Diversity of education systems in the European Union

As was mentioned in the introduction, literature on varieties or the DoC usually employs some kind analyses of education system. It is however clearly visible that those works significantly differ in terms of understanding, how the education system is defined. As a consequence, direct comparisons of those works are significantly limited. This section sums up their findings, with a respect to differences in approaches.

First notable work on the role of education in constitution of a capitalism model is one of the chapters in Hall and Soskice’s seminal book [Estevez-Abe et al., 2001]. Approach here is however very narrow. The authors use the term “skills formation” rather than education, and this institutional domain is combined with social protection. As a consequence, their discussion focuses on VET and training. Figure 1 shows the main types of (vocational) skills formation.

Figure 1

Altogether, Estevez-Abe, Iversen, and Soskice distinguished four skill profiles: “Firm/industry/occupational,” “Industry/occupational,” “Firm/occupational,” and “Occupational/general.” Countries were classified in a group on a basis of four criteria: median length of tenure, vocational training share, vocational training system, and share of population with university education. There are two major limitations in this approach: this set of characteristics is obviously too narrow to fully analyze the complexity of education systems, furthermore, at least one variable –VET system – is assessed arbitrarily. Those limitation do not undermine the significance of this work.

Another notable example of typology based purely on VET is Aventur et al. [1999], who classified European countries (old EU member states) according to employers’ roles in initial and continuing training. Countries differed from those weak in both categories (Spain) to strong in both categories (Denmark). Hannan et al. [1996], on the other hand, focused on a degrees of standardization and differentiations, with extreme groups of Germany and Netherlands (high in both degrees), and the United States and Canada (low in both).

Bruno Amable in his DoC book also devoted one chapter to an education. He managed to identify five clusters of education systems, briefly described in Table 1.

The most surprising case in Amable’s findings is Austria. In comparative education research, it is often assumed that Austrian education system is the most similar to the German one, with relatively high share of VET students in secondary education and a significant involvement of employers in education through a dual system. What is important for further conclusions, main reason for such a difference is an impact of employment variables, in particular relative employment of tertiary education graduates. On the other hand, another interesting case – single-country cluster of Finland – confirms opinions of education researchers, who widely see this country as most specific and most effective in the EU [see, e.g., Organisation for Economic Co-operation and Development (OECD), 2011].

It is also worth to note that in Amable’s work one can clearly identify LME type of education. This result is not that clear in other analyses because in Amable’s set, very important role is played by the United States. Also Japan and Korea appear to have education systems more in the US style, rather than continental European one. Therefore, this very convincing result blurs in the models focusing on European countries.

Second example of clustering education systems is the one of Farkas (see Table 2).

Farkas identified four clusters, with first two clearly separate. The border between clusters 3 and 4 is less obvious, but due to a size and heterogeneity of those countries, she decided to distinguish them. Definitely, the most surprising result here is Germany in cluster 4, altogether with Poland, Czech Republic, Slovakia, and Bulgaria. Farkas suggests that this case is an aftermath of Germany’s unification, but this explanation seems not very convincing. First of all, population of German Democratic Republic is relatively small and should not have that much impact 25 years after unification. Second, German Democratic Republic (GDR’s) education, governed for more than 30 years by Margot Honecker, was famous for its specificity, even in comparison with other communist countries. One may suppose therefore that this case is more the effect of ambiguous choice of variables, rather than a real effect of German institutional underpinnings.

The other clusters identified by Farkas seem reasonable. It is also interesting that even though Farkas included in her model labor market characteristics (such as employment and unemployment rates), only in two clusters those variables were statistically important.

Finally, similar analysis was performed by a research team from SGH, led by Ryszard Rapacki [Karbowski, 2017; Rapacki and Czerniak, 2018]. They called this institutional domain “knowledge subsystem” and identified four clusters (see Table 3). The comparisons with this work are the most difficult, since they treat this domain more broadly, with a significant role of innovation systems. Key characteristics of clusters, shown in a table confirm, that the variables on innovation played a dominant role in final results of clustering. The authors added also some effectiveness characteristics of education system, in particular results of PISA.

Clusters identified by Rapacki and Czerniak are the most coherent with clusters for all economy. Most of the results are convincing, and – what is perhaps most important – clearly show the difference between old and new EU member states. The latter are filled with Greece, Spain, and Portugal – this result is also predictable based on DoC literature. However, if we analyze the key characteristics of those clusters, we will grasp a conclusion that innovation system’s characteristics dominate over the purely educational ones. This approach better matches the goals of authors, who aimed at grasping similarities and relationships with other institutional domains, but fails to fully describe educational systems on their own.

In this work, education system is defined as an institutional domain that aims at teaching and educating people. Therefore, it is the most similar to the one of Estevez-Abe et al., who also understand education in this terms, and called it a skills formation domain. The difference is that I treat a term “skills” more broadly and assume that it is far beyond VET and on-the-job trainings.

As a consequence of those assumptions, there were no labor market characteristics analyzed, focusing directly on education system. Also a production of knowledge or innovativeness of the economies were excluded from the model. It is sometimes difficult to clearly distinguish scientific from teaching activities (in particular in tertiary education), however, for the purposes of this article such a distinction was necessary.

It should be noted that education system has few features that make it more complicated. First of all, education system in Europe is considered to be relatively heterogeneous. In that case, performing analyses based on clustering may lead to non-conclusive results.

Second, there were a number of reforms in a number of countries that hugely affected educational systems. There are also multiple initiatives to integrate educational systems within European countries. The most advanced are integrations in tertiary education. In 1999, there was established European Higher Education Area, an agreement aiming in cooperation among European universities. The most important and best known initiative was the so-called Bologna process

Please note that in Bologna process, much more countries participate than just European Union. For example, Belarus and Russia are the members of European Higher Education Area.

There are also initiatives to integrate VET; however, they are far less advanced. Main tools in this area are as follows: ECVET (European Credit System for Vocational Education and Training), a tool analogous to ECVET and EQAVET – European Quality Assurance for Vocational Education and Training. Currently, they are under development, and so far have minor influence on actual educational systems. One should also mention the European Qualifications Framework, a tool to compare different educational and qualifications systems in Europe [Cedefop, 2012, 2015; UE, 2013]. As a consequence, this article proposes to analyze higher education separately. It will help to understand uniformization tendencies in education system in the EU.

Finally, the clustering model excludes variables on employability (contrary to Amable and Farkas). First reason was mentioned by Farkas – employment and unemployment are caused by many factors and a level of education may even be not the most important one. Furthermore, even if we would be able to precisely assess the impact of education on employability, it would bring some misleading conclusions. Education system is an institutional domain that produces strongly lagged results. To make this argument more visible, at least one-third of the labor force in CEE countries are the people who most of their education spend in real-socialist schools. Therefore, their employability adds little knowledge to the features of today’s schools.

The clustering is based on PCA and hierarchical clustering on principal components (HCPC). PCA is a statistical method that allows to recognize the most and the least important variables. That opens a number of further possibilities and, in context of typologies, the most crucial is clustering based on PCA.

PCA is based on geometrical representation of variables. To do so, we create the “Cloud of individuals N_I.” We can define every country profile as a set of characteristics: {x_ik; k ∈ (1,k)}, which can be interpreted as a vector on k-dimensional space. A distance between two individuals, i and l: d2=(i,l)=∑k(xik−xlk)2{d^2} = (i,l) = \sum\limits_k {{{\left( {{x_{ik}} - {x_{lk}}} \right)}^2}} .

We can further define G_I as a center of gravity. It can be interpreted as a mean point of all the variables. Therefore, for any given individual (country profile) i, we get d2(i,GI)=∑k(xik−x¯k)2{d^2}\left( {i,{G_I}} \right) = \sum\limits_k {{{\left( {{x_{ik}} - {{\bar x}_k}} \right)}^2}} . That distance is sometimes called “the peculiarity” of the individual i, since it shows, how one differs from others. Later on, those calculations allow to define a total inertia of a cloud N_I: NIGI=∑k∑ipi(xik−x¯k)2=∑kVar[k]{{{N_I}} \over {{G_I}}} = \sum\limits_k {\sum\limits_i {{p_i}{{\left( {{x_{ik}} - {{\bar x}_k}} \right)}^2} = \sum\limits_k {Var[k]} } }

That measure is particularly useful in statistical analysis since it allows to differentiate variables and individuals with a use of variance. Further specification of PCA method can be found in Pagès [2014].

Based on PCA results, there performed a HCPC to identify clusters [Kassambara, 2017].

All the calculations were performed in R with use of the packages “FactoMineR” [Husson et al., 2018] and “Factoextra” [Kassambara and Mundt, 2017], both designed for factor and cluster analyses.

Since the goal of the article is to explore the institutional diversity of educational systems, only input variables were included. As explained in Section 4.1, variables on employability were excluded from the model (contrary to Farkas and to the lesser extent Amable), as well as those analyzing innovativeness. The model excluded also all effectiveness measures, to fully focus of direct institutional underpinnings. It is perhaps considerable to add measures such as Programme for International Student Assessment (PISA) and Programme for the International Assessment of Adult Competencies (PIAAC) results in further extensions of the model.

Obviously, vast majority of the variables epitomize the features of formal institutions; however, the informal institutions are – indirectly – represented as well. Perhaps the best example is the percentage of students in VET programs. It is both the result of formal institutions (e.g., political decisions on financing the system) and informal institutions (e.g., that creates different opinions on the prestige of this path of education).

The dataset combines three sources: the enrollment to education (in various aspects) and data on compulsory education are obtained from World Bank Data Bank, data on teacher salaries are taken from OECD. All other variables are taken from Eurostat. In all cases, the latest possible data were used, usually it is 2016 or 2017. Results for Cyprus may not be precise, since relatively high number of missing values.

The complete list of variables is shown in Table A1 in Appendix A.1.

There were 38 variables included in a model. Figure 2 shows, how they influenced the clustering, in the first factorial plane. The interpretation of this figure is following: every arrow represents one variable. If a country is high performing in a given category it follows the direction of this arrow, if low performing it goes in the opposite direction. Colors represent the importance of a given variable: those lighter are more important, and those darker have lower significance.

Figure 2

It is clearly visible that the variables that affect clustering the most is the share of the students in public institutions, at all levels of education. Furthermore, it seems interesting expenditures in relation to GDP go in another direction.

Figure 3 shows the countries’ representation in the first factorial plane. The most obvious example here is the United Kingdom, located far in the right-down quarter. This is caused mainly by the high share of students in private institutions. Other countries appear to near the middle of the factor map. One has to be careful with interpretation of this figure – proximity of countries on the map does not directly reflect their actual institutional proximity.

Figure 3

Figure 4 shows the cluster dendrogram: most important figure in this article. The dendrogram shows the institutional proximities between countries and therefore allows identification of clusters.

Figure 4

Six clusters were distinguished. They are marked on the figure in dashed boxes. The first cluster, including Belgium and the United Kingdom, is the most specific and farthest from other countries. The United Kingdom confirms here being a member of LME. Ambiguous case of Belgium should be treated carefully – this country has different education systems for Flemish and French communities (and also one for German-speaking community) so the variables for this country represent the average of two subsystems and may not in fact credibly describe any of them. Nonetheless, both Flemish and French communities are characterized by relatively high number of students in private institutions and this is the main reason, why Belgium was classified in one cluster with the United Kingdom.

Interesting case is also a cluster of Estonia and Latvia. If they were to be combined with another cluster, it would be a Scandinavian one rather than more expected CEE.

The height of the lines represents the so-called cophenetic distance between countries. It can be interpreted in a following way – the smaller is the height, the more similar are given countries. This means that the two most similar countries in the pool are Czech Republic and Slovakia. It should be noted that in general heights in this dendrogram are relatively high, with cophenetic correlation of 0.52. This number confirms the high level of heterogeneity.

Further description of clusters’ characteristics is shown in Section 6.

Figure 5 offers the comparison between New and Old Member States. This figure shows that old member states are much more diverse in terms of institutional underpinnings of educational systems. On the other hand, it shows, that in case of educational systems, it is difficult to talk about “post-communist” economies.

Figure 5

The clustering in higher education was performed analogically to the previous one. Figure 6 presents the active variables in the first factorial plane. Again, the most important aspect in case of clustering is public versus private institutions nexus. Out of all fields of study, the most important in terms of importance for clustering appear engineering, agriculture, and services.

Figure 6

Figure 7 shows the countries on the map based on Figure 6. Again, the most specific is the case of the United Kingdom.

Figure 7

Figure 8 shows the cluster dendrogram for higher education.

Figure 8

There were five clusters identified in this subsystem. The first important and interesting conclusion is that in case of higher education virtually all the New Member States are included in one cluster. As a result, cluster 5 here is very similar to Rapacki’s cluster 4.

Luxembourg constitutes the separate cluster mainly due to very high number of students abroad. It can be, however, merged with cluster 3. Another interesting conclusion is that, when higher education is treated separately, the uniqueness of Belgium disappears. This may confirm the intuition that this was caused mainly by the institutional ambiguity between Flemish and French communities.

One should also note that proximities between countries in case of higher education are in average smaller than in a clustering based on all the educational variables, with cophenetic correlation of 0.76.

Further characteristics are shown in Section 6.

Contrary to the general clustering, the model for higher education shows more diversity for New Member States. This is however to the great extent caused by Latvia and Estonia, located far in the left-down quarter (Figure 9).

Figure 9