ENTERPRISE SOCIAL MEDIA-ENABLED TRANSACTIVE MEMORY ENCODING By Michael E. Nelson A THESIS Michigan State University in partial fulfillment of the requirements Submitted to for the degree of Media and Information — Master of Arts 2018 ENTERPRISE SOCIAL MEDIA-ENABLED TRANSACTIVE MEMORY ENCODING ABSTRACT By Michael E. Nelson Transactive memory theory and empirical studies of enterprise social media suggest that virtual work-groups on enterprise social media develop transactive memory. To explore how the development of transactive memory might be operationalized in the context of enterprise social media, transactive memory was studied through counts of posts of four types of knowledge: domain knowledge, metaknowledge, non-work-related knowledge, and question. Counts of posts were observed using human coders and machine learning. Following exploratory analysis and the validation of assumptions, generalized linear modeling was used to quantify the effects of each of several different variables on counts of each of four different post knowledge types. To Carlyn, Mom, Dad, Sarah, Josiah, Alyssa, Genet, Peter, Lizzy, Daisy, Leslie, Craig, Ali, Jordan, Hunter, Sandy, David, Donald, Gloria, Alexcidius, Roderick, Pastravus, Jonny, Jessi, Juan, Kort, P4B, the Osborns, Laura, Myq, Will & the Golden Boy, Leslie, Jake, Randy, Rook, Ashley, Linda, Kurt, Joe, and Susan. iii ACKNOWLEDGEMENTS Special thanks to Dr. Van Osch, Dr. Steinfield, Dr. Coursaris, and Dr. Zhao; and to my friends Ali, Jordan, Hunter, and Jake for helping me with this project. iv TABLE OF CONTENTS LIST OF TABLES ................................................................................................................................. vi LIST OF FIGURES ..............................................................................................................................vii INTRODUCTION .................................................................................................................................. 1 THEORETICAL BACKGROUND.........................................................................................................2 Transactive Memory Theory ............................................................................................................2 Group Composition and Transactive Memory Development ........................................................ 3 Familiarity and Transactive Memory Development .......................................................................4 Transactive Memory on Enterprise Social Media ...........................................................................4 METHOD ..............................................................................................................................................6 Data Collection ..................................................................................................................................6 Measurement .....................................................................................................................................6 Group Size ...................................................................................................................................... 7 Group Visibility Policy .................................................................................................................. 7 Group Homogeneity ...................................................................................................................... 7 Group Familiarity ......................................................................................................................... 8 Group Volatility .............................................................................................................................9 Post Count for each Knowledge Type ..........................................................................................9 Analysis ............................................................................................................................................ 12 RESULTS ............................................................................................................................................. 16 DISCUSSION ..................................................................................................................................... 20 Limitations ...................................................................................................................................... 21 Future Research .............................................................................................................................. 21 BIBLIOGRAPHY ................................................................................................................................ 23 v LIST OF TABLES Table 1: Operationalization of Variables .............................................................................................................7 Table 2: Post Knowledge Type Examples ......................................................................................................... 11 Table 3: Post Counts for Each Knowledge Type Cross-Tabulated with Group Visibility Policy ................... 12 Table 4: Measures of Central Tendency and Dispersion ................................................................................. 13 Table 5: Model Coefficients – Count of Posts of Any Knowledge Type .......................................................... 16 Table 6: Model Coefficients – Count of Domain Knowledge Posts ................................................................. 17 Table 7: Model Coefficients – Count of Metaknowledge Posts ....................................................................... 17 Table 8: Model Coefficients – Count of Non-Knowledge Posts ......................................................................18 Table 9: Model Coefficients – Count of Question Posts .................................................................................. 19 vi LIST OF FIGURES Figure 1: Post Count Frequencies Fit to Selected Discrete Probability Distribution Families ...................... 14 vii INTRODUCTION Organizations possess stores of knowledge that can be put to work for the advantage of the organization (Alavi and Leidner 2001; Grant 1996). Sources include organization’s own personnel, documents, and information systems. In turn, enterprise social media (ESM) are a class of information and communication technologies (ICTs) that can mitigate this challenge for organizations by affording behaviors that support the circulation of knowledge throughout the organization (Leonardi et al. 2013; Majchrzak et al. 2013). The study aims to contribute to the development of measures transactive memory (TM), one of the explanatory mechanisms that has been used to study whether and how ESM mitigate the challenges associated with knowledge management. The study pursued this aim in the context of an ESM deployed within Global Deskcorp, an office furniture manufacturing firm. 1 THEORETICAL BACKGROUND Transactive Memory Theory Transactive memory (TM) theory explains the behavior of the transactive memory system (TMS), a “set of individual memory systems in combination with the communication that takes place between individuals” (Wegner 1986, p. 186; Wegner et al. 1985). TM research began in studies of how dyads and small groups can come to possess what often seems like a single, unitary mind. From a historical perspective, theoretic and empirical research on TM are part of a movement away from strict Popperian methodological individualism, toward a program in which knowledges—and auxiliary activities like learning, storing, and recalling—are predicable of non-human actors and aggregates of individual actors (Evans and Foster 2011; Hjorland and Albrechtsen 1995; Popper 1972). While early research tended toward lab and field studies of couples and very small groups, TM studies can currently be found in other contexts and in a number of fields, including organizational psychology, human-computer interaction, and information systems. The TMS has a network structure.1 The theory distinguishes between the structural components of TMS and the process components of TMS. The structural component of the TMS is the repository of information about the areas of expertise of each individual memory system that participates in the system. TM is the structural component of the TMS, and it is often figured as a group directory, of sorts; the set of discrete meta-memory propositions in which individual group members are mapped to their respective domains of expertise (Wegner 1986, 1995). The process components of the TMS are the mechanisms by which TM is established and maintained, and they include a) encoding, b) storage, and c) retrieval. Corresponding to these processes of encoding, storage, and retrieval are the more concrete activities of information 1 Computer networking is a common metaphor in the elaboration of TMS theory (Wegner 1995). 2 allocation, directory updating, and retrieval coordination, respectively, through which TMS members operate on and with TM (Wegner 1995). Studies of TMS at the dyadic level and the small-group level cast doubt on the feasibility of organization-level TM due to the moderating effect of group size on TMS effectiveness (Palazzolo et al. 2006; Ren et al. 2006). The question of TM at the organizational level is closely related to the field of organizational learning, in which simulation research modeling groups’ cognitive structures suggests that strong, interconnecting group-level TMS may transcend the level of the group to the level of the organization (Argote 2013; Fang et al. 2010; Ren and Argote 2011). Nevo and Wand proposed a system for supporting the extension of TM to organizations by reifying metaknowledge in an actual, computerized directory, serving as a kind of TM (Nevo and Wand 2005). This idea of externalizing or objectifying TM so that it may serve as a common resource for many people hearkens back to earlier TM theory, which proposed that objects (e.g., documents, notebooks) can serve as memory aids – often serving just as well as people. Group Composition and Transactive Memory Development Antecedents of TMS development have been observed in individual-level characteristics of group members like demographics (Hollingshead and Fraidin 2003; Wegner 1995) and skills and certifications (Bunderson 2003), and also in both group-level and organizational-level inputs (Ren and Argote 2011). Commonality in demographics like age, race, and ethnicity facilitates communication about expertise among group-members, together with factors like role, status, and credentials amounting to a positive influence on TMS development (Wegner 1995). Commonality aside, individual-level characteristics can also varyingly cue or not cue levels and/or domains of expertise, as has been observed of gender (Hollingshead and Fraidin 2003) and ethnicity (Bunderson 2003). Mutual category membership supports the development of common ground (Cramton 2001; Krauss and Fussell 1990). Where there is common ground among group-members, members are more likely to engage in the communicative processes through which TMS are established and maintained (Hwang et al. 2015). 3 Familiarity and Transactive Memory Development While the impact of member familiarity on work group performance, generally, has been consistently shown, conflicting results are found in research on the effect on TM development, specifically, of interpersonal relationships predating group formation (Goodman and Leyden 1991; Hinds et al. 2000). For work group performance on a more general level, familiarity has been observed both to relate positively to group-members’ awareness of each other’s expertise (Akgun et al. 2005; He et al. 2007; Littlepage et al. 1997) and to exert no significant effect on TMS development (Jackson and Moreland 2009; Michinov and Michinov 2009). On the other hand, in a longitudinal study of groups and TMS development, Lewis found that familiarity moderated the relationship between the distribution of expertise in early phases of the study and later emergence of a TMS (Lewis 2004). Ren and Argote reconcile the conflicting findings by theorizing that familiarity is related to the structural components of the TMS (i.e., the TM) and not to the component processes of the TMS (Ren and Argote 2011). Transactive Memory on Enterprise Social Media Enterprise social media (ESM) is the class of information and communication technologies (ICTs) that provide a platform for workplace collaboration and socialization (Leonardi et al. 2013). With TM theory’s focus on mechanisms that facilitate shared knowledge, proposed ESM affordances like visibility, persistence, editability, and association make TM theory well-suited for ESM research (Fulk and Yuan 2013; Majchrzak et al. 2013; Treem and Leonardi 2013). However, while contexts in which TM or TMS are observed to develop include work groups—both co-located (Bunderson 2003; Liang et al. 1995; Moreland and Myaskovsky 2000) and, to a lesser extent, virtual (Ariff et al. 2011; Shen 2007; Yoo and Kanawattanachai 2001)—studies of ESM as a context for the development of TM are scarce. Studies of TM on ESM that primarily observe ESM users’ perceptions of other ESM users’ expertise areas and levels yield useful findings (including the abovementioned affordances), but, proceeding mainly 4 in interviews, are by necessity limited to small samples and to questions of individuals’ perceptions based on anecdotes (Treem and Leonardi 2013, 2015). In prior studies, TM is assumed to develop, somehow, on ESM, and so the main aim of the present study is to examine whether and how TM development proceeds using quantitative methods. Following studies in which TM is observed naturalistically (i.e., in the settings in which it supposed to emerge) and measured quantitatively in verbal, nonverbal, written, and computer-mediated communications, the study aims to explore how TM can be observed on ESM. In doing so, the study offers a theory of ESM-enabled TM encoding to account for the development of TM on ESM, and also demonstrates how ESM-enabled TM encoding might be observed through ESM data. The ESM affordances of visibility, persistence, and network- informed associating2, in particular, support an ESM-based theory of TM encoding, or directory development (Yuan et al. 2010). The creation of posts and comments constitute signals of the knowledge contours of authors’ individual memory system. These communicative activities are visible to connected ESM users after creation, and are also persistent in the system such that users absent at the time of the creation of a post or comment may, even so, observe these signals at a later date (Leonardi et al. 2013). 2 Both Treem and Leonardi (2013) and Majchrzak and colleagues (2013) describe ESM’s affordance of activities that are informed by the social networks of users. The affordance named by Majchrzak and colleagues is more descriptive and thus is used hereafter in the study. 5 METHOD The study was designed as a content analysis, carried out first by human coding and subsequently by an algorithm trained on the human coding. Before turning to operationalization, the case organization and the ESM system from which data were collected will be described. Data Collection The study used an export of four years of relational log data (2012 to 2016) from Inspire, the ESM solution for intra-organizational communication and collaboration, which was obtained as part of larger, ongoing project. Inspire belonged to a multinational office furniture company called Global Deskcorp, and the Inspire dataset traced interactions of 12,894 Global Deskcorp employees. The set of all employees on Inspire was itself a subset of Global Deskcorp employees, since employees are not required to maintain a presence on Inspire. A discussion of sampling techniques is not necessary, since the dataset constituted a census of the activities of all Inspire users affiliated with at least one group. The data afforded a longitudinal view of the activities of three Inspire entities: groups, posts, and users. The Inspire data model allowed users to freely create, join, or leave groups, subject to a group’s policy. Types of groups included open, members only, private, and secret, and these policies controlled the visibility of posts and other user activities contained therein. The collected post data consisted of content and metadata fields, including but not limited to post unique identifiers, creation and modification timestamps, Inspire-feature-specific fields, and identifiers representing relationships with other Inspire entities (e.g., the user-post authorship relation). Measurement All variables were measured at the group level of analysis, using measures devised to make sense of the available data: 18851 posts occurring in 520 groups on the Inspire ESM 6 platform – posts authored by any of the 5928 users who associated in the groups as members (there were 25999 memberships). Table 1: Operationalization of Variables Name Post Knowledge Type Count Type Discrete (count) Familiarity Continuous Homogeneity Continuous Continuous Discrete (count) Volatility Group Size Group Visibility Policy Group Size Operationalization The count of posts in an ESM group at a given point in time that are labeled as domain knowledge, metaknowledge, non-knowledge, or question. The per-interval mean of the number of unique users in a group who have been co-members of a group before. The mean of the per-interval group entropy score along five user dimensions: gender, city, country, department, position. The standard deviation of the count of members added to a group over some series of intervals. The count of members in a group at a given point in time. Discrete (categorical) The type of visibility policy of an ESM group. Group size was measured as the count of members in each group at the end of data collection. Group Visibility Policy Group visibility policy was measured a categorical variable (members only, open, private, and secret). This variable was indicative of the type of a group, in terms of the visibility policy to which activities were subject. Group Homogeneity Homogeneity was measured at the group level of analysis as a continuous variable and represented the degree of uniformity of an ESM group. Group homogeneity was considered along five user dimensions: gender, city, country, department, position. Group homogeneity was operationalized as the mean of the group’s entropy score along each dimension; lower group 7 homogeneity scores were indicative of higher homogeneity, and vice versa. To measure homogeneity for a given group, a table expressing the many-to-many associations between groups and users was aggregated to yield a list of what were, in essence, rosters of users for each group at each interval. Each roster’s set of user identifiers were joined with user attributes, and entropies were calculated along the user dimensions of gender, city, country, department, position (Pedregosa et al. 2011). These entropy scores were applied to each group on the ESM, both overall and incrementally (quarterly, monthly, and daily). Group Familiarity Familiarity among Inspire users was observed at the group level of analysis to represent the degree of familiarity among members in the group. Prior measurement techniques have treated familiarity among group members largely as dichotomous. In a lab setting, Littlepage and colleagues measured group-members’ familiarity after randomly assigning participants to conditions that determined whether they would receive prior group experience or not (Littlepage et al. 1997). Similarly, applied to the context of a project-based group exercise over time, He and colleagues surveyed at group formation each member about how well they knew others, vice versa, and their rating of the group’s interpersonal knowledge (He et al. 2007). The set of groups were halved at the median, designating one group as familiar and the other unfamiliar. Since these operationalizations of group familiarity are well suited neither to capture the dynamical, networked nature of social relations on ESM, nor to tap into a broader notion of familiarity suitable for describing what happens among individuals interacting on large ESM systems, group familiarity was observed by counting a group’s users at any given point in time who were co-members of a group prior to the formation of the group under consideration. The measurement procedure to generate familiarity score involved grouping group-user associations to yield a list of user rosters for each group. Enumerating the user rosters of all groups, unique combinations of each roster’s set of users were generated. Next, the user rosters of groups 8 predating the input group were searched for occurrences of all such user combinations. A user co-membership matrix was used throughout as both a heuristic device and a sanity-check, ensuring scoring accuracy at subsequent steps in the procedure. Group Volatility Group volatility was a continuous variable that represented the degree of group-member volatility experienced by groups over the course of data collection. Here, volatility referred to volatility in the membership of a group. Groups that more frequently experienced periods in which many new users joined received higher volatility scores. The operationalization of group volatility was adapted from the annualized standard deviation of returns, commonly used for quantifying absolute volatility in securities pricing (Brien et al. 2010). Volatility was observed for each group both overall and incrementally (quarterly, monthly, and daily). Post Count for each Knowledge Type The post count for each knowledge type was a count variable observed at the group level of analysis. The posts that appeared in a group were assigned knowledge types and then counted, both overall and incrementally (quarterly, monthly, and daily). Given the division of higher and lower orders of cognition in TM theory, and in keeping with previous studies where TM encoding is the object of study, measures of post knowledge type were developed in order to serve the measurement of TM encoding (Rulke and Rau 2000; Yuan et al. 2010). Knowledges map domains, and where a domain is, itself, a knowledge and the conditions, components, contours, or mechanisms of that knowledge, it is metaknowledge.3 In the context of ESM, metaknowledge refers to awareness of some a domain knowledge instance’s quality or quantity, like breadth, availability, or location (e.g., the identity of the individual person, document, wiki, or message thread in which a piece of domain knowledge is contained on the ESM platform). 3 Further decomposition of metaknowledge into a type-subtype hierarchy is possible, but the bifurcation of knowledges into metaknowledge and domain knowledge were sufficient for the design of the study (cf. Herrmann et al. 2003). 9 Expressions of metaknowledge or domain knowledge in ESM posts and comments were counted as the codification of metaknowledge or domain knowledge held by post/comment authors. Table 2 shows model expressions of metaknowledge and domain knowledge in ESM posts and comments. In summary, counts of posts for each knowledge type were used to proxy aspects of TM encoding in the group in which the posts were produced. A coding scheme of five categories was devised to observe the knowledge types of ESM posts. All ESM content was coded as containing expressions of: 1) metaknowledge; 2) domain knowledge; 3) both domain knowledge and metaknowledge; 4) questions and/or content of a predominantly interrogative character; or 5) non-work-related knowledge and/or non- knowledge expressions. Domain knowledge and metaknowledge posts were especially important post knowledge types. TM is supposed to provide metaknowledge to participants; the more robust is the TM, the better equipped are group members to account for and subsequently act upon the knowledge of any other members. Assuming group members refrain from creating posts in pursuit of domain knowledge when they already know where to find it or how to access it (e.g., directly messaging an identified expert), one might expect an increase in encoded TM to coincide with a decrease in the creation of posts containing domain knowledge (as a function of a decrease in the creation of posts that warrant the expression of domain knowledge). ESM features (more precisely, activities afforded by the ESM) are expected to reroute demands for domain knowledge to the store of knowledge on the ESM. In short, when users are able to search a larger store of previously-asked, previously-answered questions, their ESM-enabled communicative activities should less frequently contain expressions of lower-order, domain- level knowledge and more frequently contain expressions of higher-order knowledge, or metaknowledge. The application of the coding scheme to the ESM messaging data involved a qualitative content analysis followed by an algorithmic content analysis. The former was organized as a process in which content coders started with small samples of the larger dataset, so as to enable 10 responsive gauging of the reliability of the coding scheme, with measures of reliability taken in Krippendorff’s alpha and Cohen’s kappa statistics, both widely used in the literature for evaluating content analyses, and both indicating the reliability of measures at values above .70 (De Wever et al. 2006; Lombard et al. 2002; Neuendorf 2002). For the qualitative content analysis, four content coders were recruited. Due to the subtle linguistic distinctions in the coding scheme, individuals with at least college degrees were sought out. The content coders were introduced to the scheme in two training sessions, where they were provided with definitions and examples for each post knowledge type. The process of sampling content data, tasking coders with content data, evaluating the reliability of the coding scheme, and reviewing disagreements was repeated twice before the coding scheme was ultimately found to be reliable (α > 80%). Table 2: Post Knowledge Type Examples Category Non-Knowledge 0 Code Share 37.81% Example “Hi yes I love this!!” Metaknowledge 1 12.70% “Hi Gerri - if you are willing there is some great help available in the Help Space (top, right corner in the header). Here you will find user guides (and their short, I promise) as well as FAQ's. This being said Search is a powerful thing within Inspire…chances are you will find answers to your questions in one of these areas....if not or you find it too frustrating, contact me by searching for me and sending me a direct message from my profile page (link found under my photo). You can do it Gerri!” Domain Knowledge Both Metaknowledge and Domain Knowledge Question 2 3 4 14.68% n/a 34.81% “Not yet. It will be available in March through a special anniversary section of globaldeskcorp.com” “Hi Randy - the short answer to this is yes, we do indeed, hope to use AD groups (or attributes) to populate permission groups in Inspire. I've shared this discussion with because he is testing this process now.” “Are we able to use this mobile app? App Store - Inspire SBS Mobile” Programmatic application of the categorization decisions made by human coders in the first stage of the content analysis was then made possible by tokenizing message text in the 11 dataset, excluding stopwords (viz., conjunctions, articles, and other commonly occurring parts of speech) (Pedregosa et al. 2011). Following preprocessing, labeling the entire dataset was completed through three tasks; moreover, each iteration of the labeling task was comprised of three steps: 1) feature extraction, 2) feature selection, and 3) classification. First, the task was to label messages as either originating questions (code 4) or not (codes 0, 1 and 2); question marks were found to be a useful indicator in making this division. Second, the task was to label messages as either non-knowledge related (code 0) or not (codes 1 and 2). Finally, the task was to label messages as either metaknowledge (code 1) or content-knowledge (code 2). The ultimate distribution of the population of the ESM content is provided in Table 2. In each of these classification tasks, feature extraction assigned each token a number representing the token’s probability of predicting any class. With the dataset represented as a hash table associating tokens to probability values, it was possible to select a subset of these tokens using a threshold probability as the criterion for selection. With the resultant subset, another data structure was created where each message in the dataset was assigned an actual token occurrence count. A matrix containing 70% of the dataset was used to train the classifier, and the fitted model was tested on the remaining 30% of the dataset. As mentioned above, the model was optimized by adjusting parameters of training data (e.g., adjusting the prediction- quality threshold below which tokens were excluded in feature selection) and hyper-parameters of the algorithm itself. Fitted models were evaluated using the f-score, which captures both precision and recall in the labeling of a model (as distinct from accuracy, alone). The algorithm used was a simple two-layer neural network (Pedregosa et al. 2011). Analysis Table 3: Post Counts for Each Knowledge Type Cross-Tabulated with Group Visibility Policy members only open private secret domain knowledge 1471 358 251 696 metaknowledge 1075 245 323 755 non-knowledge 3988 985 568 1599 question 2793 717 873 2154 12 The generalized linear model (GLM) technique was used to test for effects exerted on post counts by the other variables in the dataset (R Core Team 2017; Venables and Ripley 2002). GLM is appropriate where the dependent variable in a model is expected not to have been generated from the normal distribution, as tends to be the case where the dependent variable is, say, a binary outcome or a count of incidences of some event (Upton and Cook 2014). While count data commonly call for use of the Poisson distribution to supply the error term for the GLM, this distribution is not optimal where the data present a high degree of dispersion. The variance and the mean of Poisson-distributed counts are expected to be identical, but variance will exceed the mean in highly dispersed count data. Table 4: Measures of Central Tendency and Dispersion Variable Group Homogeneity (N = 520) Group Familiarity (N = 520) Group Volatility (N = 520) Any Knowledge Type Post Count (N = 18851) Domain Knowledge Post Count (N = 2803) Metaknowledge Post Count (N = 2426) Non-Knowledge Post Count (N = 7177) Question Post Count (N = 6581) Group Size (N = 520) Mean 0.71 40.36 17.66 36.25 5.33 4.61 13.73 12.57 50.00 SD 0.44 59.84 27.52 99.56 15.51 12.97 47.88 29.73 76.09 Since GLM requires specification of a probability distribution from which the dependent variable is supposed to be generated, visual inspection of observed frequency counts against expected frequency counts from a selection of probability distribution families was used to guide model specification (see Figure 1) (Delignette-Muller and Dutang 2015; Rigby and Stasinopoulos 2005). In this exploratory analysis, the counts of posts for each knowledge type were best approximated by the negative binomial distribution (Venables and Ripley 2002). For validation, however, GLMs with the random component drawn from Quasi-Poisson and Poisson distributions were also analyzed. 13 Figure 1: Post Count Frequencies Fit to Selected Discrete Probability Distribution Families Each of the post count variables – the counts of incidence for each of the four knowledge types, and one overall count of posts – were regressed onto the following predictors using GLM: group visibility policy, group size at the end of data collection, the group homogeneity score, the group familiarity score, and the group volatility score (Venables and Ripley 2002). The best- supported models for each of the five sets of GLM regressions are reported below. The regressions used the negative binomial probability distribution family with the log link function, and, in order to facilitate interpretation, incident rate ratios were generated from the logarithmic regression coefficients through exponentiation. The GLM regressions were conducted iteratively, by evaluating all combinations of the intercept and the predictor terms and ranking each of the five sets of models according to the Akaike information criterion (AIC) as a metric of relative goodness of fit (Burnham et al. 2002). Within each model, Wald’s test (producing a z value) was used to ascertain the relative importance of predictors (similar to role of the t value in summaries of ordinary linear regression models). 14 In summary, data analysis involved exploratory analysis, the validation of assumptions, and GLM regressions. On the whole, the post count data were over-dispersed. The study addressed the problem of over-dispersion among count data by 1) employing probability distributions better suited to such conditions than the Poisson; and, 2) grouping count observations in terms of the categorical variable present at the group level of analysis: group visibility policy. Measures of dispersion and of central tendency (see Table 4) were also important for selecting subsequent modeling techniques. 15 RESULTS All of the best-supported models for the five sets of GLM regressions yielded at least three statistically significant regression coefficients. Group volatility was not used as a term in any of the best-supported models. For the visibility policy term in the models, the reference category was taken to be the members-only type. Estimating the dispersion parameter of post counts of any knowledge type with maximum likelihood estimation (theta = 0.601; std. err = 0.033), the best-supported model (AIC = 4348; see Table 5) indicated that, for private groups, the rate of incidence of post counts for any knowledge type (M = 36.25; SD = 99.56) was 0.549 times the rate of incidence of post counts (p < 0.001) in the reference type of group. This model also demonstrated 2.9% and 261% increases in the rate of incidence of posts counts for any knowledge type for a one-unit increase in group size (p <0.001) and group homogeneity (p <0.001), respectively. For a one-unit increase in group familiarity, the model indicated a 2.7% decrease in the rate of incidence of the dependent variable (p <0.001). Table 5: Model Coefficients – Count of Posts of Any Knowledge Type Term β Intercept 2.315 Group Size 0.028 Group Familiarity -0.027 Group Homogeneity 0.958 Group Visibility Policy (Open) 0.061 Group Visibility Policy (Private) -0.599 Group Visibility Policy (Secret) 0.248 exp(β) 10.129 1.029 0.973 2.607 1.063 0.549 1.282 z value 14.401 6.270 -4.668 5.906 0.299 -3.877 1.598 p value <0.001 <0.001 <0.001 <0.001 0.765 <0.001 0.110 For domain knowledge post counts, estimating the dispersion parameter with maximum likelihood estimation (theta = 0.341; std. err. = 0.026), the best-supported model (AIC = 2340; see Table 6) indicated that the rate of incidence of domain knowledge post counts (M = 5.33; SD = 15.51) for private groups was 0.419 times the rate of incidence of post counts in the reference type of group (p < 0.001). This model also indicated 2.6% and 289% increases in the rate of 16 incidence of domain knowledge posts counts for a one-unit increase in group size (p <0.001) and group homogeneity (p <0.001), respectively. For a one-unit increase in group familiarity, the model indicated a 2.5% decrease in the rate of incidence of the dependent variable (p < 0.01). Table 6: Model Coefficients – Count of Domain Knowledge Posts β Term 0.438 Intercept 0.026 Group Size -0.025 Group Familiarity 1.063 Group Homogeneity 0.106 Group Visibility Policy (Open) Group Visibility Policy (Private) -0.87 Group Visibility Policy (Secret) 0.025 exp(β) 1.549 1.026 0.975 2.896 1.111 0.419 1.025 z value 1.955 4.312 -3.201 4.700 0.381 -4.018 0.117 p value 0.051 <0.001 0.001 <0.001 0.703 <0.001 0.907 Estimating the dispersion parameter of metaknowledge post counts with maximum likelihood estimation (theta = 0.301; std. err. = 0.024), the best-supported model (AIC = 2207; see Table 7) did not estimate statistically significant coefficients for any group visibility policy. However, as in the previous two models, the model for metaknowledge counts indicated 3.0% and 404% increases in the rate of incidence for a one-unit increase in group size (p <0.001) and group homogeneity (p <0.001), respectively. For a one-unit increase in group familiarity, the model indicated a 3.0% decrease in the rate of incidence of posts (p <0.001). Table 7: Model Coefficients – Count of Metaknowledge Posts Term Intercept Group Size Group Familiarity Group Homogeneity Group Visibility Policy (Open) Group Visibility Policy (Private) Group Visibility Policy (Secret) β exp(β) -0.089 0.915 0.029 1.030 0.970 -0.03 4.044 1.397 0.068 1.070 0.714 -0.337 1.408 0.342 z value -0.371 4.561 -3.66 5.804 0.228 -1.476 1.497 p value 0.711 <0.001 <0.001 <0.001 0.820 0.140 0.134 17 Estimating the dispersion parameter of non-knowledge post counts with maximum likelihood estimation (theta = 0.328; std. err. = 0.022), the best-supported model (AIC = 3042; see Table 8) indicated that, for private groups, the rate of incidence of non-knowledge post counts (M = 13.73; SD = 47.88) was 0.374 times the rate of incidence of post counts in the reference type of group (p < 0.001). The model for non-knowledge counts indicated 3.2% and 224% increases in the rate of incidence for a one-unit increase in group size (p <0.001) and group homogeneity (p <0.001), respectively. For a one-unit increase in group familiarity, the model indicated a 2.9% decrease in the rate of incidence of posts (p <0.001). Table 8: Model Coefficients – Count of Non-Knowledge Posts Term Intercept Group Size Group Familiarity Group Homogeneity Group Visibility Policy (Open) Group Visibility Policy (Private) Group Visibility Policy (Secret) β 1.433 0.031 -0.029 0.807 -0.041 -0.982 0.05 exp(β) 4.193 1.032 0.971 2.241 0.96 0.374 1.051 z value 6.522 5.17 -3.711 3.631 -0.148 -4.625 0.236 p value <0.001 <0.001 <0.001 <0.001 0.882 <0.001 0.814 Finally, for counts of posts containing questions, the dispersion parameter was estimated using maximum likelihood estimation (theta = 0.745; std. err. = 0.044), and the best-supported generalized linear regression model (AIC = 3451; see Table 9) indicated that, for groups of the private visibility policy and of the secret visibility policy, the rates of incidence of counts of posts of the question knowledge type (M = 12.57; SD = 29.73) were 0.747 and 1.598 times the rate of incidence of post counts in the reference type of group, respectively (p < 0.05; p = 0.001). Further, the model indicated 2.5% and 240% increases in the rate of incidence for a one-unit increase in group size (p <0.001) and group homogeneity (p <0.001), respectively. For a one- unit increase in group familiarity, the model indicated a 2.4% decrease in the rate of incidence of posts (p <0.001). 18 Table 9: Model Coefficients – Count of Question Posts Term Intercept Group Size Group Familiarity Group Homogeneity Group Visibility Policy (Open) Group Visibility Policy (Private) Group Visibility Policy (Secret) β 1.307 0.025 -0.024 0.876 0.1 -0.291 0.469 exp(β) 3.694 1.025 0.976 2.40 1.105 0.747 1.598 z value 8.731 6.041 -4.53 5.832 0.533 -2.028 3.269 p value <0.001 <0.001 <0.001 <0.001 0.594 0.043 0.001 19 DISCUSSION This study proposed a novel approach to observing TM in trace data generated in the context of digitally-mediated workplace activities. Moreover, the study endeavored to contribute to studies of TM and of ESM by implementing the approach and testing several theory-derived expectations of how TM should operate. Several findings resulted from the study. With regard to group visibility policy, groups with the members-only visibility policy were observed to exhibit greater rates of post count incidence than groups with the private visibility policy in four of the five models (the exception being the model for metaknowledge post counts). The rate of incidence of question posts in groups with the secret visibility policy was the only other result pertaining to group visibility policy; there were no findings of any statistical significance pertaining to open groups. In all five of the post count GLM specifications, the continuous variables of the study were significant predictors of the respective count variable. While the coefficients appeared relatively stable across models for group size and group familiarity, the group homogeneity coefficient varied a great deal more in the model of metaknowledge post counts. For group homogeneity, in particular, results ran somewhat counter to the expectation that, thanks to common ground, groups of people with high degrees of sameness in terms of their work department, job title, or job position should communicate with higher rates of metaknowledge posts (or domain knowledge posts, for that matter). One might interpret the higher rate of incidence of question posts in secret groups as consistent with the assumption that these types of groups – by virtue of their invisibility – are able to elicit open-ended dialogue in a way that is different from other types of groups. By the same token, though, one might also ask why higher rates of incidence of similar knowledge types were not observed among secret groups. While it was not explicitly hypothesized due to the exploratory nature of the study, one might have expected to find in these data higher or lower incidence of posts containing metaknowledge where groups exhibited higher or lower degrees of 20 group volatility. The influx of new group members should elicit invocations of metaknowledge as they acclimate to the group and begin to ascertain the quality and quantity of the expertise available in the group. In the same way, one might expect groups with higher group familiarity to experience less of the disruptive effect of new member influx, with familiarity serving as a sort of buffer against group volatility. These expectations should be explored in future research. Limitations The conclusions presented above bear on TM strictly as it is enacted on ESM platforms, since communications are only observed on the ESM, and not any other channels. Private, direct communications among pairs or multiple users were not observed. This however is a consequence of the study’s aim to observe transactive memory in an empirically under-studied context. The study was intentionally limited to work-related knowledge, and future research might consider how to refactor TM measurement to include measures of the other knowledge types in the ESM posts (here, originating questions and work-unrelated knowledge). The present study strictly focused on posts as resources for work-related knowledge and knowledge sharing, and not social topics or other topics unrelated to work. Finally, the study’s operationalization of group familiarity assumes a kind of automatically developing familiarity among individuals, which does not exactly square with reality; comembership in the same ESM group is not necessarily familiarity. Future Research The dynamic network analytic approach, in particular, lends itself to the sort of measurement challenges that were addressed, in part, in the present study. The dynamic network approach enables analysis of both network and sub-network state snapshots as well as the forwards and backwards paths leading from one to the other. Dynamic network analysis might also render the measurement of organization-level TM more feasible in the future. While TMS are at times theorized to exist insofar as participant individuals transact amongst themselves, and it has been argued that TMS will not likely cohere around organizational forms 21 larger than medium-sized groups (let alone a multinational corporation), the prospects for the study of the development of TM at the organizational level of analysis are exciting (Anand et al. 1998; Argote and Ren 2012; Kirchhoff 2016; Lewis 2003). Finally, future research might explore better approaches to dealing with over-dispersion in count data (e.g., hurdle models), a challenge that will very likely persist given the development of tools and techniques to process very large datasets. Lastly, insofar as studies of TM on ESM focus on the expertise directory component of TM, they intersect with studies of expertise awareness informed by other theoretical frameworks and with research on the development of expertise recommendation systems (Guy 2015; Guy et al. 2013; Zhang et al. 2007). In endeavoring to study TM in a novel context, the study offers contributions to the development of measures of TM, yielding findings of potential interest for researchers and for managers thinking about the value of ESM for an organization. Moreover, the study offers a demonstration of the use of both qualitative and quantitative research methods to study a phenomenon that most often has been studied qualitatively and/or in a smaller-scale context. 22 BIBLIOGRAPHY 23 BIBLIOGRAPHY Akgun, A. E., Byrne, J., Keskin, H., Lynn, G. S., and Imamoglu, S. Z. 2005. “Knowledge Networks in New Product Development Projects: A Transactive Memory Perspective,” Information & Management (42:8), pp. 1105–1120. (https://doi.org/10.1016/j.im.2005.01.001). Alavi, M., and Leidner, D. E. 2001. “Review: Knowledge Management and Knowledge Management Systems: Conceptual Foundations and Research Issues,” MIS Quarterly; Minneapolis (25:1), pp. 107–136. Anand, V., Manz, C. C., and Glick, W. H. 1998. “An Organizational Memory Approach to Information Management,” Academy of Management. The Academy of Management Review; Briarcliff Manor (23:4), pp. 796–809. Argote, L. 2013. Organizational Learning: Creating, Retaining, and Transferring Knowledge, Boston, MA: Springer US. (http://link.springer.com/10.1007/978-1-4614-5251-5). Argote, L., and Ren, Y. 2012. “Transactive Memory Systems: A Microfoundation of Dynamic Capabilities,” Journal of Management Studies (49:8), pp. 1375–1382. Ariff, M., Milton, S. K., Bosua, R., and Sharma, R. 2011. Exploring the Role of ICT in the Formation of Transactive Memory Systems in Virtual Teams, presented at the PACIS 2011 - 15th Pacific Asia Conference on Information Systems: Quality Research in Pacific. Brien, M. J., Panis, C. W., and Karthik Padmanabhan, M. B. A. 2010. “Volatility Metrics for Mutual Funds,” Washington, DC: US Department of Labor. Retrieved from Http://Www. Dol. Gov/Ebsa/Pdf/Deloitte2009-3. Pdf. (http://aacg.com/wp- content/uploads/2014/09/Volatility-Metrics-for-Mutual-Funds.pdf). Bunderson, J. S. 2003. “Recognizing and Utilizing Expertise in Work Groups: A Status Characteristics Perspective,” Administrative Science Quarterly (48:4), pp. 557–591. (https://doi.org/10.2307/3556637). Burnham, K. P., Anderson, D. R., and Burnham, K. P. 2002. Model Selection and Multimodel Inference: A Practical Information-Theoretic Approach, (2nd ed.), New York: Springer. Cramton, C. D. 2001. “The Mutual Knowledge Problem and Its Consequences for Dispersed Collaboration,” Organization Science (12:3), pp. 346–371. (https://doi.org/10.1287/orsc.12.3.346.10098). De Wever, B., Schellens, T., Valcke, M., and Van Keer, H. 2006. “Content Analysis Schemes to Analyze Transcripts of Online Asynchronous Discussion Groups: A Review,” Computers & Education (46:1), pp. 6–28. (https://doi.org/10.1016/j.compedu.2005.04.005). 24 Delignette-Muller, M. L., and Dutang, C. 2015. “Fitdistrplus: An R Package for Fitting Distributions,” Journal of Statistical Software (64:4), pp. 1–34. Evans, J. A., and Foster, J. G. 2011. “Metaknowledge,” Science (331:6018), pp. 721–725. (https://doi.org/10.1126/science.1201765). Fang, C., Lee, J., and Schilling, M. A. 2010. “Balancing Exploration and Exploitation Through Structural Design: The Isolation of Subgroups and Organizational Learning,” Organization Science (21:3), pp. 625–642. Fulk, J., and Yuan, Y. C. 2013. “Location, Motivation, and Social Capitalization via Enterprise Social Networking,” Journal of Computer-Mediated Communication (19:1), pp. 20–37. (https://doi.org/10.1111/jcc4.12033). Goodman, P. S., and Leyden, D. P. 1991. “Familiarity and Group Productivity,” Journal of Applied Psychology (76:4), pp. 578–586. Grant, R. M. 1996. “Toward a Knowledge-Based Theory of the Firm,” Strategic Management Journal (17), pp. 109–122. Guy, I. 2015. “Social Recommender Systems,” in Recommender Systems Handbook, F. Ricci, L. Rokach, and B. Shapira (eds.), Springer US, pp. 511–543. (https://doi.org/10.1007/978- 1-4899-7637-6_15). Guy, I., Avraham, U., Carmel, D., Ur, S., Jacovi, M., and Ronen, I. 2013. “Mining Expertise and Interests from Social Media,” in Proceedings of the 22nd International Conference on World Wide Web, ACM, pp. 515–526. (http://dl.acm.org/citation.cfm?id=2488434). He, J., Butler, B. S., and King, W. R. 2007. “Team Cognition: Development and Evolution in Software Project Teams,” Journal of Management Information Systems (24:2), pp. 261–292. Herrmann, T., Kienle, A., and Reiband, N. 2003. “Meta-Knowledge: A Success Factor for Computer-Supported Organizational Learning in Companies,” Educational Technology & Society (6:1), pp. 9–13. Hinds, P. J., Carley, K. M., Krackhardt, D., and Wholey, D. 2000. “Choosing Work Group Members: Balancing Similarity, Competence, and Familiarity,” Organizational Behavior and Human Decision Processes (81:2), pp. 226–251. (https://doi.org/10.1006/obhd.1999.2875). Hjorland, B., and Albrechtsen, H. 1995. “Toward a New Horizon in Information Science: Domain-Analysis,” Journal of the American Society for Information Science (46:6), pp. 400–426. 25 Hollingshead, A. B., and Fraidin, S. N. 2003. “Gender Stereotypes and Assumptions about Expertise in Transactive Memory,” Journal of Experimental Social Psychology (39:4), pp. 355–363. (https://doi.org/10.1016/S0022-1031(02)00549-8). Hwang, E. H., Singh, P. V., and Argote, L. 2015. “Knowledge Sharing in Online Communities: Learning to Cross Geographic and Hierarchical Boundaries,” Organization Science (26:6), pp. 1593–1611. (https://doi.org/10.1287/orsc.2015.1009). Jackson, M., and Moreland, R. L. 2009. “Transactive Memory in the Classroom,” Small Group Research (40:5), pp. 508–534. (https://doi.org/10.1177/1046496409340703). Kirchhoff, M. D. 2016. “Composition and Transactive Memory Systems,” Philosophical Explorations (19:1), pp. 59–77. (https://doi.org/10.1080/13869795.2016.1085593). Krauss, R., and Fussell, S. 1990. “Mutual Knowledge and Communicative Effectiveness,” in Intellectual Teamwork: Social and Technological Foundations of Cooperative Work, J. Galegher, R. E. Kraut, and C. Egido (eds.), Hillsdale, N.J: Lawrence Erlbaum, pp. 111– 146. Leonardi, P. M., Huysman, M., and Steinfield, C. 2013. “Enterprise Social Media: Definition, History, and Prospects for the Study of Social Technologies in Organizations,” Journal of Computer-Mediated Communication (19:1), pp. 1–19. (https://doi.org/10.1111/jcc4.12029). Lewis, K. 2003. “Measuring Transactive Memory Systems in the Field: Scale Development and Validation.,” Journal of Applied Psychology (88:4), pp. 587–604. (https://doi.org/10.1037/0021-9010.88.4.587). Lewis, K. 2004. “Knowledge and Performance in Knowledge-Worker Teams: A Longitudinal Study of Transactive Memory Systems,” Management Science (50:11), pp. 1519–1533. (https://doi.org/10.1287/mnsc.1040.0257). Liang, D. W., Moreland, R., and Argote, L. 1995. “Group Versus Individual Training and Group Performance: The Mediating Role of Transactive Memory,” Personality and Social Psychology Bulletin (21:4), pp. 384–393. (https://doi.org/10.1177/0146167295214009). Littlepage, G. E., Robison, W., and Reddington, K. 1997. “Effects of Task Experience and Group Experience on Group Performance, Member Ability, and Recognition of Expertise,” Organizational Behavior and Human Decision Processes (69:2), pp. 133–147. (https://doi.org/10.1006/obhd.1997.2677). Lombard, M., Snyder-Duch, J., and Bracken, C. C. 2002. “Content Analysis in Mass Communication: Assessment and Reporting of Intercoder Reliability,” Human Communication Research (28:4), pp. 587–604. 26 Majchrzak, A., Faraj, S., Kane, G. C., and Azad, B. 2013. “The Contradictory Influence of Social Media Affordances on Online Communal Knowledge Sharing,” Journal of Computer- Mediated Communication (19:1), pp. 38–55. (https://doi.org/10.1111/jcc4.12030). Michinov, N., and Michinov, E. 2009. “Investigating the Relationship between Transactive Memory and Performance in Collaborative Learning,” Learning and Instruction (19:1), pp. 43–54. (https://doi.org/10.1016/j.learninstruc.2008.01.003). Moreland, R. L., and Myaskovsky, L. 2000. “Exploring the Performance Benefits of Group Training: Transactive Memory or Improved Communication?,” Organizational Behavior and Human Decision Processes (82:1), pp. 117–133. (https://doi.org/10.1006/obhd.2000.2891). Neuendorf, K. A. 2002. The Content Analysis Guidebook, Thousand Oaks, CA: SAGE Publications. Nevo, D., and Wand, Y. 2005. “Organizational Memory Information Systems: A Transactive Memory Approach,” Decision Support Systems (39:4), Collaborative Work and Knowledge Management, pp. 549–562. (https://doi.org/10.1016/j.dss.2004.03.002). Palazzolo, E. T., Serb, D. A., She, Y., Su, C., and Contractor, N. S. 2006. “Coevolution of Communication and Knowledge Networks in Transactive Memory Systems: Using Computational Models for Theoretical Development,” Communication Theory (16:2), pp. 223–250. (https://doi.org/10.1111/j.1468-2885.2006.00269.x). Pedregosa, F., Varoquaux, G., Gramfort, A., Michel, V., Thirion, B., Grisel, O., Blondel, M., Prettenhofer, P., Weiss, R., Dubourg, V., Vanderplas, J., Passos, A., Cournapeau, D., Brucher, M., Perrot, M., and Duchesnay, E. 2011. “Scikit-Learn: Machine Learning in Python,” Journal of Machine Learning Research (12), pp. 2825–2830. Popper, K. R. 1972. Objective Knowledge: An Evolutionary Approach, (Revised edition.), Oxford Eng. : New York: Oxford University Press. R Core Team. 2017. R: A Language and Environment for Statistical Computing, Vienna, Austria: R Foundation for Statistical Computing. (https://www.R-project.org/). Ren, Y., and Argote, L. 2011. “Transactive Memory Systems 1985–2010: An Integrative Framework of Key Dimensions, Antecedents, and Consequences,” The Academy of Management Annals (5:1), pp. 189–229. (https://doi.org/10.1080/19416520.2011.590300). Ren, Y., Carley, K. M., and Argote, L. 2006. “The Contingent Effects of Transactive Memory: When Is It More Beneficial to Know What Others Know?,” Management Science (52:5), pp. 671–682. 27 Rigby, R. A., and Stasinopoulos, D. M. 2005. “Generalized Additive Models for Location, Scale and Shape,(with Discussion),” Applied Statistics (54), pp. 507–554. Rulke, D. L., and Rau, D. 2000. “Investigating the Encoding Process of Transactive Memory Development in Group Training,” Group & Organization Management; Thousand Oaks (25:4), pp. 373–396. Shen, Y. 2007. “Transactive Memory System Development in Virtual Teams: The Potential Role of Shared Identity and Shared Context,” in Proceedings of the 2007 ACM SIGMIS CPR Conference on Computer Personnel Research: The Global Information Technology Workforce, SIGMIS CPR ’07, New York, NY, USA: ACM, pp. 228–230. (https://doi.org/10.1145/1235000.1235054). Treem, J. W., and Leonardi, P. M. 2013. “Social Media Use in Organizations: Exploring the Affordances of Visibility, Editability, Persistence, and Association,” Annals of the International Communication Association (36:1), pp. 143–189. Treem, J. W., and Leonardi, P. M. 2015. “Recognizing Expertise: Factors Promoting Congruity Between Individuals’ Perceptions of Their Own Expertise and the Perceptions of Their Coworkers,” Communication Research, p. 0093650215605154. (https://doi.org/10.1177/0093650215605154). Upton, G., and Cook, I. 2014. “Generalized Linear Model,” in A Dictionary of Statistics, Oxford University Press. (http://www.oxfordreference.com/view/10.1093/acref/9780199679188.001.0001/acref- 9780199679188-e-662). Venables, W. N., and Ripley, B. D. 2002. Modern Applied Statistics with S, (Fourth.), New York: Springer. (http://www.stats.ox.ac.uk/pub/MASS4). Wegner, D. M. 1986. “Transactive Memory: A Contemporary Analysis of the Group Mind,” in Theories of Group Behavior, Springer Series in Social Psychology, B. Mullen and G. R. Goethals (eds.), Springer New York, pp. 185–208. (https://doi.org/10.1007/978-1-4612- 4634-3_9). Wegner, D. M. 1995. “A Computer Network Model of Human Transactive Memory,” Social Cognition (13:3), pp. 319–339. (http://dx.doi.org.proxy1.cl.msu.edu/10.1521/soco.1995.13.3.319). Wegner, D. M., Giuliano, T., and Hertel, P. T. 1985. “Cognitive Interdependence in Close Relationships,” in Compatible and Incompatible Relationships, Springer, pp. 253–276. (http://link.springer.com/chapter/10.1007/978-1-4612-5044-9_12). Yoo, Y., and Kanawattanachai, P. 2001. “Developments of Transactive Memory Systems and Collective Mind in Virtual Teams,” The International Journal of Organizational Analysis (9:2), pp. 187–208. 28 Yuan, Y. C., Fulk, J., Monge, P. R., and Contractor, N. 2010. “Expertise Directory Development, Shared Task Interdependence, and Strength of Communication Network Ties as Multilevel Predictors of Expertise Exchange in Transactive Memory Work Groups,” Communication Research (37:1), pp. 20–47. (https://doi.org/10.1177/0093650209351469). Zhang, J., Ackerman, M. S., and Adamic, L. 2007. “Expertise Networks in Online Communities: Structure and Algorithms,” in Proceedings of the 16th International Conference on World Wide Web, WWW ’07, New York, NY, USA: ACM, pp. 221–230. (https://doi.org/10.1145/1242572.1242603). 29