The Case of Impact Factor

,


Introduction
. In very broad terms, cognition can be defined by the enabling conditions for adaptive, flexible behavior, where behavior also includes what people commonly call thinking (Wheeler ).Under these lenses, the key aspect to understand cognition becomes interactivity as opposed to the more traditional view of the isolated individual and his/her processing mechanisms (Chemero ).Thus, these enabling conditions lean on people, artifacts and processes working together to ensure that tasks are accomplished.In this paper, our interest is on how, in practice, organizations and their members act, make plans and implement decisions (Hodgkinson & Healey ) providing the ground for us to examine this interactive perspective on cognition.One turns to how individuals and groups can be e ective as they identify problems and sustain, improve or set up various practices and modes of action.This systemic view is not recent in the management literature because we see it very close to the work of some of the founding fathers of the discipline (Simon ; Cyert & March ).
. Consistently with the tradition of organizational behavior research (e.g., Gavetti et al.
), we focus on cognition.However, our approach moves the boundaries of early and more recent works by interpreting brain activities and processes as part of a larger set of "equipment", enhanced and fostered by human co-operation.As we will detail in the next section, cognition is spread (better, "distributed" Hutchins ) over a diverse set of systemic resources that are internal and external to the individual.Clearly, in this view it is the process of exploiting external resources -natural, social, and artifactual (Magnani ; Secchi ) -that matters the most.While many distributed cognitive researchers have directed their interest towards natural or artifactual resources (e.g., Pedersen ; Ste ensen ), fewer have paid attention to the e ect of distributed social resources (Hollan et al. ; Secchi ).The activities that are reflected on various organizational practices seem to be the archetypical environment where these distributed processes can be studied.Building on recent work (Secchi & Cowley ), we study the case of peer-review.The case is interesting in that, while organized, it does not depend on a single institution.Further, the literature shows that, as quality control, peer-review shows several shortcomings and issues that let some scholars question its validity (see, Bornmann ).In this influential comprehensive review ( ), Bormann views it as a "social judgement process of individuals" (i.e. not as practice which arises between people in a meso scale).He finds no evidence that peer review is either reliable, valid or an e ective predictor of impact.Starting from this basis, other scholars propose peer review be rethought as a socially distributed process (Cowley ).
. This paper extends the example of peer-review to citation measures and, specifically, to impact factor (IF).While IF is susceptible to many criticisms (Vanclay ; Seglen ), the index is an important constraint on academic practice.The choice of IF was made on two grounds.One is that it is a widely used index that is employed -very controversially -to evaluate a journal's standing and the relative weight of its articles.Hence, we assume that our readers would know what this measurement index is as they encountered it in their academic life.This is to say that IF is familiar to most readers and most of our arguments may be easily understood by the average academic.The other reason for using IF is that, while widely used, it is known for several flaws.One is that it is an unreliable indicator of article quality in that a journal's IF correlates inconsistently with articles published in that same journal (e.g., Seglen ).Another is that the two-and five-year time span IF uses may be appropriate for some disciplines, but not for others (e.g., Curry ).Finally, papers with a high number of citations are very few, even in high IF journals (Colquhoun ), hence the vast majority of papers in high IF journals are not influential or less influential than other articles published in low IF journals.These issues can be anchored to at least two logical fallacies -i.e.connected claims that invoke reasoning that, while fallacious, appears sound (Woods ).IF is used to describe an academic journal as a whole, in relation to citations attracted by the articles published in there.This means that attributing characteristics of the whole (i.e. the journal) to any of its parts (i.e. the individual article) is falling on to the so-called composition and division fallacy.Another fallacy that can be evoqued is the appeal to authority or ad verecundiam, one which arises every time one uses IF to legitimize the inherent depth, sagacity, rigor, value, or strength of an article.These two have been linked to decision making (Secchi , pp. -) and are particularly useful in the study of cognition in that they help us reflect on how imprecise information can be due to mental frames and prejudices.Moreover, beliefs on IF is sometimes socially construed since it can be shared with professional associations, other faculty, members of a research group, or an institution's leadership, hence connecting to the 'organisational' side of cognition.

.
In pursuing whether IF a ects cognition as socially distributed, we use an agent-based computational simulation (Edmonds & Meyer b) of its dynamics in an organization.We show how individual preferences such as the attitude towards IF, publication outcomes, and evaluation proxies, together with more socially-nested aspects such as proximity to co-workers, group/departmental a iliation, and update of one's beliefs a ect the perception of the scientific value of a publication.In summary, one's perception of the research undertaken is a ected by individual and group attitudes to IF or, in other terms, by individual perceptions as well as social organizing.This work is oriented towards understanding if and how a theory of cognition that is not bounded by the skull (e.g., Clark ) but socially enabled (e.g., Hutchins ) applies to an organizational environment.The computational simulation considers all these aspects together and will, hopefully, help us refine the theoretical model.

.
In the following, we outline the theoretical assumptions on organizational cognition that constitute the background for the model.We then introduce the computational simulation and present its results, outlining implications for both the study of cognition in organizations and the use of IF.We conclude with limitations and a few suggestions for future research.

Theoretical Background
. In their recent work, Secchi & Adamsen ( ) draw on literature from major journals that connect 'organization' and 'cognition.'They show that, while the number of papers on cognition in organizations has risen since the s (as in many other areas), the field has never taken o .They address this failure by o ering a logical classification as the basis for reviewing the field.One approach is additive -cognition and organization can be seen as mutually illuminating yet separable.For example, organizations can be used to explore consequences of the bounds on human rationality.In the second, combinatory category, organizations are treated as open to using cognitive concepts (and vice versa).If contingencies are central (as in Simon's work; Simon ), the organization is seen to be emergent, complex and hard to predict.However, while tied to input-output models, the perspective is blind to how the social and the cultural bear on individuals-in-interaction.In the intersectional category, however, one can develop a concern with phenomena such as shared cognition, teamwork and sensemaking.Accordingly, Secchi & Adamsen ( ) adjudge this the most progressive approach and gives special weight to how Weick's work has been used (Weick ; Weick & Roberts ; Weick & Sutcli e ).Fourth, the conditional category treats cognition as a part of organizations in that it uses constructs like intelligence and motivation.
. The underlying thread in the organizational research literature with a cognition focus is the almost unanimous interpretation of cognition as information processing.This happens for two main reasons.First, scholars typically adopt Simon's early conceptualization of cognition as working like a computer (Newell & Simon ; Gigerenzer & Goldstein ).If minds are (like) machines, cognitive models are too coarse grained to capture the details of individual decision making and organizational life.Thus, while some focus on macro factors like organizational structure, change, and communication (e.g., March ), most turn to micro concerns such as individual motivation, knowledge, skills and decision making (e.g., Hodgkinson & Healey ).In considering a cognitive dimension together with standard organizational research these studies contribute to the advancement of the field (e.g., Crilly et al.
) although they very seldom attempt at covering the macro-micro gap.One aspect that is o en brought in to bridge this gap is to consider action (and, most notably interaction, see Ste ensen ) as a key feature of cognition -sometimes referred to as embodiment (Clark ; Magnani ).So, where this inherently dualistic idea is replaced by some view of embodiment, organized activity becomes inherently cognitive.The isolationist and internal perspective of the mind-machine is thus seen as having evolved in an ecology (e.g., Gigerenzer & Selten ).This paper focuses on one way of construing this claim.

.
We are aware that di erent disciplinary traditions may overlap, to some extent, with the one presented in this article.).Simon thus stressed the intelligent use of contingencies and, in his writings, o en drew on his experience in the team that designed what became the Marshall Plan.Importantly, the so-called plan self assembled around a group (including Simon) who argued that the issues be seen in terms of the balance of trade.What he calls a representation (balance of trade) thus became, in Simon's terms, a weapon and a tool.The group developed a formal organization to challenge how others saw the problem and, eventually, led to a major success in US foreign policy.Building on his experience, Simon ( ) defined communication as involving influences that shape decision premises.Far from focusing on individual expertise, he stressed how di erently people understand, and, thus, how formal and informal structures compensate for individual limits.It is from this forward-looking perspective originating from Simon that we develop our view of organizational cognition, similarly to many social simulation scholars (e.g., Edmonds & Meyer a; Conte ; Carley et al. ; Conte et al. ).
Beyond a computational theory of the mind .
Classic cognitive science was dominated by the computational theory of mind (see, Boden ).As recently as , Fodor was still calling it "the only game in town."However, the first cracks had already appeared: Dreyfus ( ) emphasized know-how -not propositional knowledge of a supposedly objective world.Searle's ( ) thought experiment in Minds, Brains and Programs picked out a major deficit in computational models.In his Chinese room thought experiment, Searle argued that if a program, or a little man in the head, carried out symbol processing, neither the man, the one in his head, nor indeed the one in that head (and so on) could possibly have any understanding.Not only was computation based on modeling propositions (see also Horst ) but neither experience nor consciousness could be clarified by programs.As the focus shi ed to understanding and know-how, interest grew in how cognition draws on experience of action.Hurley ( ) pinpointed another fatal weakness in all input-output models, the computational posits a cognitive sandwich (i.e.where 'cognition' is a filling between a slice of perception and one of action).Computational views thus reduce individuals (and groups of individuals) to mere processing systems.Much have been learned by these approaches but more current views highlight that a richer understanding can be gained through considering an extended system (Clark , ).And yet, with very few exceptions (e.g., Weick's sensemaking approach, or Ocasio's , ), management perspectives on cognition still align with these early computational views (Secchi & Adamsen ). Historically, concerns with the traditional computational approach became clearer with work in fields such as robot design and how brains and bodies co-evolved with living beings (e.g., Perico et al.
).This made researchers aware of the way individuals attune through body-based encounters with each other and the perceived world (see, Varela et al. ).For those designing computational models, the focus shi ed from symbol processing to connectionism.Cognition was seen to spread beyond the brain or, in Clark & Chalmers ( ) well known terms, opened up the hypothesis of extended mind.
. Evolution opened up flexibility long before the emergence of multi-cellular systems.While recent work traces the rise of brains and sensation to the Cambrian period (see, Keijzer ), learning is more recent and, in human worlds, adds how language co-evolved with social, technological and organizational means.Cognitive activities build on how humans learn to use interactions to discover culturally appropriate ways of acting.As adults, they develop the skills used in cognitive systems that perform tasks such as navigating a ship (see, Hutchins ) or comparing soil samples in archeology (Goodwin ).As is increasingly clear, the relevant know-how connects perception with action -there is no need to posit a cognitive sandwich.Fixing a ship's position or seeing soil color as an archeologist arise from linking bodies, central nervous systems and experience as people engage with social, linguistic and material phenomena.Humans draw on the past and imagine the future because cognition and language are distributed (see, Cowley ; Hutchins ).Given the multi-scalarity of external resources, neural functionality develops as the familiar is put to new uses (Anderson ).Systems self-simplify by using experience of perceiving-acting in an encultured world (Berthoz ).Not only are contingencies useful, but what Jaynes ( ) first called excerption enables people to break from the flow of experience to attend to do/say something unexpected (Cowley & Vallee-Tourangeau ).Each person brings forth a world (Varela et al.
) through enaction and/or the perception-action coupling known as sense-making (see, Thompson ).Though rationality is bounded, humans ceaselessly overcome their own bodily limits: they draw on the reductions of entropy associated with what Hutchins calls a cultural ecology -this simplifies procedures, choices of discourse and ways of remembering as people come to perceive similarly (Hutchins ).Rather than seek a single model of cognition, most acknowledge that, at times, thinking is habitual or procedural and, at others, it is willful and uses, for example, counterfactuals (Evans ; Cowley & Vallee-Tourangeau ).In this sense, cognition is distributed and, thus, amenable to systemic investigation.Accordingly, we pursue how IF functions in relation to, not individual judgements, but how these play out in groups.

E-cognition and the management literature .
The directions very shortly outlined above clearly align to those of the so-called distributed cognition perspective or e-cognition, as it has been referred to more recently (e.g., Menary ; Theiner ).According to this view, human "powers" are e-cological (sensitive to local resources), e-nacted (by sensing bodies), embedded (intrinsic to perceived situations), e-mbodied (link action and perception), e-xtended (use technology and organisations) or, in a word, systemic (they bind the social, material and the temporal, Cowley & Vallee-Tourangeau ).By deflating the role of the brain, systemic or e-cognition ensures that individuals and groups are able to draw on a range of computational resources (Wells ).
. Placed against the recent history of cognitive science, the findings are striking.In spite of symbol flight, the turn to embodiment and rejection of the cognitive sandwich, the organizational literature is conservative.Computational assumptions ground additive, conditional and combinatory approaches (Horst , ); in the intersectional view, whilst there is some embedding and enaction (Weick ; Weick & Roberts ), no allowance is made for other systemic aspects of cognition (the extended, the ecological or the distributed).In teamwork, for example, individuals act as a unit that accomplish tasks that presuppose shared goals.However, in contrast to the practical details of navigation described on the US Palau (Hutchins ), teams are not seen as culturally embedded, technically augmented and task-oriented systems.In our terms, little weight falls on the ecological, the embedded and the enactive.

Social organizing: A case in peer reviewing
.
In order to make the theoretical claims we chose a practical application to which they could be applied to.A peer-review system can be rather complex in its mix of cognitive and organizational elements (Cowley ) and yet it is emblematic of how modern tasks are operationalized.In fact, the nature of its activity is grounded in flexible work, voluntary and professional performance, organizational and group constraints.In an e ort to be more precise, this study limits its span to the assessment of papers (e.g., annual reviews, promotion evaluations) rather than to the process of journal double-blind peer reviewing.More precisely, using the frames from the e-cognition literature outlined above, we pursue how such a system works.

E-cological: sensitive to local resources .
The cognitive process is triggered by resources that are located in the environment surrounding the decision maker.These can be classified in material or natural resources, artifacts, and socially-bound resources (Hollan et al.

; Secchi ). It is the interplay between various mixes of resources and one's individuality (Clark & Chalmers
) that defines a cognitive process as such.In a review process, the article is clearly an artifact, together with the tools used to perform reviewing and revision -e.g., a printer, a pen, a highlighter, a computer, the desk in which one sits.These all enable cognitive processes leading to performing the given task.And, of course, the space/time combination of how these resources are exploited changes the dynamic of these processes.In the ecological assumption, also other individuals one interacts with are part of the cognitive environment.However, as shown elsewhere in this document (and also below), they are resources of a di erent kind.Finally, an ecological perspective also allows peer review to be defined in relation to the IU -that is as involving reviewer, editor, author, platform, etc. -or, in other words, to a meso area of interaction between the various players and their environments.This is particularly important due to the fact that, in standard models -especially due to the influence of Merton (e.g., ) -lies precisely in defining it as an individual-centered process that depends on an ideal agent (Small ).So it is seen in terms of the macro -as process -and the micro -what an individual does and does not do.
Consistently with some organizational scholars (e.g., Weick ; Weick & Roberts , and also Ocasio , relative to some aspects only), distributed and e-cognition keep action as an essential part of the process.This has two extremely relevant implications.First, it reminds researchers that even individual cognition happens in a time-relevant scale and that these timescales (Neumann & Cowley ) have important repercussions on the way individuals frames data.In peer review, it a ects the way one refers and interprets (feels bound) to professional standards, for example, as to represent a long-term timescale.At the same time, one is very much a ected by the particular topic of the article that, sometimes, may not perfectly fit one's expertise, hence forcing ad hoc strategies (e.g., google searches, asking another colleague).
. Second, being spread over time, any cognizant activity involves the manipulation of resources.Every time one is dealing with resources, their handling is usually more relevant than each of the resources per se.This means that, in a peer review process, having access to a computer is rather dull compared to the cognitive meaning of the word one is typing that, combined with other words and knowledge, make one think of changing it or changing its surroundings (e.g., the other words in the sentence, position, format).This continuous activity of providing new meaning to the external resource is extremely widespread in human cognition and may be referred to as re-projecting (Magnani ; Secchi & Bardone ).Moreover, any cognitive process -especially creative or knowledge-based such as assessing a manuscript -involves the possibility of tinkering, hence seeking chances and getting on something new (Bardone ).
One of the most important assumptions of this perspective is that cognition does not exist outside of the elements in which it manifests itself as it was in the so ware/hardware perspective.This is to say that there is no process without the melding of internal (one's brain and body) and external resources.In a review, the assessment of a given paper is necessarily the result of one's use (or exploitation) of the resources available, through intuition, a ordances and, more broadly (if not vaguely), perceptions.This also brings in the implication that a decision is almost never done the minute it is made.Quite the contrary, even though it is put externally once done (one could say externalized; Magnani ; Secchi & Bardone ), one can still "own" a decision, as in some sort of endowment e ect (Kahneman et al. ).Once an assessment is submitted back to the author or to the head of a committee (or to an editor, in case of journal peer review), the reviewer "owns" all the arguments it contains.

E-mbodied: link action and perception .
The social being requires a biological being.As obvious as it could sound, the role of the body, emotions, and overall perceptions were not taken into consideration in cognitive studies until recently (Wilson ; Varela et al. ).In a distributed e-cognition perspective, a reviewer would interpret and makes sense of the text by also 'feeling' it fits within one's own domain of competence or not, for example.Also, the mental (and sometimes physical) e ort required by the act of reading, understanding, interpreting, analyzing, writing, come at a cost -that of the energy dispersed to perform the task.

E-xtended: use technology and organizations
.
If cognition cannot be thought of without all the resources (or tools) that allow it to exist, then it is fair to state that one's cognition extends to include these resources (Clark ).This is a view also called of the extended mind because it seems we 'think' through the use of these resources.The assessment of a paper may result di erently depending on the tools used -this is to say that the making of one's thoughts are shaped by the interaction one has with a so ware, one or more colleagues, the strategy used to "attack" the manuscript or the writing process.

The social element .
Given the multi-scalarity sketched above, people shi between the more automatic (checking reference formatting requirements are met, printing the paper) and doing things deliberately (asking advice to a colleague on a methodological point, suggesting a twist in an argument).While a combinatorial view (as in Secchi & Adamsen ) allows the organizational to influence the cognitive, it leaves out the socially improvised-repetition based synergies -that ground intelligence in mammals (and humans).As noted, people also experience selves and others as they use excerption to shi attention, how they act, what they say, and frames of reference.Much experience is social or, in lay terms, is bound up with collective forms of meaning.Whereas a computational view treats the social as secondary, Simon prefigures systemic approaches in seeing that information di ers across persons who also vary in openness or docility (Simon ).This concept was originally introduced by Tolman ( ) and then developed within the bounded rationality framework in two stages in Simon's career (Simon , ).The idea is that humans have an attitude towards being taught (from the Latin docilis) and that this reflects on how much one listens to advice, suggestions and recommendations coming from others while making decisions.While some have elaborated on altruistic behavior stemming out of this idea (e.g., Knudsen ), others have pointed that the concept needs an organizational or, at least, a sense of community to e ectively play its e ects (Bardone ; Secchi ).Together with Secchi & Bardone ( ), we regard docility as a behavioral aspect of distributed e-cognition in that it grounds the use of external social resources to cognitive processes.By following these footsteps, we see Organisational Cognition (OC) as a field where individuals and organizations form aggregates whose functionality draws on meso scale events of social organizing.This meso scale replaces appeal to a macro-micro created by positing that Organisational Cognition is highly dependent on what has been termed social organizing.

.
In an attempt to connect the concepts reviewed in this theoretical framework to a practical case, Table offers a sketch over one phase of peer review paper assessment.In the table, we start from a sample task, then elaborate examples for each of the e-categories and then link them to a more social, or meso scale in the cognitive perspective.Finally, to further clarify the conceptually backed model, we have tried to indicate how these concepts reflect on parameters of the model as described in the next section.Here, we turn our attention to a practical example of whether OC can be framed as e-cognition to show whether and how the meso scale a ects the micro (individual) and macro scale events.We use a simulated organization where two research groups have di erent appreciation of scientific value based on acceptance or rejection of IF.

The Model .
From the above, one should expect much cognitive processing to occur as people engage in intelligence units (IU).Due to the dynamical and complex nature of the processes, we deem an agent-based simulation model (ABM) appropriate, in line with the current organizational behavior literature (Fioretti ; Secchi & Neumann ; Secchi ).In fact, due to their features ABM are especially well equipped to represent the dynamic interactions of complex systems (Miller & Page ) and of boundedly rational agents (Secchi ).

Modeling impact factor
.
In an attempt to consistently cover most aspects of the ABM simulation, this article uses the Overview, Design concepts and Details (ODD) protocol (Polhill ; Polhill et al. ).The simulation was carried over NetLogo Sample process: While reading the paper, one writes notes on methods concerns using a word processing so ware to improve technical aspects of the article and communicate to the authors via email.), an open access computational so ware specifically designed for agent-based modeling.The model is available online in the OpenABM platform at https://www.openabm.org/model/5589/version/2/view.

.
The unit of analysis is cognition as it materializes among individuals in two separate intelligence units (IUs).
Our approach has been that of following applied psychology and organizational behavior.Scholars from these fields sometimes measure cognition through psychometric scales and/or indices.These are summarized by numbers.Hence, we have assigned each agent a score for various elements (e.g., docility, perceived scientific value) in a way that mimics the results of a data collection exercise.Each agent in the simulation has a score for this attitude towards socially-bound distributed e-cognition processes.Thus, each agent is more or less prone to use information coming from the social channels surrounding it, i.e. the other agents.As explained below in further details, other aspects a ect agent 'thinking' and behavior.The simulation pursues how agents may show di erent understandings of the same content over di erent periods of time and depending on their a iliation to a specific IU.The environment can be pictured as an academic or research institute where individuals are employed and perform their work duties.As an aspect of the broader ecological and extended characteristics of one's thinking, the simulation uses impact factor (IF) to exemplify how academics may change their understanding of what grants scientific value (PSV, see below) to their work.Consistently with our framework, we assume that this understanding depends, in part, on how cognitive processes are structured by IUs.Time, evaluation of other people's publications, and professional exchanges with colleagues may also bring a diversity of contents to the same item (i.e.IF).The fact that the IF is a numerical is peculiar in that, superficially, this appears to place its face value beyond dispute.In examining this idea, the simulation investigates when and how a simple numeral can be perceived and dealt with di erently, exploring whether this is more likely to happen when individuals interact in their IU even though they all are part of the same institute, university, or organization.All parameters and notations are specified in Table .The meso scale .
The model includes two "intelligence units" (labeled IU and IU ) that each describe how a group of academics interpret their career, and especially their publications, in relation to macro-concerns.While members of IU have a less restrictive interpretation of the IF, IU treat it as the sole metric to be used in judging publication and success.The two IUs might be two departments or, indeed, two research groups/centers in the same organization.There is also a third residual group of agents who fit between the two views of career, publication, etc.They are not part of an intelligence unit but, depending on certain conditions (described below), they may join one of the units.This general attitude to career is defined by what is labeled 'perception of scientific value' (PSV) in the simulation, with values distributed normally at random to all agents, with di erences due to being a iliated with IU , IU , or no a iliation (see Table for details).We interpreted non a iliation to be reflected in a wider variation of PSV values, because of lack of any anchor, while IU and IU have the same dispersion.IU members (i.e.IF enthusiasts) have higher PSV, on average, reflected by IU average + parameter alpha (see Table ), in the sense that they 'believe' to have higher standards for science.

Agent characteristics .
To represent organizational/social cooperation and interactivity, each agent is assigned a tendency to take on social information or a so-called 'docility' level (Simon ; Secchi ; Secchi & Bardone ).This is fixed on a random-normal distribution with standard deviation fixed at 0.2 and mean taking three values [0.6, 0.9, 1.2].Thus, all agents have a level of docility that is assigned at the beginning of the simulation and is independent of their IU a iliation.Given that this parameter is important for interactions (see below) it is crucial it does not depend on the IU but it is something that agents have independent of a iliation.This attribute has been used in ABM before (e.g., Bardone & Secchi ; Miller & Lin ; Thomsen ) and it indicates the extent to which information coming from others a ects agent decision making and behavior.To some extent, it is a measure of how much one is susceptible to the interactivity in a group and it operationalizes sociallyoriented cognitive aspects (Secchi & Gullekson ).The values used to generate the three distributions have a small standard deviation such that values concentrate mostly around the mean and thus allow for clearer observation of the 'docility e ect' (Secchi & Bardone ).
. Independent of their IU, agents are also associated with attitudes towards the IF of a particular journal.The attitudes are distributed in the population of agents normally at random with fixed mean and standard deviation (see Table ).Possible alignment between IF attitudes and IU are le to the dynamics of the simulation.This is the value attributed to a scientific output by an agent that is not a iliated with any IU.∼ N (1, 0.15) This is the value attributed to a scientific output by an agent that is a iliated with IU -smaller st.dev means that values are more grouped around the mean.

Parameter
This is the value attributed to a scientific output by an agent that is a iliated with IU -the higher mean signifies stricter criteria to assess P SV .P SV di erence between IU and IU , α 0.25, 0.50 The average di erence in the perceived scientific value that members of IU have in relation to those from IU -this value a ects directly the mean of the randomnormal distribution for IU ; st.dev. is una ected.group [true, false] The tendency to be socio-cognitively closer to the other members of the IU the agent is a iliated with.change [true, false] The possibility to change IU depending on the distance of one's P SV to the mean P SV of either IU or IU .leaving, l [0, 3, 5] A random number of agents between 0 and l exits the system.joining, j [0, 5, 15] A random number of agents between 0 and j enters the system.range 4, 8 This is the value used to explore the environment that surrounds each agent.

Table : Parameter Notations and Values
.
One assumption of this model is that each published article is associated with the IF of the journal in which it appears.Another is that, we believe, certain papers do have higher scientific value than others.Both assumptions can be challenged as unrealistic.On the one hand, many journals lack such metrics and other outlets do not qualify (e.g., books, chapters, conference proceedings).On the other hand, some social constructionists would argue that all publications are of equal value qua 'texts'.While disagreeing with the latter view, we accept that the model simplifies by treating di erent IF levels as representative of various kinds of journals.However, what matters in this simulation model is how the agents perceive IF.For hard-liners (members of IU ), scientific value is a judgement about the journal and the article's scientific value is identical to the journal's IF.By contrast, for members of IU , IF is a less meaningful criterion for evaluating publications since they rely on the perceived scientific 'value' of the article -this may or may not be aligned to the IF of the journal in which it appears.
. Finally, every agent connects to other agents, preferably of the same IU through a 'proximity rule', indicated by the parameter range that takes two values [4,8].This means that each agent screens a mind space around it to see whether it is possible to connect to other agents and establish a link.The two values [4,8] for the parameter were selected to give a relatively narrow and relatively wide reach to each agent.

Publications .
Every agent is assigned a random number of publications (max 4) at the beginning and this number is recalculated every time that there is an 'evaluation event'.These events are called publication waves in the simulation and happen with fixed frequency (e.g., every year, semester, month).A er a few pilot tests, frequency is set at every 10 seconds.Every time a publication wave comes, all previous publications are archived and excluded from the pool that is available for the current round of evaluation.This mechanism is implemented to replicate assessments done by the government or by universities, where the same publication cannot be submitted for evaluation twice.
. Each publication is also assigned some intrinsic value between 0 and 1, that di ers from IF (and it is not an IF function).This is intended to imply, above all, that scientific value exists independently of metrics (Vanclay ; Seglen ).Second, it allows the value of a scientific contribution and the IF of the journal to be at odds.As already mentioned above, each and every publication has an IF; this is randomly assigned and can be any number between 0 and 5. To make the simulation mimic disruptive innovation in science, one in four low IF publications are granted high scientific value.At every assessment round, one random low IF publication has its value calculated using the transformation y = 0.8 + IF pi • 0.2, where IF pi is the impact factor of the i th publication.This procedure is implemented to make sure that some articles published in so-called "second tier" journals carry very high meaning for the scientific enterprise.This is to make sure that, in this case, IF enthusiasts (IU ) are necessarily wrong, because they would at first evaluate the paper badly even thought it has very high scientific value.

Evaluation procedures .
When a publication wave is completed, each agent evaluates papers according to the rules of their unit.IU members (hard core IF enthusiasts) evaluate papers on the sole basis of the journal's IF score.Conversely, both agents from IU and those una iliated to a IU attend to their perceived value (i.e.independently of a journal's IF).This step one of the evaluation procedure occurs as each agent evaluates one paper by another. .
Step two adds a second paper to the evaluation.Each agent selects another paper to assess which is likely to be from one of the other agents in their own IU.The evaluation rules are as for the first paper (i.e. it covaries with IF in IU and with perceived value in IU ).In this case, the paper is 'suggested' by another individual who prompts the colleague to read what it thinks is 'best'.There is one exception to the procedure.If the closer agent is from a di erent IU than the agent, then the evaluation focuses on the paper coming from that closer agent.
. Together, the two steps ensure that each agent seeks to evaluate two papers.At this point -i.e. a er two-step evaluations of, on average, two papers -a 'socialization' element is introduced.Agents will tend to match their evaluation with the rule of their own IU; however, they also use their docility and that of those around them to set evaluation to other standards (if they are so able).This is done to reflect social influences on individual thinking and behavior in organizations (e.g., Simon ; Secchi ).The mechanism passes through a function that adjusts both evaluations and their beliefs depending on how far they are from the others in the IU and in the system in general.The first phase of step three is an internal update.Agents with high docility (i.e. higher than the mean) update their paper evaluations by adjusting as a function of their docility.This is an internal update because agents with high docility second guess themselves and re-evaluate papers by slightly updating their initial judgement (Secchi & Bardone ).This step is done to allow highly docile individuals to come up with a more di erentiated evaluation than the other agents (see the online materials for details).

.
The second phase of step three is an external update.This depends on interactions between agents.If both agents have high docility, then each updates its evaluations following the other agent (i.e. if that is higher, the value goes up and, if it is lower, the value goes down).The agents also update PSV depending on whether or not the 'other' individual is from the same or another IU.If docility levels are misaligned, whereas the highly docile updates its evaluations in relation to the other, the agent with lower docility updates its evaluations only when the other individual is from the same IU.It is important to remember that all the evaluation phases in step three arise between agents that are connected through a link and in a given range.

Model dynamics .
The initial set up starts with a fixed number of agents (100) that randomly appear on a two-dimensional space.
Their position is intended to be a 'mental space' -not just a physical location.The initial number can be altered with those leaving and joining, to mimic retirements, employee turnover, hiring, and other organizational processes of this kind.Where a number of agents leave the organization permanently they are removed from the simulation right a er a publication wave.The number can range randomly between 0 and the upper limit that is fixed in the simulation runs to [0, 3, 5] (see Table ).The number of agents that join the organization creates new agents right a er a publication wave and can be of any number between 0 and the upper limit (randomly fixed at [0, 5, 15]).These two conditions of leaving and/or joining start from publication wave 3 and add more dynamism to the simulation, with rapid expansion, when joining tends to 15, and leaving ranges between 0 and 3, or decline, when joining is 0 and leaving can be up to 5.This is in line with Simon's characterization of IU (Simon ).
. Soon a er their appearance, agents characterize themselves as members of IU , IU ; a minority will have lack of membership.Agents then initially connect together with other same-IU members based on proximity (or range in Table ) and start paper evaluations as soon as the new publication wave arises.Agents also connect with other agents from outside their own IU by using a subset of range -i.e.only if these are closer than members of their own IU.This is done under the assumption that it takes a bit more 'likemindedness' to be willing to connect to someone from outside of one's own group.
. Once relationships are established, the three-step evaluation can take place.One of the implications of this is that, at some point, una iliated agents may decide to join one of the two IU groups.These agents join groups depending on whether PSV is closer to the average of IU or that of IU .
. In the group condition, agents from the same IU stay clustered together and thus limit interactions with other agents outside their own group.This condition is implemented to separate the two IUs so that it is easier to observe emerging coalitions.This is the case when an organization prompts groups (departments, centers) to develop an independent climate or, roughly speaking, to live the organization in di erent ways.The change condition allows for all agents to shi from their initial IU placement.The mechanism whereby this happens is fixed by PSV such that, where one's PSV is closer to the mean PSV of the other IU, this agent 'feels' closer to the others and shi s to their IU.There is no limit as to how o en an agent changes a iliation.

Findings
. A er a few pilot runs to study convergence and sensitivity, we settled on the set of conditions described in Table and framed as a factorial design of 2 7 •3 3 .This triggers the question of how o en the simulation should run per each configuration of parameters (Ritter et al.
; Lee et al. ).Provided that computational experiments are assimilated to real life experiments (e.g., Hoser ), statistical power analysis can be used for this purpose (Secchi & Seri ).Using a smallest e ect size of interest (SESOI) approach (Seri & Secchi ), we calculate a number that satisfies the conditions for power 1 − β = 0.95 at the 0.01 significance level for an e ect size of 0.1 (Cohen ), in an attempt to be conservative.The factorial design leads to 3456 di erent configurations of parameters and power calculations indicate that 11 runs are su icient.However, in striving to make sense of the four conditions (group × change= 2 • 2) separately, we also need to know how many runs are su icient for 3456/4 = 864 configurations.This leads to 21 runs per configuration, given the SESOI approach, we set on 25 times per configuration of parameters, leading to an estimated expected power of 0.98.

.
As far as time for each run is concerned, the simulation spans 24 evaluation periods.If we assume that evaluations happen every year and that every academic may publish between 0 and 4 peer reviewed articles a year , then we have 24 years in the life of a hypothetical organization.At every evaluation round (i.e.publication wave) we take a snapshot of agent's characteristics and a iliation.

Analytical strategy .
The data produced by the simulation was computed on the Abacus .supercomputer, available through the Danish e-infrastructure cooperation (DeIC) to universities and research centers in the Europe.The information in the file amounted at approximately 1.05 Gb of data.Specific simulation results are calculated with the aid of graphics and the generic impact of parameters is presented in fixed e ects panel regression models as reported in the Appendix.All calculations are performed using R, an open source platform for statistical computing.
. As usual with ABM simulations, the results presented here are a selection from a set of findings that would require more than one paper to be properly introduced.While overall results are presented in the Appendix, we deem those that follow to be most relevant in pursuing our enquiry on whether how meso scale events and interactions contribute to organizational cognition .
A range e ect .
The scope of the analysis is to unveil the circumstances (if any) under which cognitive processes involve a wider social context that is located in between the macro and the micro scales, specifically in what we call meso scales or social organizing.Figure -graphs (a) to (f) -pictures regression lines that estimate how PSV for members of IU (blue lines) and IU (red lines) evolves as average IF increases due to the variation of range, leaving, and joining the organization.All six graphs compare regression lines where agents are allowed (dotted lines) or not allowed (solid lines) to change intelligence unit. .
The parameter range indicates the approximate reach of each agent; the higher its value, the more likely it is that agents interact with others outside their own IU.This should have significant repercussions on agents' PSVs because, conditional on their docility levels, they might update their evaluations more o en.In the graphs presented in Figure , mean docility is held constant at 0.6, while range varies.Actually, between graphs (a) and (d), (b) and (e), and (c) and (f), range is the only parameter that changes from 4 to 8.This leads to an increase in the e ects that take place, on average, for both IU and IU .Under all conditions, IU 's PSV seems to be little a ected by either range values or change and, accordingly, there is only a slight strengthening of the registered e ect.Being able to change IU leads to an increase in the agents' PSVs; this is minimal for IU and very strong for the IU members which are also a ected by an increase in the range of interactions.

Strengthening group ties
. Figures to build on  From the figure we can observe that the overall tendency is for an increase of PSVs, except when fewer agents leave the system (leaving = 5).Here too, the group e ect does not seem to be particularly strong or detectable at all and changes in PSV appear mild.As more agents leave the organization, IU 's PSV slows the pace of its increase.This seems not to be proportional to the declining number of agents that belong to IU (Figure ), hence suggesting that the variation in PSVs is to be associated with group e ects and numbers of agents leaving the organization.However, there might be an e ect of agents moving away from IU because of higher requirements for PSV.leaving is set to wither 0 or 3).However, it shows that PSVs decline towards the end of the evaluation periods when leaving can be up to 5. A er examination of the data, we found that the number of IU agents is down to zero at round 19 under that condition (Figure ), meaning that all agents are a iliated with IU (non-IU are di icult to find under change = true).So, a higher degree of docility seems to wipe out IU in an organization facing decline (i.e.leaving = 5), because some leave while all remaining others move to IU . .These findings are confirmed by data in Figure where the organization faces high turnover, with many agents leaving but, on average, more coming in (i.e.joining = 15).While the group e ect does not impact the shape of the curves, there is a decrease in PSV -i.e. a more consistent evaluation over time -when more agents leave.
The e ect of docility on IU seems to be that of the less radicalized view, as if IU agents would abandon a rigid interpretation of the IF and converge towards a more stable PSV (more like IU ).Less variability in numbers radicalizes the positions instead.Highly turbulent organizational environments (higher agent turnover rates) seem to flatten the curve while more static organizational environments appear to increase PSV, as if the IF-based evaluations need to always be better at every round.

Implications and Conclusions
. We interpret the ABM model of IF and PSV in relation to two main implications, each of which draws on a scale from the perspective of organizational cognition defined in the theoretical framework above.
The e ect of macro scales .These results are partially consistent with what we might expect from members of the two IUs.On the one hand, under the specific conditions analyzed above, members of IU are not a ected by the value of IF because they tend to base PSV on other parameters and, as shown, this is what happens.Even when IU interact more with non-IU agents (i.e. when range takes higher values), PSV increases slightly and it seems not a ected by an increase in IF values.On the other hand, there is a positive impact of IF on PSV for IU ; this applies especially when there are opportunities to change a iliation status (i.e.IU).In this case, it seems that interactions strengthen how IF positively a ects PSV for IU members .The parameter change fine tunes the IU, making them more orthodox so that IU becomes populated by more IF-indi erent agents and IU is populated by IF-enthusiasts.In general, however, IU members are less a ected by institutional or macro pressures, such as the opportunity to change a iliation.IU , by contrast, tend to increase PSV as IF increases; this becomes more radical as change of IU becomes possible (i.e.bears no institutional costs) and in relatively stable organizational environments (i.e.agent turnover is low).This leads us to reflect on how meso scales filter macro or institutional scales in that they make agents process them di erently.By implication, not only are macro aspects perceived di erently by di erent individuals but how they are perceived is a function of the IU that constitutes the medium through which these elements come to the agent's knowledge.
Figure : Perceived Scientific Value (PSV) as it varies according to average journal publication's IF (d = 0.6, group = false, alpha = 0.25) findings from Figure and specify what happens to PSV when IU members intensify their relations and group ties (parameter group = 'true' or 'on') as opposed to when their ties weaken (parameter group = 'false' or 'o ').Figures and show the impact of the group e ect as a selection of up to 0, 3, or 5 agents leave the organization at every evaluation round, with docility = 0.6, joining = 0 and change = 'f alse'.

Figure : Figure :
Figure : Conditional plotting of regression curves fitted on Perceived Scientific Value (PSV) as varies according to evaluation/publication waves for IU members (d = 0.6, change = f alse, alpha = 0.25, joining = 0).

Figure : Figure :
Figure : Conditional plotting of regression curves fitted on Perceived Scientific Value (PSV) as it varies according to evaluation/publication waves for IU members (d = 0.6, change = true, alpha = 0.25, joining = 0).

Figure : Figure : Figure :
Figure : Conditional plotting of regression curves fitted on Perceived Scientific Value (PSV) as it varies according to evaluation/publication waves for IU members (d = 1.2, change = true, alpha = 0.25, joining = 0).
Figure presents what happens to IU 's PSV when individuals leave and join the organization (with change = true, group = true, docility = 1.2).
For this reason, in declaring that our starting point is that of distributed e-cognition (see below Hutchins

Short description Sample resources Sample meso aspect(s) Model's how-to
They can be intended as the baseline cases, i.e. those figures that are used as a benchmark for the others.In Figure, as we get closer to the end of the rounds (i.e.24), IU 's PSVs su ers a very slight decline, moving from 1.0 to ca. 0.9, where a group e ect is not detectable at all (see also Model in Table , in the Appendix).Figure shows values for IU 's PSVs.