An internationalised Europe and regionally focused Americas: A network analysis of higher education studies

The study on which this article reports investigated the internationalisation of higher education studies by examining collaborations in the form of international co-authorships. We analyse how network-based mechanisms, related to structural relationship between authors (preferential attachment, i.e., higher tendency to collaborate among the most productive ones) and node level features (homophily, i.e., tendency to collaborate with similar others), affect higher education co-authorship networks. We build a bipartite co-authorship network based on 17,262 publications from 33 specialised higher education journals indexed in Scopus from 1996– 2018. Scientific collaboration in higher education mainly occurs within national borders. We found that higher education is not an internationally oriented field of research, with around 90% single-country publications. A geographical divide was observed between the two largest communities (Europe, Asia and Oceania vs. the Americas) which was also reflected in the research themes addressed by these communities, structured around the known divide between (1

are exogenous to the co-authorship network, i.e., they do not consider the role of the position or relationships among actors (scholars/universities), which have thus far been mostly overlooked in the higher education literature.Therefore, this article contributes to filling this gap by adopting a social network theory perspective to study the internationalisation of higher education studies.
Social network theory (SNT) states that the set of direct and indirect relationships among actors provide opportunities and constraints that affect their behaviours (Granovetter, 1973).Following SNT, the characteristics and internal dynamics of the co-authorship network-namely, the connections among actors (ties) and the position of actors within the network-should be considered explanatory factors in the internationalisation of science (Acedo et al, 2006;Wagner & Leydesdorff, 2005;Wagner et al, 2015).Based on this theoretical lens, we articulated three research goals, one foundational for our study and two about the internationalisation of higher education studies: 1. Disambiguation of scientific entities' names (e.g., names of authors and institutions), have often been taken for granted, while the association of a scholar with a specific institution has paramount effect on the dis-connectivity of co-authorship networks (D'Angelo & van Eck, 2020;De Stefano et al, 2013).We investigated the level of improvement in the accuracy of descriptions of network constructions obtained by the correct identification of institutions and author names, i.e., disambiguation.We wish by this to contribute to the higher education literature in methodological terms.
2. We have sought to identify the main features of higher education co-authorship networks, specifically, to describe international co-authorship networks.
3. We have investigated how network-based mechanisms explain the dynamics and growth of co-authorship networks for international higher education studies through a bipartite exponential random graph model (ERGM) and bipartite community detection.We have carried out a statistical analysis of hypotheses on specific network mechanisms such as preferential attachment, homophily and the significance of network sub-communities.These techniques have rarely been applied to higher education co-authorship (Biancani & McFarland, 2013;Kezar, 2014).
We first review literature regarding the internationalisation of higher education studies.We then present the main network-based mechanisms to be tested using our statistical model.This model is illustrated in the fourth section along with the data.Findings will then be presented and discussed in the last two sections.

| LITER ATURE ON THE INTERNATI ONALISATI ON OF HI G HER EDUC ATION S TUDIE S
There is a dearth of papers that have empirically examined the internationalisation of higher education studies (Avdeev, 2019;Jung & Horta, 2013, 2015;Kuzhabekova et al, 2015;Kwiek, 2020;Tight, 2007Tight, , 2014)).While most of them focus on either an entire continent or regional area, few present a global perspective.Tight's analyses (2007Tight's analyses ( , 2014) ) focus on a sample of North American and non-North American higher education journals by revealing a remarkable division between the North American community and the rest of the world.
North American journals are dominated by authors based in the Americas, both in terms of authors' country of origin (Tight, 2007) and citation patterns (Calma & Davies, 2017;Tight, 2014) open towards North American and Asian authors over time (Kwiek, 2020;Tight, 2014).This low level of collaboration is claimed to be the result of several factors, including North American methodological nationalism and a longer history of higher education studies in that region.Furthermore, this division is also a consequence of the nature of higher education studies, which is largely driven by emerging specific national and local policy problems (Kehm, 2015;Tight, 2007Tight, , 2014)).Liu et al (2019) examined how the research themes of higher education studies vary among countries.The authors found that Australia and the United Kingdom play a central role in the international community because several countries share a significant research theme similarity with them.Similarity is heavily influenced by the policy problems that countries are facing.The authors argue that this high similarity in research themes may be considered a key factor for explaining the probability for collaborations among scholars.The European case is, in this sense, explicative, due to an increasingly shared policy framework and the growth of EU-level funding for collaborative research, as pointed out by Kwiek (2021).
In terms of the degree of international collaboration among higher education scholars, while Kosmützky and Krücken's (2014) analysis reveals a steady trend over time, Kuzhabekova et al (2015) present a slowly increasing tendency.However, this share is still only 11.3% of the publications analysed whereas the majority of articles (80.6%) were produced by authors from a single country.Avdeev (2019) confirms that international higher education collaboration is progressing, even if by a slow increase, since 2000.Similar modest levels of international collaboration were noted in Jung andHorta's (2013, 2015) analysis of Asian higher education research networks, which highlights a substantial isolation of scholars (43% single-author publications).Domestic co-authorship patterns are definitly more frequent (see similar results in Lovakov and Yudkevich (2021) for post-Soviet countries).Jung and Horta (2013) also find that productivity and co-authorship patterns rely on a handful of scholars within a few universities, whereas the remaining publications are authored by part-time researchers (Kwiek, 2020).
Moreover, by examining a sample of prestigious higher education journals, Kwiek (2020) shows that the distribution of country affiliations is changing over time, with the diminishing influence of American scholars compensated by the growth of Continental Europe and East Asia.Finally, Avdeev (2019) uses spatial scientometric techniques to demonstrate that the intensity of international co-authorship in higher education is still dependent on geographical and linguistic proximity between authors.
These papers explain the internationalisation of higher education studies by a variety of elements that include historical circumstances, geographical proximity and linguistic commonality (Avdeev, 2019).Other explanations address policy-related dynamics like funding mechanisms, academic mobility, maturity of higher education systems and similarities in research themes (Jung & Horta, 2015;Kim et al, 2017;Liu et al, 2019;Tight, 2007Tight, , 2014)).Among these factors, the nature of the relationships between authors and universities embedded in the coauthorship network-which is discussed in the next section-has been overlooked (Biancani & McFarland, 2013;Kezar, 2014).

| A SO CIAL NE T WORK THEORY PER S PEC TIVE ON FAC TOR S AFFEC TING INTERNATIONAL RE S E ARCH COLL ABOR ATIONS
Studies on international research collaborations have considered a diverse set of explanatory factors (Kwiek, 2021;Wagner & Leydesdorff, 2005).Luukkonen et al (1992) underline the role of geographical proximity and linguistic similarities among countries.The role of supranational policies (e.g., EU funding initiatives on research collaboration) are argued to also support international research collaborations (Kwiek, 2021;Wagner et al, 2015).Countries with less developed scientific capacity lean towards advanced countries in terms of collaboration preferences (Wagner & Leydesdorff, 2005).Although these elements contribute to explain the growth of international research collaborations, empirical studies have produced controversial evidence (Wagner & Leydesdorff, 2005;Wagner et al, 2015) suggesting that other explanatory factors can shed light on the dynamics of international research collaborations.Social network theory (SNT) provides, in this sense, a complementary perspective.Dynamics at the level of co-authorship networks are considered explanatory factors.These consist of the connections among actors (ties) and the position of actors (nodes) within the network (Acedo et al, 2006;Biancani & McFarland, 2013;Wagner & Leydesdorff, 2005;Wagner et al, 2015).SNT identifies (1) endogenous and (2) exogenous explanatory factors.Based on these factors we have developed and tested hypotheses using a statistical model.

| Endogenous factors
Endogenous factors are structural recurrent processes based on the presence of ties, which explain the formation of other ties.One main endogenous mechanism is preferential attachment (Biancani & McFarland, 2013).
Preferential attachment is the propensity to collaborate with scholars who already present many previous collaborations (co-authorships).(Barabási et al, 2002).This mechanism emphasises the power of popularity in research networks, as also articulated in the theory of cumulative advantage in science, i.e., the Matthew effect (Merton, 1968).Several empirical works demonstrated the role of preferential attachment in driving co-authorship networks (Akbaritabar et al, 2020;Dahlander & McFarland, 2013;Moody, 2004;Newman, 2001a;Sciabolazza et al, 2017;Wagner & Leydesdorff, 2005b;Zhang et al, 2018).The first hypothesis (H1) that we tested was the following: scholars with several previous collaborations present a higher probability of establishing new co-authorships.

| Exogenous factors
Exogenous factors are the attributes of nodes that influence the formation of new ties.Homophily is a major explanatory mechanism in networks (Akbaritabar et al, 2020;Sciabolazza et al, 2017;Zhang et al, 2018).Homophily is the tendency of nodes to start new ties with whom they present similarities on one or more attributes (Biancani & McFarland, 2013).Empirical studies have demonstrated the effects of different homophily effects in coauthorship networks, including same gender, language, ethnicity (Dahlander & McFarland, 2013) and similarity of research themes (Akbaritabar et al, 2020;Zhang et al, 2018).However, differently from preferential attachment, homophily can be tested in empirical analyses only if a node's attributes can be retrieved from publication data.
Consequently, we will concentrate on three homophily effects that can be tested based on our data, and are considered significant exogenous factors in literature.
Firstly, research productivity has been largely associated with scientific collaboration.Studies have proved that collaboration affects productivity positively (Biancani & McFarland, 2013).However, how research productivity influences the formation of new collaborations has been less studied.We include it as a homophily effect, testing whether those with a similar level of research productivity tend to collaborate more with each other (Zhang et al, 2018).
Secondly, we consider academic seniority, intended as the length of time that a scholar has been publishing.This effect tests both whether scholars tend to collaborate with those who present a similar seniority and whether collaboration stems from a relationship such as between students and mentors (e.g., PhD students or early career researchers and supervisors) (Floyd et al, 1994).
Thirdly, scholars might collaborate several times on different occasions leading to multiple publications, while some co-authorship events might happen only once.This repeated collaboration relevantly affects the turnover of the authors in the sample.Some may enter the sample, publish only a few times and exit the pool of actively publishing authors, while some authors stay for longer time and continue to publish actively (Akbaritabar et al, 2020;Palla et al, 2007).Controlling for homophily based on the latest year of publication evaluates this effect.We formulated three additional hypotheses based on these homophily effects.Our second hypothesis addressed homophily for research productivity (Hypotheses 2): scholars with similar levels of research productivity have a higher tendency to collaborate with each other.The third hypothesis addressed homophily for academic seniority (Hypotheses 3): scholars with similar academic seniority have a higher tendency to collaborate with each other.
The fourth hypothesis addressed homophily for publication activity (Hypotheses 4): having a similar date of latest publication (e.g., publishing until recently) increases the tendency to collaborate among scholars.
Finally, SNT scholars emphasise the presence of communities within networks (Akbaritabar et al, 2020;Palla et al, 2007;Sciabolazza et al, 2017).Algorithms are used to examine whether collaboration patterns persist between or within denser areas of the network (groupings).These communities significantly influence the overall connectivity of the network and were for this reason addressed in our fifth hypothesis on social closure among groups (H5): there are cohesive communities in the co-authorship network with a higher tendency to collaborate among community members.

| DATA
There is not an uncontested list of higher education journals, thus, we followed previous studies by adopting as a selection criterion journals whose name contains the terms higher education or tertiary education (Jung & Horta, 2013;Kim et al, 2017;Liu et al, 2019;Lovakov & Yudkevich, 2021).We did not consider professional journals (e.g., The Chronicle of Higher Education and Times Higher Education) or discipline-specific journals (e.g., Arts and Humanities in Higher Education), instead concentrating the analysis on the core of the higher education studies by avoiding discipline-based influences (Tight, 2018).We selected 33 specialist higher education journals (Table 1) consistent with previous studies (Jung & Horta, 2013;Kim et al, 2017;Liu et al, 2019;Lovakov & Yudkevich, 2021).We then queried Scopus 2019 from the German Bibliometrics competence centre (Kompetenzzentrum Bibliometrie, 2020) for all articles and reviews published in the 33 journals from 1996 to 2018.The dataset includes various metadata for each publication, including publication year, title, author name(s), affiliation addresses and journal name.

| Disambiguation process and methodology
Data delivered by Scopus are prone to errors, and there is a strong need for disambiguation of institution and author names (D'Angelo & van Eck, 2020) without which the constructed networks will have multiple representations of the same actor and an artificially higher level of dis-connectivity (De Stefano et al, 2013).Therefore, we used affiliation addresses to disambiguate institutions using the Research Organization Registry (ROR) API (Research Organization Registry, 2020) which uses data from the Global Research Identifier Database (Global Research Identifier Database, 2020), International Standard Name Identifier, Crossref and Wikidata.ROR takes the different name spellings and misspelled words, acronyms and multiple languages into account and uses the full address strings that Scopus delivers.To disambiguate author names, we used Scopus's author identification numbers, which offer a somewhat improved data quality though they are still prone to error (Aman, 2018).Using this data, we constructed two-mode (bipartite) co-authorship networks.One mode is publications and for the other mode we constructed two sets of networks for authors and institutions.
Table 2 presents descriptive metrics for non-disambiguated and disambiguated networks at the levels of institutions and authors and illustrates the effect of disambiguation procedures.This is relevant to our first research goal to contribute to methods for disambiguation.The networks without any disambiguation present a higher disconnection, indicated by the number of connected components (e.g., 12,876 disconnected groups and 8,703 disconnected groups).Lower numbers of connected components are present in the disambiguated versions of the network (e.g., 786 disconnected groups and 7,998 disconnected groups) and point to the improvement achieved by disambiguation.Similarly, the reduced number of institutions (Number of institutions/authors), from 25,860 to 2,788 in the disambiguated version, indicates that out of every nine institution IDs, eight were redundant and only one should be considered.We exclude non-disambiguated institutions and their publications from the sample, which reduces the Number of papers, indicating that after disambiguation, our coverage of the publications is reduced from 17,262 to 13,875 (−20%) since we keep in the sample only publications for which all collaborating institutions are successfully disambiguated.The number of papers covered is higher at the authors level (99%), as Scopus numeric author IDs have a higher coverage; therefore, fewer publications are excluded (i.e., the publications with non-disambiguated authors without a Scopus author ID).Looking at the number of connected components, it is evident that disambiguation by Scopus author ID does not offer a perfect improvement, since each author can have multiple IDs (Aman, 2018).Networks built using each of these data inherit different qualities (De Stefano et al, 2013).At the institution level, the share of nodes in the giant component, which is the largest connected group of nodes (i.e., % of bipartite nodes in G) is increased from 11% to 88%, which is closer to expected rates.While Scopus author IDs offer some improvement, indicating an increase in share of nodes in the giant component from 19% to 23%, the result is not comparable to the accuracy achieved through institution name disambiguation.
In order to respond to our second research goal pertaining to features of co-authorship networks, we have computed the shares of single-country and multi-country (international) co-authored higher education publications over time.For our third research goal, pertaining to how network-based mechanisms explain the dynamics and growth of co-authorship networks, we developed Bipartite Exponential Random Graph Models (Bipartite ERGMs) (Wang et al, 2013) to test the aforementioned hypotheses (Hypotheses 1-Hypotheses 4).This statistical model presents the probability of tie existence (dependent variable) resulting from the effects of both endogenous and exogenous factors (Zhang et al, 2018).
In order to identify communities of co-authorship-in line with our third goal and fifth hypothesis-we used bipartite community detection, specifically, the Constant Potts Model (CPM) (Traag et al, 2011(Traag et al, , 2019) ) implemented in Leidenalg library in Python.CPM uses a resolution parameter (constant), leading to communities where the link density between the communities is less than and where the link density within communities is greater than (Traag et al, 2011).We set to 6 × 10 − 5 in the case of institutions.We chose this parameter after comparing the number of communities detected against the number of institutions and publications included in each community to arrive at a consistent distribution without too many or too few communities or community members.
Community detection at the level of authors was not performed due to lack of disambiguated and high-quality data.
TA B L E 2 Higher education studies co-authorship networks using non-disambiguated and disambiguated data at the institution and author levels (G = giant component of the network, SID = Scopus numeric IDs)

Institution level
Author level Moreover, we extracted the most frequent research themes and topics from the abstract and titles of the identified studies on higher education using VOSviewer (Liu et al 2019;van Eck & Waltman, 2010).VOSviewer is a software tool for constructing and visualising bibliometric networks.It uses the VOS algorithm and places terms that more frequently appear together in documents at a closer distance on a term map.Each publication can be the product of a single collaboration or a collaboration between multiple institutions.We then calculated relevance scores based on the composition of these contributing institutions for each publication, which allowed us to investigate whether and how the network communities, detected through our community detection procedure, also vary in terms of research themes addressed and whether there is a thematic focus among different communities.We set the number of term clusters to be detected to two, to check alignment with the distinction between learning and teaching versus policy-based studies highlighted in Tight (2004).We are aware that this dichotomy might appear reductionist in light of the more recent and detailed classifications of higher education themes (Daenekindt & Huisman, 2020;Liu et al 2019), but some empirical studies have demonstrated its functionality for understanding the thematic focusses of the higher education communities (Horta & Jung, 2014;Kim et al, 2017), which is consistent with our purposes.

| FINDING S
Looking at the trend of publications authored by institutions from a single country versus multiple countries we found that most of the publications (90%) were authored by institutions within a single country, which aligns with Kuzhabekova et al (2015) and papers with regional foci (Jung & Horta, 2013;Lovakov & Yudkevich, 2021).
Table 3 presents the results of the bipartite ERG models used for testing our first four hypotheses (Hypotheses 1-Hypotheses 4) on structural (preferential attachment) and homophily effects.It presents two sets of models at the level of authors (Models 1, 2, 3) and institutions (Models 4, 5, 6).Models 3 and 6 are the full models at each level, including all structural and homophily effects.

| Preferential attachment (Hypotheses 1)
Results indicate that there is a significant tendency towards lower preferential attachment at the authors level and higher preferential attachment on the institutions level (Models 1, 4), confirming our first hypothesis (H1) for institutions while for authors, due to disambiguation issues, we won't expect the results to be reliable.In other words, a few key institutions conducting research on higher education attract several co-authorship ties, while the same cannot be observed for authors.This effect stays similar when considering homophily effects.However, since Model 3 (full model, authors level) did not converge due to the high complexity of effects, we cannot claim that its results are as reliable as those for Model 6 on the institutions level.

| Homophily for research productivity (Hypotheses 2)
Similarity in the total number of publications is modelled as a continuous variable in our ERGMs in absolute difference, meaning the effect is interpreted in reverse.This similarity increases the probability of co-authorship tie formation at the authors and institutions levels, while this effect is statistically significant only at the authors level (Models 2, 5).These effects stay similar in the full models (Models 3, 6), which indicates the presence of an exclusive club (a small group with several co-authorships) of authors and institutions who tend to collaborate more among themselves.
TA B L E 3 Bipartite ERGMs' results (giant component of the higher education studies networks)

| Homophily in academic seniority (Hypotheses 3)
Similar first publication years, here a proxy for closer academic cohorts, present a higher probability of engaging in co-authorship at both the authors and institutions levels.Although we cannot control for authors' age due to the data, this finding might signal that early career researchers do not regularly publish with their mentors.The homophily in academic seniority effect was not significant at the institutions level (Model 5) until when the preferential attachment effect is entered in the full model ( 6), indicating a higher probability of collaboration among institutions with similar levels of seniority.In other words, this higher probability emphasises that institutions that have been historically active in higher education studies collaborate more among themselves.

| Homophily in publication activity (Hypotheses 4)
Having a similar date of latest publication (e.g., publishing until recently) decreased the probability of a tie at the authors level but increased it at the institutions level.However, this effect was significant only at the authors level, indicating that co-authorship events among authors with similar publication activity can be one-off incidents that do not occur regularly.

| Social closure among groups (H5)
Our community detection procedure identified 36 communities from the giant component of the co-authorship network at the institutions level which confirms our fifth hypothesis (H5).Although the giant component is connected, it consisted of 36 denser areas that are cohesive subgroups presenting a higher number of coauthorship ties within themselves than between these different groups.We analysed the geographical composition of 36 communities that emerged from the data; specifically, to better understand who were members of the communities.We analysed the communities both in terms of size (number of institutions) and productivity.We observed communities of similar sizes that presented drastically different rates of productivity.Community 0 emerged as a community with members predominantly from Europe, Asia, Oceania, and Africa and a small share of the Americas.In contrast, higher education research institutions from the Americas mostly populated community 1.Another interesting observation is the split of Asian and European research institutions on higher education across multiple communities of smaller sizes that have differing productivity rates.Oceania is an interesting case where a small number of research institutions conducting studies on higher education, mainly located in community 0, have a high rate of productivity and collaborate with different institutions globally (see Figure 1).
Moreover, we investigated whether the identified communities, especially the two largest (communities 0 and 1), display a particular focus on higher education studies themes.We used term maps for identifying research theme foci; specifically, by identifying research topics in the focus of higher education research publications based on a co-occurrence of terms in title and abstract.Our analysis highlights two main clusters of research themes, namely, the learning and teaching approach and policy-based studies.Our findings are consistent with previous studies (Kim et al, 2017;Tight, 2004).Term maps were constructed for each community.Community 0 is dominated by Europe, Asia, Oceania and Africa.Community 1 is dominated by North, Central and South America, i.e., the Americas.Community NA stands for Not Available and represents those institutions who were not members of the giant component of higher education studies network and hence not included in community detection.The ics such as college choice and college experience, with only the privatisation topics coming from the policy-studies domain.The focus of community 0 instead bridges comparatively more the policy themes with the learning and teaching themes.A stark contrast is evident in comparison to themes covered by the American higher education community (i.e., community 1).While the policy versus learning and teaching themes divide is much more evident within the American community, it is equally evident that policy-based studies are comparatively more prominent for higher education authors in community NA who were not part of the giant component (see Figure 2).

| D ISCUSS I ON AND CON CLUS I ON S
The present article investigated the internationalisation of higher education studies through a quantitative analysis of co-authorship networks.We articulated the disambiguation of research networks as a foundational research goal.Two additional goals pertained to developing methods and understanding the growth of co-authorship networks.The latter two goals were specific to higher education studies: in methodological terms, how the accuracy of network construction may improve as a result of disambiguation of institution and author names; to what extent higher education studies are international and how network-based mechanisms explain the growth of higher education co-authorship networks.
In response to our first goal, regarding the differences between disambiguated and non-disambiguated networks, our findings suggest that the method we have used improves the identification of co-authorship networks particularly on the level of institutions (Aman, 2018;D'Angelo & van Eck, 2020;De Stefano et al, 2013).While at the author level, the accuracy was lower and affected the results to a certain extent.This shortcoming applies to empirical research that takes data from bibliometric databases as uncontested truth without probing into problematic aspects of the data.
In response to our second research goal concerning the internationalisation of studies on higher education, we found that studies on higher education are not internationally oriented, as also confirmed by previous research (Avdeev, 2019;Horta & Jung, 2014;Jung & Horta, 2013;Kwiek, 2020;Lovakov & Yudkevich, 2021;Kuzhabekova et al, 2015).Publications by authors from a single country remain more prevalent (90%).Internationalisation in research on higher education is significantly lower compared to internationalisation in the social sciences (Acedo et al, 2006;Akbaritabar et al, 2018).To reverse this trend, a stronger involvement of English-speaking countries (especially the US, Australia and the UK), which are the most productive countries of higher education studies (Calma & Davies, 2017;Liu et al, 2019) seems crucial to build a more integrated and international community.
The second part of the analysis, corresponding to our third research goal, employs network-based mechanisms to investigate the nature of collaborations in higher education studies (Biancani & McFarland, 2013;Wagner & Leydesdorff, 2005).Preferential attachment (Hypotheses 1) emerged as a significant explanatory mechanism at the institutions level (this is lower for the authors due to disambiguation issues).Therefore, new co-authorships are formed primarily with institutions who already present many co-authorships, similar to other disciplines (Barabási et al, 2002;Moody, 2004;Sciabolazza et al, 2017;Wagner & Leydesdorff, 2005;Zhang et al, 2018).Regarding homophily effects (Hypotheses 2-Hypotheses 4), seniority among authors and institutions was associated with collaboration among peers.Prolific authors also show a higher tendency to collaborate with other prolific authors, hence forming an exclusive club of highly prolific authors and institutions co-authoring among themselves.Other homophily effects are associated with gender, language and PhD granting university, to name a few (Biancani & McFarland, 2013;Zhang et al, 2018).However, due to limitations of the data, we did not test these homophily effects in the study on which this article reports.
Community detection allowed us to confirm our fifth hypothesis regarding the occurrence of cohesive communities.These cohesive subgroups present a higher number of co-authorships among community members than between different groups in the network.These communities differ both in terms of geographical location and

1
Distribution of HE studies institutions from different continents versus their membership in 36 HE studies communities detected from the giant component of network (left panel shows size of communities based on institutions, right panel shows the productivity based on publications).Source: Authors count in continents vs communities of HE Number of HE Publications

F
I G U R E 2 Term maps showing the research theme focusses of HE studies publications of community 0, community 1 and those institutions not a member of the giant component (community NA).The yellower colour shows the research theme focus (except top left, L&T (red), policy-based studies (green)).Source: Authors focus of community 1 was largely skewed towards learning and teaching, research themes revolving around top- TA B L E 1 Higher education research journals, total number of articles and reviews published 1996-2018