The closer the members of a group are embedded in networks of social relations between them, the better are the prospects that group solidarity can be attained and stabilized. This proposition is almost a truism for analysts of group solidarity. However, my recent dissertation research (Flache 1996, Flache and Macy 1996) demonstrates that often a dense network may undermine group solidarity. The reason is that particularly in a dense network individuals may have positive relations not only to those who contribute to the group interest but also to deviants who violate group solidarity. As a consequence, actors may be reluctant to impose social control for fear to loose relations with deviants. This possibility is important both for researchers and group organizers. Accordingly, the substantive research problem is to identify and test conditions under which network embeddedness fosters group solidarity. To tackle this problem, the proposed research will relax three restrictive assumptions of the original study, 1) perfect information, 2) unrestricted interaction networks and 3) social relations consisting exclusively of reward exchange. In this way it will be possible to address new factors that shape the relationship between network embeddedness and group solidarity. For example, testable hypotheses can be generated about the effects of subdivisions or hierarchy in an organization. The related methodological research problem arises from the comparison of ‘forward-looking’ (anticipating) and ‘backward-looking’ (learning) models of individual behavior in my previous experimental work (Flache 1996: ch.4-5). In that comparison some conditions for the relationship between networks and group solidarity were better predicted by a backward-looking micro foundation, other conditions were more accurately described by a forward-looking analysis. This led beyond the literature, which favors the ‘forward-looking’approach. In the proposed research we now aim to identify the precise mix of forward-looking and backward-looking micro assumptions that is appropriate for the analysis of a particular macro condition. For this purpose, we develop and compare formal decision models representing both approaches separately, and models that systematically integrate the approaches. The empirical part of the research relies in a first step on laboratory studies that test selected propositions under highly controlled conditions. Subsequently, available data sets about organizational networks will be used to test hypotheses robust in the laboratory. Empirical tests are applied for two purposes. First, we test substantive propositions about conditions that shape the relationship of network embeddedness and group solidarity. Second, we assess the relative predictive accuracy of different mixes of forward-looking and backward-looking decision making for the analysis of the macro conditions that shape success or failure of group solidarity.

Model development and hypothesis generation will mainly take place in the first two years of the research. The theoretical analysis departs from my previous exchange theoretical approach (Flache 1996: ch.1-3). In a first step, the extended model is applied to analyze reward and monitoring mechanisms separately. In a second and third step, the network mechanisms of reward exchange and monitoring are integrated successively. The theory of repeated games is used to represent strategically rational forward-looking actors. For backward-looking learners, previous computer simulation models of stochastic learning are adapted. These models may turn out to be too simple, however, when various network mechanisms are combined in the analysis. At that point, we select and adapt suitable learning models from the literature. These are in particular neural network models and genetical algorithms. Models integrating forward-looking and backward-looking approaches will be developed drawing on game theoretical models of rational learning and on simulation models of short term anticipation. In this research I will continue three lines of my present activities. This requires research visits within the Netherlands, to the U.S.A. and to Germany, respectively. I will collaborate with researchers in the NWO-Pionier program ‘The management of matches’ at the University of Utrecht to combine my previous game theoretical work with their recent analyses. With Michael W. Macy at Cornell University, I work together in extending my previous computer simulations of learning models. With Rainer Hegselmann at the University of Bayreuth, I will further develop recent analyses of group solidarity by means of cellular automata representations of social networks. Empirical work will start in the second year both with the development of experimental designs and with inspection of data sets of networks in organizations that are available within the ICS research context. These are in particular data on informal networks in organizations and data on so-called local exchange trade systems, LETS. The third year of the study will be mainly devoted to experimental research and analysis of field data sets. In the course of the research, results will be published in the form of articles in international scientific journals and, eventually, a book on ‘social networks and group solidarity’ is planned.

The closer the members of a group are embedded in networks of social relations between them, the better are the prospects that group solidarity can be attained and stabilized. This is almost a truism for analysts of group solidarity. For example, consider a neighborhood where members try to raise money to improve the public playing ground for their children. The literature overwhelmingly suggests that the prospects for success of the campaign are the better, the more people in this neighborhood meet each other, talk to each other, and have friends in the neighborhood (e.g. Opp and Gern 1993, 673; Gould 1993, 734-740; Marwell and Oliver 1993, 104). However, my own recent work questions this consensus. It showed both theoretically and empirically that often a dense network may undermine group solidarity rather than sustain it (Flache 1996, Flache and Macy 1996). In a nutshell, that analysis focuses at the desire of actors to obtain social rewards from other group members, e.g. their overtures of friendship. Previous work overlooked that this desire may often compromise actors’ willingness to exert social control. Particularly in a closely knit network of exchanges of social rewards, actors may be reluctant to sanction ‘deviants’ who fail to act in the group’s interest, because actors fear to loose rewards from those deviants. The possibility of this ‘double edge of networks’ clearly is important both for researchers and organizers of group solidarity. For example, firms often create ‘company towns’, to stimulate the emergence of networks between employees (Coleman 1994, p. 173). The double edge of networks suggests that it is important to identify the conditions under which such a measure may backfire against the organizations’ interest. However, the original study employs assumptions that are too limited for application to most situations of group solidarity. Accordingly, the proposed research will successively relax the restrictions of 1) perfect information, 2) unrestricted interaction networks and 3) social relations that consist exclusively of reward exchange. In this way new testable predictions will be generated about conditions under which group solidarity and network embeddedness come together. These predictions will then be confronted with empirical data.

Scientific background and theoretical approach. The key concept in our analysis is social control, because we assume that individual behavior is purposive and guided by self-interest – following the ‘individualistic perspective’ on group solidarity (e.g., Coleman 1990, Hechter 1987, Homans 1974, Olson 1965). This perspective points out that group solidarity often is precarious, because individual actors may deviate from group obligations whenever this furthers their self-interest. At the same time, group solidarity may thrive when social control between group members directs ‘selective social rewards’ to contributors and sanctions to deviants. We focus in our research on small and intermediate size groups, like small firms or villages, because here social control primarily relies on interaction in interpersonal networks rather than on central authorities. In our original analysis we assumed that network relations shape social control exclusively as conduits of social rewards. However, the literature points out that social control requires both monitoring of deviant behavior and selective rewards (e.g. Hechter 1984). Network relations sustain monitoring, because they allow people to keep an eye on one another and to gossip about others’ conduct. Flache (1996) could not take monitoring into account, because his analysis assumed that actors were always perfectly informed on others’ behavior, regardless of network contacts. Moreover, he ignored that interaction possibilities for reward exchange and communication often are restricted in a specific manner, as by subdivisions in an organization. We propose to move beyond the original research in assuming 1) imperfect information and 2) restricted interaction networks. In this way, the assumption can be relaxed that social relations consist exclusively of reward exchange and 3) the combined effect of reward and monitoring mechanisms can be systematically analyzed. This promises new insights, because the possibility of a double edge of networks implies that the two mechanisms may often generate conflicting implications for social control. More and denser network relations may make actors more reluctant to loose rewards from deviants. At the same time more and denser network relations deter free riding because they make deviations more visible. Obviously, a combination of the two mechanisms promises to yield new testable hypotheses about conditions under which a particular effect prevails. This leads beyond insights in the literature, where both mechanisms are assumed to imply that close embeddedness in social networks fosters group solidarity (for ‘reward arguments’ see, e.g., Coleman 1990, 269-282; Holländer 1990; Homans 1974, pp. 156; for monitoring arguments see, e.g., Raub and Weesie 1990; Ben Porath 1980, 6). Consequently, this literature has never looked at conditions under which negative effects of network embeddedness on group solidarity might dominate.

The substantive research problem. The substantive research problem of this study is to identify conditions under which network embeddedness fosters group solidarity. To address the problem, our formal models of network effects in group solidarity will be successively extended. In this process we collaborate with researchers in the NWO-Pionier program ‘The Management of Matches’ at the University of Utrecht, to integrate our own work on reward effects with their recent research about monitoring in networks (Raub and Weesie 1990, Buskens 1995,1997a). Testable hypotheses will be generated for particular network features. Previously addressed in this area are network density (e.g. Raub and Weesie 1990), network centralization (e.g. Marwell and Oliver 1993, ch.5), or, on the individual level, outdegree of an actor (Buskens 1997b). ‘Structural holes’ (Burt 1992) will also be addressed, because these may constitute strategically critical positions in reward exchange and communication processes in a group. Selected hypotheses will be confronted with empirical data both of experiments and field studies. Further below, we describe in more detail the theoretical and empirical approach of our research. Before this, we discuss the methodological research problem of the study.

The methodological research problem. Our methodological research interest focuses on the comparison and integration of different theoretical micro foundations. More in particular, we compare three approaches to model individual decision making, the ‘forward-looking model’, the ‘backward-looking model’ and an integration of these approaches, respectively. The forward-looking model is the micro economists’ rational man who anticipates consequences of his actions and, accordingly, chooses the action alternative that will maximize his self-interest (e.g. Coleman 1990, 13-19). The backward-looking model instead assumes that individuals optimize adaptively, driven by success and failure in the past (e.g. Macy 1993) or by imitation of successful role models (Heckathorn 1996). The ‘integration model’, finally, assumes that learning and anticipation are related elements of decision making. Until recently the forward-looking model was seen as the appropriate micro heuristic for the analysis of macro phenomena in sociology and economics (Wippler and Lindenberg 1987, Coleman 1987). However, it has been argued that learning mechanisms at the micro level need to be taken into account to appropriately understand collective phenomena (e.g. Flache and Hegselmann 1998, Flache 1996, Roth and Erev 1995, Macy and Flache 1995, Macy 1993, Winter 1987). To give an example, empirical analyses of buyer-seller relations show that trust between business partners may be established under a wider range of conditions than predicted by a straightforward rational anticipation model, because actors learn from positive experience with previous transactions (e.g. Batenburg et al. 1997, Buskens 1997b, see also Raub 1997). A second example derives from my own experimental work (Flache 1996, ch. 5). I found that groups achieved higher production of a common good, when restricted information facilitated coordination of individual actions by ‘try and error’. This was predicted by a learning model, but not by a corresponding forward-looking model. At the same time, both in the analysis of buyer-seller relations and in my experiments, a number of effects were described correctly by the forward-looking model. These studies show that both micro assumptions of learning and of anticipation may be needed. However, there is little insight into the precise mix of backward-looking and forward-looking assumptions that is useful to predict effects of particular macro conditions. Accordingly, our methodological research problem is to identify conditions - such as the presence of coordination problems in common good production - under which a particular micro foundation is suitable to predict factors that shape the relationship between network embeddedness and group solidarity. To tackle the problem, we develop comparable formal models drawing on the different micro foundations. We then apply the models to generate and test competing hypotheses for our substantive research problem.

In the following, we give a broad outline of model building and the empirical approach. Specific details of the research methods are described in the toelichting bij 19.

Model building. Formal models will be developed to analyze the joint effect of various conditions on the relationship between network features and group solidarity. We inspect conditions that shape actors’ interest in peer approval, such as the feasibility of alternative sources of social rewards outside a group, because these conditions affect reward mechanisms. Furthermore, monitoring arguments lead us to address conditions that shape the amount of ‘noise’, i.e. the degree to which group members lack information on others’ contribution to the groups’ interest. This is mainly realized through the introduction of a so-called interaction network. Broadly, this network describes ‘who can meet with whom’ in the group, which is an important argument in the literature on monitoring. In organizations, for example, the interaction network often is influenced by subdivisions and hierarchical structures (Lazega and Van Duin 1997, Han 1996). The interaction network shapes monitoring, because its structure affects how fast information on observed deviations spreads in a group. We model group solidarity in terms of a repeated exchange process, in which individuals decide both whether to make contributions to a collective good and whether to give selective rewards to peers (Flache 1996:1-3). Only actors who are related in the interaction network can exchange rewards. Imperfect information is introduced by the assumption that members learn only irregularly or unreliably whether others contributed to the group effort, unless they are related in the interaction network. Finally, the model assumes that knowledge of observed deviations spreads through communication in network relations. Only actors related in the interaction network can communicate. For simplicity, we assume in the initial phases of model building that information spreads as byproduct of network contacts (cf Raub and Weesie 1990). Later we take into account that actors may use information strategically. For example, actors may offer deviants to deliberately conceal observed defections. Similarly, they may threaten to spread lies that others defected (cf. Blumenberg 1997, 207-211). To analyze ‘forward-looking’ individual decision making we use the analysis of conditionally cooperative equilibria in repeated games (Friedman 1971). This approach incorporates both our own game theoretical models of reward effects (Flache 1996, ch. 3) and the analyses of monitoring effects that we adapt from Buskens (1995,1997a) and Raub and Weesie (1990). To analyze learning behavior, we start out with a simple reinforcement model of reward mechanisms from Flache and Macy (1996). When necessary, more sophisticated learning models will be selected from the increasing body of social science applications of formal learning models (for an overview, see Bainbridge et al. 1994). Approaches to integrate anticipation and learning, finally, will be adapted from game theoretical analyses of ‘rational learning’ and of short term anticipation on basis of experience (see below for more details).

Illustrative hypotheses. To give a flavor of the results we aim to obtain, we present illustrative hypotheses. Previous research suggests that centralization of informal network relations is conducive for group solidarity. However, a theoretical combination of monitoring and reward mechanisms may reveal conditions under which this positive effect of centralization turns into a negative one. To explain, the positive effect derives from the faster information flow in centralized networks (Yamaguchi 1994). The more centralized networks are, the more readily deviations may be sanctioned, because ‘free riders’ become soon known in the group. But high dependence on peer rewards may revert this mechanism. The more dependent actors are on peer rewards, the more they may refrain from sanctions for fear to loose future social rewards from deviants in the group. Network centralization then exacerbates the problem. In a centralized network those actors who are in central positions are also those who are very vulnerable to the loss of social rewards from deviants, because they interact particularly frequently with potential deviants. However, it considerably reduces the punishment a deviant may face, when particularly central actors do not sanction. Hence, the larger actors’ dependence on peer approval, the more centralization of a network may turn from an asset for group solidarity into a liability. There is also a related potential for the assumption that actors use information strategically. This assumption may imply that central positions in a network are particularly susceptible to collusion with deviants when actors highly depend on peers’ approval. Collusion then consists of a mutually profitable exchange between peripheral deviants and a central actor, where the central actor conceals observed deviations from the rest of the group and the deviant ‘bribes’ the central actor with social rewards. Empirical collusion phenomena have been described by organizational sociologists (Crozier 1963, 52-56). Corresponding game theoretical models have been provided in the literature on the so-called ‘principal-agent’ relation in organizations (Tirole 1986). However, hitherto there is no systematic application of this collusion argument to group solidarity. We aim to contribute to such an application.

Empirical research. We proceed in two steps to test selected hypotheses derived from the models. In the first step, we adopt our previous experimental paradigm (Flache 1996, ch.5). In this experiment subjects trade two exchange commodities, their contribution to a ‘group good’ and selective ‘social’ rewards. The approach will be elaborated to incorporate the theoretical extensions of 1) imperfect information, 2) monitoring possibilities, and 3) a restricted network of interaction opportunities. To test hypotheses, we manipulate between groups conditions like the centralization of the interaction network, and the unit value of the exchange commodity ‘social reward’. We use hierarchical linear models to test model hypotheses about effects of these manipulations on the macro outcomes ‘group solidarity’, ‘reward network’ and ‘communication network’. Recently developed ‘actor oriented’ statistical models (Snijders 1996) will also be employed to compare micro assumptions. In the second step, we use field evidence to further test hypotheses that were confirmed by the experiments. For this purpose we will closely inspect two suitable data sets that are available within the ICS research context at the time of the study. One dataset comprises information on the dynamics of networks between members of three different organizations together with information on the their ‘contribution’ to the organizational good. The second data set contains information on the participation of individual members of LETS (Local exchange trade systems), as well as their position in the exchange network.

The following consists of two parts. First, the work plan of the study is described in detail. Second, specific details of the research methods are elaborated.

1. Work plan:

In the first two months of the study, substantive literature on networks in group solidarity will be analyzed to 1) identify relevant conditions shaping reward mechanisms and monitoring mechanisms 2) to collect empirical information on characteristics of interaction networks in processes of group solidarity and 3) to collect empirical information on sources of imperfect information in groups.

In the entire following year models will be developed and applied for hypothesis generation. The first half year will be invested for the first phase of model building. Separate baseline models of reward and monitoring mechanisms will be developed. In the first two months of this time we use the knowledge acquired in the previous phase to integrate restricted interaction in the original reward exchange models. Subsequently, a research visit of two months at the Pionier program in Utrecht will be used to develop corresponding models for monitoring effects under imperfect information. The last two months are devoted to development of model versions that integrate the forward-looking and backward-looking decision models derived before. The next half year focuses on the second phase of model building, the combination of monitoring and reward mechanisms. Two months each will be devoted to modeling and analysis of forward-looking, backward-looking and models that combine the mechanisms of monitoring and reward, respectively. Subsequently, we again aim to integrate the two micro foundations. Related to this, further exchange with the Pioneer Program takes place in order to join our work with Pioneer’s effort to integrate forward-looking and backward-looking decision models in the analysis of reputation effects in two party relations (Buskens 1998). Accordingly, another month will be spent at Utrecht University.

Subsequently, four months are invested to analyze implications of more complex learning mechanisms drawn from the literature. In particular, my own work with Michael W. Macy at Cornell University will be continued with a visit of three months. In this time we further develop our present neural network models of clique formation in group solidarity (see Kitts et al., submitted). I will extend these models so that also imperfect information and restrictions of the interaction network can be taken into account. In this way, results of the previous analyses in the present research will be compared to predictions of the neural network approach.

The model building phase will be accomplished with a third phase of six months in which models are developed that assume strategic information exchange. Furthermore, my previous computer simulations with cellular automata will be adapted to study network effects in a spatial context. For this purpose, two months of this period will be spent at the University of Bayreuth. Here I will continue my work with Rainer Hegselmann to incorporate both monitoring and reward exchange mechanisms in cellular automata representations of interaction networks.

Simultaneously with model building, the last 6 months of the second year are used to prepare empirical research. In this phase the experimental approach of my dissertation will be adapted and fine tuned in pilot studies. For this purpose, the computer network software used for the original experiments is adapted to the new experimental design. Standard experimental software available at that time will be inspected for its suitability to implement the experiment. When necessary, a new implementation of the system with state of the art software tools is carried out. This period will also be used to inspect the two field data sets and search for other potential sources of field data. In particular, it will be discussed with the researchers responsible for the data collection how exactly network variables and variables about individual contribution to group interests are operationalized. Furthermore, it will be clarified when and how access can be given to the data of interest.

The last year of the research focuses on three main activities. First, experiments will be carried out and analyzed. This requires recruitment of subjects, experimental sessions and analysis of data. When necessary, statistical models will be adapted for the analysis of micro level data. The second activity is analysis of field data. This analysis will draw on the statistical techniques developed for the experiments. The third activity is preparation of a book that presents results of the study. This book will draw on a series of articles that will have been prepared in the preceding phases of the research. In particular, it is planned that the research visits in Utrecht, Cornell and Bayreuth each result in a theoretical article in a scientific journal. Further articles are planned about confrontation of hypotheses with empirical data.

Modeling imperfect information. The game theoretical literature suggests two approaches for modeling imperfect information. First, a number of game theoretical analyses employ noisy observables (for a brief overview see Kreps 1990, ch. 14). In this approach, actors observe each others’ contribution after every iteration of the repeated game, but due to ‘idiosyncratic’ random errors observations may differ from the true contribution. Second, in the line of Raub and Weesie (1990) a number of studies assume that actors have no knowledge at all on others’ behavior, except for network contacts that provide direct observations or communication about third parties’ actions in the past. To adapt this approach to collective good situations, we assume that actors have even without network contacts at least some information on others’ contributions, for example in terms of a random subsample of contributions that they observe every iteration.

Construction of interaction networks. To derive predictions about effects of characteristics of interaction networks, we draw on simulation techniques from Yamaguchi (1994) and Buskens (1997a). Simulation will be used to generate a sample of interaction networks with sufficient variation in the independent variables of interest (e.g. density, centralization, actors’ outdegree). Our models will then be applied to every network in the sample, in order to compute predictions for the dependent variables (contribution to the collective good, network of reward exchanges, network of communication). Finally, this artificially created data set will be analyzed with standard regression techniques to derive hypothesized relationships between variables.

Modeling forward-looking actors. To model conditional cooperation in repeated games we apply so-called trigger strategy equilibria (Friedman 1971). In such an equilibrium every member of a group ‘cooperates’ perpetually, because he knows that otherwise every other member will eternally ‘punish’ the deviant as soon as he learns of the deviation. For the present research we adapt the trigger strategy analysis of my dissertation (Flache 1996, ch. 3). Following that analysis we consider three possible forms of cooperation, ‘only contribute’, ‘contribute and reward every other member’, and ‘only reward every other member’. Predictions for group solidarity and reward network are then derived from the restrictiveness of the conditions under which forward-looking actors comply with the respective behavior. Technically, these are the conditions under which the behavior is supported by a subgame perfect equilibrium (Selten 1975). The assumption of restricted exchanges enters into the analysis in terms of the incentive of actors to participate in reward exchange. Broadly, the more relations an actor has, the larger the long term losses from being sanctioned, but the larger are also the short term gain from unilateral deviation from the exchange. The assumption of restricted observation and communication enters into the analysis in terms of the duration actors face between a unilateral deviation and the ‘detection’ and punishment by others. The longer this duration, the larger the incentive to reap short term benefits of deviation. To assess this effect, we draw on analytical and computer simulation techniques recently developed to analyze how fast information spreads in a particular network structure (see Buskens 1995,1997a). For the analysis of strategic use of information, finally, the criterion of subgame perfectness (Selten 1975) will be particularly relevant. In this phase of the analysis we will address the question whether in a trigger strategy equilibrium the threat to spread information about others’ deviation is always credible. We expect to find that this is not the case. For example, in a network position where a potential deviant can be observed only by one other group member the observer may prefer to withhold his knowledge, in order to avoid a series of mutual punishments in the group. In terms of subgame perfectness, the observer enters a subgame in which he faces an incentive to deviate from his equilibrium strategy and to not carry out the threat to spread information.

Modeling backward-looking actors: stochastic learning and beyond. The computer simulations of Flache and Macy (1996) implement reinforcement learning in terms of a Bush-Mosteller stochastic learning model (Bush and Mosteller 1955). This model can be straightforwardly adapted for the assumption of restricted reward exchanges. Additional assumptions need to be specified for imperfect information, because the original model assumes that actors receive stimuli in terms of information on others’ contribution after each iteration. Model behavior will be analyzed by computer simulation and, where possible, with analytical results from the theory of ‘learning in games’ (e.g. Fudenberg and Levine 1998). The model is primarily designed for simple decision problems with few alternatives. More sophisticated learning models will be required in the third phase of model building. A possible approach are so-called „attractor neural networks„. In an attractor neural network an optimization rule is applied so that individuals simultaneously adapt various actions and relations between them in order to minimize „tension„ (for a similar approach, see Zeggelink et al. 1996). Kitts, Macy and Flache (1998) combined this approach with our previous work to simulate clique formation in group solidarity. A second possible approach are genetical algorithms. This approach models how in a decision making process new random combinations of behavioral strategies are optimized by ‘survival of the fittest’ (e.g. Axelrod 1987, for a similar approach with so-called finite automata see Probst 1996).

Integrating forward-looking and backward-looking decision making. The literature offers two approaches we will inspect. First, game theoretical ‘rational learning’ models assume that actors take strategic decisions under incomplete information, i.e. they lack knowledge about the exact preferences or beliefs of their opponents. In the course of interactions, actors then observe others’ actions and adapt both their own behavior and their beliefs about others accordingly. ‘Rational learning’ models analyze the properties of the equilibrium strategies of this process. In such an equilibrium, actors’ beliefs about one another and their actions are mutually consistent (e.g. Friedman 1986, 129-136; Fudenberg and Tirole 1991; for a similar approach in a bargaining context, see Stokman and Stokman 1995). The second approach of ‘short term anticipation’ assumes that actors anticipate consequences of their actions in the near future on basis of experience with the effects of their actions in the past. This is for example the case in so-called ‘mutual adjustment’ processes, where players simultaneously respond optimally to each others’ actions in the past (e.g. Fudenberg and Tirole 1991, 23-29; Fudenberg and Levine 1998).

Experiments. The experimental approach of my dissertation research will be adapted as follows. A network software will be developed through which participants anonymously interact in the course of the experiment. The software guides participants to simultaneously take decisions about their contributions and the rewards they give to others. The software imposes the interaction network on the group and restricts information flow and reward exchange possibilities accordingly. To test hypotheses of the first two phases of modeling, a manipulation is to be elaborated that spreads knowledge about observation of others’ actions as a byproduct of ‘interaction’. We expect that the number of members per group needs to be at least twice as large as in the original experiment (N=5), because otherwise it is not possible to attain sufficient variation in network structures.

Field data. The first data set consists of three samples that have been collected in a hospital, a computer firm and a paper factory, respectively (joint research of G. van de Bunt, H. Hangyi and R. Wittek, ICS Groningen). Employees were interviewed in a longitudinal design about social relations at the workplace. In two organizations, information on the interaction network was collected in terms of data about bilateral task interdependence. In addition, items were measured that allow to assess individuals’ interest in organizational success (e.g. formal training), their contribution to organizational success (e.g. colleagues’ assessment of willingness to cooperate). The second data set (Local Exchange Trading Systems) will be collected in a NWO financed project about ‘The future of community’ (coordinated by A. Nieboer, ICS Groningen). Here, items will be measured that allow to assess individuals’ positions in the exchange network and their contribution to the LETS organization as a whole. In addition to these two datasets, we will inspect the extensive data base that meanwhile has been accumulated in the NWO Pionier program in Utrecht on factors shaping cooperation in two party relations (e.g. Raub 1997).

Data analysis. We use ‘hierarchical linear models’ (Bryk and Raudenbush 1992) because this approach takes into account the interdependence of subjects within one and the same group in a between group experimental design (for some background see Flache 1996, ch. 5). Steglich (1998) has used the dataset generated by the experiments of my dissertation to develop and apply statistical techniques for testing learning assumptions. The corresponding models will be adapted to analyze individual decision making in the new experimental design. Finally, the statistical models developed for the experimental data will be adapted for testing hypotheses with field data.

Axelrod, R. 1987. The Evolution of Strategies in the Iterated Prisoner’s Dilemma. In: Davis, L. Genetic Algorithms and Simulated Annealing: 32-41.

Bainbridge, W.S., E.E. Brent, K.M. Carley, D.R. Heise, M.W. Macy, B. Markovsky, J. Skvoretz 1994. Artificial Social Intelligence. Annual Review of Sociology 20:407-436.

Batenburg, R. S., C. Snijders, W. Raub 1997. Contacts and Contracts: Temporal Embeddedness and the Contractual Behavior of Firms. ISCORE 107. Utrecht University.

Ben-Porath, Y. 1980. The F-Connection: Families, Friends and Firms and the Organization of Exchange. Population and Development Review 6: 1-30.

Bryk, A.S., S.W. Raudenbush 1992. Hierarchical Linear Models. Applications and Data Analysis Methods. London: Sage.

Burt, R. 1992. Structural Holes. The Social Structure of Competition. Cambridge: Harvard University Press.

Bush, R.R., F. Mosteller 1955. Stochastic Models for Learning. New York: Wiley.

1995. Social networks and the effect of reputation on cooperation. ISCORE paper 42. Utrecht University.

Blumenberg, B. 1997. Das Management von Technologiekooperationen. Amsterdam: Thesis Publishers.

Crozier, M. 1967. The Bureaucratic Phenomenon. Chicago: University of Chicago Press.

Flache, A. 1996, The double edge of networks: An analysis of the effect of informal networks on cooperation in social dilemmas. Amsterdam:Thesis Publishers.

Flache, A., R. Hegselmann 1998. Rational vs. adaptive egoism in support networks: How different micro foundations shape different macro hypotheses, in: Leinfellner, W, Köhler, E. (eds.). Game Theory, Experience, Rationality. Foundations of Social Sciences, Economics and Ethics. In Honor of John C. Harsanyi. ( Yearbook of the Institute Vienna Circle 5/97); Dordrecht-Boston-London: Kluwer.

Flache, A., M.W. Macy 1996. The weakness of strong Ties: Collective action failure in a highly cohesive group. Journal of Mathematical Sociology 21:3-28.

Fudenberg, D., D. Levine 1998. The theory of learning in games. Unpublished manuscript (available at http://levine.sscnet.ucla.edu/papers/contents.htm).

Fudenberg, D., J. Tirole 1991. Game Theory. Cambridge, Massachusetts. MIT Press.

Gould, R.V. 1993. Trade Cohesion, Class Unity and Urban Insurrection: Artisanal Activism in the Paris Commune. American Journal of Sociology 98:721-754.

Han, S. 1996. Structuring relations in on-the-job networks. Social Networks 18:47-67.

Heckathorn, D. D. 1996. The Dynamics and Dilemmas of Collective Action. American Sociological Review 61: 250-277.

Hogarth R.M., M.M. Reder 1987 (eds.) Rational Choice. The Contrast between Economics and Psychology. Chicago: University of Chicago Press.

Holländer, H. 1990. A Social Exchange Approach to Voluntary Cooperation. The American Economic Review 80:1157-1167.

Homans, G.C. 1974. Social Behavior. Its Elementary Forms. New York: Harcourt Brace Jovanovich.

Kreps, D.M. 1990. A Course in Microeconomic Theory. New York: Harvester.

Kitts, J., M.W. Macy, A. Flache 1998. Structural Learning: Conformity and Social Control in Task-Oriented Groups. submitted to: Computational and Mathematical Organization Theory.

Lazega, E., M. van Duin 1997. Positions in formal structure, personal characteristics and choices of advisors in a law firm: a logistic regression model for dyadic network data. Social Networks 19: 375-397.

Lindenberg, S. 1998. Solidarity: Its Microfoundations and Macro Dependence. A Framing Approach. To appear in: Doreian, P.; T. J. Fararo (eds.). The Problem of Solidarity: Theories and Models. Gordon and Breach.

Macy, M.W. 1993. Social Learning and the Structure of Collective Action. In: E. Lawler et al. (eds.). Advances in Group Processes 10:1-35. Greenwich: JAI Press.

Macy, M.W., A. Flache 1995. Beyond Rationality in Models of Choice. Annual Review of Sociology 21:73-91.

Marwell, G., P. Oliver 1993. The Critical Mass in Collective Action. A Micro-Social Theory. Cambridge: Cambridge University Press.

Olson, M. 1965. The Logic of Collective Action. 2nd edition. Cambridge, Massachusetts: Harvard University Press 1971.

Opp, K., C. Gern. 1993. Dissident Groups, Personal Networks and Spontaneous Cooperation: the East German Revolution of 1989. American Sociological Review 58:659-680.

Ostrom, E. 1990. Governing the Commons. The Evolution of Institutions for Collective Action. Cambridge: Cambridge University Press.

Petersen, T. 1992. Individual, Collective and Systems Rationality in Work Groups: Dilemmas and Market-Type Solutions. American Journal of Sociology 9:469-510.

Probst, D.A. 1996. On Evolution and Learning in Games. Bonn: Universität Bonn.

Raub, W. 1997. Samenwerking in duurzame relaties en sociale cohesie. Amsterdam: Thesis Publishers.

Raub, W., J. Weesie 1990. Reputation and Efficiency in Social Interaction: an Example of Network Effects. American Journal of Sociology 96:626-654.

Roth, A.E., I. Erev 1995. Learning in Extensive-Form Games: Experimental Data and Simple Dynamic Models in the Intermediate Term. Games and Economic Behavior 8:164-212.

Selten, R. 1975. Re-examination of the Perfectness Concept for Equilibrium Points in Extensive Games. International Journal of Game Theory 4:25-55.

Snijders, T.A.B. 1996. Stochastic Actor Oriented Models for Network Change. In: Doreian, P, F.N. Stokman. Evolution of Social Networks. Gordon and Breach Publishers: 185-208.

Steglich, C. 1998. Learning Models for Social Dilemmas: a Case Study. Paper presented at the ICS forum day, Groningen.

Stokman, F.N. et al. 1997. Solidarity 2000+. Groningen, ICS.

Stokman, F.N., J. V. Stokman 1995. Strategic Control and Interests. Its Effects on Decision Outcomes, Journal of Mathematical Sociology 20: 289-317

Tirole, J. 1986. Hierarchies and bureaucracies: on the role of collusion in organizations. Journal of Law, Economics and Organization 2:181-211.

Wasserman, S., K. Faust. 1994. Social Network Analysis. Cambridge: Cambridge University Press.

Winter, S. 1987. Comments on Arrow and Lucas. In: Hogarth, Reder 1987: 243-250.

Wippler, R., S. Lindenberg 1987. Collective Phenomena and Rational Choice. In: Alexander, J.C. et al. (eds.). The Micro-Macro Link:135-152. Berkeley: University of California Press.

Yamaguchi, K. 1994. The Flow of information through Social Networks: Diagonal-Free Measures of Inefficiency and the Structural Determinants of Inefficiency. Social Networks 16: 57-86.

Zeggelink, E., F.N. Stokman, G.G. van de Bunt 1996. The Emergence of Groups in the Evolution of Friendship Networks. In: Doreian, P, F.N. Stokman. Evolution of Social Networks. Gordon and Breach Publishers: 45-71.