To create an up-to-date and state-of-the art overview of knowledge transfer mechanisms and tools, a literature review was done using several databases, including Scopus, Web of Science, Science Direct and others. As the Scopus and Web of Science databases are seen as two of the most authoritative databases (Aghaei Chadegani et al. 2013), these two databases were mainly used. To gain an overview of the main concepts, several keywords and search terms were used and listed in Table 1.

More than 150 papers were identified and retrieved from the different databases. These papers were first briefly scanned in order to determine applicability to the topic. Suitable papers were imported into the ATLASti Qualitative Data Analysis software tool. Within ATLASti, each paper was coded according to the conceptual attributes that could be identified from the manuscript. These codes were later grouped and analysed to obtain a list of knowledge transfer mechanisms, as well as knowledge transfer capability characteristics of the mechanisms to be used as dimensions for analysis. The next two sections address the identified transfer mechanism, as well as knowledge transfer mechanism characteristics in more detail.

In the context of projects, knowledge transfer mechanisms can be defined as the formal and informal means of transmission for sharing, integrating, interpreting and applying know-what and know-how embedded in individuals and groups that will aid in the performance of project tasks (Boh 2007; Jasimuddin, Connell & Klein 2014). From the literature including the work of various scholars (Brachos et al. 2007; Duffield & Whitty 2015; Frank, Ribeiro & Echeveste 2015; Jasimuddin et al. 2014; Jeon, Kim & Koh 2011; Ortega 2001; Patriotta, Castellano & Wright 2013; Prencipe & Tell 2001; Sáenz, Aramburu & Blanco 2012; vom Brocke & Lippe 2015; Wickramasinghe & Widyaratne 2012), a variety of transfer mechanisms were identified that are in principle applicable to project-to-project knowledge transfer. In the literature, some mechanisms were discussed in some detail, while others were only mentioned by name. From the selected papers, 73 mechanisms where initially identified and after carefully evaluating each mechanism’s description, those mechanisms with a highly similar description were grouped together under a common name. This reduced the number of mechanisms to a total of 59 to be included in the study. The final list of the 59 tools and mechanism, including a brief description is provided in alphabetical order in Table 1-A1. The next section addresses the knowledge transfer mechanism characteristics in more detail.

In the previous section, a total of 59 knowledge transfer mechanisms were identified. These mechanisms need to be described by a number of dimensions (more than two), which will enable a fine-grained categorisation. The knowledge transfer mechanisms, therefore, need to have a specific ability, fitness or quality that is necessary to facilitate knowledge transfer across projects. In order to find dimensions for knowledge transfer mechanisms, the original literature review also included a search on the characteristics of knowledge flow and the capability of transfer mechanisms to transfer knowledge of a specific knowledge characteristic. The literature review was able to identify a number of characteristic elements, namely, knowledge explicitness (Snider & Nissen 2003), knowledge richness (Boh 2007), knowledge stickiness (Cummings & Teng 2003), knowledge reach (Nissen 2007), knowledge flow time (Snider & Nissen 2003) and timeliness (Cummings & Teng 2003), knowledge formality (Boh & Wong 2013; Patriotta et al. 2013) and knowledge distortion (Hansen 2002). As knowledge explicitness, richness and the stickiness characteristic element relate to each other these three are combined into a single element namely knowledge explicitness and together with the remaining four make up the five knowledge transfer mechanism characteristic dimension that will be used in this study and are discussed in the following sub-sections.

Knowledge explicitness can be defined as how codified a particular piece of knowledge is if one would place this knowledge on a continuum scale. The two extremes of this continuum would be tacit knowledge on the one hand versus explicit knowledge on the other (Snider & Nissen 2003). According to Nonaka (2007), new knowledge is often tacit and always starts within the individual. Tacit knowledge is personalised knowledge and consists partly of technical skills or so-called know-how, which is hard to write down, but can also consist of mental models, beliefs and perspectives that are so engraved that they can be taken for granted (Nonaka 2007). At the other end of the continuum explicit knowledge is found, which is formal and systematic and can be easily codified, communicated and shared. Typical examples include product specifications, operating procedures, instructions, scientific formula and computer programmes and databases (Nonaka 2007; Snider & Nissen 2003). Hansen et al. (1999) maintained that firms can follow two knowledge management strategies, namely, a personalisation strategy (tacit knowledge closely tied to the individual that developed the knowledge) and a codification strategy (explicit knowledge codified and routinised and revolve around ICT-based technologies). Companies can, therefore, manage this knowledge in essentially the same way as part of either a knowledge personalisation strategy, or a knowledge codification (technology-based) strategy or even a combination thereof (Ambos & Ambos 2009). For the purpose of this study, the explicitness dimension is categorised into three groupings as depicted in Table 2.

Knowledge reach is defined as the extent to which knowledge is shared with others in and across organisation units, for example, team members, project, departments and organisations, as well as the number of people affected by the knowledge. Snider and Nissen (2003) and Nonaka (1994) define four organisation-based aggregations of people, namely the individual, the work group, the organisation and the inter-organisation aggregation. In a more recent meta-analysis on research publications on the topic of knowledge brokering in project management, Holzmann (2013) identified four organisational levels across in which knowledge can flow, namely, the team, the project, the organisational level and industry. By combining the different levels as defined by previous authors, this study defines knowledge reach on project-based organisations as aggregations of people across various types of organisational boundaries, namely the individual, the project team, organisational project teams and inter-organisational project teams. The reach dimension can be used for categorising knowledge transfer mechanisms in the sense that certain mechanisms are more suited to transfer knowledge from person-to-person, that is, coaching, while other mechanisms are more suited for transferring knowledge over larger distances and addressing more people, that is, training sessions.

For this study, knowledge reach is illustrated in the following reach levels, namely the individual, the project team, organisational project teams and inter-organisational project teams as described in Table 3.

Flow time pertains to the length of time that is required to transfer knowledge from a sender to a receiver or from one state to another, for example, tacit to explicit (Nissen 2007). The flow time dimension can be seen as short (i.e. fast moving), as well as long (i.e. slow moving), for the knowledge to move between these states and can vary by several orders of magnitude. To operationalise the flow time dimension, Snider and Nissen (2003) divided the flow time in a corresponding scale using highly-granular temporal markers, namely hours, days, months and years. The flow time dimensions can be used for categorising knowledge transfer mechanisms in the sense that certain mechanisms are more suited for quick and in-time transfer of knowledge, that is, face-to-face discussion and email, while other mechanisms are more suited for longer term transfer, that is, post project reviews and best practices.

For this study, the flow time dimensions of Snider and Nissen (2003) are used, namely hours, days, weeks, months and years as described in Table 4.

The synchronicity mediation dimension pertains to networks through which knowledge flows, as well as the effect of intermediaries between the sender and receiver of knowledge flows. Hansen (2002) argues that business units that interact and share knowledge with each other in a direct way, thus where the knowledge network path is short, the receiving unit obtains more knowledge and performs better than units where the network paths are longer and more indirect. According to Hansen (2002), longer network paths lead to information distortion, making the search for useful information more difficult, while shorter network paths with direct interaction can more easily share knowledge but are more costly to maintain because of the effort of interaction. Within project teams, time spend on interacting and sharing knowledge is less time spent on project-related activities. Understanding the knowledge transfer mechanism’s mediation dimension can aid in the selection of transfer mechanisms that has less knowledge distortion but also require a larger effort to manage.

For this study, the synchronous effect of each knowledge transfer mechanism will be measured according to two options, namely, direct or synchronous mediation versus indirect or asynchronous mediation between the sender and the receiver as described in Table 5.

The formalisation dimension pertains to whether the transfer mechanism is capable of transferring knowledge in a formal or informal way and whether the mechanism can ensure that the knowledge that is transferred can be institutionalised in a formal way or not. Informal knowledge transfer refers to ad-hoc and unstructured knowledge transfer that is typically done in an unplanned, need-to-know and social manner and occurs naturally without external intervention, while formal knowledge transfer is designed to enable the transfer of learning and knowledge in a planned way as typically addressed through the formal structure and routines within organisations and are endorsed by these organisations (Boh 2007; Kitimbo & Dalkir 2013; Klimkeit 2013). Both formal and informal transfer can take place between persons, as well as though documentation (Boh & Wong 2013).

For this study, three dimensions are used to evaluate the extent to which a transfer mechanism is capable to transfer the knowledge in a formal way. The three dimensions are described in Table 6.

To find knowledge transfer mechanism groupings, it had to be researched to what extent each of the identified transfer mechanisms scores on the five characteristic dimensions that were discerned in the previous section. The justification and value of the scores cannot solely depend on the opinion of a single person or researcher; therefore, a group of experts is required to provide a combined viewpoint to justify the association.

It seems sensible to apply the Delphi technique in this case. Per definition, a Delphi technique is a structured communication technique that relies on a panel of experts. The experts answer a questionnaire in two or more rounds. After each round, the answers, including a motivation, are summarised in an anonymous way. Through a next round, the experts can review their initial answers in light of the summary of replies from the previous round’s panel members. By doing several rounds, it is believed that answers to the questions will converge towards a final answer where after the process is stopped. This full convergence is actually what was not aimed for in this case, as it would diminish variation to a large extent making our next analytical step (categorisation) virtually impossible. In other words, variation was deliberately allowed for. In the next section, we explain the approach taken in this study.

In this study, an independent expert group was used as reaching of high consensus through multiple rounds would be a highly undesirable outcome especially considering the latent clustering analysis technique for which variation in the data is a necessity. The word ‘group’ instead of ‘panel’ is used as the authors would like to stress that to avoid social pressure for consensus, the group members did not physically meet and came to their judgements individually and independently.

A questionnaire was designed using Survey Monkey as a web-based survey tool. A supplement document was also prepared, which listed the 59 transfer mechanisms, as well as a short description of each mechanism to ensure that there would be a common understanding of each mechanism by every expert. Each knowledge transfer mechanisms had to be evaluated against the five characteristic dimensions as described in the previous section. This resulted in a questionnaire comprising of five sections, one for each characteristic dimension, and 59 knowledge transfer mechanisms in each section. Each member of the expert group, therefore, had to individually and independently answer 295 questions related to the knowledge transfer mechanisms and their transfer characteristics. To reduce the possible effect of lost concentration and the emergence of patterns in the answers because of the large number of questions each participant had to answer, the questions in each section of the questionnaire were randomised.

Stitt-Gohdes and Crews (2004) mention that members of an expert group should be knowledgeable about current information and perceptions regarding the topic under investigation, should have experience in it and should be working in such an environment. To meet these criteria, the members of the expert group were identified and selected based on their working knowledge and experience into the behaviour of individuals, teams and organisations and have a good understanding of project management processes including events in projects where knowledge transfer takes place. In addition, the members had to have some insight in how knowledge is transferred within and across projects, as well as knowledge and experience in the selection and use of knowledge transfer mechanisms. Members of the expert group were, therefore, selected according to the following criteria:

Work experience: Members had to have at least 5 years working experience in projects;

By using the above criteria, 20 experts were identified from different organisations and invited to participate in the study. The questionnaire and the supporting document were sent to each of the identified experts and followed up by email, telephone and instant messaging. From this group, 17 experts responded and individually completed the questionnaire. The demographics of the expert group are depicted in Table 7. From the table, it is evident that the selection of the group did meet the criteria and provided a good representation.

The data collected through the questionnaire were analysed by evaluating the median score, as well as the standard deviation of every item. Although reaching consensus on the individual item level was not a requirement of the research process, it nevertheless showed that on 85% of the items consensus was reached as indicated by low standard deviation values on the ordinal response data. As such this is an interesting outcome as only the capturing every expert’s viewpoint was aimed for. As experts reached consensus on individual items independently from each other, but hopefully not between the items, is a comforting result as fine-grained categorisation otherwise would become very difficult to accomplish empirically for the reason that within variation would become too large.

In the previous sections, a state-of-the-art overview of knowledge transfer mechanisms and tools is provided, several dimensions of these tools identified and data collected. The next step is selecting a statistical technique that is suited for multi-dimensional categorisation and gives objective measures to decide on the number of categories fitting the data well. Latent class cluster analysis is the statistical technique selected (Haughton, Legrand & Woolford 2009). A discussion of the characteristics of this technique is provided below.

Latent classes are unobservable (latent) subgroups or segments (Magidson & Vermunt 2002). Ideally, variables (in this case, characteristics of knowledge transfer mechanisms and tools) within the same latent class have to be similar on certain criteria, while transfer mechanisms grouped in other different latent classes are dissimilar from each other in certain important ways (Vermunt & Magidson 2005). Latent class analysis is a statistical analysis technique for finding subtypes of related cases (latent classes) from multivariate numeric or categorical data on the basis of maximum likelihood estimation. Latent class analysis may be used to analyse combinations of dichotomous, nominal, ordinal and count variables and uses appropriate distributions for each scale type (Magidson & Vermunt 2004a). Latent class analysis differs from more traditional approaches such as factor analysis, structural equation models and random-effects regression models as these techniques are severely limited in practice because they require all variables to be continuous, and it uses the assumption of multivariate normality to justify a linear model (Magidson & Vermunt 2004a).

In latent class analysis, a maximum likelihood algorithm classifies cases (knowledge transfer mechanisms and tools) into clusters based on membership probabilities estimated from a parametric model. The procedure seeks to identify the smallest number of latent (unobservable) classes that is sufficient to account for the relationships among the observed variables (mechanism transfer capability dimensions). The latent variable is to explain all the relationships among the observed variables of its class. The analysis typically starts by fitting a baseline model for but one latent class. If this one-class solution does not fit the data well, the analysis incrementally adds latent classes to the model until the model fit is adequate. The model is evaluated in two ways, namely, model fit and model usefulness. Model fit is determined by statistical means on the basis of a log likelihood criterion, in this case, the Bayesian information criterion (BIC). In short, the BIC provides information about the explanatory power of a model relative to the number of parameters employed. The lower the BIC, the better the fit of the model (Magidson & Vermunt 2004b). Besides the BIC, the p-value (p > 0.05, indicating adequate fit) and the number of parameters used (fewest number indicating parsimony). Model usefulness is determined by substantive interpretation of the latent classes and whether they are meaningful to address the underlying questions at hand and through verifying the classification quality by use of classification tables or the use of an entropy criterion to estimate the number of clusters from a model (Celeux & Soromenho 1996). For this study, a multi-level latent class model prediction was done to determine the best cluster model for the provided data. The higher level in the latent class model consists of the 17 respondents from the expert group and the lower level consisted of the respondent’s responses on the different dimension of the transfer mechanisms. By using a multi-level model, the analysis corrected for differences in responses because of a lower consensus on some of the dimensional scoring of the mechanisms.

For the identification of the latent clusters, the software application Latent Gold 4.0 was employed (Vermunt & Magidson 2005). Summary fit statistics of the models are depicted in Table 8. By using the BIC value as model selection criteria, it can be concluded that a five cluster model has the lowest BIC value of 9930.4362. As there was no other cluster model with the same BIC value, the lowest log likelihood (LL) value, number of parameters (Npar) and Classification error (Class Err) were not needed to as an additional verification of the best cluster fit.

A profile table for the identified five cluster model is depicted in Table 9. The first row of the table indicates the size of each cluster as a percentage of the respondents. In addition, the other cells in the profile table contain (marginal) conditional probabilities that show how the clusters are related to the indicators, for example, the dimensions of the knowledge transfer mechanism characteristics. For the indicators of an ordinal nature, for example, Explicitness, Reach, Flow Time and Institutionalisation, the mean value of the indicator categories is also provided. The probabilities for each indicator sum up to one: (1) within each cluster. The probability value is interpreted as the respondents chance to choose that specific indicator category within that cluster. For example, respondents in Cluster 1 have a 66.89% chance of responding that the knowledge transfer mechanisms grouped in this cluster transfer mainly explicit knowledge. The higher the probability value, the more the indicator category belongs to that cluster. In Table 9, the dominant probabilities are shown and indicate the indicator categories belonging to each cluster. The characteristic of each knowledge transfer mechanism cluster will be defined according to these indicator categories, as well as the transfer mechanisms that belong to each identified cluster.

From Table 9, the prevailing characteristics of each cluster are identified by selecting the dominant probability values, as well as evaluating the mean value for every dimension in every cluster. For example, in Cluster 1, the dominant probability is 0.6689 and the mean value is 2.6434 (»3) which corresponds with mainly explicit; therefore, mainly explicit knowledge is transferred by this cluster of mechanisms. The identified clusters, their characteristics as well as associated transfer mechanisms are discussed in more detail in the next section.

From the multi-level latent class analysis, a frequency count of each knowledge transfer mechanisms belonging to a specific cluster was determined, where the maximum value for the frequency count equalled to 17 (the total number of participants in the expert group). By making use of a Pareto criterion, where the transfer mechanisms in each cluster are sequenced in descending order from highest count to lowest count and then selecting the transfer mechanisms that make up the first 80% of the total count, the dominant knowledge transfer mechanisms for each cluster were identified with very little overlap of transfer mechanisms between the different clusters. Cluster labels were given to each identified cluster according to the cluster characteristics, as well as the dominant knowledge transfer mechanisms, which form part of each cluster. These clusters labels are discussed in the next sub-sections.