More than meets the eye: uncovering the evolution of the OECD’s institutional priorities in education

ABSTRACT Over the last 20 years, since the launch of its flagship study the Programme for International Student Assessment (PISA), the Organization for Economic Co-operation and Development (OECD) has become a behemoth of transnational influence on education policy making. To better understand the evolution of the OECD’s perspective on education and the importance of PISA within the OECD’s overall publications, I analysed over 900 OECD publications on education with a topic modelling approach. The analysis revealed that contrary to popular belief, the OECD is a multi-centric organization with various agendas in education policy making, focusing on economic changes, higher education, management, planning, and budgeting. PISA made up only 12% of the overall text corpus and was therefore far less discussed than expected. This is in sharp contrast to the general impression held, namely that the OECD’s work in education mostly relates to PISA. Therefore, this study reveals that the OECD has a far broader agenda than anticipated. This paper acts as a systematic review of OECD literature to gain insights into its institutional objectives.


Introduction
The Organization for Economic Co-operation and Development's (OECD) most wellknown study, the Programme for International Student Assessment (PISA), launched its first results in 2001. Since then, there has been an ongoing debate among the public and policy makers regarding what it means to be both well-educated as an individual and to have a good education system as a country. In a globalized world, having the right kind of education to satisfy the needs of the labour market, while retaining individual freedom and one's choice of profession seems more important than ever. In light of this debate and the OECD's behemoth-like governing power in this deliberation, I ask what the overall focal point of the OECD's education publications is and how this emphasis has developed over time. In this context, I also investigate what place PISA occupies within the OECD's overall work on education policy.
Providing high-quality education has always been regarded as the modern welfare state's responsibility toward its inhabitants (Meyer, John Boli, and Ramirez 1997;Weymann 2014). A problem arises in determining an education systems' ability to equip students with the tools necessary to overcome inequality and adversity so that they can succeed in a globalized labour market. Therefore, assessing the status quo of students' educational 21 st century skills, and a state's capacity to educate its citizens is the premise under which the OECD publishes countless evaluations on the state of countries' education systems and governance tools. Some studies suggest that the OECD's tireless work with PISA and its results might even be the cause of a global increase in education policy reforms, affecting not only OECD member states but also other countries (Breakspear 2012). It certainly has increased the number of studies on education, and the public debate around educational quality has gained new traction (Hopfenbeck et al. 2018). One could say that this development has contributed to a shift in power for education policy making; not only has education been moved to the forefront of policy making for the betterment of society, but a major shift of responsibility from national governments to international organizations (IOs) has occurred at the same time.
However, the OECD and especially their primary assessment tool, PISA, are heavily criticised. Aside from analytical limitations, the OECD's focus on preparing students for a specifically Western-oriented labour market along with its limited assessment of only math, science, and reading courses -which potentially overstates the importance of these skills in everyday life -are recurring issues. Moreover, the focus of OECD comparisons, reports and publications is often seen as too narrow to represent the skillset generally considered to be important for life (Addey 2017;Fischman et al. 2019;Zhao 2020;Sjöberg 2019). Other studies however suggest that the OECD has no influence in limiting curricula at all, and country comparisons do not influence education reforms (Rautalin, Alasuutari, and Vento 2018;Fischman et al. 2019;Baird et al. 2016). This contradictory evidence regarding the OECD's agenda and influence, as well as scepticism concerning the OECD's objectives for curricula and otherwise, necessitate a closer examination of the organization's overall educational activities, both historically and presently.
A bird's-eye view of education policy in general reveals that policies are diffusing and education systems all around the globe are becoming increasingly similar, mostly aligning with OECD standards. This development illustrates that the balance of power IOs could exude over states has tilted considerably in their favour.
The most frequently voiced criticism, however, is not directed toward the diffusion of policies in itself, but rather at the specific type of policies that might be diffusing due to increased awareness through PISA and reports. In other words, there is concern that the perspective of the OECD and its underlying ideas on education are too narrow and economically oriented but nevertheless inform education policy makers (Morgan and Volante 2016;Sjöberg 2019). Education systems might be increasingly focused on teaching knowledge and skills that directly transfer to labour market purposes rather than on counteracting social inequalities (Addey 2017;Fischman et al. 2019). Conversely, the OECD negates this perspective and repeatedly underlines its focus on equity and equality (OECD 2018b(OECD , 2018a. Due to the scepticism toward PISA, the OECD is already being studied by scholars; however, most studies concerned with the OECD's activities are limited on the impact of PISA (e.g. Hopfenbeck et al. 2018;Domínguez, Vieira, and Vidal 2012;Breakspear 2012), the applicability of OECD standards and benchmarks in different contexts (e.g. Bulle 2011), or the OECD's role as a knowledge broker for evidence-based policy making (e.g. Niemann and Martens 2018). Several publications discuss the OECD's orientation and perspective on education (Ydesen and Bomholt 2020;Bürgi 2017;Tröhler 2013;Sellar and Lingard 2013;Ydesen 2019).The internal complexity of the OECD as an organization and its agenda-setting procedures are only scarcely studied (e.g. Centeno 2018).
Nevertheless, none of the aforementioned studies could answer the question: What are the OECD's institutional priorities with education over time and how does PISA fit within this body of work? The underlying model, form, and function of education, which the OECD is testing for and recommending is still unclear, despite increasingly frequent publications on the organization.
In this paper, I am therefore inspecting the OECD's overall emphasis on education and how this has evolved. I study most of the OECD's publications on education with a topic model to detect what the OECD regards as important in education, which topics are most discussed, and how the belief of what matters most has changed over time. The overall goal is to determine the OECD's changing priorities in education policy. In addition, I will pay special attention to the prevalence of PISA and related publications within the entire portfolio, mainly due to the overwhelming attention and criticism PISA has gained.

Contextualization
It is important to clarify how the OECD has gained influence in educational policy making, and how the IO exercises its power to justify the in-depth analysis of this paper. It is also imperative to understand the OECD as a complex bureaucracy and what this implicates for the IO's internal structure, process, and consequently its output. In the following section, I will therefore, introduce the theoretical framework and previous research necessary to put this paper and its results in the proper context. To this end, I will introduce the general mechanism of soft power through which IOs influence national policy making. Then, I will discuss the development of the OECD's institutional form with Barnett and Finnemore's understanding of IOs as complex bureaucracies. Lastly, I will focus on the OECD's historical and ideological development to stress how the OECD's institutional structure might contribute to the OECD's agenda setting.

Understanding IO influence
To understand the importance of knowing the OECD's overall agenda, it is helpful to discuss the general mechanisms behind IOs' influence. This section, therefore, reviews how IOs gain their legitimacy, to clarify how IO authority can facilitate the diffusion of policies, standards, and norms. This is important especially in the case of the OECD, an IO without any 'hard power', which is still able to shape education policies, even influencing non-member states.
Formally, IOs are associations of member states, in most cases established to mediate issues concerning multiple countries simultaneously. These institutions are based on internationally agreed upon principles, standards, and rules that harmonize the behavioural expectations of the actors involved (Rittberger et al. 2019). However, as bureaucracies, IOs are additionally equipped with a distinct form of social authority and the expectation of rational behaviour (Barnett and Finnemore 2004). Consequently, IOs can appear as actors themselves, which allows them to steer their members' behaviour. These expectations of rationality and rule-bound behaviour, coupled with the analytical expertise, not only gives IOs like the OECD their legitimacy and authority but also their agency.
What sets the OECD apart from other IOs such as the United Nations Educational, Scientific and Cultural Organization (UNESCO), the Word Bank, or the European Union (EU), is the fact that the OECD has no prescriptive mandate over its member states, and thus cannot coerce members to adhere to standards and norms; nor can it distribute or cut monetary capital, as for example the World Bank can. The OECD exercises its so-called 'soft power' through the well thought out broadcasting of its analytical findings (Bieber and Martens 2011;Atkinson 2010;Addey and Sellar 2018;Rutkowski 2007). Through its reports and evaluations, the OECD greatly contributes to the overall content of discussion regarding education policy. Since IOs -like the OECD -are perceived as rational actors it is generally seen as beneficial for governments to accept their recommendations as it improves states' legitimacy as rational political entities (Meyer, John Boli, and Ramirez 1997). Thus, to oversimplify and overstate this concept, one could claim that whoever controls the global discourse, controls the world (at least to some degree). The reliance on soft power and the incredible influence the OECD has despite this, underlines how important it is to thoroughly understand the IO's output.

The OECD's internal structures
The bureaucratic structure of any IO shapes its distinctive internal logic and behaviour (Barnett and Finnemore 2004). Complex bureaucracies are designed on efficiency and tend to specialize and strongly compartmentalize with increasing complexity of the field. This can result in a multitude of organizational units or departments within an IO, which is the case for the OECD. Each department has its 'social' components, with slightly different views, goals, and dynamics. Being highly dependent not only on the cultural context, but also staffs' expertise and training, IOs are also products of these factors (Barnett and Finnemore 2004;Meyer and Rowan 1977). In addition, the OECD is a mandate-based IO and different countries might request different services, further scattering the OECD's education agenda. This internal structure may foreshadow some of the results of this analysis. Overly complex internal structures may result in issues such as unresponsiveness to input from national delegates (Centeno 2018), as the IO's structure can be stiff and inflexible (Barnett and Finnemore, 1999). Centeno (2018, 93) describes three general decision-making levels in the OECD: the summit level (governing and executive power) with the council and the Secretary-General, the organizational level (tactical and formal power) with the secretariats, and the sector specific activities level (official activities-related) with committees. Semi-autonomous working groups can be established across all levels. While the council consists of member-states' envoys and has political and administrative power over the IO, it delegates its strategic power to the Secretary-General, who then sets the agenda.
The Secretariat of Education itself consists of directorates, which often have subdepartments and divisions, responsible for planning and preparing different projects and publication series. The committees (manned by country envoys) can intervene in all activities, however, they are frequently overruled by the Secretariat, as Centeno states (2018, 95). There are several important semi-autonomous bodies to mention, such as the Centre for Educational Research and Innovation (CERI) founded in 1968. Its governing board consists of member countries' envoys, it focuses on research, innovation, and indicator development (Papdopoulos 1994;Schuller 2005). Another important unit is the Indicators of Education Statistics (INES) program, established in 1988 as a precursor of today's PISA department -responsible for collecting quantitative data (Ydesen and Grek 2020). The founding of INES marked a very important transition in the location of responsibility within the IO; the OECD started taking control over national indicators. In fact, through the INES program, indicators were established as providers for reliable data for benchmarking. This eventually led to the transnational competitiveness in education we see today. The program facilitated the OECD's move from knowledge production to policy making and helped establish the IO as key actor in the education policy landscape (Grek and Ydesen 2021).

Ideological developments
Since the beginning, there have been debates between a US market liberalism approach and a more European social democratic tradition (Lingard and Grek, 2007). During the first few years, there were intense disputes concerning the relationship between democratization and economic advancement as well as the idea of public education being a prerequisite for both these concepts. Placing education in the hands of economists caused some discomfort (Spring 2015). Additionally, concepts such as governance from a distance, delegating accountability, and the focus on resource management were grounds for debate (Ydesen 2019;Tröhler 2014).
The introduction of privatization and social equity as necessity for human capital and innovation under the label 'New Public Management' (NPM) during the 1980s laid the foundation for the introduction of test-based accountability and the OECD's outcome orientation (Lingard and Sellar 2016). Accountability had become the key mechanism for the state to regulate and maintain control while not being directly involved in education provision (Verger and Parcerisa 2018). During the 70s and 80s, the IO still entertained philosophical and ideological debates about the nature and application of performance measures. Then, during the 90s, the IO moved its focus from input to output orientation (Papdopoulos 1994). This indicates a significant change in the overall approach to education, as the output orientation made way for neoliberal governance, where the local level is held accountable for the implementation of reforms and the overarching governing structure steers from a distance (Verger and Parcerisa 2018). The (re)definition of lifelong learning as a tool for a more flexible adjustment of peoples' knowledge and skills caters to the needs of changing labour markets (Rizvi and Lingard 2006). In 1997 the 'Definition and Selection of Competences' (DeSeCo) program was initiated, starting a new classification and standardization strategy in education (OECD 2002). Thus, the OECD went from 'philosophical doubt to statistical confidence; from covering some countries to covering most of the world; from a focus on inputs to a focus on outputs; and from occupying an experimental status to being a central part of the Organization's educational work.' (Henry et al. 2001, 90).
Today, the OECD is especially proud of PISA's 'real life application', the transfer of academic knowledge to the 'real world', and its assessment of 21 st century skills -and according to Andreas Schleicher, this agenda will be expanded in the upcoming years (Schleicher 2014;OECD 2018a). In fact, aside from 'PISA for Development', a study designed for countries without the capacity to implement 'regular' PISA, multiple other studies allowing single school assessments (PISA for Schools), teachers' use of OECD assessments (PISA4U), and even a 'Baby PISA' for toddlers are all at different stages of development (Lewis 2019).
Thus, even though the OECD's current ideological perspective is very much a neoliberal one, all publications contain in some form or another the assurance that equity, inclusion, and social cohesion are incredibly important to the IO (OECD 2018a). Papadopulous describes this ongoing dualism as 'tension between efficiency and equity' (Papdopoulos 1994).
It becomes clear that there was always internal conflict brewing: The debate of equity against effectiveness, the internal struggle over the discomfort with the economization of education, the public declaration of an equity-focused agenda while the publication output demonstrated otherwise, and so on. This might be a reason for, or a consequence of the different working groups and managerial changes within the OECD.
Despite the aforementioned studies on the OECD's education agenda and development, the overall thematic output over time it is still unclear. The institutional priorities of the OECD are ambiguous. To fully understand the consequence of these struggles in the education agenda of the OECD, it would be necessary to read a large amount of OECD publications, which seems nearly impossible given the sheer size of the yearly output. Although, the complexity of the issue would suggest the necessity to do so (Sellar, Thompson, and Rutkowski 2017). With this paper, I therefore intend to provide an alternative approach to examine the OECD's entire body of publications on education in hopes of gaining insight into how the organization's understanding of education has evolved over time. Important to note is that I do not examine what impacts policy making, but rather provide an exhaustive overview of the OECD's output.

Topic modelling
In recent years, the production of data in various forms has skyrocketed. However, the new quantity of data produced every day also demands new methods of analysis to keep up with production speed. Qualitative and even some forms of quantitative text analysis require time-intensive reading as well as the coding and preparing of documents, an endeavour not always feasible for unstructured textual data. Topic modelling offers an alternative solution to classify, label, and reduce these unstructured masses of textual data (Wesslen 2018). Topic modelling as a quantitative text analysis method trains an unsupervised machine-learning algorithm to discover a set of topics in a corpus of documents. This method allows researchers to detect topics in large corpora of textual data, reducing the data, without having to read all of it, a process and result similar to clustering methods. This represents an alternative to qualitative text analysis, as it is a comparatively fast procedure and can create an overview over the corpus. It is also a method that is specifically helpful to create a coarse overview, rather than showing nuanced differences within the discourse. As stated before, the overall thematic emphasis of the OECD in education is somewhat ambiguous. Therefore, topic modelling is a great method to gain a full picture of the OECD's work as preparation for more nuanced qualitative analyses.
Topics are defined as a multinomial distribution over words, each term has a probability of belonging to a topic. A document can be understood as a distribution over all topics. Consequently, a document is a representation of all topics with varying proportions. This means that the algorithm groups words to form a topic depending on how often they appear together in the text corpus. It is a bag of words approach; the sentence structure, therefore, does not matter. A probability for the word combinations in each topic is then calculated, showing how often a topic appears in the corpus.
The Structural Topic Model framework (STM) by  allows for the joint estimation of topics and document-level metadata. STM is implemented in the statistics environment R (R Core Team 2019) and will be used for the following analyses. I will include a non-linear measure of time as a covariate and let the topics vary over time in order to detect the topic prevalence, i.e. how much a topic is discussed at a given point in time.
In general, topic modelling can be viewed as speed-reading large quantities of textual data. The results can therefore vary between models and interpreters. Although not all information in a corpus will be extracted, the most frequently used topics will be discovered even in repeated analyses. The clear advantage of topic modelling to qualitatively coding documents lies in the simultaneous analysis of extensive amounts of documents, an otherwise nearly impossible endeavour. Thus, the contribution of this paper is its analysis of the entire corpus of documents on education produced by the OECD thus far, which has never before been analysed to this extent.

The OECD's education publication output
Over 7,297 publications in multiple languages can be found under the keyword 'Education' in the OECD's online library. 1 This includes chapters, single graphs, and datasets along with full publications, working papers, series, reports and books. Over 1,000 publications were initially collected for the analysis in this paper, both from the online library in December 2019 and older publications from the OECD archives in Paris. I excluded single chapters published separately, graphs, datasets, policy briefs, and working papers. Even though the latter are published by the OECD, they often include a disclaimer, stating that they do not represent the organizations' views but only the authors'. Finally, 941 publications from 1955 until 2019 were included in the final analysis. There might be new publications available upon the date of publication of this paper. 50% of the publications were published after 2001, which shows the drastic increase in publications on education in recent years. Some of the documents are bilingual and include pages in French that were not removable. The publications could be sorted into 90 different publication series, such as 'Education Policy Outlook', 'OECD Reviews of Migrant Education', and so forth.
The publications were processed with R and cleaned irrelevant sections such as tables, references and hyperlinks. On average, a document contained 165 pages with 6,678 lines of which 18% have been cleaned of references, graphs, or white-space. The removed parts were stored separately and revised thoroughly.
The documents were treated with the standard procedure for topic modelling, lemmatization, stop-word removal, and removal of non-readable words. Often co-occurring words and country names were linked together, such as 'education system', 'labour market', and 'policy analysis'. The final corpus contained 941 documents and a dictionary with a total of 308,734 words.

Results
In the following section, I will first elaborate on the model evaluation and quality then I will present the chosen model and the topics themselves. For this I will first show the proportion of the topics in the corpus, contrast this result with a reduced corpus, analyse the correlation between the topics, and lastly, interpret the topics' development over time.

Evaluation
Topics developed from a machine learning algorithm are resulting from counting word co-occurrences and, therefore, often lack coherence for the human interpreter.
Consequently, a few model evaluations need to be undertaken. All following models are calculated with the year as a b-splined prevalence covariate, meaning the occurrence of the latent topics is estimated depending on time (year of publication), while the effect of time is not considered to be linear.
Models with a large number of topics are often used for detailed analysis, whereas models with few topics serve for a better overview (Asmussen and Møller 2019).  recommend around 5 to 50 topics in a corpus with a few hundred to few thousand documents, such as this. Figure 1 shows two parameters for model evaluations of multiple models with different numbers of topics: the residuals and the semantic coherence. Semantic coherence describes the internal consistency of words within topics and the residuals represent the unexplained variance within each model. Accordingly, semantic coherence should be maximized and the residuals minimized.
Based on the fact that the residuals are lowest at 27 topics while the semantic coherence is highest at 20, 24 and 30 topics, a model with either number of topics could be chosen. The model selection can be difficult, which also explains the reluctance of social scientists to use topic modelling to gain an overview of text corpora. Contrasting the semantic coherence and the exclusivity of the different models together can share more insights into the models. Exclusivity, in contrast to semantic coherence, describes how selective words are in certain topics and how rarely words in one topic occur in another. The list of the 30 most prevalent words in each topic can be found in the appendix materials (Table 1).
In next step, the topics were evaluated qualitatively with a procedure loosely based on the evaluation task by Mimno et al. (2011) and Chang et al. (2009). The topics were evaluated for their coherence: Do the top 30 words in one topic represent one topic or multiple, and are there words not fitting the general theme? What are the top 10 documents of each topic discussing? Are the concepts of the topics consistent with the documents? Consequently, the topics were labelled according to the words and the documents.
In general, some mismatch of words and documents is to be expected. Interestingly, most topics in this model are found in one publication series respectively, and there seems to be a time-dependence of the topics as well, seeing as most documents representing a topic were often published in the same decade. Several topics did not seem to consist of a coherent topic (topics 5, 10, 19, and 29), but the documents belonged to one series which in turn covers multiple different issues such as Education at a Glance. A few series covered more narrow issues than the topic suggests (topic 3, 14, 16, 25). Two topics seemed to contain subtopics (11,20). One topic contained all French sections in the corpus and was found in the earliest documents published (22). To conclude the evaluation, most of the topics can be interpreted and will be discussed in further analysis. Interestingly, most topics depend on the decade and the publication series, demonstrating that each decade and publication series has its own vocabulary. Even series like Education at a Glance, which cover different aspects of education systems, are clustered to contain one topic, further confirming this observation. Figure 3 shows that the most prevalent topics are Changing Society (18), System Quality (16), Higher Education Management and Planning (9), Regional Governance in Higher Education (27), and Labour Market Skills (21), making up about a third of all topics. The least prevalent topics are concerned with Education System Classification (1), Primary and Secondary Education (22), and PISA Data Analysis (7).

Prevalence
Interestingly, contrary to popular perception of the OECD's work in education, the majority of topics in the entire corpus are not concerned with PISA. Only four topics primarily consist of documents on PISA: PISA Results (28), PISA Results and Disadvantaged Students (13), PISA Analytical Framework (6), and PISA Data Analysis (7). In total, only 9.72% of the full corpus are topics involving PISA, which is much less than expected, especially since PISA has almost become a synonym for the OECD itself. To make sense of this finding, the model was refitted to include only the documents published after 2000 to reduce the bias of time. Figure 4 shows the topic proportions of the full model and the refitted model. Naturally, in the reduced model, PISA has a higher prevalence. In this model, System Quality (16) and Labour Market Skill (21), followed by School Resources (3), are the most prevalent topics covering 10.06%, 8.51%, and 8.29% of the corpus, respectively. The four topics directly connected to PISA, collectively, cover 17.01% of the corpus from 2000 to 2018. So, even when controlling for the fact that PISA only started in 2000, the number of documents discussing PISA is relatively low compared to other topics.
Many of the topics include aspect of higher education, retraining, management, or include another connection to the labour market, relating to different groups on the labour market or working women with families. The strong connection to the labour market can be somewhat explained by the original mandate of the OECD: rebuilding Europe's economy. Also, the perpetuation of the economization of education is somewhat explicable, given the OECD's history. The management aspect has been a priority since 1980 (Lingard and Sellar 2016). Both primary and secondary education are discussed much less frequently than expected. The same goes for equity and equality. The concepts of 'equity' and 'equality' appear in the top 50 most probable words of three topics (23, 28, 29), while 'economy', 'productivity', and 'labour market' appear in 7 (13,20,21,23,25,26,27). These results hint toward the OECD's main target in education: ensuring a productive labour market with skilled workers.

Topic correlations
To further inspect how the topics relate to one another, I built a network of topic correlations. Topics are positively correlated when they appear in the same documents. The graph was clustered with a Louvain Clustering method for community detection to make the correlations more apparent. The node size depends on the topic prevalence; the edge width represents the strength of the correlation between the topics, showing document overlap.
There are five clusters of topics sharing documents seen in Figure 5. The cluster First Classifications consists of topics found in the early documents. The second cluster Innovation incudes topics related to innovations as a reaction to changing society and working population. Two topics seem out of order here: Higher Education Management and Planning (9), as well as Cost and Return of Education (12), both of which would also fit just as well in the cluster on Higher Education, Development and Governance. Cluster three contains the PISA topics and cluster four topic concerning Standards and the trends thereof. The fifth cluster, Higher Education, Development and Governance is most diverse, as it contains topics on all levels of education with a strong emphasis on tertiary education, policies, and development. The PISA cluster is only correlated with the TALIS topic, Understanding Learning and the Brain, and Country and Indicator Trends in the other clusters. Topic 5, Innovation and Planning is correlated to most other clusters, while topic 16, System Quality is most central in its own cluster; many documents are therefore dealing with System Quality.

Regression
As a last step, a final regression analysis was performed. The regression coefficients are usually not interpreted as such, but rather introduced to analyse the ratio of topics over time while controlling for metadata. The following Figures 6, 7, and 8 show the prevalence of topics over time with a 95% confidence interval.
The topics vary a lot over time and have different amplitudes and confidence intervals. The topics that are most prevalent in the early years show a much greater amplitude, as there are fewer documents and therefore a smaller range of different topics that could be detected. Most topics follow the expected distribution. Training for Scientific and Technical Personnel (14) and Higher Education and Manpower (11) are among the earliest topics, while Innovation and Planning (5), Higher Education Management and Planning (9) and Changing Society (16) have the highest amplitude between 1970 and 2005. Recent spikes in prevalence include PISA topics, but are also seen in System Quality (16), Labour Market Skills (21), and School Resources (3). These time-dependencies can be attributed to the main agenda at the respective point in time: Initially, new scientific personnel were needed; after the initiation of the CERI and INES, a new focus for governance through management arose and, more recently, a PISA-related shift to include secondary education as a priority, instead of higher education as seen in the earlier decades.
Especially interesting is that topics related to indicators and standards (2, 8, 10) are consistently present but with a very low prevalence. The OECD claims to be the standard-setting institution in education, yet the discussion of set standards is a relatively low concern in terms of publication volume. Another interesting result is the recent decline of planning related topics (5,9,11), System Quality (of early childhood education institutions), however, is steadily more discussed, along with Labour Market Skills. What is surprising here is not only the short life-cycle, but also the low prevalence and amplitude of the PISA topics. The prevalence of topics concerned with equity is relatively low as well, despite the recurring assurance of its priority as described earlier.
PISA Results of Disadvantaged Students and Children at Risk are the only obvious topics concerned with equality of opportunity. In addition, of the 50 most prevalent words for all topics, only two (28: PISA Results and 29: Reviews of National Higher Education Programs) contain the word equity. None of the topics include equality, and only one (23: Working Women and Families) includes inequality. This is a strong contrast to the OECD's repeated affirmation of its emphasis on equity. Additionally, a rather large amount of the topics is concerned with management in higher education. The organization and management of formalized education seems to be of much higher concern for the OECD than originally anticipated. The most prevalent topics can be described as discussing the reaction of institutions to a changing economy and society, by planning and managing as well as developing strategies for regional governance.

Discussion
As seen in the previous section, the OECD has a large area of interest and influence in education. The topics with the highest overall prevalence were concerned with Changing Society, System Quality, and Higher Education Management and Planning. What is interesting here are not only the topics that are detectable but also the topics that are not. The topics detected by the algorithm paint a fairly different picture of the OECD's main concerns and interests in education than expected. While the overall themea recurring discussion of education fitting to the labour market -is not surprising given the OECD's founding mission, the mismatch between the common perception of the OECD's work and the topics actually detected in the corpus is quite revealing. For example, policy advice -something the OECD has taken on as its mission -seems to come wrapped in a statistic about the current trends and challenges, as it cannot be detected as a unique topic as it is not clearly distinguishable in the documents. The low prevalence of topics more explicitly showing policy reviews or advice hints toward a more covert transmission of policy recommendations by being embedded in other topics. Similarly, despite the overall view that assessments and system evaluation are a main objective of the OECD, there are surprisingly few topics dealing directly or specifically with school assessment aside from the topic Assessment and Evaluation (30). Contrary to popular belief, the assessment of the status quo, therefore seems to play a minor role.
As discussed earlier, the analysis yielded surprising results regarding the OECD's stance on equity in education. The OECD communicates social cohesion and equal opportunities as being distinct and of utmost importance for their work but the low prevalence of these topics is surprising. Papadopulous (1994) named this the 'tension between efficiency and equity', a tension that seems to be leaning toward efficiency given the fact that there are more topics concerned with management and planning rather than equity or inequality.
Probably the most interesting results are the multitude of topics and the varying percentages of the topics within the corpus. Given the recent preoccupation of educational research with criticism related to PISA, the multitude of other topics the OECD is concerned about is much less discussed. Additionally, PISA-related topics, something the OECD is mostly credited for, make up around 10% of the entire corpus, which is a relatively small percentage given the extensive media coverage PISA receives. Even when only a subset of documents is included in the analysis, PISA still makes up only around 17% of the corpus -a surprisingly low percentage. A considerable amount of the corpus is concerned with topics outside of formal schooling such as higher education, human resources, or management and planning. Education, seems to be comprised of more than 'just' the compulsory part of formal schooling, rather it is viewed as a collection of skills dependent on economic demands. These skills are distributed across all age groups and their relevance requires regular evaluation, innovation, and management. This is an aspect which is not often discussed when it comes to evaluating the OECD's work. Given the previous elaboration of the OECD as a bureaucracy and the resulting tendency for specialization and compartmentalization this should, however, not come as a surprise. The OECD's simultaneous discussion of similar issues (management, planning, and finance) across different topics could be attributed to the internal complexity and the multitude of units within the OECD -all of which have varying publications. Naturally, as shown by the results, these different publication series contain varying topics. However, some of the topics overlap between series, so does the vocabulary. The internal struggle of the OECD and the publication of multiple topics with similar wordings, such as 18 and 19 (both concerning societal changes) or 5 and 8 (both concerning planning), could be derived from the struggle of internal units within the OECD. The duplication of topics in different units as well as the separation from the overall IO agenda could be a result of different staff and orientations in these units. It is difficult to discern which organizational unit is responsible for the different publication series on education. However, based on the historical development of the OECD, and the internal ideological disagreements that have been discussed above, the development of conflicting or duplicate agendas within different publication series and organizational units is an expected outcome. In addition, the OECD is attempting to unite the needs of a wide range of culturally different member countries, a struggle as demonstrated in their efforts on taxation (Sharman 2012).
Only one question remains: why is PISA is the OECD's norm-giving tool and overall legacy in education? Even though PISA and system assessments are important to the OECD and their prevalence in later publications is relatively high, the IO is still bound by the limits of its bureaucratic nature. And even though it can act independently due to the high concentration of expertise within the IO, it is still bound to its member states demands. Given the OECD's general authority in education policy and their status as 'rationalized other', policy makers and scholars should be aware of the fields the OECD is potentially influencing, aside from secondary education.

Conclusion
In this paper I analysed over 900 documents of the OECD on education with a topic modelling approach to determine the overall focus of the OECD for the field of education. Additionally, this paper explored, if and how this focus changed over time, and what role the PISA-study plays in this framework. The detection of certain concepts indicates that these concepts are increasingly important for the organization. They are only detected if they are discussed relatively often and can be distinguished from other concepts in the corpus through their exclusive vocabulary.
The publications' contents were sorted into 30 topics, showing a broad range of interests of the OECD. The most prominent topics are concerned with changes in social and labour market structure, planning and budgeting in higher education, skill development specifically for the labour market, and innovation. In its early years, the OECD was mostly focused on education system classification, training technicians and the cost of education. Issues such as the changing society, higher education management and planning, as well as innovation; were much more important during the 80s. During the late 90s, higher education, city development, and tertiary policies were more prevalent. In 2000, PISA and TALIS brought assessment and evaluation, school resources, and indicators to the centre of attention. Disadvantaged students' results in PISA, skills for the labour market, and the quality of early childhood education programs (System Quality) are the newest topics to be discussed extensively. The OECD, therefore, went from openly planning to educate for the economy (training scientific and technical personnel), to managing and planning higher education (increasing the focus on regional governance), and lastly to PISA, early childhood education and care programs, and labour market skills. It is striking how much higher education and innovation but also regional development dominate the topics. Innovation as a reaction to a changing world could be expected. However, it is not what the OECD is known for. The orientation toward regional development can be explained through the output orientation and especially through the New Public Management regime introduced in the 80s (Lingard and Sellar 2016), which directed the responsibility and accountability for education to the regional level. The multitude of different topics, sometimes intersecting heavily and other times outright contradicting each other, hint toward the internal complexity of the IO. As discussed before, the OECD has a multitude of internal units dealing with education, all potentially competing for funding and recognition, leading to a variety of changing 'agendas' in education.
Hence, even though the OECD is so active in education, it is helpful to keep its origins and organizational structure in mind when discussing its influence on education. Its overall goal was and always will be the advancement of its member states' economies (Bürgi 2017). This is reflected in the topics discussed. This aspect is often the source of disapproval among researchers concerned with the OECD's impact, almost polarizing the debate between the importance and effect of education on individual freedom versus the economic advancement of society as a whole (Zhao 2020). While it seems that most of the publications are indeed focused on higher education, planning, and ultimately the labour market, there are still publications concerned with disadvantaged students, different groups within the labour market, and societal changes. To what respect the reduction of inequality is targeted through economic advancement is a different issue.   My overall conclusion on the OECD's body of educational works is as follows: Surprisingly, the OECD discusses a multitude of topics over time, aside from PISA. In fact, PISA does not hold a prominent role within the OECD's publication output, despite the singularity of its success. In addition, there is a relative fluctuation of topics through time, with many concepts reoccurring or being discussed in multiple publication series simultaneously. This variety of topics reflects the complexity and the bureaucratic structure of the organization itself. Previous research already suggests what the analysis confirms: Multiple organizational units might result in multiple perspectives. Ultimately, the overall focus of the OECD and its publications in education seems to be integrating all citizens into the labour market after they have been 'managed' through different forms of state-regulated but locally organized education. When studying the OECD or utilizing the OECD's policy tools, whether it be PISA itself, its reports, or just the mere presence of its recommendations as justification for policy making, it is important to think of the OECD as an economic organization and to keep the breadth of its agenda in mind.