Study inclusion and exclusion criteria

Data for the number of peer-reviewed research publications between 1963 and 2017 across countries was gathered using Elsevier Developer Application Programming Interface (API) with an HTTP connection to the Scopus database. The data was downloaded using an Elsevier API (dev.elsevier.com) together with code written in Python using the Python library script found at github.com/ElsevierDev/elsapy.

The classification for each publication in its respective discipline was grouped according to the subjects defined by Elsevier's SciVerse Scopus All Science Journals Classification (ASJC), which is based on the article abstract, scope, and similar contents that the journal publishes. Each of these ASJC codes was used to filter the number of publications in the queries to the API.

Only articles and conference papers were included in the dataset. Reviews, book chapters, viewpoint articles, and letters were excluded from the analysis. All of the listed journals in the Scopus database were selected. However, journals that stopped publishing during the sample periods were not included, since full information was not available.

Countries for which the data was incomplete or unreliable for any year or discipline at the time of data extraction, such as Brazil or Argentina, were excluded from the study. Countries which underwent significant transformations, such as Russia or Germany were also excluded from the study to avoid artificial sharp changes at the time of their geographical or geopolitical redefinitions (integration or disintegration). This led to a study universe comprising 62 countries.

Regarding the country classification of wealth, World Bank classifications [23] were used to map countries to low, low-medium, high-medium, and high income for the purpose of sub-group analyses. The only political entity explicitly excluded was Taiwan, since the World Bank gross domestic product (GDP) report [23], used in this manuscript for income classification, did not include Taiwan in their database. Tables 1 and 2 provide lists of research disciplines and countries, respectively, included in this study. While the scope is not comprehensive, it includes all countries with full history across all time periods.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Table 1. Abbreviations for the subject area classifications included in the analysis, as described in Elsevier’s All Science Journal Classification (ASJC) scheme.

https://doi.org/10.1371/journal.pone.0232458.t001

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Table 2. Universe of countries meeting the study inclusion criteria and included in the analysis, as well as manuscripts per country.

https://doi.org/10.1371/journal.pone.0232458.t002

Research products are assigned to each of the host countries of the contributing authors. By assigning in full each article to each researcher’s host country, the research contribution and productivity per country is more fairly reflected. Assigning a manuscript to a single country (e.g., that of the first author) could penalize work by researchers in lower-productivity countries who may disproportionately rely on joint research with colleagues in higher-productivity countries. Splitting the weight of each manuscript (evenly or otherwise) among countries of the co-authors would also be inadequately reflecting the contributions of each country or unevenly assessing the research efforts per discipline. For example, medical manuscripts oftentimes have a dozen or more co-authors in multiple countries, while mathematical articles tend to have much fewer (oftentimes in a single country). A split of the weight of articles per author/host country could unfairly over-estimate the research productivity in the mathematical discipline relative to the medical discipline. The raw data used in this manuscript is available at https://doi.org/10.3886/E118746V1.

Analysis

The number of publications per country and discipline over the period of analysis is likely upward trending and non-stationary, with growing numbers of publications per year for most countries and disciplines. This is likely to be especially prevalent for incipient research disciplines, growing with increased maturity of the disciplines and penetration within higher education institutions (e.g., computer sciences), as well as country’s population and wealth [24]. To remove the effect on the analysis of this uptrend, we define P_t,c,f as the share of publications for country c relating to research field f in year t, where publications are normalized so that 1^TP_t,⦁,f = 1 for each vector of country publication shares P_t,⦁,f (where ⦁ denotes all possible categories of a given index–in this example, all possible countries).

Country shares of publications for each research field evolve over time, not only depending on the absolute level of research produced by the country, but also in relative terms to the rest of the world’s research in that discipline. If two research fields, f₁ and f₂, are similar, they are expected to define similar vectors of country shares , as a reflection that research productivity in one discipline does not occur in isolation from research in the other.

Let d(Ω, Ϛ, f₁, f₂) be a distance measure between research fields f₁ and f₂ over a given period Ω (which can be single or multi-year) for a subset of countries Ϛ. The distance function d is defined in Euclidean Distance Matrix form [25,26] and is cumulative over the period Ω, with non-diagonal elements:

Research field f₁ is closer to field f₂ than to field f₃ during period Ω for the subset of countries Ϛ when d(Ω, Ϛ, f₁, f₂) < d(Ω, Ϛ, f₁, f₃). Therefore, a weight function of linkages between research fields can be defined as w(Ω, Ϛ, f₁, f₂) = d(Ω, Ϛ, f₁, f₂)⁻¹. An undirected graph representing the relationships between research fields for a subset of countries Ϛ, during period Ω is composed of pairs of fields f₁ and f₂ (vertices in the graph) connected through edges with weights w(Ω, Ϛ, f₁, f₂). This is performed using the graph.adjacency routine within the igraph package in R.

Since all pairs of vertices are connected through edges with non-zero weight, all graphs are effectively complete. However, most edges may have low (relative) weights, indicating lower levels of relative connectedness between research fields. In order to extract the backbone structure (set of relationships that remain after a statistical network reduction approach is implemented), we follow [22] and apply their proposed disparity filter using the disparityfilter package in R. This approach extracts the network with edges that are statistically significantly (α = 0.05) away from the null model (complete network), without penalizing research fields that may be proportionally smaller or less connected. This approach produces significantly more sparse graphs that represent only statistically significant edges between connected vertices. Note that the value of the significance level (α = 0.05) is pre-specified ahead of the analysis to avoid data snooping. Higher values of α provide more dense networks, while smaller values will result in more sparse networks. As with most statistical approaches, the value of α = 0.05 is both conventional and arbitrary.

Upon constructing the backbone network, a fast greedy algorithm [27] is used to define the optimal community structure of the graph using the cluster_fast_greedy algorithm also within the igraph package in R. These communities allow for dense groups of vertices that share commonalities to be grouped together, forming natural clusters of research fields based on their relationships with other fields. This is represented graphically by splitting the edges into intra-cluster edges (black) and inter-cluster edges (red). Intuitively, nodes connected through black edges belong to a cluster that shares similar features, while those linked through a red edge represent vertices that, while belonging to different clusters, are still similar. For the purpose of graph simplicity, only vertices with at least one connecting edge upon performing the disparity filter are included as part of the resulting graph.

Graphical representations of countries based on their similarities across research disciplines are also possible through the adaptation of the distance function to measure the distance between pairs of countries across research discipline productivity shares. This denotes countries that are similar in both absolute and relative research productivity across all fields (though subsets of fields are also possible). For any given pair of countries c₁ and c₂, the distance is defined across fields as follows: and the corresponding weights w* are defined as the inverses of the non-diagonal components of the new Euclidean Distance Matrix. While other distance measures were possible (e.g., Manhattan, Cosine, etc.), we chose the Euclidean distance for interpretability.

Finally, country research productivity shares are expected to be largely driven by country size. A measure that further normalizes the research productivity shares (scaling them to add to one across disciplines for each country) is needed for country size-adjusted comparisons of inter-disciplinary shares across countries when using the Euclidean distance. This can be achieved by normalizing the research productivity shares P over all possible countries c. This leads to the final, normalized version of the networks, introduced in this study, where countries of differing sizes are compared.

Note that, while the study is cross-sectional for each 5-year period defining each of the aforementioned probability vectors, there is an inter-temporal component in the analysis of the differences of those analyses over time.

Discussion

The nature of scientific contributions in the form of published research offers a window into the relationships among the disciplines, as well as their association with the economic status of a country. Among some of the key features unveiled in this manuscript are: (1) countries with high (per capita) income have similar structures of their research productivity in relative terms, forming a strong network with identifiable geographical, linguistic, commercial, and geopolitical links; (2) some disciplines have strong relationships that remain across all country income levels (such as Computer Sciences and Mathematics), while others are more dependent on the income level of the country and their reliance on the primary sector (Agriculture and Veterinary); (3) China appears more recently within the network with similar research productivity profiles to high-income Asian countries, though it shows a history of swings that aligns with an erratic view of their research community; (4) countries outside the high-income group tend to form small isolated clusters mostly driven by geographical proximity, rather than income status; (5) middle-income countries show more dynamics in the discipline network than countries in lower and higher income brackets. This could be due to either funding deficiencies hindering dynamics (low income countries) or consolidation of relative importance of research disciplines for high-income countries and rigidities to modify the status quo (e.g., politics around budgets and departmental expansions); and (6) disciplines such as Psychology and Arts are over-weighted (in relative terms) in high-income countries, while disciplines such as Agriculture and Immunology are more relevant and remain over-weighted for low-income countries. This is reflective of the different concerns that wealthier countries face (mental health, leisure) versus those of lower income countries (food security, survival) [28]. Diminishing interdisciplinary relations, such as those mentioned regarding Economics and Business, can be concerning given the need to tackle ever more complex problems [29].

Higher dynamics in middle-income clusters could be due to those countries exposed to less pressure to maintain the status quo than other income clusters. Low-income clusters could be more bounded in their evolution by their constraints (scientific and financial) to shift resources between disciplines. High-income clusters could be more bounded in their evolution by structural constraints (e.g., inter-departmental budget fights and shared governance may induce a higher propensity to maintain the inter-disciplinary status quo).

The method used in this manuscript allows for identification of clusters of countries that may be considered ‘aspirational’ for developing countries for economic or geopolitical purposes, allowing for identification of long-term distributions of resources among disciplines to match those in targeted clusters. The present work did not assume any clusters of scientific fields a priori, but identified the clusters based on the data. This is different from studies which assume a priori a particular clustering of disciplines [30], which may explain why those studies found that national wealth was not associated with tradeoffs between altruistic and economic motives. While finer methods for clustering areas of research based on publication data have been performed [31,32], the work presented in this manuscript focused on a more course approach to extract more robust relationships and their dynamics over time.

This study does not engage in causality arguments. The causal relationship between research productivity and economic advancement has largely been discussed in the literature [6–11]. On that basis, this study provides a benchmark for developing countries to define their research ecosystem, as a means to pursue economic advancement. Hence, this study broadens the window on the dynamics of the research ecosystem among countries. It helps pinpoint critical aspects of this ecosystem in each country and helps to plan for future developments.

Study limitations and future research

The scope of this study is bounded by the number of countries for which the data was: (1) complete and reliable across years and disciplines; (2) not affected by country expansions/territorial redefinitions; and (3) available from the World Bank with regards to GDP per capita country classification. This led to a universe comprised of 62 countries with complete history, and has the potential to affect some regions more than others. While the study largely expands work done in this area, it is not comprehensive and does not cover all countries throughout the world. Countries that experienced major geographical integration or disintegration processes (e.g., Germany or Russia) were not included, to avoid artificial jumps mostly driven by geopolitical events rather than scientific dynamics. However, while partial in nature, this study expands the literature on relationships among research disciplines and their relationship with economic clusters over time.

Expanded or sub-group analyses are feasible for a larger list of countries if constrained solely to more recent periods. This would allow for geographical or geopolitical sub-analyses. For example, data for all countries in South America would be available for more recent years, and a network comprising only these countries could be constructed to explore in more depth the intra-region relationships. Such networks would be less affected by the strength of the similarities found within higher income countries, which dominate the analyses in this manuscript. However, while possible through the methodology introduced in this manuscript, this type of more granular analysis is out of the scope of this study.

Throughout this manuscript, a link between country wealth and research profile is explored. When those associations exist, causation cannot be claimed within the scope of this manuscript. Also, there is no evidence that those associations are not driven by a set of other common/latent factors (e.g., political factors, geographical factors, individualism, level of penetration of capitalism, access to national/transnational public/private funding of research, perceptions regarding the relevance of higher education, etc.).

Finally, the mapping of the research literature to countries or disciplines is arbitrary and does not follow any validated or commonly-agreed structure. There are multiple ways in which each piece of research can be assigned to countries and disciplines, all of which would show different angles to the same questions. A comparison of all those forms, while interesting, remains outside the scope of this manuscript.