin solved problems. The program was not just triggering production rules; it was following a well-dened pattern of inference that involved (1) feature abstraction, (2) heuristic match, and (3) solution renement. Even if the developers of mycin had talked about their work as the elicitation and engineering of hundreds of production rules that related directly to the rules expressed by infectious-disease experts, those rules merged into a singular, consistent problem-solving method when an observer could obtain sufcient distance. More important, Clancey identied dozens of other contemporaneous knowledge-based systems that also seemed to be using heuristic classication as their principal problem-solving method. For example, Elaine Rich had built a system called Grundy that served as a kind of automated librarian, making recommendations to potential borrowers of books to read; Grundy would interview the borrower, perform feature abstraction to conclude some generalizations about the client, then use heuristic match to identify classes of books in which the borrower might be interested (e.g., mystery novels, science ction, biography), and then use additional knowledge to be able to recommend specic books. Workers at sri International had built a system called Prospector that would identify geographic areas in which to drill for minerals; Prospector would elicit features about a given geographic area, use feature abstraction to determine appropriate generalizations, and match those generalizations heuristically to classes of minerals that might be present (Duda
171 Building Intelligent Systems & Shortliffe, 1983). The systems that Clancey (1985) studied were designed for different application domains, but all of them used a shared problem-solving paradigm, one based on feature abstraction, heuristic match, and solution renement. In subsequent years, developers have identied and written computer programs to implement several reusable problem-solving methods that can be applied to a variety of tasks (McDermott, 1988). These contributions include methods such as heuristic classication and establish-rene, where the problem is one of classication, in that the solution set is preenumerated and the problem solver selects one or more solutions from that set. Other problem-solving methods, on the other hand, require the problem solver to construct a solution on the y. These constructive methods encode algorithms for planning, constraint satisfaction, simulation, and so onproblems where it is impossible (or impractical) to preenumerate the solution set in advance. Problem-solving methods themselves can be viewed as having three components: (1) a denition of the algorithm that includes the problem-solving steps that the method carries out (often in some formal language) and the relationships that must hold between the methods input data and the methods output data once the algorithm has nished running (e.g., in the case of heuristic classication, the input data must be mapped to some classication), (2) a description of the format of the methods input data and output data (often referred to as a method ontology; Gennari, Tu, Rothenuh, & Musen, 1994), and (3) software that actually implements the method as a computer program that can execute on some machine. Problem-solving methods can be implemented in virtually any computer language. What is essential is that the method must be programmed in a manner that allows the data on which the methods operate to be provided in terms of an appropriate domain ontology.
172 modeling complex systems responding application-neutral requirements dened for the method (e.g., a case to be classied, input data). Developers can relate the entries of an ontology of microorganisms to the abstract solution set evaluated by the heuristic-classication method; they can relate denitions (such as that of compromised host) to the inferential knowledge required for feature abstraction. Thus, the domain ontology denes the propositional knowledge of the application area; the problemsolving method denes a generic strategy for obtaining a solution for a certain class of problems; the manner in which the developer relates elements of the domain ontology to the data requirements of the problem-solving method determines how the method actually will be used to solve particular cases in the application domain (Crubzy & Musen, 2004). Imagine that a developer were to re-create the mycin system using these types of building blocks (Figure 7). The domain ontology would include concepts such as patient, organism, and antibiotic. Attributes of these concepts would specify the signs and symptoms that patients might demonstrate; the laboratory tests that physicians might request; the relationships between signs, symptoms, and test results; possible classes of infecting organisms; and so on. The reengineered mycin system would use the heuristic-classication problem-solving method to address the task of identifying the most likely infecting organisms (as in Figure 6 above). A separate method would be needed to compute the most parsimonious set of antibiotics that can provide protection against all the potential microorganisms named by the organism-identication task. This latter method needs to ensure (1) that the physician will administer at least one antibiotic that will treat each of the possible pathogens, (2) that the patient receives the fewest number of antibiotics possible, and (3) that the patient does not receive any antibiotic to which he or she is allergic or that may interact with any other drug that he or she is taking. Thus, the mycin antibiotic-selection task might be best solved using a constraint-satisfaction problem-solving method. The system builder would need to relate the data requirements of both the heuristic-classication method and the constraint-satisfaction method to the classes of data described in the domain ontology. In practice developers specify this wiring of domain ontologies to problem-solving methods by declaring explicit mappings between these two kinds of building blocks (Crubzy & Musen, 2004; Gennari et al., 1994).
Figure 7. Reengineering the mycin system with modern components. Propositional knowledge would be represented as a domain ontology. The problem-solving knowledge required for the organism-identication task would be encoded using heuristic classication. The system would require an additional constraint-satisfaction problem-solving method to address the task of constructing a minimal set of antibiotics that can cover for all likely pathogens.
The knowledge-based-systems community has gained considerable experience over the past decade building intelligent systems using explicit domain ontologies, knowledge bases that add detailed content specications to those ontologies, and reusable problem-solving methods (Musen, 2004). These classes of building blocks have enabled a wide range of developers to create extremely complex systems and, most important, to maintain those systems over time as needs evolve and situations change. The evolution in thought has taken place rather quietly. There has been no great hype about any of this. There has been no overbearing excitement, as there was when rule-based systems debuted in the early 1980s. Although the vast majority of small-scale knowledge-based systems continue to be built using the rule-based approach, knowledge-engineering groups are increasingly turning to the use of ontologies and problem-solving methods to construct systems in a way that makes it clear (1) what knowledge about the world actually has been encoded in the systems (what is the ontology) and (2) how the systems go about solving the tasks with which they are confronted (what is the problem-solving method). Ontologies and problemsolving methods exist at two different levelsboth as conceptual entities (abstract models of concepts and of generic problem-solving approaches) and as software artifacts (implemented object systems
174 modeling complex systems that store concept descriptions and implemented problem solvers). When thinking about the design stages depicted in Figure 3 above, it is clear thatabstractlydevelopers can use ontologies and problem-solving methods as building blocks for purposes of conceptual modeling andconcretelydevelopers can use these same building blocks to implement actual computer programs.
175 Building Intelligent Systems before the human brain will be able to understand its own cognition or to manage in an unaided fashion the complexity of computer systems that attempt to simulate human reasoning. Despite our best intentions, cognitive models do not lead to practical computational artifacts that are either highly scalable or easily maintainable. During the 1980s there was a surge of excitement that diverted the attention of some workers in ai away from modeling psychological phenomena to the enticing prospect of modeling basic neurological structures. Articial neural networks came on the scene with tremendous fanfare and the promise that they would overcome the limitations of many of the cognitive models of intelligence (Hinton, 1991). These connectionist architectures consist of layers of nodes (sometimes called neurons) linked together in a manner that is inspired by neural circuitry in the cerebral cortex but that is not representative of any specic neural pathway. Particular values of input data trigger specic nodes in the bottommost layer, with propagation of signals upward through the network. This activity ultimately causes the nodes in the topmost layer to signal an appropriate outputa combination of nodes that can be inferred to represent the classication of the input signal into some category. Articial neural networks have proved to be extremely successful in a large number of classication tasks. They are incapable of performing constructive problem solving, however. The networks can only transform input signals at the bottom layer to output signals at the top layer, and, thus, the solution space (represented by the top layer of nodes) must, by denition, be preenumerated. Although articial neural networks continue to be used to address multivariate classication problems, they do not solve the general problem of engineering computer systems that demonstrate intelligent behavior. The major lesson of the past ve decades is that the most pragmatic route to the creation of intelligent systems is, not through psychology or through neuroscience, but rather through engineering. This lesson has been learned the hard way, by building a large number of systemsmany of them successful, many of them not. Our collective experience has resulted in our ability both to engineer intelligent systems from reusable software components and to maintain and enhance those systems in a maximally efcient manner. Although there were no traditional engineers invited to the Dartmouth conference in 1956, it is the ability to dene a formal engineering discipline for the cre-
176 modeling complex systems ation of intelligent systems that may, ultimately, have had the greatest effect on the eld. The construction of intelligent systems is a problem in constructing artifacts. The importance of the engineerwho cares about the design, implementation, and testing of novel manufactured entitieshas eclipsed that of the psychologist.
177 Building Intelligent Systems Protg assists in the creation, selection, adaptation, and use of domain ontologies and problem-solving methods and aids developers in linking the two types of building blocks together. From the perspective of Protg, a domain ontology must dene a structure for an application area in a manner that system builders and system users can agree on. There is no need for either the users or the participants to have innate beliefs about the particular ontology in advancethe ontology is merely a social convention that species the concepts in the application domain and the attributes of those concepts. Similarly, the choice of a particular problem-solving method to solve a given task is often pragmatic. There is no need for the problem-solving method in any way to approximate the approach that a person might take in solving the given task. The building blocks that Protg users manipulate to create intelligent systemslike the intelligent systems themselvesare simply artifacts. Protg is used regularly by thousands of developers worldwide. It is built in a modular fashion that facilitates the development of plug-ins that can extend the functionality of the basic system. The growing community of Protg users (more than 10,000 developers at the time of this writing) have implemented a wide range of plug insmany of which address functions that our original development group could never have anticipated. Like the Internet itself, where software engineers have augmented the original Web-browser metaphor with countless extensions that go well beyond the original vision for the World Wide Web, Protg has provided a framework for building knowledge-based systems where much of the ownership belongs to the community of developers at large, rather than to the original programming team. Building intelligent systems is inherently a creative process, and the ways in which users choose to support their creativity by augmenting the basic Protg system seem to be almost limitless (Musen et al., 2000). Protg enables developers to work with domain ontologies and problem-solving methods as reusable building blocks for system development without the need to buy into any kind of psychological theory. For readers of this volume, it may, indeed, come as a disappointment that psychology does not play an obvious, central role when developers use Protg to build intelligent systems. Where psychology does take center stage, however, is in studying the way in which developers might interact with systems such as Protg and in
178 modeling complex systems evaluating how alternative design choices in the workbench might affect the ways in which system builders design ontologies and knowledge bases and map their ontologies to selected problem solvers. The modern study of human-computer interaction (Norman & Draper, 1986) and computer-programming behaviors (Weinberg, 1985) has its origins in cognitive psychology; there is an essential role that psychologists can play in analyzing the human problem solving required to build and debug intelligent systems from reusable building blocks. Developers of intelligent systems must manage complexity, visualize intricate relationships, and assess system performance in ways that lend themselves to formal psychological study. We are already at the stage where system builders may incorporate into their work ontologies that contain hundreds of thousands of concepts and problemsolving methods of enormous computational complexity. For our development methods to continue to scale up, there is an urgent need to provide more guidance to system builders, who may need to examine and assemble these intricate building blocks. We are beginning to see the emergence of new methods to help in the visualization of complex knowledge bases and ontologies (Storey, Noy, Musen, Best, Fergerson, & Ernst, 2002), but there is clearly much more work to be done.
179 Building Intelligent Systems The Internet and the Web are about to change qualitatively the nature of intelligent systems. Just as users can translate print documents into hypertext markup language (html) and post them as Web pages anywhere on the Internet, developers now can publish ontologies and knowledge bases on the Web and can make them available to any computer anywhere in the world. Problem-solving methods such as heuristic classication can be made available as Web services, and, suddenly, intelligent systems can be fashioned out of components distributed throughout cyberspace (Benjamins, 2003). The next generation of intelligent systems will comprise the same kinds of components that have become popular as building blocks for creating stand-alone systems, but now those building blocks will be situated throughout the Internet. The result will be what investigators have come to call the Semantic Web, a World Wide Web designed for access, not by humans reading html pages, but instead by computers that will surf the Web in search of appropriate ontologies, knowledge bases, and problem-solving methods (Berners-Lee, Hendler, & Lassila, 2001). The vision is that intelligent agents will be able to locate on the Internet the right building blocks to solve realworld tasks, to access those building blocks, and to use them in their problem solving. Although, at present, creating the Semantic Web remains a research goal, the scientic community has become captivated by the possibility of extending the scope of intelligent systems throughout cyberspace and by the potential of intelligent systems that can draw on vast, virtual libraries of resources that have no physical limitations and to which a global community of developers can contribute (De Roure & Hendler, 2004). There are no psychological correlates for intelligent problem solvers that are based on building blocks distributed across the Internet. There is no model for human cognition that matches the behavior of an intelligent agent that determines dynamically how to optimize the processing of huge amounts of data in parallel and that takes advantage of arbitrary memory stores and processing units that may happen to be available anywhere in the world. Like Deep Blue, intelligent systems that will use the Semantic Web and the computing grid will be successful primarily because of their ability to exploit particular computational technology, not because they will model known cognitive processes. Nevertheless, this new generation of intelligent systems will still use the same kinds of ontologies,
180 modeling complex systems knowledge bases, and problem-solving methods that have become important in the engineering of current intelligent systems, such as those now developed using Protg. Expanding the scope of automated intelligence from programs that run on stand-alone computers to applications that will draw on the resources of the Semantic Web will require a new class of development tools and engineering methods. The principal issue in building intelligent systems is one of managing complex models and their interrelations. As those models individually increase in scope and collectively become distributed in cyberspace, developers will face new challenges in conceptualizing ontologies and navigating their mappings to the various Web services that will provide problemsolving support. The need to visualize the dynamic ow of information across networks and to debug systems when their components are widely distributed will stimulate the creation of a new generation of development tools and techniques. The availability of the entire Internet as a frontier for the creation of new kinds of intelligent applications is extraordinarily inspiring. The Internet alone is not enough to guarantee our success, however. The Semantic Web will become a reality only with the advent of appropriate engineering principles, of new tools that support those principles, and of new methods that can make the use of those tools cognitively tractable.
Conclusion
Early developers of knowledge-based systems viewed their work as a problem in applied psychology. Their belief was that, the better ones model of an actual cognitive process in humans, the better would be the computational artifact that would attempt to solve the same task in the world. Rule-based systems offered appealing computational architectures because much of human problem solving could be modeled in terms of production systems. In practice, however, it was difcult for rule-based systems to scale up to the requirements of many real-world tasks. As rule-based systems were reaching the peak of their popularity, psychologists were strongly questioning the claim that such systems could even be viewed as appropriate cognitive models in the rst place. It is most helpful to view the construction of intelligent systems
181 Building Intelligent Systems as a problem in engineering, rather than in psychology. Intelligent systems can be created from reusable building blocks that include (1) domain ontologies, (2) knowledge bases that provide content knowledge structured by those ontologies, (3) problem-solving methods that automate stereotypical tasks such as heuristic classication and constraint satisfaction, and (4) mappings that relate the concepts in domain ontologies to the knowledge and data requirements of problem-solving methods. Currently, Protg is the most widely used workbench for building intelligent systems from reusable components and demonstrates the utility of this approach. The advent of the Semantic Web will entail the creation of new kinds of intelligent systems from the same kinds of components that we use nowbut the components will be distributed in cyberspace. There is an exciting challenge to build a new generation of usable development tools that will address the enormous complexity that will result when systems can be assembled dynamically from computing resources that are scattered across the Internet. Psychologists will have an invaluable role to play in helping design and evaluate new methods and tools for piecing together the widely distributed building blocks.
Note
The Protg system is supported as a national biotechnology resource by the National Library of Medicine under grant lm007885. Development of Protg has been supported by contracts from the National Cancer Institute and the Defense Advanced Research Projects Agency and by grants from the National Library of Medicine and the National Science Foundation. The system may be freely downloaded from our Web site under an open-source license (see http://protege.Stanford.edu).
References
Anderson, J. A. (1987). Skill acquisition: Compilation of weak-method problem solutions. Psychological Review, 94, 192210. Bachant, J., & McDermott, J. (1984). r1 revisited: Four years in the trenches. ai Magazine, 5(3), 2132. Benjamins, V. R. (2003). Web services solve problems, and problem-solving methods provide services. ieee Intelligent Systems, 18(1), 76-77.
Fostering Team Effectiveness in Organizations: Toward an Integrative Theoretical Framework Eduardo Salas, Kevin C. Stagl, C. Shawn Burke, and Gerald F. Goodwin
University of Central Florida; University of Central Florida; University of Central Florida; U.S. Army Research Institute Organizations seeking to foster teamwork in uid environments are increasingly turning to teams as a preferred performance arrangement. The ongoing proliferation of teams is due in part to the rise of global market opportunities. Multinational corporations facing unprecedented domestic and foreign competition are utilizing teams as a means to organize production and service around a 24-houra-day process. At one time in history the sun never set on the British Empire; today, however, the sun never sets on Intel as software engineers collaborate around the globe in order to meet organizational nancial forecasts. The spread of teams is also due in part to the ongoing technological revolution. Once restricted to collocation, technology now makes it possible and protable for team members to be distributed in space-time. As organizations become more comfortable with, and dependent on, technology, we believe that opportunities for team-based collaboration will abound.
186 modeling complex systems Unfortunately, however, history has repeatedly shown that team performance is an elusive, dynamic, and complex phenomenon (Salas, Stagl, & Burke, 2004). Indeed, many recent tragic incidents (e.g., the uss Vincennes, the uss Stark, Occidentals Piper Alpha platform, the space shuttle Columbia) have been attributed to breakdowns in teamwork. These failures are due in part to the often uid and sometimes chaotic environment in which teams operate. Thus, developing a thorough understanding of how teams interact in a synchronized fashion to achieve shared goals is critical to meeting short-term tactical objectives and to cultivating long-term strategic success. The proliferation of teams and the concomitant focus on orchestrating team effectiveness so that team members, teams, multiteam systems, and organizations can reach their full potential is supported by a growing body of theoretical and empirical research (see Belbin, 2000; Bell & Kozlowski, 2002; Beyerlein, Johnson, & Beyerlein, 2003; Cooper & Robertson, 2004; Edmondson, 2003; Hackman, 1990; Ilgen, Major, Hollenbeck, & Sego, 1993). In fact, during the past century a plethora of research initiatives have been undertaken with the goal of more fully illuminating the complexities of teamwork and, thereby, advancing the science of teams. Therefore, it is no surprise that, in a recent literature review, we identied more than 800 articles and chapters that present empirical evidence addressing some aspect of team effectiveness. We suspect that this body of empirical evidence will only continue to expand exponentially in step with the ongoing proliferation of teams. In addition to the growing number of empirical investigations noted above, there is a concomitant increase in team-effectiveness theory building. Those concerned with developing a deeper understanding of teamwork in order to promote team synchronicity have proposed a wide variety of theoretically grounded models and frameworks that address both team development and team effectiveness. Later in this chapter we extract a representative sample and expound on how each of these initiatives has taken a different path to framing team effectiveness, including midpoint effects (Gersick, 1988), episodic team processes (Marks, Mathieu, & Zaccaro, 2001), temporal entrainment (Ancona & Chong, 1999), and team maturation (Morgan, Salas, & Glickman, 1994). Although dozens of intriguing conceptualizations of teamwork have surfaced in the past 25 years, there has not been a compre-
187 Team Effectiveness in Organizations hensive review of team-effectiveness models and frameworks conducted since the early 1990s (see Campion, Medsker, & Higgs, 1993; Tannenbaum, Beard, & Salas, 1992). Since that time, organizations have changed in many ways, and much has been learned about teams (see Campbell & Kuncel, 2001; Edmondson, 1999; Hackman, 2002; Ilgen, 1999; Kozlowski & Bell, 2003; West, Tjosvold, & Smith, 2003). This state of affairs suggests that there are a growing number of unincorporated team-effectiveness research ndings, models, and frameworks in the current body of relevant literature. If there are, indeed, a number of recent research initiatives undertaken to frame teamwork that were not already addressed by prior comprehensive reviews, then it is likely that they include constructs, construct operationalizations, and proposed construct relations that incrementally add to our understanding of the complex picture of team effectiveness. Furthermore, there are several problems associated with having such wide diversity in the denition and measurement of team constructs. Likewise, there are clear positives associated with maintaining a common language when engaging in team research (Salas & Cannon-Bowers, 2000). Also, accumulating and integrating what is known to date are challenges confronting both team researchers and industrial organizational psychology (Campbell, 1990b). Given the above arguments, we believe that undertaking a comprehensive review and integration of team-effectiveness models and frameworks will be benecial to team researchers by revealing common ground on the substantive team constructs of interest. Therefore, in this chapter we present a multilevel, multidisciplinary synthesis of team-effectiveness models and frameworks advanced during the last 25 years. This endeavor was undertaken to collect and integrate current comprehensive models and frameworks of team effectiveness in order to present a more coherent picture of the factors constituting and impinging on teamwork. Specically, in this chapter we begin by dening the terms team, teamwork, team performance, and team effectiveness. Once these preliminary issues are addressed, we describe a set of criteria utilized to guide our literature review. Our review of the team literature resulted in the identication and subsequent integration of 138 initiatives that modeled or framed aspects of team performance or effectiveness. A representative sample of 11 models and frameworks that
188 modeling complex systems typify cutting-edge team advancements are reviewed in detail. On the basis of this research initiative, we advance an integrative, albeit preliminary, multilevel framework of team effectiveness. The advanced framework was designed as a simple yet rich heuristic that reects the current state of the art within the team domain. Furthermore, we expect that our framework will serve as a departure point for team researchers concerned with addressing the salient features and conditions that constitute and affect teamwork in specic types of teams. After describing the novelties of our approach to framing team effectiveness, we discuss a few practical applications (i.e., measurement, training, stafng) and how they can be more fully appreciated when viewed through the lens afforded by a theoretically grounded framework. This chapter concludes with a call for the generation and, ultimately, widespread adoption of a set of short-, mid-, and long-term goals that can serve to guide future team research. We believe that team research has reached its teenage years and that tackling the substantive issues that remain to be addressed by team researchers will help curtail the haphazardness that has characterized the selection of constructs investigated up until this point in time. Much work remains to be done, and a collective, coordinated effort among all team researchers is needed to accomplish it.
teams
The term team is dened in the online Merriam-Webster dictionary as a number of persons associated together in work or activity. In addition to this conceptualization, a plethora of other perspectives have been advanced during the past 25 years to dene team (Guzzo & Dickson, 1996). For our purposes we dene team as follows: it is a complex entity consisting of (1) two or more individuals (2) who interact socially and (3) adaptively, (4) have shared or common goals, and (5) hold meaningful task interdependencies; it (6) is hierarchically structured and (7) has a limited life span; in it (8) expertise and roles are distributed; and it is (9) embedded within an organizational/environmental context that inuences and is inuenced by ongoing processes and performance outcomes (Salas, Dickinson, Converse, & Tannenbaum, 1992). This denition exemplies action teams that perform some of the most critical work in the global economy, including emergencymanagement teams, medical teams, command and control military teams, cockpit crews, control-tower teams, oil-rig crews, submarine teams, sports teams, forensic-science teams, gene-sequencing teams, archaeological teams, and space-exploration teams. It also typies many types of teams appearing in more traditional business settings, such as nancial-analyst teams, research-and-development teams, construction teams, marketing teams, and fast food teams. All these team exemplars often perform in uid environments typied by (1) rapidly evolving and ambiguous situations, (2) no optimal answers, (3) information overload, (4) intense time pressure, and (5) severe consequences in cases of error (Orasanu & Salas, 1993). In fact, modern operational environments are characterized by a historically unparalleled accelerating rate of change that mandates team exibility, adaptability, and resilience. We suspect that operational environments will increase in uidity in step with emerging technologies.
teamwork
Teamwork is dened in the online Merriam-Webster dictionary as work done by several associates with each doing a part but all subordinating personal prominence to the efciency of the whole.
190 modeling complex systems Although this denition is elegant, a more detailed one has been advanced in Salas, Sims, and Klein (2004): Teamwork is a set of exible behaviors, cognitions, and attitudes that interact to achieve desired mutual goals and adaptation to the changing internal and external environments. Teamwork consists of the knowledge, skills, and attitudes (ksas) that are displayed in support of ones teammates, objectives, and mission. Essentially, teamwork is a set of interrelated thoughts, actions, and feelings that combine to facilitate coordinated, adaptive performance and the completion of taskwork objectives (pp. 497498). This denition illustrates the core of what we mean when we refer to teamwork throughout the remainder of this chapter. In the previous paragraph, we dened teamwork as adaptively enacted processes undertaken to achieve shared goals. Furthermore, we noted that a set of knowledge, skills, attitudes, and other characteristics (ksaos) underpins and enables the enactment of the steps constituting a given process. However, the question remains as to which ksas are central to teamwork. In other words, what constitutes teamwork? The answer to this question, not surprisingly, is highly contingent on the type of team under consideration. As noted in the previous section, this chapter emphasizes a particular type of team (e.g., one that is hierarchically structured, with tight interdependencies). Therefore, the ksas presented in the remainder of this section are most appropriate for teams with characteristics similar to those mentioned above. However, this caveat is not meant to imply that some of the ksas advanced below do not generalize to all types of teams, as it would be difcult to imagine any interdependent collective that did not on occasion require communication or coordination. There have been many systematic initiatives undertaken to illuminate the ksas that constitute teamwork (see Cannon-Bowers, Tannenbaum, Salas, & Volpe, 1995; Dickinson & McIntyre, 1997; Fleishman & Zaccaro, 1992; Marks et al., 2001; Marks, Zacarro, & Mathieu, 2000; McIntyre & Salas, 1995; Salas, Sims, & Burke, 2005; Smith-Jentsch, Zeisig, Acton, & McPherson, 1998; Stevens & Campion, 1994). For example, McIntyre and Salas (1995) delineated teamwork skills and performance norms from their programmatic work with navy tactical teams (e.g., guided-missile teams). The four teamwork skills advanced by McIntyre and Salas are (1) mutual performance moni-
191 Team Effectiveness in Organizations toring, (2) feedback, (3) closed-loop communication, and (4) backup behavior. Also identied were two performance norms: (1) a teams self-awareness and (2) the fostering of within-team interdependence. Another approach to understanding teamwork was taken in Cannon-Bowers et al. (1995). These researchers collected and synthesized prior research addressing teamwork. This effort resulted in the advancement of the ksa competencies proposed to underlie teamwork in organizations. Specically, Cannon-Bowers et al. suggest that teams require several knowledge competencies, including (1) cue-strategy associations; (2) teammate characteristics; (3) role responsibilities; (4) shared task models; (5) team mission, objectives, norms, and resources; (6) relation to the larger organization; (7) task sequencing; (8) team role-interaction patterns; (9) procedures for task accomplishment; (10) accurate task models; (11) accurate problem models; (12) boundary-spanning role; and (13) teamwork skills. Furthermore, they suggest that teams require dozens of skills, skills that can be clustered into eight higher-order dimensions, including (1) adaptability, (2) shared situational awareness, (3) performance monitoring and feedback, (4) leadership/team management, (5) interpersonal relations, (6) coordination, (7) communication, and (8) decision making. Finally, they propose that teams require a variety of attitudinal competencies, including (1) team orientation, (2) collective efcacy, (3) shared vision, (4) team cohesion, (5) interpersonal relations, (6) mutual trust, (7) collective orientation, and (8) a belief in the importance of teamwork. While McIntyre and Salas (1995) and Cannon-Bowers et al. (1995) made great strides in advancing our collective understanding of the core content constituting teamwork, the more recent research of Dickinson and McIntyre (1997), Marks et al. (2001), and Salas et al. (2005) has made a signicant contribution to our collective understanding of the dynamic interdependencies existing among the components of teamwork. These three particular initiatives are critical enough to warrant closer inspection, and, thus, a thorough description is provided in a later section reviewing extant theoretical models and frameworks of team effectiveness. It will satisfy for now to state that teamwork ksas operate, not in isolation, but dynamically, simultaneously, and recursively as they unfold over time to emerge as team performance.
team performance
Performance is dened in the online Merriam-Webster dictionary as either (a) the execution of an action or (b) something accomplished. What is noteworthy in this denition is the inclusion of both cause (i.e., action) and effect (i.e., accomplishment). This mingling of antecedent and consequence is rejected by those researchers concerned with modeling human performance. In fact, Campbell (1990a, p. 704) states: Performance is not the consequence(s) or result(s) of action; it is the action itself. Following this line of thinking, performance has been dened within the domain of industrial organizational psychology as goal-directed behavioral or cognitive action (Campbell, 1990a; Campbell, Dunnette, Lawler, & Weick, 1970; Dunnette, 1963). While Campbell, Dunnette, and their colleagues assertions provide a basis for understanding human performance at a conceptual level, in practice the timing of performance measurement often leads to confusion. This is because operationalizations of performance are contingent on the timing of measurement. Thus, prior research has often indiscriminately conceptualized performance as either a series of behavioral or cognitive actions that unfold over time or the outcomes resulting from those actions (Kozlowski & Bell, 2003). Specically, when performance is examined longitudinally, it is often conceptualized as a series of actions. However, cross-sectional research is typied by the measurement of performance outcomes or the result(s) of expended energy (e.g., number of products produced). Our point is that, regardless of whether longitudinal or cross-sectional research is conducted, team performance remains either a behavioral or a cognitive action and outputs of these actions are best conceptualized as performance outcomes. In both longitudinal and cross-sectional team research, a record can be compiled of actions and outcomes at multiple levels in the conceptual space. This speaks to the issue of what constitutes team performance. We assert that team performance consists of the display by one or more team members of taskwork competencies (e.g., running a computer program), teamwork competencies (e.g., backup behavior), and integrated team-level action (i.e., coordination, adaptation). This line of thinking leads to the conclusion that team performance is a multilevel phenomenon, a point echoed elsewhere in recent scientic literature (Kozlowski & Klein, 2000).
193 Team Effectiveness in Organizations Essentially, team performance can be conceptualized as a bottom-up emergent process unfolding from individuals to teams. In other words, team performance begins with team members and emerges upward to teams. However, team performance is also simultaneously a top-down process. This line of reasoning suggests that higher-level context has a direct or moderating effect on team performance. However, this does not preclude the foundation of team effectiveness residing in the cognition, behavior, and affect of individual team members. In fact, team members thoughts, actions, and feelings via social interaction, dyadic role exchanges, and amplication have emergent properties that manifest at higher levels (i.e., bottom-up emergence). The assertion that team performance is both a top-down and a bottom-up multilevel phenomenon that emerges over time is consistent with the guiding tenets of general systems theory (von Bertalanffy, 1956) and its variants, which collectively acknowledge that teams are hierarchically nested, interdependent, open systems (Kozlowski & Klein, 2000).
team effectiveness
Effectiveness is dened in the online Merriam-Webster dictionary as producing a decided, decisive, or desired effect. Thus, effectiveness is, not the outcome produced from team performance, but rather the result of a judgmental process whereby an output is compared to a subjective or an objective standard. Similarly, effectiveness has been dened within industrial organizational psychology as the evaluation of the results of performance (Campbell, 1990a). While this conceptualization of team effectiveness seems straightforward, Cohen (1994) disagrees: Most group effectiveness researchers think about effectiveness as a multidimensional construct, but do not agree as to the criteria of work group effectiveness (p. 71). For example, team effectiveness has been operationalized as (1) archival records of productivity, (2) self-report measures of employee satisfaction, and (3) manager judgments of effectiveness (Campion et al., 1993). Subsequent research reported in Campion, Papper, and Medsker (1996) operationalized team effectiveness as (1) immediate managers judgments at two points in time, (2) senior managers judgments, (3) peer managers judgments, (4) subordi-
194 modeling complex systems nates judgments, (5) archival records of satisfaction, and (6) archival records of performance appraisals. A wide variety of other operationalizations of team effectiveness have also been advanced. For example, Cohen (1994) adopted a tripatriate perspective asserting that the variables contributing to team effectiveness can be clustered into three broad categories, including (1) team performance, (2) team members attitudes about quality of work life, and (3) withdrawal behaviors. According to Cohen (1994), each of these three global factors encompasses a number of effectiveness-related variables. Specically, the performance factor includes (1) controlling costs, (2) increasing productivity, and (3) increasing quality. The team members attitudes factor incorporates (1) job satisfaction, (2) team satisfaction, (3) satisfaction with social relationships, (4) satisfaction with growth opportunities, and (5) organizational commitment. The withdrawal behaviors include both (1) absenteeism and (2) turnover. Hackman (1987) also asserts that group effectiveness can be conceptualized in terms of three components. The rst component consists of a judgment by those stakeholders who review the work of teams in terms of whether it meets their standards for quality and quantity. The second is whether the needs of group members are satised by their team participation. The third is whether group interaction has served to maintain or strengthen the groups ability to work together at some future date. Finally, Sundstrom, DeMeuse, and Futrell (1990) suggest that team effectiveness is composed of (1) managers and customers judgments about the acceptability of performance and (2) team viability, where team viability is dened as commitment on the part of team members to continuing to work together. We suspect that some of the ambiguity underlying the nature of team effectiveness lies in an incomplete specication of effectiveness across levels and time. To date, most team research has operationalized and measured effectiveness at both the individual (i.e., team members self-reports of satisfaction/commitment) and the team (i.e., ratings of team coordination) levels. In other words: Team effectiveness is different from individual effectiveness (Yetton & Bottger, 1982, cited in Tannenbaum et al., 1992, p. 117). Furthermore, individual effectiveness and team effectiveness are distinct from, but interdependent with, multiteam-system and organizational effectiveness. Certainly, a comprehensive discussion of team effective-
195 Team Effectiveness in Organizations ness must extend beyond the individual and team levels of analysis typically addressed in research in order to incorporate effectiveness factors at the multiteam system and organizational levels. This is because the criteria relied on by stakeholders to form effectiveness judgments are contingent on whether the system under consideration is the team member, the team, the multiteam system, or the organization. In addition, team effectiveness judgments are intertwined with temporal factors such as team maturation. As a team and its members mature, a given criterion, once heavily relied on by senior stakeholders in forming their judgments about the effectiveness of a particular aspect of the team, may no longer be salient. Moreover, the type of criteria adopted may change as teams pass through the early stages of formative development en route to achieving process gains. For example, during the norming stages of team development, judgments of effectiveness may be directed at the accuracy of team-member mental models. However, as the team matures, the emphasis may shift from individual mental models to shared and compatible mental models that serve as a cognitive reservoir for fueling coordinated adaptation to novel challenges. As noted above, the criteria adopted by stakeholders to form their effectiveness judgments change over time. Specically, we noted that temporal dynamics such as team maturation inuence what criteria are relevant and how those criteria are weighted. The criteria utilized by stakeholders to form their effectiveness judgments may also vary over time as a function of the raters level in the organization. In this example, the stakeholders organizational level is meant to convey expertise or systems thinking (see Senge, 1990). Certainly, a strong argument can be made that, as incumbents navigate a given organizational hierarchy and develop a history with the target unit under consideration, there is a concomitant evolution in their perspective on what and how much of a given set of qualities constitute effectiveness. The above-noted line of reasoning is supported by empiricalresearch evidence collected from the naturalistic decision-making paradigm suggesting that expert decision makers (e.g., reghters, chess masters, police ofcers) emphasize different criteria than novices do when forming judgments about stimuli (see Klein, 1993). Thus, it may be plausible to conceive of an interaction between the
196 modeling complex systems target unit being rated and the level of the stakeholder (i.e., rater) when predicting both (1) the types of effectiveness criteria attended to and (2) how chosen criteria are weighted to form global effectiveness judgments. If these arguments are valid, they have implications for the design of performance-appraisal systems and for rater training. Thus, these assertions warrant closer attention in future theory building, research, and practice. Essentially, effectiveness is a value judgment that is inuenced by a number of factors. In this subsection, we noted how a particular criterion or set of criteria can change as a function of (1) the target units level in the conceptual space, (2) the target units level of maturation, and (3) the idiosyncratic characteristics of raters such as systems thinking. Although these three factors are undoubtedly important in understanding stakeholder judgments of team effectiveness, a number of other pertinent issues also exist. For example, a host of individual (e.g., latent cognitive resources, available cognitive resources, biases), team (e.g., performance phase), situational (e.g., time pressure, novelty), and organizational (e.g., culture, reward systems) factors also inuence judgments of effectiveness. In sum, the substantive similarities and differences of the content constituting and the process underlying effectiveness judgments across system levels and time remain to be addressed in future research endeavors.
Doolen, Hacker, & Hackman & Oldham (1980) Van Aken (2003) Beck (2002) Drexler, Sibbet, & Forrester (1988) Hall (2001) Bolman & Deal (1992) Driskell, Radtke, & Salas (2003) Haward et al. (2003) Brandes & Weise (1999) Druskat & Kayes (1999) Heinemann & Zeiss (2002) Brodbeck & Greitemeyer (2000) Durham & Knight (1997) Helmreich & Foushee (1993) Bunderson & Sutcliffe (2002) Edmondson, Roberto, & Hertel, Konradt, & Watkins (2003) Orlikowski (2004) Burke, Stagl, Salas, Pierce, & Entin & Serfaty (1999) Higgins & Routhieaux (1999) Kendall (2006) Campion, Medsker, & Erez, LePine, & Elms (2002) Higgs & Rowland (1992) Higgs (1993) Campion, Papper, & Fleishman & Zaccaro (1992) Hinsz, Tindale, & Vollrath Medsker (1996) (1997) Choi (2002) Flemming & Monda-Amaya (2001) Hoegl, Praveen, & Gemuenden (2003) Cohen (1994) Flood, Hannan, Smith, Turner, West, Hogan, Raza, & & Dawson (2000) Driskell (1988) Cohen & Bailey (1997) Furst, Reeves, Rosen, & Hollenbeck et al. (1995) Blackburn (2004) Cohen, Ledford, & Gersick (1988) Hollingshead & McGrath (1995) Spreitzer (1996) Janz, Colquitt, & Noe (1997) Millitello, Kyne, Klein, Shea & Guzzo (1987) Getchell, & Thordsen (1999) Jung, Sosik, & Baik (2002) Mischel & Northcraft (1997) Sheard & Kakabadse (2001) Kahai, Sosik, Avolio (1997) Morgan, Dailey, & Kulisch Sheehan & Martin (2003) (1976) Karau & Kelly (1992) Morgan, Salas, & Smith-Jentsch, Zeisig, Acton, Glickman (1994) & McPherson (1998) Kelley (2001) Navarro (1994) Sonnentag, Frese, Brodbeck, & Heinbokel (1997)
Tubbs (1994) Walker & Vines (2000) Werner & Lester (2001)
West, Borrill, & Unsworth (1998) Schippers, Den Hartog, Koopman, Yeatts & Hyten (1998) & Wienk (2003)
our criteria for inclusion (see Table 1). Specically, inclusion was dictated by whether a particular initiative addressed team performance or effectiveness and described the relations between three or more constructs or construct categories. The latter of these two criteria was designed to screen out the thousands of research initiatives that addressed only a few of the myriad constructs currently believed to constitute the complex nomological network of teamwork. Although our literature review, which included searches of both electronic databases and our own extensive collection of team literature, resulted in an achieved population of 138 models and frameworks, only 11 of the identied theories are reviewed here. A strong case can, undoubtedly, be made that the remaining 127 sources are
199 Team Effectiveness in Organizations signicant in their own right. Unfortunately, space constraints preclude a thorough examination of all identied models and frameworks. Space limitations also restrict the detail with which any given theory can be addressed. Thus, the inclusion of the 11 theories expounded on in this section is not intended as a covert endorsement of a particular line of research. However, with this caveat in mind it should also be noted that we believe that the sample of models and frameworks selected for review in this chapter is quite representative of the larger population of theoretical research prevalent within the team domain. In fact, the 11 models and frameworks reviewed below were deliberately sampled because they illustrate some of the key advancements contributing to our collective understanding of teams over the past 25 years. The particular initiatives selected for review in this section were vital in shaping our thinking about the key constructs constituting and contributing to team effectiveness. Thus, this section provides critical background information for the reader interested in gaining a deeper appreciation of why a given model or framework was strongly inuential in the development of the conceptually integrative team-effectiveness framework presented later in this chapter. Prior to launching into a full discussion of the salient features of the 11 models and frameworks, we begin by describing the criteria utilized to select these particular research efforts from the larger population of initiatives identied via our literature review.
200 modeling complex systems Model Form Perhaps the most prevalent theme within the team-effectiveness literature reviewed was the widespread use of ipo (inputprocess-output) models as a guiding framework (McGrath, 1964). The epistemological foundation for ipo models originates within general systems theory and its many derivatives (see von Bertalanffy, 1956). Thus, the use of ipo models to frame team effectiveness is consistent with derivatives of general systems theory, including both sociotechnical systems theory (Emery & Trist, 1960) and open systems theory (Katz & Kahn, 1978). Although ipo models are predominant in the teams domain, it should be noted that there is ongoing debate about the adequacy of this approach (Hackman, 2002; Ilgen, Hollenbeck, Johnson, & Jundt, in press; Sundstrom et al., 1990). Generally speaking, ipo models highlight the importance of throughputs as mediators or moderators of the relations between input factors (e.g., team and individual characteristics) and outputs (e.g., team satisfaction, performance). Thus, the use of an ipo model to frame team effectiveness is particularly important for capturing the dynamic interactions and emergent states that constitute teamwork and illuminating the nomological network of these processes. Describing effectiveness through an ipo lens not only emphasizes the importance of the interactions between inputs, processes, and resulting outputs but also recognizes that team outputs can be fed back into input variables and that teamwork does not occur in a vacuum (Tannenbaum et al., 1992). The use of ipo models reects the current state of the art within the teams domain, as witnessed by the fact that the preponderance of models and frameworks identied by our literature search adopted this format. Thus, describing teamwork through an ipo lens has advanced our collective understanding of the factors that constitute and promote team effectiveness. However, there have been other advances with regard to team effectiveness that are even more intriguing, and these innovations will be reviewed in the following subsections describing model dynamism and model focus. Model Dynamism A second critical difference between the models and frameworks of team effectiveness identied in our literature review is that they tend to diverge on the degree to which they explicitly consider how team processes may differ across the dynamic task cycle of teams. While initial attempts by researchers to describe
201 Team Effectiveness in Organizations teamwork focused primarily on team inputs as a mechanism for manipulating team outcomes (Goldstein, 1993; Guzzo & Dickson, 1996), more recent models and frameworks of team effectiveness have served to illuminate the black box of team process. For example, Campion and his colleagues research (Campion et al., 1993; Campion et al., 1996), although inclusive of many of the key constructs believed to relate to team performance, is relatively static in nature. Conversely, the research advanced by Dickinson and McIntyre (1997) adopts a more dynamic position. The movement to model uidity reects a growing recognition within the teams community that collective task performance requires adaptive moment-to-moment interteam and intrateam interaction (Dickinson & McIntyre, 1997; Fleishman & Zaccaro, 1992; Marks et al., 2001; Shiett, Eisner, Price, & Schemmer, 1985). Modeling the dynamic nature of interdependent action blurs the distinction between traditionally held notions of predictors and criteria and, therefore, mandates the building of increasingly sophisticated theories in order to capture the inherent complexities. In addition to raising a number of intriguing directions for future endeavors, research such as that presented by Dickinson and McIntyre (1997) and Marks et al. (2001) has implications for human-capital-management systems. For example, if particular processes take precedence over other processes at specic phases of teamwork, then performancemeasurement, -feedback, and -training systems should be appropriately tailored to reect these contingencies. Model Focus A third primary difference between the research efforts identied in our literature review concerns the overall focus of the work. The focus factor pertains to whether a particular model or framework denes performance as being driven primarily by internal team factors or by external context. A similar approach was utilized by Ancona and Chong (1999) to map existing group theory. These authors assert that the key to understanding team behavior lies in a shift in focus from internal factors, such as midpoints and stages of team development, to the temporal role of team context. These assertions are theoretically supported by a branch of mathematics known as coupled oscillations (i.e., entrainment). The stream of thought concerning entrainment is grounded in the premise that team members behavior becomes entrained to
Figure 1. Conceptual model of team performance (Nieva, Fleishman, & Reick, 1978).
naturally occurring contextual cycles. For example, entrainment is witnessed in the recurring urry of goal-directed activity that takes place within a top management team as the scal quarter comes to a close. As the management teams cycles become entrained to the quarterly business cycle, the environment serves as a pacer, rhythm setter, creator of opportunity, source of interrupts, and context for the meaning of time. Efforts to understand teamwork via entrainment suggest that performance is largely determined by contextual factors in the teams external environment. However, many of the models and frameworks delineated in the material covered in our literature search conceptualize performance as a phenomenon that is actively driven by members and events within the team. The differences between these perspectives have important implications for framing team effectiveness. For example, interventions meant to enhance teamwork have traditionally targeted variables inside the team for change. Given the compelling argument supporting entrainment, it seems that traditional approaches emphasizing internal factors should be augmented by an understanding of key external factors. A simultaneous internal and external focus could illuminate insights into critical processes such as the timely delivery of scarce resources. In fact, some initiatives have already been undertaken to examine the joint inuence of both internal and external forces (e.g., Kozlowski, Gully, Nason, & Smith, 1999; McGrath, 1991).
sition variables (e.g., skills, heterogeneity) and group structure (e.g., formal leadership, work norms). In addition, it includes organizational-level input factors such as resources available (e.g., training, consulting) and organizational-structure variables (e.g., rewards, supervisory control). The relations between individual- and organizational-level input factors and team effectiveness are proposed to be mediated by group processes. The model also indicates that group task complexity, uncertainty, and interdependence moderate the relations between group processes and outcomes such as satisfaction. Hackman (1987) also advanced a model of team effectiveness (see Figure 3). This model illustrates how a broad range of variables can inuence teamwork. There are a number of propositions that ow from this model, including that input factors (e.g., organizational context, group design) are related to team processes, which, in turn, are related to team effectiveness (e.g., customer satisfaction with product, member satisfaction, capability of team members to work together over time). More specically, the model highlights the importance of fostering an organizational context that supports and reinforces teamwork via rewards, education, and availability of information. Group design is a second input factor illustrated that is proposed to relate to team processes. According to Hackman, group
design consists of such things as (a) task structure, (b) group composition, and (c) appropriate group norms regarding teamwork. Hackmans (1987) model also species process criteria of effectiveness that can serve to guide the diagnosis of team weaknesses. These process criteria include (a) level of effort, (b) amount of knowledge and skill, and (c) appropriateness of task-performance strategies. Finally, the model proposes two variables that may moderate the depicted relations. First, the relations between team inputs and team processes are moderated by the ability of the group to minimize process losses (i.e., gain group synergy). Second, the relations between team processes and team effectiveness are moderated by the material resources available to the team. Thus, no matter how well team members interact with one another in terms of effort, skill, and performance strategies, if there are not enough material resources, the task may not be completed. A fourth model of team effectiveness reviewed was advanced by Gersick (1988) and is known as the punctuated-equilibrium model (pem) (see Figure 4). The pem is empirically grounded in the ndings of research conducted with a sample of eight diverse teams. The evidence accumulated from Gersicks research suggests that teams determine an initial method of performance during their rst meeting and adhere to this method until the midpoint of the target objective is reached. Gersick argues that, at the midpoint, team members become aware of the time left to completion and switch their strategy accordingly. Although fairly simple in its depiction, this research makes an important contribution by showing that teamwork is dynamic and that teams adapt their performance strategies in accordance with temporal contingencies. In Tannenbaum et al. (1992), which reported one of the rst integrative efforts to frame team effectiveness (see Figure 5), the researchers reviewed many of the then-current ipo models of team effectiveness and developed an integrative framework that built on prior initiatives. This framework is more complex than the pre-
viously described Gladstein (1984) model but includes some of the same variables. Specically, it illustrates four distinct types of input variables, including (1) task characteristics, (2) work characteristics, (3) individual characteristics, and (4) team characteristics. It suggests that these input factors affect each other and also serve to affect both team members and team processes (e.g., backup behavior, coordination, adaptability) that occur over time. Not surprisingly, both the individual team member and team processes are proposed to affect team-performance outcomes (e.g., quality, quantity, time, errors). The framework also depicts system feedback, resulting from team performance and performance outcomes, cycling back as subsequent system input. Furthermore, it recognizes that training or teambuilding interventions may moderate the relations between inputs and processes as well as those between processes and performance outcomes. A nal addition of this framework over previous models is the recognition that organizational and situational characteristics affect team effectiveness, not just at the input stage, but throughout the entire ipo process. A sixth model of team effectiveness reviewed herein was advanced in Campion et al. (1993) (see Figure 6). Synthesizing all ve team-effectiveness models already illustrated, Campion et al. devel-
Figure 6. Synthesized model of group effectiveness (Campion, Medsker, & Higgs, 1993).
oped a metamodel of team effectiveness. This hybrid model incorporates only those constructs proposed to directly affect team effectiveness, thus ignoring key mediators and moderators of the relations between team inputs and outputs. Campion et al.s model parsimoniously frames the complexities of team effectiveness. Although it may oversimplify the dynamic, recursive nature of team performance, this programmatic theoretical formulation has received empirical support from two diverse samples of teams (Campion et al., 1993; Campion et al., 1996). Campion et al.s (1993) metamodel describes ve categories of variables that are proposed to affect team effectiveness: job design, interdependence, composition, context, and process. Specically, job
Figure 7. Team-evolution and -maturation model (Morgan, Salas, & Glickman, 1994). * Gersick (1988). ** Tuckman (1965).
209 Team Effectiveness in Organizations design subsumes self-management, participation, task variety, task signicance, and task identity. Interdependence is proposed to encompass task interdependence, goal interdependence, and interdependent feedback/rewards. Composition is proposed to subsume heterogeneity, exibility, relative size, and preference for group work. Context is proposed to cover training, managerial support, and communication/cooperation between groups. The process category includes potency, social support, workload sharing, and communication/cooperation within groups. Building on the work of Gersick (1988) and Tuckman (1964), Morgan et al. (1994) developed a model that illustrates the stages that teams progress through before, during, and after task performance (see Figure 7). The model proposes that task-orientated teams progress through a series of developmental stages at varying rates (see Tuckman, 1965). The specic stage at which a given team begins and how quickly the team progresses through the proposed stages depend on such characteristics as (a) members experience as a team, (b) individual expertise, (c) task characteristics, and (d) environmental context. Morgan et al.s model also proposes that, as a team progresses through these stages, there are two tracks of skills that must be mastered before it can perform effectively: taskwork and teamwork. Taskwork represents the task-orientated skills that the members must understand and acquire for task performance (Salas et al., 1992, p. 10). Conversely, teamwork skills reect the behavioral interactions, cognitions, and attitudinal responses that must be mastered before a team can work together effectively. The advantages of the Morgan et al. model over previous theoretically grounded models are twofold. First, it illustrates that there are two sets of skills that must be mastered in order for a team to be effective, both task and teamwork. Subsequent research has provided empirical support for this assertion (Glickman et al., 1987). Second, by illustrating behaviors that occur within each developmental stage, the model serves to frame how teamwork develops during task performance and training. More recently, Dickinson and McIntyre (1997) proposed a model that describes the interrelations between essential teamwork processes such as communication, team orientation, team leadership, monitoring, feedback, backup behavior, and coordination (see Figure 8). In this model communication acts as the glue linking together all
other teamwork processes. Another aspect of the model is the integration of both team skill competencies (e.g., team leadership) and team attitude competencies (e.g., team orientation). Together, team leadership and team orientation are proposed to facilitate a team members capability to monitor his or her teammates performance. Furthermore, the model proposes that performance monitoring drives both the content of feedback and timely backup behaviors. It also suggests that, when all the aforementioned teamwork competencies are occurring in unison, they synergistically serve as a platform for team coordination. In turn, the feedback resulting from team coordination serves as input back into team processes. Although Dickinson and McIntyres (1997) model incorporates many of the skill competencies underlying teamwork, it fails to model many of the critical antecedents and outcomes of team process. The ninth initiative reviewed in this section, undertaken to illuminate the complexities of team effectiveness, is the research of Marks et al. (2001) (see Figure 9). This systematic research has markedly advanced our understanding of the content constituting team process and the uid nature of teamwork as processes unfold over time. Marks et al. advance a temporally based framework of team effectiveness that extends recent notions of team process by categorizing throughputs into recurring phases. Specically, this episodic model of team process consists of a series of recursive ipo loops proposed to occur sequentially and simultaneously during both a transition stage and an action stage of performance. Distinct competencies characterize the action (e.g., mission analysis, goal specication)
Figure 9. The rhythm of team task accomplishment (Marks, Mathieu, & Zaccaro, 2001).
and the transition (e.g., systems monitoring, coordination) stages, suggesting that certain ksas take precedence depending on the timing of performance. Interpersonal processes are proposed to occur during both stages. As previously noted, Marks et al.s (2001) effort to model teamwork is theoretically as well as practically intriguing. The implications for those charged with observing, assessing, and rewarding team performance include the need to develop a new appreciation of the timing of their measurements as well as new human-capitalmanagement systems that account for which processes are predominate at a given time. There is nothing as practical as a good theory (Lewin, 1951, p. 169), and this maxim is nowhere more applicable than when utilized to describe meaningful research that offers a more robust specication of the dynamic nature of teamwork. Another recently advanced theoretical initiative reviewed was the Big Five model, proposed by Salas et al. (2005). This model was developed in an effort to highlight the essence of teamwork by illustrating the relations between the processes that they argue constitute the core of interdependent interaction (see Figure 10). Specically, this attempt to model teamwork highlights the centrality of ve core teamwork processes, including (1) team leader-
Figure 10. The Big Five teamwork model (Salas, Sims, & Burke, 2005).
ship, (2) team orientation, (3) mutual performance monitoring, (4) backup behavior, and (5) adaptability. Furthermore, the Big Five model also illustrates the importance of three ancillary team products and processes, including (1) shared mental models, (2) closedloop communication, and (3) mutual trust. Taken together, these eight constructs are dynamically related to one another and collectively form teamwork. At a broad level the Big Five model of teamwork proposed in Salas et al. (2005) adopts a similar approach to that of the above-described theory advanced by Dickinson and McIntyre (1997). However, the Big Five model proposes a set of constructs and construct interrelations that differ somewhat from Dickinson and McIntyres attempt to describe the dynamic nature of teamwork. For example,
Figure 11. ito model of team adaptation (Burke, Stagl, Salas, Pierce, & Kendall, under review).
214 modeling complex systems while Dickinson and McIntyres theory primarily highlights the centrality of skill competencies, that proposed by Salas et al. moves beyond modeling teamwork behavior to incorporate affective competencies (i.e., mutual trust), cognitions (i.e., shared mental models), and emergent team phenomena (i.e., adaptability). The 11th and nal programmatic effort undertaken to investigate team effectiveness reviewed here was advanced by Burke, Stagl, Salas, Pierce, and Kendall (2006). These researchers proposed a model of team adaptation within an ipo framework (see Figure 11). The model advanced by Burke et al. emphasizes the centrality of an adaptive process that unfolds over time to emerge as team adaptation. Specically, this applied research initiative denes team adaptation as an emergent phenomenon that coalesces over time from the unfolding of an adaptive process whereby one or more team members utilize their resources to functionally change current behaviors, cognitions, or attitudes to meet expected or unexpected demands. Essentially, team members draw from their individual and shared resources to detect, frame, and act on a set of cues that signal the need for functional change. As this adaptive process is carried out, feedback is generated that subsequently serves to revise shared cognition and adaptive input factors. Thus, the adaptive process is recursive by nature. The team-adaptation model illustrated in Figure 11 is the only initiative reviewed in this chapter that has radically departed from the traditional predominant focus on either task or team performance. Certainly, this state of affairs is indicative of an area ripe for exploration, and, therefore, it is our opinion that additional theory building should be undertaken to develop meaningful models of (1) task, (2) team, (3) contextual, and (4) adaptive performance. Perhaps one fruitful avenue for future exploration involves framing these unique but interrelated aspects of performance in a nomological network of lawful relations. For example, research could be carried out to develop a deeper understanding of the similarities and differences between team and contextual performance, which seem to share somewhat similar content and emphasis. We believe that the similarities and differences between contextual performance (see Borman & Motowidlo, 1993) and team performance warrant closer attention in future research endeavors. It may be particularly interesting to examine whether indexes of contextual performance
215 Team Effectiveness in Organizations capture a practically meaningful amount of unique variance in substantive criteria beyond that already explained by indexes of teammember performance.
217 Team Effectiveness in Organizations thermore, we drew heavily on what we considered intriguing constructs and ideas from the broader domain of team literature. The outcome of this inquiry is a preliminary integrative framework of team effectiveness (see Figure 12). In the remainder of this section we review some of the critical components of the framework while emphasizing how it expands prior programmatic efforts to understand team effectiveness. The multilevel integrative framework depicted in Figure 12 offers researchers and practitioners a simple but meaningful heuristic that illustrates some of the most important aspects of team performance. On a broad level, the framework highlights the role of team inputs (i.e., individual characteristics, team characteristics, task characteristics, work structure) in promoting teamwork and team performance. In addition, the framework illustrates the moderating role of individual-level cognition (i.e., expectations about roles and requirements) on the relations between team inputs and throughputs. Essentially, team inputs are actively interpreted by team members, who, via this ongoing process, form stable yet malleable expectations regarding the nature of their obligations. Team members with accurate expectations are more likely to know which team processes to engage in and when particular activities should occur. As the processes constituting teamwork are dynamically, simultaneously, and episodically enacted over time, they lead to shared cognition. Shared cognition (e.g., shared mental models, team-situation awareness, psychological safety) accrues as teamwork occurs and, in a recursive fashion, inuences subsequent teamwork activities. The framework also suggests that team performance results in individual- and team-performance outcomes. Performance outcomes produce system feedback that, over time, can serve to change both the organizational inputs and the teams inputs available to subsequent performance episodes. The organizational environment is characterized by a cue stream that is also interpreted by team members, thereby contributing to both individual and shared cognition. Furthermore, team leadership inuences, and is inuenced by, the accuracy of individual and shared cognition. In the remainder of this section, the various aspects of the framework are detailed in greater specicity. The integrative framework of team effectiveness illustrated in Figure 12 incorporates four categories of input factors, including individual characteristics, team characteristics, task characteristics,
218 modeling complex systems and work structure. Each of these four broad categories is an umbrella for multiple specic constructs. For example, the category of individual characteristics covers a wide range of phenomena, such as task ksas, motivation, team orientation, mental models, and personality. In turn, each of these constructs is related to team processes and team-performance outcomes to varying degrees. For example, the results of empirical research have repeatedly supported the importance of a team orientation (e.g., Driskell & Salas, 1992; Eby & Dobbins, 1997; Goodwin, OShea, Driskell, Salas, & Ardison, 2004). Likewise, research results suggest that personality variables such as extraversion, openness to experience, and adjustment are essential for coordinated performance (Driskell, Hogan, & Salas, 1987; LePine, 2003). The category of team characteristics also includes myriad constructs that may either have been formally established (e.g., power structure, performance arrangements) or have emerged upward from the unfolding of coordinated dyadic role exchanges of team members over time (e.g., team-level openness to experience, teamlevel team orientation). For example, empirical work has supported the relation between team cohesion and team performance (Dailey, 1980; Mullen & Cooper, 1994). Given the paucity of attention afforded team characteristics to date, additional research should be undertaken to investigate other critical factors such as team climate and culture. The task-characteristics category includes task organization, task type, and task complexity. Both theoretical research and empirical research suggest that task characteristics are, indeed, critical to fostering team effectiveness (Herold, 1978; McGrath, 1984; Steiner, 1972). Interestingly, the task-characteristics category reminds us that not all teams are alike and that task variations leading to different levels of interdependence can have a profound effect on how teams interact. These differences will be important to future research initiatives undertaken to develop team typespecic models and frameworks of team effectiveness. One implication of our emphasis on work structure is that social-network analysis (see Krackhardt & Brass, 1994) can be applied to analyze communication patterns and that the ndings of this process can be utilized to model, predict, and manage team-member and team performance. The fourth category of input variables, work structure, encom-
219 Team Effectiveness in Organizations passes work assignment, team norms, and communication structure. Work structure is critical to team performance because, as open systems, teams consist of both a formal work structure and a unique but interdependent social structure. Together, the socio and technical systems serve to shape both what cues team members attend to and how they react to those cues. Furthermore, work characteristics such as communication structure dictate who has access to what information and when, which is essential for negotiating what one wants to accomplish. Also, team norms have a considerable inuence on what behaviors are deemed appropriate (Hackman, 1990). It should be noted, however, that the particular constructs exemplied in the above descriptions were largely adopted from the programmatic theoretical framework advanced by Tannenbaum et al. (1992). This is important because, while there exists a general consensus about the nature of the broad categories of input variables, the specic constructs proposed to be encapsulated within these categories vary from research program to research program. For an equally impressive yet somewhat different set of team-input variables, see Campion and colleagues (Campion et al., 1993; Campion et al., 1996) metamodel of team effectiveness as reproduced in Figure 6 above. Our framework draws heavily on many of the recent advancements appearing in the body of team literature and perhaps most of all from cutting-edge research in cognitive psychology. For instance, while we frame many of the same inputs (i.e., individual characteristics, team characteristics, task characteristics, work structure) that are traditionally advanced within the team-effectiveness domain, our framework suggests a novel purpose for these components. The hybrid framework depicted in Figure 12 reects our belief that the myriad relations between input variables and teamwork processes are moderated by team-member cognition. Thus, team-member expectations, established and changed by team inputs, ongoing teamwork, team leadership, and the organizational context, moderate the relations between inputs and processes. Essentially, team members with accurate, exible expectations about their roles and requirements will be better positioned to engage in appropriate team-member and team processes at the optimal point in time. Shared cognition also plays a prominent role in the hybrid framework. Shared cognition such as shared and compatible mental mod-
220 modeling complex systems els and shared situation awareness serves to form templates that are drawn on by team members during teamwork. Shared templates or shared frames of reference imply processing objectives that constitute the social reality that team members share (Hinsz, Tindale, & Vollrath, 1997). Specically, this set of instantiated cognitive structures enhances teamwork processes because it provides team members with the insight required to understand (1) what they should and should not be doing, thinking, and feeling, (2) how to go about accomplishing stated objectives, and (3) when to enact ksaos in support of processes in order to meet tactical and strategic objectives. Also, as teamwork occurs, the organizational environment changes, and/or team leaders intervene, shared mental models and other forms of shared cognition are revised. In the absence of this cognitive reservoir, teamwork processes will still ensue; unfortunately, however, process losses will likely abound. In addition to having a distinctive cognitive avor, our framework also incorporates advancements made in modeling the dynamic nature of teamwork (Dickinson & McIntyre, 1997; Fleishman & Zaccaro, 1992; Marks et al., 2001; Shiett et al., 1985; Salas et al., 2005). Specically, our hybrid framework reects the cyclic, episodic nature of activity in high-performance teams by illustrating a sample of core teamwork processes in a cylinder. The point we are trying to communicate by framing teamwork processes with a cylinder is simple: that both team members and team-level competencies must be displayed in a dynamic, simultaneous, and episodic fashion in order to foster individual and team effectiveness. In fact, several revolutions may be needed to meet a single objective. Furthermore, as multiple objectives are interlaced or additional interrelated but unique projects undertaken, the teamwork processes displayed in a single revolution may not be homogenously directed at the accomplishment of a single goal. Our integrative framework of team effectiveness also acknowledges the centrality of team leadership throughout the life span of the team (see Marks et al., 2000; Stagl, Salas, & Burke, in press). While the construct of team leadership is a complex phenomenon with a variety of conceptualizations (e.g., functions, roles, traits, competencies, implicit perceptions, negotiated dyadic role exchanges) most would agree that team leaders and the leadership processes that they enact are essential to promoting team performance, adaptation,
221 Team Effectiveness in Organizations and effectiveness. In fact, recent meta-analytic evidence highlights the importance of team-leadership behaviors in achieving team outcomes (Burke et al., in press). It is through the leadership process that team leaders act to synchronize task and developmental cycles. Moreover, they institute the conditions that teams and their members draw on before, during, and after task episodes (Hackman, 2002). Consequently, team leaders act as major drivers and maintainers of team development, performance, and effectiveness over time. Another unique aspect of our integrative hybrid framework is the fact that it illustrates, albeit at a broad level, the interdependencies between team members competencies and team competencies. In our framework, both individual-level (e.g., interpersonal) and team-level (e.g., adaptive, coordination) processes are highlighted. Furthermore, the processes displayed in our framework are meant as a small sample of the number of competencies that must be uidly interlaced as team performance unfolds over time. Essentially, teamwork/team performance is composed of the dynamic display of (1) team-member taskwork ksas, (2) team-member teamwork ksas, and (3) team ksas. As the competencies constituting these three sets unfold over time, they serve to yield team performance. While prior programmatic efforts to frame teamwork have included both individual-level team members competencies and team-level competencies, the distinction between these ksa sets, as well as their symbiotic interdependencies, has seldom been explicitly considered. For further clarication of this issue, the inquisitive reader is directed to Campbell and Kuncel (2001), who provide a more complete discussion of this topic. As noted above, our framework illustrates both individual- and team-level processes; therefore, it is intuitive that it should also include both individual- and team-level performance outcomes. The individual- and team-level performance outcomes depicted in the model are meant to include both traditional output indexes such as the number of goods produced as well as emergent states (Marks et al., 2001) or psychosocial traits (Cohen & Bailey, 1997) that result from a teams experiences while navigating its operational challenges. These emergent states (e.g., cohesion, collective efcacy) are simultaneously proximal products of interdependent interaction and serve as subsequent inputs to new performance episodes. Incorporating multiple performance outcomes mandates the need for multiple ef-
222 modeling complex systems fectiveness evaluations. Thus, our framework acknowledges that the particular level of effectiveness achieved by a given team member or team will vary depending on the dimension(s) under consideration. Furthermore, the performance-outcome dimensions that are most relevant to the evaluation process will change as a function of time. In essence this suggests that teams and their members can be described in terms of effectiveness patterns. The degree to which variables constituting these patterns can compensate for one another when contributing to a global index of effectiveness that can subsequently be utilized by organizations for practical decision-making purposes remains an area ripe for exploration.
summary
The multilevel integrative framework described in this section offers researchers and practitioners a useful, yet simple, heuristic that can be called on to quickly understand the most important aspects of interdependent performance. The integrative framework illustrated in Figure 12 above is intended to highlight often-ignored issues in team-effectiveness research. Specically, our hybrid framework advances the science of teamwork by (1) adopting a cognitive approach to framing the dynamic and multilevel nature of team effectiveness from an ipo perspective, (2) illustrating some the interdependencies between team members competencies and team-level competencies, and (3) describing key antecedents to team performance and effectiveness. On further conceptual renement, we expect that our hybrid integrative framework will be useful in (1) dening a nomological net of teamwork, (2) formulating empirically testable propositions, and (3) designing interventions to facilitate human-capital management. With regard to this latter point, in the next section we review the practical implications of our hybrid framework for performance measurement, training, and stafng teams.
Practical Implications
During the last 25 years, hundreds of research studies have been guided by the team-effectiveness theories presented in this chapter.
223 Team Effectiveness in Organizations In some cases, the accumulated ndings from this research have been shaped into principles and guidelines (e.g., Salas, Burke, & CannonBowers, 2000; Salas, Burke, & Stagl, 2004). The delineated principles and guidelines offer essential insight to those concerned with designing, developing, and delivering human-capital-management interventions. Thus, for those individuals charged with the application of measurement tools, instructional strategies, and task-based simulations, it seems that, as we have seen, there is nothing as practical as a good theory. Keeping with this line of thinking, the use of theoretically grounded team-effectiveness models for measuring performance, conducting developmental interventions, and stafng teams will be addressed in this section.
224 modeling complex systems factors, including the purpose of measurement, the nature of incorporated stimuli, target ksa competencies, measurement timing, and anticipated costs (McIntyre & Salas, 1995). While measures can be designed to assess any part of the ipo system, capturing diagnostic information mandates measuring processes as raw materials are transformed into nished products and services as well as team-performance outcomes.
225 Team Effectiveness in Organizations must also be taken when designing interventions to help ensure that both horizontal transfer and vertical transfer ensue (Kozlowski, Brown, Weissbein, Cannon-Bowers, & Salas, 2000). Together, theories of team effectiveness and principles delineated from their application can be a powerful source of information for fostering team performance and effectiveness.
staffing teams
There is a growing awareness of the importance of managing human capital at multiple levels (see Campbell & Kuncel, 2001; Huselid, 1995; Ployhart & Schneider, 2005; Schneider, Smith, & Sipe, 2000; Stevens & Campion, 1999). Once dominated by a largely individualistic emphasis, private- and public-sector organizations now recognize that the maxim the people make the place (Schneider, 1987, p. 450) holds true at all levels in the conceptual space. Essentially, all teams must import energy in the form of human capital in order to avoid entropy and enjoy the prosperity that interdependent interaction can produce (see Katz & Kahn, 1978). Thus, the traditional concerns of person-job t and person-environment t must be balanced with equal attention directed at person-team t (Hollenbeck et al., 2002). Although, a few endeavors have already been undertaken to delineate guidelines for stafng teams (see Driskell et al., 1987; Jackson & Ruderman, 1996; Klimoski & Jones, 1995), additional research is required in order to more fully rene current processes as well as to develop new stafng strategies that simultaneously enhance both individual and team effectiveness. Following this line of thinking, researchers have begun to explore the importance of individual differences for both team members and team performance. Particularly noteworthy are those investigations that offer compelling theoretical rationales for how individual-level characteristics synergize and emerge as team-level phenomena (see Barrick, Stewart, Neubert, & Mount, 1998; LePine, 2003; Neuman & Wright, 1999). Specifying the mechanisms whereby individual-level characteristics emerge upward to collective phenomena within various team types characterized by varying levels of interdependencies is an issue warranting closer attention.
227 Team Effectiveness in Organizations Latin while cognitive psychologists are versed in the queens English? We believe the answer is twofold. Foremost, the current state of affairs characterizing team research nds the area largely devoid of collectively set, overarching short-, mid-, and long-term goals. Agreed-on tactical and strategic objectives must be articulated if the domain of team research intends to continue maturing in an integrative, productive fashion. After all, differentiation ultimately results in chaos without integration. A second reason for the rampant disjointedness is the lack of an integrative structure that, if present, could serve to house what is currently known, guide future endeavors, and be seamlessly updated by distributed stakeholders to reect accumulating ndings. For example, O*Net has radically reshaped the outmoded Dictionary of Occupational Titles. Perhaps it is time for T*Net to emerge. A failure to systematically encode and integrate is far from the only problem contributing to the frailty of research investigating teams. Unfortunately, team research is also, on occasion, methodologically weak. For example, far too often, in their haste to advance the eld, team researchers settle for single-source expediency at the expense of scientic rigor. Specically, increased use of triangulation methodology through the use of multiple measures is needed. Transparent Likert scales, subject to a host of contrived response protocols, continue to dominate the landscape of team members attitude assessment when seemingly superior measurement techniques such as Thurstone scaling and the conditional-reasoning approach are readily available. Is the extra effort and cost of developing a conditional-reasoning instrument not offset by the more accurate picture of a respondents standing on an underlying attribute that would be provided? Keeping with the above line of thinking, a similar issue warranting considerable attention concerns the measurement of team cognition (Salas & Fiore, 2004). Perhaps there is no single greater concern, as it is widely accepted that shared cognition is the foundation of the coordinated dyadic role exchanges that underlie much of team performance. However, for all the fanfare, there is little known about measuring cognition in teams (Cooke, Salas, Cannon-Bowers, & Stout, 2000; Cooke, Salas, Kiekel, & Bell, 2004). What is known suggests that team research would substantially benet from the renement of proved techniques for measuring shared cognition, such as
228 modeling complex systems card sorting, concept mapping, textual analysis, relatedness ratings, laddering interviews, and multidimensional scaling. Perhaps the situation is not so bleak or the outcome of theoretical chaos as of yet predetermined. However, the increasing diffusion without the proper structure in place to dene, guide, and integrate forthcoming advancements will, in our opinion, only contribute to the fragility that currently characterizes the teams domain. In order to curb the ongoing diffusion and other substantive problems noted above, we assert that, where appropriate, new initiatives should build on what is enduring from the past. Future endeavors should be informed by the insights afforded by the models outlined in this chapter. Furthermore, we argue for a return to basics because it is only by beginning anew that we can hope to set the substantive goals and structure that will thrust us forward in an organized, meaningful manner.
Concluding Comments
The complexity of contemporary work environments compels a recognition that it is no longer economically viable to navigate current challenges via an exclusive reliance on individual workers (West, Borrill, & Unsworth, 1998). In fact, an accelerating rate of change, driven by the ongoing technological revolution, is sweeping away the last remnants of a business landscape once dominated by an emphasis on colocated individuals (Priest, Stagl, Klein, & Salas, 2006). The lesson learned is clear: organizations as open systems are always growing in differentiation and integration. Thus, organizations must adapt their vision, structure, and human-capital practices in order to avoid entropy and enjoy prosperity in the 21st century. It is our hope that the research effort described in this chapter serves to further the understanding of teams, teamwork, team performance, and team effectiveness by providing common ground or a shared mental model of the science of teams. Furthermore, we eagerly anticipate the day when team researchers everywhere step up to the challenges laid at their feet herein and overcome their differences to collectively dene the future of team research.
Note
The views expressed in this work are those of the authors and do not necessarily reect ofcial army policy. This work was supported by funding from the Army Research Institute under sbir topic #osd02-cr01 and under baa #dasw01-04-r-0001. We would also like to acknowledge Dana Kendall for conducting a thorough literature review.
References
Aldag, R. J., & Fuller, S. R. (1993). Beyond asco: A reappraisal of the groupthink phenomenon and a new model of group decision processes. Psychological Bulletin, 113, 533552. Alper, S., Tjosvold, D., & Law, K. S. (2000). Conict management, efcacy, and performance in organizational teams. Personnel Psychology, 53, 3, 625642. Ancona, D. G., & Caldwell, D. F. (1992). Bridging the boundary: External activity and performance in organizational teams. Administrative Science Quarterly, 37, 634665. Ancona, D., & Chong, C.-L. (1999). Cycles and synchrony: The temporal role of context in team behavior. In R. Wageman (Ed.), Groups in Context, 2 (Research on Managing Groups and Teams, Vol. 2, pp. 3348). Amsterdam: Elsevier. Annett, J., & Cunningham, D. (2000). Analyzing command team skills. In J. M. Schraagen, S. F. Chipman, & V. L. Shalin (Eds.), Cognitive task analysis (pp. 401-415). Mahwah nj: Erlbaum. Argote, L., & McGrath, J. E. (1993). Group processes in organizations: Continuity and change. In C. L. Cooper & I. T. Robertson (Eds.), International review of industrial and organizational psychology (Vol. 8, pp. 333389). New York: Wiley. Arrow, H., McGrath, J. E., & Berdahl, J. L. (2000). Small groups as complex systems: Formation, coordination, development, and adaptation. Newbury Park ca: Sage. Avolio, B. J., Kahai, S., Dumdum, R., & Sivasubramaniam, N. (2001). Virtual teams: Implications for e-leadership and team development. In M. London (Ed.), How to evaluate others in organizations (337358). Mahwah nj: Erlbaum. Baldwin, T. T., & Bedell, M. D. (1997). The social fabric of a team-based M.B.A. program: Network effects on student satisfaction and performance. Academy of Management Journal, 6, 13691397. Balkundi, P., & Harrison, D. (2004, August). Networks, leaders, teams and time: Connections to viability and performance. Paper presented at the meeting of the Academy of Management, New Orleans. Barrick, M. R., Stewart, G. L., Neubert, M. J., & Mount, M. K. (1998). Relat-
In the early 1970s, when clinical psychology was undergoing its cognitive revolution, a young psychologist named Michael Mahoney became a voice for paradigmatic change. One of this volumes editors, then a graduate student, attended a debate between Mahoney and some of the old-guard gures of behaviorism. With a style and wit that became legendary, Mahoney showed a homemade lm to illustrate his position. The rst scene opened on a bleary-eyed graduate student, working at a typewriter, who paused to stagger to the refrigerator and quaff a sodauniquely human self-reinforcement? The scene faded to a dog pawing a typewriter. Suddenly the dog stopped, and the camera followed him to the refrigerator, which he opened to lap up one of several puppy treats. He then returned to his task at the typewriter. This was repeated until all the treats had been consumed. As the scene faded, the camera panned to the typewriter to reveal what the dog had been writing: I type, therefore I think I am. For the next 30 years, Mahoney delighted the psychology community with such demonstrations and parables. He became a leading gure in the constructivist movement, a later expression of his early inclinations toward integrative and holistic thinking. His work on chaos theory and related ideas in development and holistic psychotherapy led to his participation in this volume of the Nebraska Symposium on Motivation. A broad scholarly community was deeply saddened when he passed away on May 31, 2006. This volume of the Symposium is dedicated to his memory.
Change and stability are the twin engines of human consciousness. The story gets complex, but so, too, do humans. That story has fascinated many over the course of human history and the short span of our individual lives. Personally, human change processes have long been at the heart of my passions as a scientist. How do people change? What is it that changes when a person changes? Why do some of us seem to change more easily than others? Why is change so difcult for so many? And, ultimately, what should we do?
246 modeling complex systems I shall not presume to offer more than reverent gestures in the direction of responses to such questions. Before offering those gestures, however, let me explain the nontechnical tone of what follows. The growing interest in complexity is reassuring to those of us who have been following its scientic scent for some time. A common comment made by colleagues, students, and practitioners, however, is that the literature on complexity is too complex. They nd it difcult to understand. Some of the difculty may have to do with the subject matter, but some of it is also due to the use of technical terms that are unfamiliar. My goal is to use language and concepts that are familiar and accessible. This means, of course, that I will be simplifying, andas Alfred North Whitehead put itall simplication is oversimplication. Technicalities of theory and research evidence can be found among the resources cited in the reference list, a list that has been considerably winnowed. To better serve those whose primary interests may be more practical than theoretical or evidential, I will conclude with a brief discussion of how appreciations of dynamic complexity might inform professional life counseling. Let me begin, however, at the beginning.
247 Complexity and Human Change Processes the base concept of Western rationality. Thanks to Pythagoras and Plato, Western civilization has been steeped in three assumptions: that stability (order) is more permanent, real, and beautiful than change (disorder); that ideas, numbers, and the intellect are better sources of knowing than experiences, sensations, and the body; and that the mind and the body are separate. These three assumptions are being challenged by complexity studies.
248 modeling complex systems The fourth and most recent leap made by life on earth has been the creation and exchange of symbols. Instead of exchanging only their nuclear material, symbol-using creatures exchange information and expressions of their own experiencing. Symbol use occurs only among a few life-forms, all of whom are social. Sociality and symbolic processes seem to be intimately connected. All that we think of as modern is an expression of our accelerating exchange of symbols (e.g., civilized community, arts, science, technology). To summarize this brief history of biological evolution, life as we know it originated in contrasts. Boundaries continue to serve critically important roles in our lives. So do centers. We are life-forms with a legacy of coordinated centering, and we continue to seek and create order in our lives. We diversify and exchange, and it is through our diversity and connectedness that we continue to develop. We become complex, and we change at accelerated rates.
249 Complexity and Human Change Processes Darwins two greatest contributions were his emphasis on mutability (change) and emotionality. In mammalsand particularly in social primates like usthese dimensions are related. Motivation and emotion are based in movement. The words themselves reect this legacy. We are moved to changeput into literal motionby a complex interaction of processes. Our emotions reect our embodiment, and this realization has resulted in the erosion of mind-body dualism. The new look in the cognitive sciences and psychotherapy is one that vigorously embraces embodiment and emotionality. Body and mind can no longer be meaningfully separated. Reason and experience need not be cast only in terms of conict. And order and chaos are generative tensions, not separate species. Before elaborating much more of this story, it might be wise to recall an aphorism attributed to Arthur Schopenhauer. He suggested that there are two kinds of people in the world: those who believe that there are two kinds of people in the world and those who do not. This means, among other things, that we are categorical beings. We tend to classify. Just look at psychology. Why do we classify? Because it is a way of creating order out of chaos. Aristotles earliest work was on categories. He spent much of his life classifying (things, beings, areas of study, etc.). He formalized logic, and we use his logic to classify and organize our lives. How do we classify? By creating contrasts. Contrasts and the tensions they reect or create are also at the heart of our planetary life.
250 modeling complex systems millennia. A popular term for connecting these perspectives is constructivism. In the interest of focus, I shall forgo a digression into its rich legacy in Eastern and Western philosophy and science. In terms of name croppings, however, the following individuals have played roles in the history of constructivism: Lao-Tzu, Buddha, Heraclitus, Giambattista Vico, Immanuel Kant, Arthur Schopenhauer, Johann Herbart, William James, Hans Vahinger, Lev Vygotsky, Alfred Adler, Jean Piaget, Viktor Frankl, Friedrich Hayek, George Kelly, Jerome Bruner, and Albert Bandura. What is constructivism? Succinctly, constructivism is a term now commonly applied to a family of theoriesa metatheorythat share a view of human beings as active, complex, and connected lifelong learners. Throughout a fascinating diversity of expression, constructive theorists share ve basic themes: (1) the active agency of the human organism; (2) the importance of order or meaning in human experience; (3) the centrality of an embodied identityoften called a selfin the organization of experiencing; (4) the relational matrix of experiencing, most apparent in our immersion in social and symbolic networks; and (5) lifelong developmental unfolding. With such a breadth of themes, it is, perhaps, not surprising to nd that former theoretical factions have encountered one another on common conceptual ground. Thus, representatives of cognitive, behavioral, existential-humanistic, and psychodynamic perspectives have all used the term constructivist to describe some of their most recent developments. Given this remarkable conceptual concordance, a closer examination of the ve themes is warranted.
active agency
The active agency of the person is a key feature of constructivism. Humans are not simply reactive to the forces of their environments. We are proactive as well. We choose. We change our environments in ways that change us. Bandura has termed it reciprocal determinism. The emphasis is on reciprocity and the efcacy of an agent who is participating in life by inuencing the world by which it is being inuenced. This consideration of the person as an active agent is an important contrast to former portrayals that viewed humans as passive pawns in a mechanical universe.
ordering processes
The second theme of constructivism is that much of human activity is devoted to organizing experience. We are all order freaks, some more so than others. We seek to nd, establish, and maintain a stable order in and from which to live our lives. Recurring patterns in nature have helped. Within their rst year of life, for example, human infants organize their cycles of sleeping and being awake in congruence with cycles of light and dark (day and night). Biological self-organization is a demanding task of infancy. One of the most powerful expressions of this organization is apparent in early emotional development. Emotions are expressions of life organization that made their rst appearance with the emergence of mammals about 165 million years ago. Among other things, emotions have contributed to our exibility in adapting to new challenges and in deepening our capacities to relate to one another. Parental care of offspring is a common example. Most relevant here is the role of emotions in organizing our styles of being. As extensive evidence now reects, what we call personality and personality traits are fundamentally emotional patternsstylized expressions of our literally feeling our way through life.
253 Complexity and Human Change Processes There is an ironic paradox in human complexity. We long for order, yet we liveliterally and gurativelyin the essential tensions that mark the edges of chaos. Complex-systems scientists tell us that this is a necessary fact of our existence. This existential fact needs to be addressed with practical compassion. What does all this chaos and complexity mean for us in our everyday lives? Are we doomed to live hypertense, off-balance lives? I do not believe this to be the case. Permit me to translate my limited understanding of these matters in the direction of some practical suggestions for how we counsel professionally.
254 modeling complex systems A second-order change is often called a transformation or a personal revolution. In some instances people may actually change their name, their occupation, or their life partner and community. When a second-order change happens suddenly, it may be called a quantum shift. But suddenness is not common, and the distinction between kinds of change is more complex than I have here presented it. Perhaps a translation into clinical practice will help clarify that complexity.
255 Complexity and Human Change Processes requests for explanations are requests for order. This level of inquiry might be called a pattern level. Clients are searching for meaning. The quest for meaning is a natural, common, and healthy expression of our essence as order-seeking creatures. But this is whereI reluctantly confessmuch of psychologys theorizing has failed to serve constructively.2 To be optimally helpful in developmental processes, order seeking must reect and serve activity. Activity is at the heart of the third level of focus in psychotherapy. I call it the process level. We are, of course, always in process, even when we do not acknowledge or label that fact. Process-level work in psychotherapy tends to be more spontaneous, embodied, and emotional than the solution-focused planned interventions of problem solving. Work at the level of process tends to be experiential, exploratory, and experimental.3 Although formal exercises (in session and as homework) can add structure to these ventures, their stochastic movements cannot be completely anticipated or predicted. Process-level work can, therefore, be among the most challenging and rewarding (for both client and therapist). Such work appeals to a wisdom that cannot be localized in the left hemisphere of the neocortex. Note that I am using contrasts to structure (organize) my remarks: two kinds of people, two kinds of change, three levels of focus. Let me introduce another contrast that may further serve an understanding of how we construct orderings to cope with complexity.
256 modeling complex systems order that emerges in complex open systems. Teleology is direction (orderly movement) dened by a specic destination; teleonomy is direction (orderly movement) without a single or specic destination. The two most common examples of teleonomy are biological evolution and human personality development. Viewed backward historicallyboth reect a remarkable orderliness. Their patterned orders reect the operation of principled processes. But the orders reected in biological evolution and a persons life-span psychological development are different from planned, rational, teleological orders. An appreciation of that difference is critical to our being able to provide responsible professional counsel on human lives and their conduct (Mahoney, in press). Complex interactive systems exhibit complex expressions of chaos and both kinds of order (teleological and teleonomic). Early Greek political philosophers liked to distinguish between the laws of nature and laws designed by humans to govern their own conduct. We have inherited the burdens and blessings of the Greeks early love of rationality and its applications to the design of our social systems. There is, however, a signicant limit to what can be rationally designed and institutionalized without causing harm. This was one of the central insights of Bernard Mandeville, Adam Smith, Adam Ferguson, and David Hume. They recognized that there is a second kind of order that appears in human conduct. It is the result of human action, but not of rational design. I here call it ow to emphasize the contrast with mechanistic force. Flow reects the order that spontaneously emerges from within the ongoing activity of its constituent agents. Open, complex systems express endless, dynamic exchanges. They follow rules of order that can never be completely specied, predicted, or controlled. Hayek called this the primacy of the abstract (abstract implies tacit). He distinguished between the level of principles and the level of particulars. Concrete particulars are the consequences of multiple interacting forces in a widely distributed and ever-changing network. One can approximate a description of the principles by which complex systems operate, but one can never hope to perfectly predict the particulars. An example in psychotherapy is that of language production. Neither we nor our clients can accurately predict the exact words that may emerge from either of us in the coming moments. Those words are the particulars. But there are undeniable principles that constrain and organize what
257 Complexity and Human Change Processes does emerge (e.g., rules of grammar, the languages we speak, our current concerns). Process-level work in psychotherapy honors this second, more complex and teleonomic kind of order. It respects a kind of wisdom in the distributed yet connected systems that are the developing person. Such work reects an appreciation for evolutionary processes within the individual. Process-level work respects the operation of three essential processes in developmental change: variation, selection, and retention.4 All learningand, therefore, all developmentrequires variation or novelty. New forms are a necessary part of adaptive change (whether expressed as organisms, behaviors, or ideas). Variation (diversity) is key. Most new forms will not be viable. They will not be selected (naturally or otherwise) as keepers by the environments in which they are expressed. A few, however, will offer novel solutions to adaptive challenges. With a little luck these variants will be retained (through genetic coding, neural structuring, habitual use, social custom, epistemic authorization, etc.). Karl Popper and Don Campbellboth pioneers in evolutionary epistemologynoted that the rst and third processes in complex developing systems are in dialectical contrast. The variance generators live in essential tension with the order protectors. One family of processes is bent on liberal protection of degrees of freedom in the expression of new forms. This family of liberation is at odds with another, more conservative family. Retention processes strive to conserve life orderthe very order that is being challenged. We like to believe that we are being rational in our selection processes, but, as David Hume noted, our emotions are often far more powerful than our intellect would like to admit. What does all this mean, practically speaking? It means, among other things, that we have apparently internalizedliterally, incorporated (brought into our bodies)the very processes from which we (and other earthly life-forms) have emerged. Every person is a live-in developmental laboratory. Yes, we are bounded beings, and we are quite protective of our boundaries. Our insides are teeming with activity. Much of what is inside has become organized into systems and centralized structures. It is worth remembering Sechenovs discoverypublished in Darwins timethat the primary activity of the nervous system is inhibitory. We are teeming with excitatory impulses. Their control, voluntary and otherwise, is exercised pri-
258 modeling complex systems marily through inhibitory processes. We live as and within essential tensions (Mahoney & Mahoney, 2001). Level on level, we can see development as a dialectical (contrast-generated) development of lifelong gestures of balance. I believe that our gestures can be rened. Let me therefore turn to some practical strategies that can be useful in psychotherapy (Appendix B).
259 Complexity and Human Change Processes contract and expand. Centering is not just a preliminary emphasis; it is a recurrent skill to be revisited and rened throughout the course of our work together.
260 modeling complex systems movement (e.g., walkingif only in circlescan positively affect the course of therapeutic process). The standing-center exercise, which may last for only a few minutes, can also illustrate some lessons that are both literal (embodied/concrete) and gurative (metaphoric/abstract). I may ask clients to close their eyes while standing in order to focus their attention inward and on the sensations in their feet. Eye closure will also increase postural swaythe slight rocking motion by which the body continually nds and adjusts its balance. We are usually unaware of this movement because it is so basic and habitual. Becoming aware of it can serve to illustrate more general lessons about balance in life: e.g., we are most aware of our center (or balance) when we lose it; we are always moving (even when we are unaware of that fact); our body-brain processes intuitively know well how to nd their way home; and home (or center) is always there, even when we momentarily cant nd it.
261 Complexity and Human Change Processes stability, I emphasize the development and practice of orderly routines. For clients whose current needs are primarily to break out of a rutted order, I emphasize exploration with new possibilities. Do something different is a frequent homework assignment heard by my clients. Take a different route home, rearrange your closet, tune in to different channels on your radio or television. The only constraints are that these ventures into novelty be self-caring and socially responsible. The idea is to introduce more variance into their life order. With such individuals a constructively challenging interpersonal style can be balanced with the more comforting style required when they feel destabilized. From this dynamic systems perspective, psychotherapy is a special form of human relationship in which and from which clients can nd compassionate safety andeach when they are readybegin to explore alternate ways of experiencing themselves, their worlds, and their possible development. Exercises for exploring and experimenting with new skills and different possibilities for experiencing cover the entire spectrum of techniques employed in psychotherapy and other forms of developmental life counseling (Mahoney, 2003a). Among my favorites are personal journaling, movement meditation (which literally encourages clients to move toward new strengths, skills, and enjoyments), and mirror time. Mirror time is a self-relational exercise that asks the individual to spend brief regular periods of time in front of a mirror. A common format is to suggest alternate minutes with eyes open (looking in the mirror) and eyes closed (focusing internally on bodily sensations, thoughts, and feelings). The fundamental strategy is to encourage selfcompassion, but the process must be paced and structured according to individual starting points and learning styles. Not all clients will take kindly to the exercise or themselves, but laboratory, clinical, and eld research have suggested that self-relational skills are important components in life quality and well-being.
262 modeling complex systems forth through an ever-changing base of viability. The same is true of our interactions in relationships. Episodes of openness and connection may alternate with episodes of closure and distance. Part of the unique challenge of being a mental health professional is the demand to develop an operating center that is large and exible enough to safely hold all the destabilized centers being served. Psychotherapy can be a destabilizing experience for therapists, and this should not surprise us. We therapists are changed by our work. The life-span psychological development of psychotherapists is often accentuated and accelerated. Therapists should, therefore, be encouraged to prioritize self-care and to explore contexts and experiences that serve their own unique developmental path and pacing.
Concluding Remarks
Appropriately, I hope, I shall close with an opening. I began with a reference to my passions as a scientist fascinated with human change processes. Science and passion have been in a strained relationship for too long. The sources of the strain are less important here than are the means by which we might bring them back into creative balance. As a gesture in that direction, I shall conclude by taking a risk. One nds very little creative writing in the literature of scientic psychology, yet it is often through creative expressionsgestures toward novel experiencethat clients and therapists best learn their dance. I encourage my clients to write, draw, sing, and sculpt their hearts out. Occasionally, they share some of their private creations with me. One client wrote a poem expressing her feelings as we moved toward the conclusion of our work together. She asked if I would write something she could take home with her. The poem that emerged for her is what I now leave with you: Its a season of transition and youre on the move again On a path toward something you cannot disown; Searching for your being in the labyrinths of heart And sensing all the while youre not alone. Yes, you seem to keep on changing for the better and the worse
263 Complexity and Human Change Processes And you dream about the shrines youve yet to nd; And you recognize your longing as a blessing and a curse While you puzzle at the prisons of your mind. For as much as you seek freedom from your agonies and fears And as often as youve tried to see the light, There is still a trembling terror that your liberation nears As you struggle with the edges of your night. For your Reason is a skeptic and rejects what it desires, Playing hard to get with miracles and signs; Till a Witness gains momentum and emerges from within To disclose the patterns well above the lines. Then a window has been opened and youve let yourself observe How the fabric of your Being lies in wait; And you want to scream in anger and you want to cry for joy And you worry that it still may be too late. For the roller coaster plummets with a force that drives you sane As you tightly grasp for truths that will abide; Never fully understanding that your need to feel secure Is the very thing that keeps you on the ride. You survive the oscillations and begin to sense their role In a process whose direction is more clear And you marvel as your balance point becomes a frequent home, And your lifelong destination feels like here. So with gentleness and wonder, with questions and with quests You continue on the path that is your way; Knowing now that you have touched upon the shores of inner life, And excursions deeper cant be far away.
264 modeling complex systems There will be so many moments when an old view seems so strong And you question whether you can really change; And yet, from deep within you, theres a sense of more to come And your old view is the one that now seems strange. Take good care, my friend, and listen to the whispers of your heart As it beats its precious rhythm through your days; My warm thoughts and hopes are with you on your journeys through it all . . . And the paths of life in process nd their ways. Do be gentle, Process Pilgrim; learn to trust that trust is dear, And the same is true of laughter and of rest; Please remember that the living is a loving in itself, And the secret is to ever be in quest . . .
265 Complexity and Human Change Processes quickly. Such resistance reects basic self-protective processes that serve to maintain the coherence of the living system. 6. Ordering processes always operate in relation to their own contrasts, which are disordering processes. Order and disorder are facets of the same dynamic diamond. 7. The ongoing interaction of ordering and disordering processes creates a simultaneous interplay of both familiar and novel experiences. 8. Novelty is necessary for learning and development. Novelty involves contrast. 9. Familiarity and consistency are essential to life support, systemic coherence, and human well-being. 10. The dynamics between familiarity and novelty lie at the heart of human change processes. 11. Order is necessary for development. 12. All learning and development involve transitional disruptions or perturbations in systemic functioning. At all levels, development feeds on disruptions and then digests them into familiarity. 13. It is in the context of disorder that a living system exhibits both its greatest rigidity and its greatest variability. When rigidity reigns, stereotyped behavior or frozen passivity is common. Variability may emerge, particularly in a safe relationship that encourages ventures into exibility. From this exibility in being and these expressions of variability, more fullling and functional activities may emerge and vie for potential selection in the expression of that persons life. 14. When novel experiencescontrastsare decient relative to an individuals capacities and developmental needs, stagnation and hardening of the categories are likely. 15. When novel experiences greatly exceed the individuals capacities to balance, feelings of being overwhelmed are common. Episodic or chronic disorder and breakdown may result. 16. When new experiences are well timed and suited to the individuals current developmental capacities and edges, developmental breakthroughs may emerge and effect whole-system transformations in experiencing. 17. Although disorder may be experienced and expressed in highly patterned processes of human activity, it is diverse, individually unique, and systemic; we shall advance in our attempts at con-
266 modeling complex systems ceptualization and classication only as we are willing to embrace the limits of symbol systems to capture human uniqueness and the ultimate ineffability of complex system dynamics. 18. Persons can become trapped in disorder. Many psychiatric disorders are, in fact, rigidly ordered patterns. 19. Change is a nonlinear process. It is neither continuous nor cumulative. Rather, change processes reect many small shifts punctuated by occasional sudden leaps and frequent returns to earlier patterns of activity. 20. Change is often experienced in waves or multirhythmic oscillations. Anchors for dimensions of oscillation are often abstract conceptual polarities, such as life/death, right/wrong, good/bad, real/ fake, and sacred/profane. Descriptions of experiences of oscillation often include references to opening and closing, expansion and contraction, loosening and tightening, or approach and withdrawal. 21. Change emerges from a shifting matrix of competing possibilities. Old, tenured patterns of activity compete with new and experimental possibilities. Like all other forms of evolution and revolution, this competition is never nished. New patterns, when they gain dominance in the competition, themselves become the old and familiar order in contrast with which newer patterns emerge and compete. 22. Ongoing competitions in development are neither won nor lost in reference to allegedly absolute criteria. Some competitors (i.e., impulses of activity) are selected to assume temporary positions in the drivers seat of the body. The old patterns remain as contenders, and they may win occasional episodes of ascendancy in future situations. Old habits are not eliminated completely, but they can be effectively displaced by new ones. 23. Selection processes always include human agency. What decides the ongoing competition among activity patterns is a complex dynamic system that emerges out of and expresses a human will. 24. Selection processes must mature into retention processes if a chosen activity is to become an inuential pattern in ongoing self-organization. A changeto become a change that makes an enduring and, therefore, signicant differencemust be actively practiced. 25. Successful (adaptive) change is facilitated by a rhythmic orchestration of exploratory, selective, and perpetuating activities (i.e., experiments in living, evaluative choices regarding which experi-
267 Complexity and Human Change Processes ments are working, and action patterns that serve to perpetuate and elaborate valued experiences). 26. Self-relationships, which emerge in social and symbolic relationships, powerfully inuence life quality and resilience under stress. Awareness, acceptance, and celebration are common quests in self-relating. 27. Interpersonal relationships involving strong emotions are powerful contexts for psychological development. Safe and loving intimacy is an expression of trust, which lies at the core of optimal contexts for development. 28. Symbol systems including language and the arts may offer valuable structure and welcome stimulation in personal development. 29. Conscious practices, both spiritually and secularly pursued, are central to qualities and directions of life experiencing. Intention and action are key (even when the goal is stillness). 30. Psychotherapy should reect an appreciation of the history and motivational power of personal realities, the role of interpersonal, symbolic, and self-relational processes in the maintenance and change of personal realities, and the complex existential agency and responsibility of the socially embedded individual. 31. Love is basic to life. What is basic to life is basic to psychotherapy. Spiritual and wisdom traditions that embrace these insights may be precious sources of comfort, companionship, and direction in the complex processes of life-span personal development.
268 modeling complex systems 4. Let the clients set the pace, and honor their process. Some clients will want to move very quickly; some will not want to move at all. Pushing often results in pushing back or digging in. Expansions and contractions tend to alternate. Recognize and respect the importance of timing. 5. Encourage (but do not force) emotional expression. Allow your clients to feel, and to feel freely, but do not convey the message that they must be emotionally expressive if they are to make progress or win your respect or caring. When feelings emerge, invite elaborations and explorations. Seek to understand beyond familiar words and summary labels. Locate felt emotions in bodily sensations. 6. Allow and invite yourself to feel emotional in the process of counseling. Let your heart lead your helping. Open yourself to feeling. Learn your patterns around pain. Cultivate compassion, loving kindness, balance, and trust. Recognize that a primary responsibility of your role as a professional helper is to maintain a spirit of centeredness (balance) large enough to accommodate the combined energies of your client and your self. You will be challenged in your balancing abilities, particularly if you allow yourself to be emotionally alive to your interactions with clients. When you are overwhelmed (e.g., by intense emotions, whether your clients or your own), take a moment to breathe deeply and to refocus on your intentions to help the person you are serving. 7. Trust that your clients can endure their pain and be strengthened by the process. You cannot take their pain away, although you might often wish that you could. But you can hold a steady course of condence in their capacities. 8. Emphasize safety, and offer as much structure as your client needs. Use routines, rhythms, and rituals to create a sense of order and familiarity. Give your clients freedom to create their own path, yet be ready with suggestions and illustrations when they request direction or structure. 9. Afrm and encourage experimentation and exploration. Invite clients to experiment with safe and socially responsible excursions into new patterns of experiencingespecially new patterns of action, interpretation, explanation, or meaningmaking. Afrm the processes and feelings involved; express genuine respect for the challenges of changing, and offer generous encouragement of your clients active engagement with those challenges.
269 Complexity and Human Change Processes 10. Teach compassion, forgiveness, and self-care. Help your clients develop compassion for themselves and others. Encourage forgiveness. Teach self-care: set a good example.
Notes
These remarks are elaborated in Mahoney (1991, 2003a, 2003b, 2004). To preserve textual ow, additional resources are included in the reference list without being specically cited in the text. For their contributions to the development of the ideas here presented, the author is grateful to Albert Bandura, Donald T. Campbell, Vittorio F. Guidano, Friedrich A. Hayek, Thomas S. Kuhn, and Walter B. Weimer. 1. For further information, see www.constructivism123.com, the Web site of the Society for Constructivism in the Human Sciences, or the journal Constructivism in the Human Sciences. 2. Until recently, psychology and psychiatry have aided and abetted strong, culturally transmitted inclinations to name and blame. Emphases have been placed on pathologies, deciencies, and culpabilities rather than resources, skills, and appreciated capacities. Problems tend to be associated with pain, and pain ranks high as a motivator of human inquiry. Naming things sometimes helps reduce their frightening or puzzling aspects. Naming people associated with the creation of pain is often involved in a blaming process. We do it every day on both individual and collective levels. Some clients are well served by having a name for their problem and a sense of community in their struggles and suffering. Some clients are temporarily relieved to blame their parents, their partners, and their brains for the pain they are experiencing. Psychology is full of labels and theories that encourage naming and blaming. I have elsewhere elaborated my concerns about the categorical preoccupations of our profession. Most relevant here is the practical consequence that categorical names and no exit/no action explanations often disserve a client. When a diagnostic label results in a clients identifying with the problem, change may become more (rather than less) difcult. Identication with ones problem(s) results in an added challengeto change, one must change, not only the problem(s), but also ones core sense of identity. Likewise, when an explanation suggests that people are a certain problematic way because of their brain chemistry, personality, developmental history, or the like, clients may resign themselves to insight without action. Activity, please recall, is at the heart of our being alive. It is also at the heart of our potential capacities for constructive, developmental change. 3. Historically, mid-20th-century behavior therapists liked to distinguish themselves from the experiential humanistic-existential tradition, which they viewed as less scientic and experimental. However, both camps were and are radically empirical in William Jamess sense (Mahoney, n.d.)
References
Anderson, W. T. (1990). Reality isnt what it used to be. San Francisco: Harper & Row. Anderson, W. T. (Ed.). (1995). The truth about the truth: De-confusing and reconstructing the postmodern world. New York: Putnam. Anderson, W. T. (1997). The future of the self: Inventing the postmodern person. New York: Tarcher/Putnam. Arciero, G., & Guidano, V. F. (2000). Experience, explanation, and the quest for coherence. In R. A. Neimeyer & J. D. Raskin (Eds.), Constructions of disorder: Meaning-making frameworks for psychotherapy (pp. 91118). Washington dc: American Psychological Association. Bandura, A. (1997). Self-efcacy: The exercise of control. New York: Freeman. Bruner, J. (1990). Acts of meaning. Cambridge ma: Harvard University Press. Bruner, J. (2002). Making stories: Law, literature, life. New York: Farrar Straus Giroux. Collins, R. (1998). The sociology of philosophies: A global theory of intellectual change. Cambridge ma: Harvard University Press. Damasio, A. (1999). The feeling of what happens: Body and emotion in the making of consciousness. New York: Harcourt Brace. Ford, D. H. (1987). Humans as self-constructing living systems: A developmental perspective on behavior and personality. Hillsdale nj: Erlbaum. Freeman, W. (1995). Societies of brains. Hillsdale nj: Erlbaum. Gergen, K. J. (1991). The saturated self. New York: Basic. Gergen, K. J. (1994). Realities and relationships: Soundings in social construction. Cambridge ma: Harvard University Press. Gergen, K. J. (1999). An invitation to social construction. London: Sage. Goldberger, N., Tarule, J., Clinchy, B., & Belenky, M. (Eds.). (1996). Knowledge, difference, and power: Essays inspired by womens ways of knowing. New York: Basic. Guidano, V. F. (1987). Complexity of the self: A developmental approach to psychopathology and therapy. New York: Guilford. Guidano, V. F. (1991). The self in process. New York: Guilford. Hayek, F. A. (1952). The sensory order. Chicago: University of Chicago Press. Hayek, F. A. (1964). The theory of complex phenomena. In M. Bunge (Ed.), The critical approach to science and philosophy: Essays in honor of K. R. Popper (pp. 332349). New York: Free Press. Hayek, F. A. (1976). Law, legislation, and liberty: Vol. 2. The mirage of social justice. Chicago: University of Chicago Press.
In keeping with the theme of balancing theory, research, and modeling approaches with potential practical applications, this, the nal chapter of this volume is a brief overview of one application of modeling complex processes associated with the practice of rehabilitation. Several levels of analysis, molar to molecular, are addressed in contributions in this volume. Hence, the collection may in some senses be seen as quite eclectic. However, from the perspective of a multifaceted system like that seen in the world of practical clinical rehabilitation for injured and/or ill individuals, it may also be seen as representing complementary areas that together have a great deal to offer with respect to the design and delivery of rehabilitation and to rehabilitation-related research. To illustrate the point, a specic model developed in Nebraska, the Madonna modelmeant to model the quite molar complex process of clinical rehabilitationis presented. Developed as a framework to systematically design and deliver comprehensive, coordinated rehabilitation programs, it is also a model to support design, conduct, and analysis of contextually based, clinical rehabilitation research with an explicitly practical focus. While the specic model presented here is not derived from other sources, it is to be acknowledged immediately that the individually designated aspects of the model together constitute a set of
276 modeling complex systems foci that would probably be endorsed as relevant to the design and operation of comprehensive rehabilitation programs by most experienced rehabilitation providers and recipients.1 The individual aspects of the model are discussed only in broad outline below, along with brief excerpts drawn from the present collection illustrating the relevance of these areas of research and thought to research and practice in clinical rehabilitation. Most, if not all, facets of the rehabilitation model presented here are addressed in a typical rehabilitation treatment plan or program. The model (see Figure 1) is represented in a circumplex form, intended to emphasize the interdependence of the various facets in the design and delivery of individualized treatment programs for individuals requiring comprehensive rehabilitation. In practice, by carefully evaluating the relations that existor are necessary to establishamong these aspects in each unique clinical situation, interdisciplinary rehabilitation teams are most likely to work successfully with the range of stakeholders to create a plan that will optimize functional outcomes, minimize unanticipated complications, and maximize efciency. When these facets are addressed in isolation, outcomes are much more likely to be suboptimal. Although using the model to guide treatment planning naturally introduces some degree of complexityrelative to a simple critical path, for exampleit is crucial to recognize and respond accordingly to the linked, interdependent nature of these foci. Moreover, the pattern of relations between the aspects of the model, and the relative importance of addressing each aspect, differs among diagnostic categories and, indeed, across patients within the same diagnostic group, owing to differences in patient characteristics and life circumstances. Examination of these relations and incorporation of them in the initial planning process is an important ingredient in designing and delivering a coherent, maximally successful program. Invoking and using the model at the outset of treatment is also a useful vehicle for beginning the process of educating patient and family about the nature of their circumstances and some dimensions of their prospects in the coming months or years. The model can also serve as a coherent guide for research. A rehabilitation research agenda can be expressed as a study of the quantitative and qualitative relations between elements of the model and treatment outcomes. The distinctive focus of such research is the
Figure 1. The Madonna model: an integrative focus guiding treatment and research.
examination of the relative contributions on important functional outcomes of the many separately delivered, but interacting, interventions that a rehabilitation patient receives. For example, many studies have been conducted examining the utility of a range of physical-therapy evaluation and treatment techniques for addressing patients mobility outcomes. However, many patients receive comprehensive programs; they are treated by multiple clinicians and receive many different modalities. It is the interaction of all these interventions, along with their own organismic vitality, that produces the nal result. However, comparatively little is known in specics about how these independent variables interact to produce the outcomes of interest. Therefore, to maximize functional outcomes, minimize complications, and increase efciency, one needs to look at the combined, joint inuence of these many inuences. As we learn more, we will be able to furnish clinical treatment teams with the
278 modeling complex systems information they need to ensure inclusion of crucial ingredients in the treatment plan and omission of non-value-added components. Ideally, treatment and research endeavors are reciprocal elements of one integrated process.
patient/family engagement
This aspect of the model refers to application of processes and procedures to optimize patient and family involvement, motivation, and engagement in the rehabilitation process, without which the effects of the implementation of subsequent components will be suboptimal. From the psychological perspective, this segment of the model includes assessments and treatments designed to minimize trauma associated with injury or illness and to promote positive adaptations to associated challenges. Identication of patient and family strengths and resources early in the rehabilitation process is crucial to a successful outcome. Attention to this dimension includes management of a treatment settings psychosocial milieu to promote an environment communicating optimism and encouragement, inviting maximum independence for patients and families. Identifying efcient and reliable ways to discover and tap these strengths should be a major research focus. Themes in psychology currently associated with variants of positive psychology and strengths-based psychology (Buckingham & Clifton, 2001; Clifton & Nelson, 1992; Fredrickson, 2001; Lopez & Snyder, 2003; Seligman, 2002) are promising approaches to systematic and proactive development and application of strategies to optimize this aspect of the rehabilitation model. Michael Mahoneys chapter in this volume, with its emphasis on constructivismexplicitly recognizing human beings as actively complex, socially-embedded, and developmentally dynamic self-organizing systems (as the tagline on the cover of Constructivism in the Human Sciences describes his journals mission)offers many avenues to explore in the process of promoting and maintaining patient, family, and social-network engagement. These ideas are explored more fully elsewhere (Mahoney, 2003). Lajoies insights and ndings as presented in her chapter in this volume would seem to be as applicable to educating patients and families as to training professional personnel.
therapeutics
This element of the model refers to the various evaluation and treatment techniques available for use in rehabilitation programs. Instru-
280 modeling complex systems mentation and equipment, range and delivery of therapies, medications, and so on, all fall within the scope of this facet of the model. From a research perspective it represents a focus on improvement and application of specic forms of patient evaluation and treatment interventions. Clinicians and clinical teams identify promising new approaches or identify particular challenges; research is designed and trials implemented to rene and validate evolving resources and procedures. In addition to individual therapeutic modalities, other research foci should include determination of: Optimal timing of treatments following an accident or illness Optimal intensity of treatments Optimal duration of treatments Optimal mix of the interventions provided to a patient, including determination of the relative contributions of the constituent interventions.
All the Symposium speakers provided valuable insights in this area of focus. Each paper suggests innovative approaches that can serve to enrich clinicians repertoire of tools to enhance rehabilitation practice, whether in the psychiatric- or the physical-medicine realm.
technology
The technology component of the model places an emphasis on development and application of the rapid advances in engineering sciences and derivative technology. Progress in the elds of electrical engineering, mechanical engineering, industrial engineering, civil engineering, computer sciences, and other domains has tremendous promise with respect to helping individuals with disabilities successfully meet the challenges confronting them in their homes and communities. The rapidly developing area of biomedical engineering reects the increasing importance of technology in rehabilitation as well as in medicine generally. For example, computerized augmentative and alternative communication devices are available now to help individuals with speech, language, and/or cognitive impairments participate more effectively in the activities of daily life and life roles. Laser technology is being employed to help individuals with signicant motor impairments control their environments and
281 Editors Postscript participate more fully in the important domains of their lives. Socalled smart environments that recognize specic individuals are now being engineered and rened, via the application of electrical sensors, laser-sensitive surfaces, computers, wireless communication technology, global positioning technology, and so on. These environments are designed to be responsive to differences in individuals capacities (e.g., cognitive, physical, and communicative) in a given environment (e.g., home, school, workplace), automatically altering the environment in ways that best conform to an individuals unique strengths and needs, thereby increasing prospects for optimal participation in important life activities, enhancing productivity, and maximizing positive quality of life. Implementation of emerging technologies is an important emphasis in the work reported in the chapters by Salas et al., Lajoie, and Musen in particular.
living setting
The living-setting facet of the model reects the importance placed on individualized, discharge-destination-specic rehabilitation-program design. To foster independence and quality of life for many individuals following disabling illnesses or injuries, living settings need to be redesigned, incorporating structural modications and/or installation of assistive and alternative technologies. Increased research in this area will facilitate development of rehabilitation and engineering practices that are effective and cost-efcient and that ensure that patient reintegration into his or her living setting is optimized. This aspect of the model is not represented with marked emphasis in this years Symposium volume but should be addressed in any comprehensive research or practice efforts, reecting the salience of contextual factors in human experience and functioning. This would seem to be a very useful area to explore in a future Symposium.
participation
This element of the model emphasizes the signal importance of maintaining a rehabilitation focus on identication of effective and efcient ways of linking individuals and families to resources in
282 modeling complex systems their community, in the effort to optimize reintegration into home, school, vocational, and recreational settings: that is, life roles. This aspect reects the strong emphasis on activities and participation articulated in World Health Organization (2001). This dimension of the rehabilitation model is understood to include the individual and his or her family, social network, and community as well as the environment as a system, taken at the broadest of levels. Research in this area, though often appropriately focused on individual change and empowerment, also includes development, examination, and implementation of effective strategies to ensure that community resources are fully accessible as well as efforts targeting timely and adequate education of community members concerning strengths and needs of their disabled members, again with the overarching aim of promoting a successful resumption of signicant and productive life roles. Operationalization of this facet of the model underscores the importance of rehabilitation-oriented input into community-planning activities. Perspectives and traditions resonant with this component of the model include ecological psychology (e.g., the work of Roger Barker and his colleagues on the concept of behavior settings and the implications of how such settings are manned) and community psychology (e.g., Heller & Monahan, 1997; Rappaport, 1977; Sarason, 1988).
program evaluation
The program evaluation component of a rehabilitation model is crucial to maintenance of an evidence-based culture, providing timely informational feedback to team members and administrators concerning the outcome of the activities of the program/system. Information regarding improvements in patient functioning, number and pattern of short- and long-term complications associated with recovery processes, and relative efciency of service delivery is developed and made available on a regular basis. Typical measures include Patient outcomes; Program improvement; Program integrity (qa); Cost efciency; Quality of collaborations.
283 Editors Postscript Evolving team/program activities include the integration of new protocols as well as the successful implementation of relevant research ndings. Regular dissemination of this information with other organizational teams as well as with colleagues at other rehabilitation settings is an indispensable mechanism for improving rehabilitation practices.
285 Editors Postscript systems. The chapters by Neufeld, Salas et al., and Musen are all quite relevant to consult when contemplating this aspect of the decision-making process.
model summary
The model described here represents an approach to focused rehabilitation treatment and research that is probably compatible with a general rehabilitation philosophy that has been elaborated over the years in a number of settings and that is comprehensive and true to the multifaceted nature of the rehabilitation process. Using this model as a guide, one can work to build and elaborate a successful research program aligned with the essential elements of an effective rehabilitation program. Many of these dimensions are well addressed in the chapters in this volume. As indicated in the schematic of Figure 2, there are a number of elements involved in the rehabilitation process. Indeed, the ow chart presented here is an obvious oversimplication. However, one can see many areas of application for the thoughts presented in the chapters in this volume that can be brought to bear on various aspects of the rehabilitation process. Space does not allow description of these applications, but readers are urged to reect on the potential application of these ideas to rehabilitation and psychological practice in general as they read and reect on the wealth of stimulating ideas offered in this volume.
Note
1. The notion of comprehensive rehabilitation is meant to apply to clinical situations in which afictions and impairments are substantial, often involving multiple organismic systems. For example, rehabilitation following stroke frequently requires a comprehensive approach because multiple capacities are affected. At the other end of the spectrum, a bruised shoulder sustained in a softball game might require a few brief sessions of ultrasound therapy, along with a home exercise program, but would not require the use of assistive technology, computer-based clinical decision support, modication of the home environment, or a community-reintegration program.
References
Buckingham, M., & Clifton, D. O. (2001). Now, discover your strengths. New York: Free Press. Clifton, D. O., & Nelson, P. (1992). Soar with your strengths. New York: Dell. Fredrickson, B. L. (2001). The role of positive emotions in positive psychology: The broaden-and-build theory of positive emotions. American Psychologist, 56, 218226. Heller, K., & Monahan, J. (1997). Psychology and community change. Homewood il: Dorsey. Lopez, S. J., & Snyder, C. R. (Eds.). (2003). Positive psychological assessment: A handbook of models and measures. Washington dc: American Psychological Association. Mahoney, M. J. (2003). Constructive psychotherapy: A practical guide. New York: Guilford. Rappaport, J. (1977). Community psychology: Values, research, and action. New York: Holt, Rinehart & Winston. Sarason, S. B. (1988). The creation of settings and the future societies. Brookline ma: Brookline. (Original work published 1972) Seligman, M. E. P. (2002). Authentic happiness. New York: Free Press. World Health Organization. (2001). International classication of functioning, disability and health. Geneva.
Contributors
C. Shawn Burke is a research scientist at the University of Central Florida, Institute for Simulation and Training. Her expertise includes teams and their leadership, team adaptability, team training, measurement, evaluation, and team effectiveness. Dr. Burke has presented at 66 peer-reviewed conferences, published 33 articles in scientic journals and books related to the above topics, and assisted organizations in evaluating aviation-related team-training programs and reducing medical errors. She is currently investigating team adaptability and its corresponding measurement; issues related to multicultural team performance, leadership, and the training of such teams; and the impact of stress on team process and performance. Dr. Burke earned her doctorate in industrial/organizational psychology from George Mason University and is a member of the American Psychological Association, the Society for Industrial and Organizational Psychologists, and the Academy of Management. She serves as an ad hoc reviewer for Human Factors and Quality and Safety in Healthcare. Gerald F. Goodwin is a research psychologist at the U.S. Army Research Institute for Behavioral and Social Sciences, assigned to the Leader Development Research Unit (ldru). He received his ms and PhD in industrial/organizational psychology from Pennsylvania State Uni-
288 modeling complex systems versity. Dr. Goodwins current research focus is on leader and team effectiveness issues, particularly with regard to the joint, interagency, and multinational context. He also provides statistical and methodological support within ldru. He was previously employed at the American Institutes for Research, where his project work included test development, employment-litigation support with regard to statistical analysis, training evaluation, and performance modeling. He is a member of the Society for Industrial and Organizational Psychology, the American Psychological Association (apa), and apa Division 19 (Military Psychology). He currently serves as an ad hoc reviewer for the Journal of Applied Psychology and Human Performance and has recently joined the editorial board of Human Factors. Zeno Kupper received his PhD in clinical psychology from the University of Fribourg, Switzerland, in 1998. He was trained in clientcentered and cognitive-behavioral psychotherapy and has extensive clinical experience in the rehabilitation of patients with severe mental illness. He is a research and clinical psychologist at the University Hospital of Psychiatry, University of Bern, Switzerland. In his research he has been engaged in developing analytic techniques to identify and measure complex patterns in the course of mental disorders, for example, using variants of time-series analysis and cluster analysis to identify characteristic patterns of schizophrenia patients response to treatment and rehabilitation and modeling stable and unstable states in the course of psychotherapy. This kind of analytic technology aims to translate the insights of dynamic systems theory into clinical research. Susanne P. Lajoie received her doctorate from Stanford University in 1986. In addition to being the James McGill Professor she is chair of the Department of Educational and Counselling Psychology at McGill University. Dr. Lajoie has engaged in a wide array of innovative research and scholarly activities where she applies cognitive theories in the design of computer-based learning environments for classroom and real-world applications. She uses a cognitive approach to skill identication and applies her research to the design of computer-coached practice environments in the areas of science, statistics, and medicine. She has numerous publications, including two volumes on computers as cognitive tools published by Erlbaum.
289 Contributors Michael J. Mahoney was a pioneer in the cognitive-behavior-therapy movement in the 1970s, holding public debates with Joseph Wolpe and other classical behaviorists in the newly formed Association for the Advancement of Behavior Therapy. His analyses of the role of verbal representations in behavior acquisition and change were compelling demonstrations of the limitations of cognition-free behaviorism. His seminal 1974 Self-Control: Power to the Person (with Carl E. Thoresen) heralded the contemporary age of cognitively oriented, social-learning-theory-based therapies. As cognitive constructs became acceptable and then mainstream in behavior therapy, Mahoney moved on to even more complex models of the psychotherapy enterprise, derived from chaos and self-organizing systems theories. Most recently, Professor Mahoney was a leading thinker in conceptualizing behavior change from a systemic perspective. The motivation to change plays a key role in these conceptualizations, but with the systemic view that motivation is a complex interactive process that includes both organismic and environmental components. Mahoneys application of systemic models to psychotherapy was at the frontier of this area, an endeavor to systematize one of the most complex and difcult-to-quantify decision processes in the human repertoire. At the time of his Nebraska Symposium presentation, Dr. Mahoney was a professor at the University of North Texas. He was also distinguished consulting faculty at Saybrook Graduate School and Research Center, San Francisco. A recipient of many honors, Dr. Mahoney was a fellow of the American Association for the Advancement of Science, the American Psychological Association, and the World Academy of Art and Science. He was also the executive director of the journal Constructivism in the Human Sciences. A collection of his recent poems, Pilgrim in Process, was published in 2003 by Kinder Path. Mark Musen is professor of medicine (biomedical informatics) and computer science (by courtesy) at Stanford University, where he is head of the Stanford Center for Biomedical Informatics Research. Dr. Musen conducts research related to intelligent systems, the Semantic Web, reusable ontologies and knowledge representations, and biomedical decision support. In 1989, he received the Young Investigator Award for Research in Medical Knowledge Systems from the American Association of Medical Systems and Informatics. He received a
290 modeling complex systems Young Investigator Award from the National Science Foundation in 1992. In 2006, he received the Donald A. B. Lindberg Award for Innovation in Informatics from the American Medical Informatics Association. Dr. Musen sits on the editorial boards of several journals related to biomedical informatics and computer science. He is the coeditor of the Handbook of Medical Informatics (Springer, 1997) and the coeditor-in-chief of the journal Applied Ontology. Richard W. J. Neufeld is a professor in the Departments of Psychology and Psychiatry, University of Western Ontario, where he also is a core faculty member of the Program in Neuroscience. He has received the Joey and Toby Tannenbaum Schizophrenia-Research Distinguished Scientist Award, being the rst psychologist recipient, the Ontario Mental Health Foundation Senior Research Fellowship, and the University of Western Ontario Faculty of Social Science Research Professorship. Professor Neufeld is a past associate editor of the Canadian Journal of Behavioral Science and of Psychological Assessment. Professor Neufeld has authored or edited 7 books and journal special sections. His 155 publications and 21 technical reports have appeared in journals ranging from the Journal of Mathematical Psychology, the British Journal of Mathematical and Statistical Psychology, and the Psychological Review to the Journal of Abnormal Psychology, the Journal of Consulting and Clinical Psychology, and Psychological Assessment. Jeffrey Poland received an ma in clinical psychology from the Southern Connecticut State University in 1982 and a PhD in the philosophy of science from the Massachusetts Institute of Technology in 1983. He is the author of Physicalism: The Philosophical Foundations (Oxford University Press, 1994) and a coauthor (with William Spaulding and Mary Sullivan) of Treatment and Rehabilitation of Severe Mental Illness (Guilford, 2003). He has held academic positions at Colgate University and the University of NebraskaLincoln, and he currently teaches in the Department of History, Philosophy, and Social Science at the Rhode Island School of Design and in the Department of Psychology and the Science and Society Program at Brown University. Eduardo Salas is Trustee Chair and professor of psychology at the University of Central Florida. He also holds an appointment as program
291 Contributors director for Human Systems Integration Research Department at the Institute for Simulation and Training. Previously, he was a senior research psychologist and head of the Training Technology Development Branch of navair-Orlando for 15 years. During this period, Dr. Salas served as a principal investigator for numerous researchand-development programs focusing on teamwork, team training, advanced training technology, decision-making under stress, learning methodologies, and performance assessment. Dr. Salas has coauthored over 300 journal articles and book chapters and has co edited 15 books. He currently edits the annual series Advances in Human Performance and Cognitive Engineering Research for Elsevier. He is currently designing tools and techniques to minimize human error in aviation, law enforcement, and medical environments. Bill Shuart is the director of the Institute for Rehabilitation Science and Engineering at Madonna Rehabilitation Hospital in Lincoln, Nebraska. He received his PhD in clinical psychology from the University of Nebraska in 1985, where he retains a clinical afliation with the Department of Psychology. His research interests include modeling team decision making in rehabilitation, with an emphasis on identifying effective applications of information technology to improve medical rehabilitation evaluation and treatment services for children and adults with disabilities. Will Spaulding received his PhD in clinical psychology from the University of Arizona in 1976. He was a postdoctoral fellow in mental health research and teaching, under the mentorship of Rue L. Cromwell, in the Department of Psychiatry of the University of Rochester. After his postdoctoral work, he joined the faculty of the University of NebraskaLincolns Clinical Psychology Training Program, where he is currently a professor. In addition to training clinical psychologists, he conducts research on the nature of mental illness, clinical decision making, and the outcome of treatment and rehabilitation. He also practices as a clinical psychologist in rehabilitation- and recovery-oriented services for people with severe and disabling mental illness. He has coedited two previous volumes of the Nebraska Symposium on Motivation and is the author, with Mary Sullivan and Jeffrey Poland, of Treatment and Rehabilitation of Severe Mental Illness (Guilford, 2003).
292 modeling complex systems Kevin C. Stagl is an organizational consultant with Assessment Technologies Group and formerly served as a research scientist and research assistant at the University of Central Floridas Institute for Simulation and Training. Dr. Stagls research investigates team performance, leadership, development, and adaptation. The lessons learned and practices distilled from this effort have appeared in scholarly outlets such as the Journal of Applied Psychology, Leadership Quarterly, Organizational Frontiers Series, and International Review of Industrial and Organizational Psychology. Wolfgang Tschacher studied psychology at the University of Tbingen, where he received his PhD in 1990. He also received psychotherapy training in systemic therapy at the Institute of Family Therapy, Munich. After his habilitation in psychology, he received the Venia legendi in 1996 at the University of Bern, Switzerland, where he was awarded a professorship in 2002. He works at the University Hospital of Psychiatry and is currently head of the Department of Psychotherapy. His main interests are in empirical psychotherapy research and experimental psychopathology, with an emphasis on dynamic systems approaches and phenomena of cognitive self-organization.
Subject Index
Page numbers in italics refer to illustrations. active agency, 250 activity, 255, 257258, 269n2 adaptation, team, 212214, 213 Adler, Alfred, 250 algorithms, clinical, 165166 algorithms, numerical, 8, 73n2, 176 algorithms, problem-solving, 171 Amazon.com, 164 American Association for Articial Intelligence, 153 American Psychological Association, 3 analytic derivations, 8 anova, 2, 10, 12 apprenticeship, 130, 136138 Aristotle, 163, 246, 249 articial intelligence (ai): A. Newell and H. A. Simons research on, 151153; boom and winter of, 153154; in eld of cognitive science, 86; history of, 147; Mark A. Musens study of, xii
xiii; overview of, 145147; and performance models, 128129. See also computer-based learning environments (cbles); computer software; technology articial neural networks, 175 associative stage of learning, 156 attention skills, 259260 attitudinal competencies, 191 attractors, 9499, 95, 115 autonomous stage of learning, 156 avionics trouble-shooting, 126, 127 Bacon, Francis, 248 balance skills, 259260 balance theory, xxxxii Bandura, Albert, 250 base distributions, 4953, 5051, 58, 61, 6972 base rates, 3942, 68 Bateson, Gregory, 253 Baum, F. L.: The Wizard of Oz, 124125 Bayesian theory: on bold conjecture and falsiability, 8; current
Author Index
Page numbers in italics refer to illustrations. Acton, B., 190 Aguirre, G. K., 43 Aleong, P., 132 Alexander, P. A., 124, 125 Alley, W. E., 127 Almond, R. G., 129 Ancona, D., 186, 201 an der Heiden, U., 117 Anderson, J. A., 156 Anderson, J. R., 125, 130 Andreasen, N. C., 92 Ardison, S., 218 Arnold, M., 90 Arpaia, J. P., 94 Ashby, F. G., 5, 11, 18, 19, 27, 52, 56, 69 Averill, J. R., 45 Azevedo, R., 126 Bachant, J., 174 Bacon, Francis, 1 Bailey, D. E., 221
Baker, A. J., 5 Baker, T. B., 5 Balakrishnan, N., 11 Bamber, D., 60 Bandettini, P. A., 43 Bandura, A., 73n4, 123, 250 Banerjee, S., 105 Barab, S. A., 131 Barker, Roger, xiv, xv, xxxivxlii, 282 Barrick, M. R., 225 Batchelder, W. H., 9, 19, 40, 60, 68 Batsell, R. R., 10 Baum, F. L., 124 Baur, N., 102 Beard, R. L., 187, 206 Blair, J., 117 Belbin, R. M., 186 Bell, B., 227 Bell, B. F., 186 Bell, B. S., 187, 192 Benjamins, V. R., 179 Benn, K. D., 59 Berger, J. O., 19 Berners-Lee, T., 179 Best, C., 178