Changing the topics: the social sciences in EU-funded research projects

Abstract

How has EU research funding changed over time? This article shows a shift from research on the structural challenges arising from European integration toward research that focuses on providing immediate solutions for stakeholders from enterprises and public institutions. Using data from the EU CORDIS database and combining topic models with multiple correspondence analysis, we show how the relation between the EU and social science research can be seen in the topics of the research projects. Firstly, we find two principles of division: first between welfare state social science vs. stakeholder-driven innovation and second between cultural heritage and digital platforms vs. governance of society. Secondly, we show that, over time, there is a clear shift toward stakeholder-driven innovation. This shift in the goals of EU-funded research projects—from social integration to market innovation—illustrates a shift in for whom and for what the EU funds social science research.

Appendices

Appendix

Appendix A: stability and selection

Data selection

For each framework programme we identify the calls, funding schemes or groups of funding schemes that are specifically for the social sciences or with a clear majority of SSH projects. The selection criteria for each framework programme are based on reports that collect and monitor the inclusion of SSH in the framework programmes. Note that we did not harvest the projects directly from the reports but used the reports to navigate the CORDIS register (Table

Table 2 Selection criteria

2).

Topic model selection

Here we present the selection of the number of topics, the threshold for dichotomization and the stability of our interpretation of the space if we had chosen differently. The selection of the number of topics in a Latent Dirichlet Topic Model (LDA) is an important yet relatively arbitrary step in the process. This step is especially critical when the topics are the main unit of analysis and manually interpreted. Following (Baier and Gengnagel 2018), we depart from conventional procedures and choose a number of topics that would be infeasible for human interpretation. In Fig. 

Fig. 4

Number of topics and dichotomization threshold

4 each dot represents an MCA and the share of adjusted variance that is explained on the two first dimensions. We select the threshold for dichotomization for different numbers of topics by choosing the MCA solution with the highest level of explained variance. For each iteration, topics with fewer than 5% of the active projects above the cut-off are dropped and projects with fewer than four active topics are also dropped. This explains why the line for 200 and 400 topics is shorter—after a threshold of 0.03, the analysis collapses.

Interpretation stability

In order to assess the stability of our interpretation across different solutions and thresholds we propose a novel approach to the analytical comparison of MCA spaces. By condensing the interpretation of a plane to a structure of four opposing triads of supplementary points we can see whether the analytical interpretation of several planes is similar. The method relies on the following three definitions:

A valid triad is a set of three supplementary categories positioned in the same quadrant in a plane with a distance from its center to origo greater than 0.25 and preferably more. Either closely together or in the same angle from origo. A triad is no longer valid if the points are dispersed in multiple quadrants or too close to origo. Analytically the three points in the triad should summarise the quadrant in the space—and should be of importance to the interpretation of the space.

A valid opposition is two triads in opposing quadrants in the plane with a distance between their centers of more than 0.5. The opposition is no longer valid if the two triads overlap. Analytically the two opposing triads should be indicative of the main oppositions in the space and if the opposition was to disappear the interpretation of the axis in the plane should change.

A valid structure of oppositions is two oppositions made up of twelve points and four triads that sum up the interpretation of the space. Ideally any plane with the same structure of oppositions would have similar interpretations of their axis. The structure is no longer valid if one of the oppositions or triads are no longer valid and with that the interpretation would change.

A lot of care should be taken when selecting the points for the triads and when constructing the structure of oppositions. The analytical relevance of the triads and oppositions needs to be ensured—this means that just selecting the most extreme points from a quadrant is not the best way forward. It is more fruitful to think of the triads as mini-profiles that are likely to co-exist and that are positively correlated internally. Ideally the terms are not too rare, as is often the case for extreme points, as they then are poor representatives of the quadrant.

The strength of this approach is that it does not require that each of the compared analyses have the same set of active rows and columns. This lets us explore different numbers of topics and different cutoffs for dichotomization. As a result we should be cautious in our comparisons between points—as they are not on the same dimensions and with the same scales. This approach only ensures that the interpretation of the plane would be roughly similar—we cannot track the movement of individual points or triads, nor say if an opposition is slightly stronger or weaker.

This approach requires that the same supplementary categories are available across all planes that are compared. In the present analysis we luckily have an almost unlimited supply of supplementary variables in the presence or absence of a given term in a project's written objective. In order to make it easier to compare the spaces, they are rotated so that a selected point (in this case gender) is always in the same quadrant.

In the two figures above we select 4 triads that sum up the main oppositions in the space. ABC exemplify the northwest quadrant, with the mean point of the projects that mention, gender, minority and enlargement. These three terms indicate engagement with a specific version of welfare state social science typical for sociology and anthropology. The opposing quadrant in the southeastern corner is represented by DEF– which sums up a quadrant that is preoccupied with capital, innovation and business. Terms that are typical of finance and business economics. These are also terms that have become more common in H2020 – unlike ABC. The other opposition is an opposition between terms frequently used in macroeconomics (GHI) vs. terms used in digital humanities or in projects that integrate development of ICT tools or apps (JKL). These oppositions sum up the analysis of the space and is highly related to the change in topics between H2020 and the former frameworks programmes.

In Fig. 

Fig. 5

Interpretation stability at different dichotomization thresholds

5 we see the relative stability of the interpretation of the plane across different solutions. The structure of oppositions is similar to the 100 topic solution in all but the solution with only 10 topics where the structure of oppositions is somewhat deformed but still present. This speaks to the stability of the plane and its interpretation with regards to different iterations, cutoffs and number of topics.

In Fig. 

Fig. 6

Interpretation stability and framework programmes

6 we turn to the importance of the H2020 projects and whether we would get the same structure of oppositions depending on their inclusion. If we exclude the H2020 projects from the analysis we see that the structure is somewhat deformed and tilted to the one side. The structure of the H2020 projects alone finds the oppositions in the original analysis. Which in sum indicates that the structure of the plane has changed with the H2020 projects and that we find a somewhat different structure in the previous framework programmes.

Contributions and eigenvalues

Here we present the contributions of the most contributing categories for the reported analysis. It is important to note that the topics generated by LDA are stochastic which means that they change slightly each time you run the analysis. Topics are named after their three most distinctive terms – but those terms are again, slightly different for each run of the analysis. So, when reading these tables emphasis should be on a broader reading and on the relative contributions of categories as well as the frequencies (Tables

Table 3 Variance and eigenvalues

3,

Table 4 Categories contributing above average to dimension 1

4,

Table 5 Categories contributing above average to dimension 2

5).

Appendix B: Coded terms

See Table

Table 6 Rise and decline in coded terms in H2020 projects objectives

6.