The Persistence Puzzle: Bibliometric Insights into Dropout in MOOCs

Abstract

Massive Open Online Courses (MOOCs) are a relatively new educational model that provides free access to educational content regardless of location or time. Despite these benefits, MOOCs encounter significant challenges, such as low completion rates, high dropout rates, and inconsistent participant comprehension, often due to the absence of simulations and practical activities. Incorporating sustainable education principles into MOOCs could bring benefits to the long-term effectiveness of the learning process, ensuring conscious learning practices. To address the issue of MOOC dropout rates and assess the scientific interest in this area, a bibliometric analysis was conducted on a dataset of 193 papers sourced from the ISI Web of Science database, spanning from 2013 to 2023. Papers were selected based on relevant keywords for the study. The analysis highlights key academic institutions, leading authors, and publication trends within this field. It reveals a strong and growing interest in MOOC dropout rates, with an annual growth rate of 40.04%. Research trends were identified by analyzing n-grams from keywords, titles, abstracts, and keywords plus, supplemented by a detailed review of the most cited papers globally. A collaborative network analysis was performed to explore author collaborations, their global distribution, institutional affiliations, research objectives, and study directions. The findings from the most cited papers show an increasing focus on understanding the factors contributing to MOOC dropout rates and developing strategies to address these issues.

1. Introduction

Massive open online courses (MOOCs) are online courses designed for a large number of participants [1]. MOOC platforms, such as Coursera, edX and Udacity, offer MOOCs in various disciplines in partnership with universities and colleges. They allow access to education for anyone with an internet connection, integrating multimedia elements for an immersive experience. MOOCs have become popular by breaking down geographical barriers and providing global access to quality educational content [2]. Since their inception, MOOCs have been divided into two types: Connectivist MOOCs—which focus on knowledge creation and generation, and Extended MOOCs—which focus on knowledge dissemination [3,4]. These platforms have come to be used to deepen a large number of courses that are part of different fields such as, but not limited to, the following: medicine [5], demography [6] and supply chain management [7].

1.1. Setting the Scene

MOOCs are characterized primarily by their open accessibility and capacity for unlimited participation, and they have significantly transformed the landscape of corporate education within the university setting [8]. Even by “corporate education”, one can infer the diversity of their use, going from the school level, even reaching the level of including educational processes or programs made available to employees by their employers to help them to instill different knowledge and to support them to have continuous professional training. Beyond the traditional videotaped lectures found in course materials, MOOCs offer forums designed specifically to foster community interactions, both among students and between instructors and students [9]. A crucial aspect of the long-term adoption if MOOCs relies on their degree of alignment with the sustainable education principles. By fostering inclusive, lifelong learning opportunities and integrating environmentally responsible digital practices, MOOCs can contribute to a more equitable and resource-efficient education system. Sustainable MOOCs emphasize accessibility, digital equity, and pedagogical strategies that promote engagement while reducing digital waste and unnecessary resource consumption [10].

As a brief overview of the history of MOOCs, they have evolved from the earliest methods of distance learning, starting with correspondence courses in the 1700s. The 1990s saw a major change with the advent of personal computers and the internet, allowing universities to offer online courses. These courses, initially intended for a restricted audience, gradually became accessible to the general public. Open learning initiatives in the 1970s and 1980s further widened access to education [11]. The first MOOC, “Connectivism and Connective Knowledge”, was offered for free by Stephen Downes and George Siemens at the University of Manitoba in 2008, and in 2011, Stanford University released “Introduction to AI” by Sebastian Thrun and Peter Norvig. The New York Times called 2012 the “Year of MOOCs” thanks to the rise in platforms such as Udacity, Coursera and edX [12,13].

Over the past two decades, distance learning has undergone rapid evolution. A pivotal moment in this transformation occurred in 2019 with the outbreak of the COVID-19 pandemic, which compelled countries worldwide to implement home learning measures, thereby significantly accelerating the adoption and advancement of distance learning technologies and methodologies [14]. MOOCs have provided free access to seminars by renowned professors and ongoing support through various events and educational materials during the COVID-19 pandemic. They ensure curriculum sharing, adaptive outcomes and open enrolment [15].

However, learner engagement in MOOCs tends to be self-directed and highly inconsistent, contributing to the high dropout rates frequently observed in these courses. This issue has been the focus of numerous research studies [16,17,18,19,20]. According to the researchers’ findings, even though MOOC enrollment requests were high, recorded completion rates were below 10% [21]. Low completion rates have created numerous problems, affecting benefits for learners, providers and MOOC platforms, which have lost both time and profits. Given the large number of interested learners, researchers have sought to understand the underlying motives that lead learners to drop out of MOOCs [22]. Celik and Cagiltay [23] summarize in their study that although the participants wanted to complete the courses in order to receive a certification, their intentions changed along the way, the main factor being lack of time. Aldowah et al. [24] identify six main factors that influence student dropout from MOOCs: feedback, academic skills, social support, course design, social presence, and prior experience. Secondary factors include engagement, time, interaction, motivation, course difficulty and family or work circumstances. Other studies involve, but are not limited to, the following: predicting both dropout and learner behavior [25], identifying techniques to reduce dropout and avoid the lack of interaction found among MOOC participants [26], identifying ways for tutors to intervene and reduce dropout [27] and finding ideas to make it safe to incorporate virtual manipulatives into MOOCs [28]. Zhu et al. [29,30] conducted a systematic review related to MOOC research trends and techniques in 2018 and 2020, respectively.

As MOOCs expand, incorporating sustainability-focused approaches will be critical to maintaining their effectiveness and durability. Sustainable MOOCs can act as a bridge to ongoing, adaptable learning while also addressing digital inclusiveness, environmental responsibility, and long-term student engagement. Exploring these characteristics can provide useful insights into optimizing MOOCs for global educational requirements.

1.2. The Objective of the Study

In this paper, a bibliometric analysis was conducted to identify and analyze the most influential papers in the field of MOOC dropout. As one of the most complex phenomena in contemporary online education, MOOC dropout remains a difficult dilemma to solve, especially from the perspective of identifying and tackling the factors that cause it. While research over the years has provided valuable insights, the above-mentioned studies confirm that drop-out persists and that its causes are interlinked and varied. These reasons can be likened to pieces of a jigsaw puzzle, each representing an essential element, which, when put together, form a pattern that favors dropout. Thus, the persistence of this phenomenon highlights the need for a more in-depth approach and customized solutions to increase retention in MOOCs.

In addition to providing readers with a comprehensive overview of the publications, authors, countries and affiliates that have made significant contributions to academic research in this field, this paper aims to observe trends in the analysis, to highlight the main themes studied, to highlight the temporal evolution, the relationships between concepts, theories and applicability, to uncover gaps and possible future directions of research, and to provide a comparison of both the results and the techniques applied. Furthermore, as far as the field of education is concerned, this paper can be used as a point for the identification of various barriers and the assessment of reasons for participation or dropout in MOOCs.

To support the above-mentioned main objective, the following research questions were formulated:

• How has MOOC abandonment research evolved over time?
• What are the characteristics of the articles that stood out in the field of MOOC dropout?
• Who are the top authors in MOOC dropout research?
• Which journals are the researchers’ favorite journals for publishing articles in this field?
• Which are the leading universities in publishing work on MOOC dropout?

Bibliometric analysis offers a different approach to contributing to the research, by analyzing, from a scientometric point of view, papers belonging to a specific field through using, in general, large amounts of data, relying on tools such as VOSviewer or R and developed databases such as Scopus or Google Scholar [31]. Due to its increasing popularity, bibliometric analysis has been used in this study due to the availability, accessibility and progress in scientific software, which has enabled us to handle large volumes of data; thus, this method of analysis has a significant impact on research due to the totality of the indicators it involves [32,33,34,35,36]. Moreover, it has been observed over time that it helps researchers to discover patterns of collaboration, the structure of a research field and emerging trends, using targets and massive data, facilitating the identification of different gaps in knowledge, the generation of new ideas and the positioning of contributions in a field [37,38,39,40]. In addition to the importance that this method of analysis gives to quantitative data, an increased interest in qualitative data has been observed over time, and one can observe the implications that bibliometric analysis has had and still has on research evaluation [41].

In the paper by Passas [42], the researcher made a comparison between bibliometric analysis, systematic reviews and meta-analysis. The author noted that although both bibliometric analysis and meta-analysis are quantitative analyses, they differ in focus. Meta-analysis can be described more as a tool that helps to extend theory by examining the relationships between variables, whereas bibliometric analysis is easier to approach by exploring both the intellectual and bibliometric structure of a field via the investigation of the relationships between institutions, authors or topics. Moreover, the researcher argues that the three techniques are complementary, each offering unique benefits: bibliometric analysis is suitable for providing general reviews, thus being used for large datasets, but not beneficial for research that includes small datasets, since it is able to quantify the scientific literature; meta-analysis is more of a summarization technique at the statistical level, being beneficial for studies with similar contexts and methodologies; systematic reviews are conducted on a specific topic in a structured way, being useful for specific reviews, not broad ones.

1.3. Manuscript Contribution

This paper aims to analyze the evolution of the MOOC dropout research, as stated above, by considering the elements related to a bibliometric analysis, and to further identify which are the authors’ areas of interest in terms of MOOC dropout, so that this can be explored further in future research. With the passage of time, as can be seen in

Table 1. Summary of previous bibliometric and review research papers on MOOC dropout.

Reference	Focus
Zong et al. [43]	Outdoor education, highlighting the thematic transition from environmental governance to environmental education.
Hallinger et al. [44]	Developing research on education for sustainable development in East Asia between 1991 and 2023.
Basheer et al. [45]	Engaging higher education institutions in achieving the SDGs.
Dönmez [46]	Sustainability in education, highlighting the increase in the number of publications and the shift in emphasis from environmental education to education for sustainable development.
Alghamdi et al. [47]	Limitations of traditional models for predicting dropout in MOOCs and exploration of the use of advanced artificial intelligence methods to improve prediction accuracy and support effective interventions in online education.
Alsuhaimi and Almatrafi [48]	Applying a transferable deep learning method for automatically classifying MOOC forum posts based on confusability indicators, thus improving student support and reducing dropout rates through early and personalized responses.
Luo and Li [49]	Examination of the impact of academic and emotional support on the sustainable use of MOOC platforms in English language learning.
Swacha and Muszyńska [50]	Use of demographic data to predict early student dropout in MOOCs, identifying factors such as age, educational level, student status, nationality and disability as predictors of dropout.

, the interest of researchers in the debated topics and the analyses carried out by other authors have led to the use of bibliometric analysis in the field of education.

Zong et al. [43] conducted a bibliometric analysis to investigate the development of outdoor education over time. The authors observed a gradual increase in interest by changing themes from environmental governance to environmental education. Hallinger et al. [44] observed the increasingly significant development of research (1991–2023) on education for sustainable development in East Asia.

Basheer et al. [45] analyzed how higher education institutions are involved in achieving the Sustainable Development Goals (SDGs). The authors started from existing research both on the integration of the SDGs into the educational curriculum and on how these efforts are evaluated and tracked. Drawing from a total of 83 articles, the findings focus on the increasing importance of sustainability and on the need to integrate the SDGs into all academic areas of higher education institutions.

With a similar theme, Dönmez [46] analyzed research on sustainability in education using bibliometric analysis, highlighting the increase in publications and the shift in focus from environmental education to sustainable education with the SDGs entering them mainstream. The study emphasizes the role played by education in sustainability and the need for further research for the more effective implementation of educational efforts, laying a foundation for future research.

In addition to these mentioned examples, several of the other authors’ studies and their focus are presented in Table 1.

Considering the papers listed in Table 1, it should be stated that the present paper distinguish which of the studies from the previous researchers mainly focus on education research in general, which focus on higher education in particular, and which focus on various issues related to education such as the evolution of education in a particular area, assessment practices, dropout prediction methods, learners’ confusion, the involvement of academic and emotional support, etc., by conducting a quantitative analysis of MOOC dropout research based on an overview of the publications, citations, authors, and institutions. Furthermore, this paper explores how research themes in the area of MOOC dropout have evolved and which are the themes that have attracted the interest of the research from the field by taking into account the most cited papers.

1.4. Paper Roadmap

The first part of this work involves the extraction of a database comprising selected articles, chosen using specific filters that will be detailed in the subsequent section. Part 2 follows with an explanation of the methodology and materials that form the basis of the research. The selection process, guided by these filters, was meticulously carried out to identify the most relevant articles in the field of MOOC dropout. Part 3, the most comprehensive section, presents the bibliometric analysis of the selected papers. This analysis covers articles published between 2013 and 2023 and includes a detailed examination of the authors, their journal preferences, existing collaborations, and their most globally cited articles. Additionally, a rigorous analysis of the most frequent word groups and factorial analysis were conducted, while two thematic maps and an evolution map were created. Part 4 presents the conclusions drawn from the bibliometric analysis of MOOC dropout and discusses the limitations of this study.

2. Materials and Methods

To identify the most relevant papers, key journals, and influential authors in the study of MOOC dropout, this paper employs bibliometric analysis. Bibliometric analysis, also known as scientometrics, utilizes mathematical and statistical methods to quantify scientific activity and assess its impact over a specified period [51,52]. A review of the specialized literature indicates that this method is frequently employed by researchers across various fields to conduct scientometric analyses [53,54,55,56,57,58].

In order to perform a bibliometric analysis, according to the literature [59,60], it is necessary to divide the analysis into four parts, which can be seen in Figure 1:

• Dataset extraction: Downloading the dataset from the selected platform, Clarivate Analytics’ Web of Science Core Collection, also commonly known as Web of Science (WoS) database, in our case;
• Analysis: Conducting analyses through the use of a specific software—Biblioshiny 4.2.3 library in R version 4.4.1 in our case;
• Discussions;
• Conclusions and limitations.

2.1. Dataset Extraction

In this subsection, the process of collecting and creating the database will be described. First, the platform from which the database is extracted is presented, followed by the specific indexes that were used, as well as the filters that were used in order to extract only those articles that we considered relevant for this analysis.

For the first part, involving the selection of the platform from which to extract the database, the choice of WoS is based on its extensive coverage of different disciplines [61,62,63,64,65] and its restrictive selection of journals, which gives it a solid reputation in the scientific community [66], as argued by Bakır et al. [67]. Although other databases such as Google Scholar, Scopus, PubMed, IEEE, and the Cochrane Library are available, the Web of Science (WoS) platform was selected for this analysis due to its compatibility with Biblioshiny, the bibliometric tool employed in this study. WoS allows seamless data import into Biblioshiny, which is crucial for the analysis process. While both VOSviewer and Biblioshiny are among the most widely used bibliometric analysis tools, each of them has limitations regarding data import capabilities. Biblioshiny supports raw file imports from Scopus, WoS, Lens, Dimensions, and Cochrane Library, while VOSviewer can import raw files from Dimensions, Lens, WoS, and PubMed [68,69].

It is important to emphasize that the extraction of data from the WoS dataset is conducted on a subscription basis. As noted by other researchers [70,71], the dataset obtained from the Web of Science (WoS) is closely tied to the specific subscription level utilized, making it essential to specify the indexes accessed. In our case, as detailed in

Table 2. Indexes used for WoS dataset.

Index Name	Period
Science Citation Index Expanded (SCIE)	1900–present
Social Sciences Citation Index (SSCI)	1975–present
Emerging Sources Citation Index (ESCI)	2005–present
Arts & Humanities Citation Index (A&HCI)	1975–present
Conference Proceedings Citation Index—Social Sciences and Humanities (CPCI-SSH)	1990–present
Conference Proceedings Citation Index—Science (CPCI-S)	1990–present
Book Citation Index—Science (BKCI-S)	2010–present
Book Citation Index—Social Sciences and Humanities (BKCI-SSH)	2010–present
Current Chemical Reactions (CCR-Expanded)	2010–present
Index Chemicus (IC)	2010–present

, we employed all ten available indexes within WoS for extracting the dataset. While the indexes are available for the mentioned periods, depending on the field of analysis, the period in which the paper belonging to a specific field has been written can be shorter.

To achieve the objectives of the proposed research, a dataset was extracted using specific filters to ensure alignment with the research goals. The process involved four distinct steps to generate the dataset, as highlighted in

Table 3. Dataset selection.

Exploration Steps	Filters on WoS	Description	Query	Query Number	Count
1	Title/Abstract/Keywords	Contains specific keywords related to MOOCs	((TI=(MOOC*)) OR AB=(MOOC*)) OR AK=(MOOC*)	#1	8443
		Contains specific keywords related to dropout	((TI=(dropout)) OR AB=(dropout)) OR AK=(dropout)	#2	32,480
		Contains specific keywords related to MOOCs and dropout	#1 AND #2	#3	455
2	Language	Limited to English	(#13) AND LA=(English)	#4	432
3	Document Type	Limited to Articles	(#14) AND DT=(Article)	#5	212
4	Year published	Excludes 2024	(#15) NOT PY=(2024)	#6	193

.

First, filters were applied to retain only those articles that included the most significant terms in their titles, abstracts, or keywords. The initial filter selected papers containing the keyword “MOOC*”, identifying a number of papers equal to 8443. It should be stated that the “*” has been in order to facilitate the search of both singular (“MOOC”) and plural (“MOOCs”) forms of “MOOC” keyword. The second filter was applied to select research containing the keyword “dropout”, which resulted in a significantly larger dataset of 32,480 papers compared to the previous filter. Since our objective was to focus specifically on the dropout rates of individuals enrolled in online courses, the third filter was used to refine the dataset further by retaining only those papers that contained both the terms “MOOC*” and “dropout”. This refinement reduced the dataset to a total of 455 papers meeting both criteria. Comparing the selection criteria, it was observed that Billsberry and Alony [72] used the abbreviations “MOOC” and “MOOCs” to form the database used in their bibliometric analysis, which was mainly aimed at mapping the courses and identifying the most significant themes. Querying the Scopus database, the researchers came up with a total number of 3114 published journal articles, 597 book chapters, 145 reviews and 53 editorials from the period 2009–2022. Furthermore, in order to ensure that the database is an appropriate one for the purpose of analyzing dropout in distance education, much like in this analysis, Kurulgan [73] used the keyword “dropout” and its derivatives to filter the database extracted from WoS.

The second stage involved applying a filter to include only articles written in English, reducing the dataset from 455 to 432 articles. English was selected because it is an internationally recognized language and the most commonly used language for academic publications, as noted by Raman et al. [74]. Researching other bibliometric analyses carried out, it was found that this criterion has also been applied to observe the following: the increase in interest in MOOCs by researchers over time [75], the awareness of students regarding their education on the mismanagement of toxic waste [76] and the understanding of both the integration and use of virtual reality in the educational process [77].

The third step involved filtering the dataset to include only documents classified as “articles”, resulting in a total of 212 papers. According to Donner [78], categorizing a paper as an “article” denotes a report of original research, regardless of length, and may include meta-analyses. This classification technique aligns with the indexing practices used by WoS and Scopus datasets. Donner further argues that, in bibliometric analyses, it is crucial to limit the dataset to articles presenting primary research findings. Additionally, distinguishing between different document types is essential in scientometrics due to their varied content and purposes, which can lead to different citation patterns. Other researchers have chosen to include this filter related to articles among the papers selected in the bibliometric analysis, in order to identify research literature on initial teacher training in relation to inclusive education [79] or to observe how the field of gamification has evolved over a certain period of time from an educational point of view [80].

In the final stage of generating the dataset, a filter was applied to limit the publication years by excluding 2024. This exclusion was made because 2024 is not a full year and there is a possibility that papers submitted to WoS may not have been indexed by the time the dataset was extracted. Moreover, if articles published in 2024 would have been included in the analysis performed, different indicators such as those related to citations would have been biased as these papers would not have had enough time to be cited. Thus, indicators such as the number of citations per article would have been affected and would not have accurately reflected their real impact, given that they would have been at the beginning of the citation process and would not yet have had significant traction in the literature. A similar approach was used by Basheer et al. [45], where the dataset was limited to publications from 2018 to 2023. After applying this final filter, the dataset was refined to a total of 193 articles.

2.2. Analysis

This sub-chapter represents the most complex part of this study, describing the totality of the analyses that were applied to the dataset. For this part, the program RStudio version 4.4.1 was used, particularly the package Bibliometrix version 4.2.3. together with the command biblioshiny(). R serves as both a language and an environment for statistical and graphical computation, providing a diverse array of graphical and statistical techniques. It is highly extensible and supports functional and object-oriented programming, making it a powerful tool for conducting complex analyses such as bibliometric studies [69].

In Figure 1 the central part of the analysis is divided into four distinct sections, with the final section further subdivided into four subcategories. The first section, “Dataset Overview”, provides details on the key indicators of the selected dataset, setting the context for the research domain. These indicators include the time period during which the papers were published, the total number of papers and their sources, the average number of citations per paper, the annual scientific output, the number of references, the annual evolution of average article citations per year, the total number of authors, and whether authors preferred to publish individually or in collaboration.

The second section of the analysis focused on evaluating the sources to identify the most relevant journals over the analytical period. This was performed by examining the total number of documents published in the field, applying Bradford’s Law [81], and assessing the journals’ impact through their H-index. Discovered in 1934, Bradford’s law is one of the three major bibliometric laws, alongside Zipf’s law and Lotka’s law. Bradford’s law posits that scientific papers on a given topic are distributed according to a mathematical function. As the number of articles on a topic increases, the number of journals publishing those articles also increases. According to this law, if we cluster journals producing roughly the same number of articles, the number of clusters is proportional to $1:\mathrm{n}:{\mathrm{n}}^2$, where n represents a multiplier [81,82]. The H-index is a measure that captures scientific output by considering both the total number of publications and the number of citations. It assesses productivity and provides an overview of an individual’s research performance, but it is not ideal for direct comparison between researchers [83].

The third section provides a more detailed analysis of the authors, identifying those with the most significant scientific contributions to MOOC dropout research. This analysis was further broken down by year, examining the authors’ affiliations, the countries with which they collaborated, and the creation of a collaborative network to visualize how authors grouped together. Additionally, this section highlighted the countries that recorded a high number of citations.

The fourth section, titled “Analysis of Literature”, is subdivided into four subcategories: the first two subcategories (“Most cited papers—overview” and “Most cited papers—review”) focus on the most cited works, providing readers with insights into the research approaches and objectives of authors who have studied MOOC dropout. Each paper is accompanied by a comprehensive summary, including details such as the number of authors, the journal in which studies were published, the total number of citations, the total number of citations per year, and the normalized citation counts. The third subcategory, “Word Analysis”, concentrates on identifying the most frequently used keywords, including author keywords, “keywords plus”, and common word groups (n-grams) found in the abstracts and titles of the selected papers. The final subcategory features two three-field plots, which illustrate the connections between countries, authors, and journals, as well as between affiliations, authors, and keywords. The N-grams are actually the bigrams and trigrams found, which represent groups of words (two or three words as the case may be) that had a significant number of occurrences in the identified works. In addition, the last sub-category includes a thematic map of the authors’ keywords, used to describe the current state of the research field and to highlight possible future research directions. A thematic evolution map is provided to show how research techniques have evolved up to 2020 and beyond. Finally, a factorial analysis is conducted to explore and understand the structure of relationships between the variables in the dataset.

2.3. Discussion

The discussion part of this article will analyze the significance of the results obtained from the bibliometric analysis of the evolution of MOOCs and the interest of researchers in this topic. It will discuss general observations related to the significant growth of publications in the field of MOOCs and discuss the relevance of publications in major journals, as well as significant contributions by renowned authors or regional contributions. Moreover, the discussion will expand on the main research directions identified and the identified prediction techniques will be compared; the main causes of dropout, personalized interventions to prevent dropout, and the impact of cultural and social network factors on dropout rates will also be discussed.

2.4. Conclusions and Limitations

In the conclusions section, we will summarize the main results of the bibliometric analysis of MOOC dropout based on articles indexed in ISI Web of Science. We will also summarize the interest in this field, which were the most cited articles and their countries of origin, which were the more frequently used analysis techniques, which were the most active institutions in this field, and which are the limitations of this analysis.

3. Dataset Analysis

This section focuses on analyzing the papers selected through specific filters, with particular attention given to the number of authors, the existing collaborations between them, and the insights they were able to uncover through their analyses.

3.1. Dataset Overview

Table 4. Main information about data and document contents.

Indicator	Value
Timespan	2013:2023
Sources	101
Documents	193
Average citations per documents	19.68
References	6560
Keywords plus	255
Author’s keywords	573

reveals that our dataset comprises 193 articles, published across 101 different journals between 2013 and 2023. According to Dhawal Shah [84], 2013 was a pivotal year for MOOCs, marked by a surge in student enrollments. While there were approximately 100 MOOCs in 2012, this number skyrocketed to nearly 700 in 2013. The selected articles collectively contain 6560 references, a substantial number reflecting the complexity of the education field, which necessitates a wide range of sources for comprehensive analysis across various studies.

In addition to the key statistics discussed, the final two indicators in Table 4 provide insights into the content of the documents. The dataset contains 573 “author’s keywords”, resulting in an average of approximately 3 (2.97) keywords per document. The “keywords plus” indicator, totaling 255, represents keywords extracted from the titles of articles that were cited, with an average of approximately 1 (1.32) keywords plus per document.

Regarding the annual scientific output in the area of MOOCs dropout, determined based on the extracted dataset, Figure 2 shows that the highest number of articles in the analyzed field—34—were published in 2022. Additionally, there is a noticeable increase in the number of articles from 2016, with only 4 articles, to 2019, when 29 studies were published. Throughout the analyzed period, interest in this field has followed an upward trend, with an annual growth rate of 40.04%. However, a decrease is observed between 2022 and 2023, with five fewer articles published in the latter year.

According to Shah [84], by the end of 2019, more than 900 universities worldwide had either launched or announced a total of 13,500 MOOCs. It can also be seen that, in 2019 alone, approximately 2500 courses were launched by 450 universities. Furthermore, it was concluded that the increased interest in MOOCs in 2020 was a side effect of the COVID-19 pandemic [85].

Figure 3 shows the average citations per year for the selected articles in the dataset with regard to MOOC dropout. Until 2017, this indicator steadily increased, starting at 0.33 in 2013 and reaching 6.46 in 2017. During the same period (2013–2017), the “average citations per article” indicator also showed rising values, starting at 4 and reaching 51.7 by 2017. This suggests a growing interest in scientific research in this field during that time.

While the early years showed a steady increase in the average number of citations per year, from 2017 to 2023, this indicator fluctuated, peaking in 2020 with a value of 4.82. These fluctuations may be attributed to heightened curiosity during the initial period, when the field was new and internet resources were less developed, leading to greater reliance on forums and early research.

At the author level,

Table 5. Authors.

Indicator	Value of the Indicator
Authors	566 authors
Authors of single-authored documents	20 authors
Authors of multi-authored documents	546 authors

indicates that there were a total of 566 authors—determined through the analysis of the extracted dataset. Although this number may seem relatively low, it aligns with the limited number of papers in the dataset. Of these authors, only 20 were published as sole authors, while the remaining 546 collaborated on various research projects. Given the dataset includes 193 papers, this suggests a strong tendency toward teamwork in MOOC abandonment research. The complexity of the field, requiring in-depth analysis and expertise from various disciplines, makes collaboration more advantageous. On average, each document involved approximately three (2.83) authors.

Table 6. Author collaboration.

Indicator	Value of the Indicator
Single-authored documents	21 documents
Documents per author	0.34 documents/author
Authors per document	2.93 authors/document
Co-authors per documents	3.45 co-authors/document

provides a clearer view of author collaboration, and the information in it was obtained using the data in the extracted dataset. By dividing the “authors of single-authored documents” by “single-authored documents”, it reveals that, on average, single authors have nearly one paper each (0.95). Additionally, there are, on average, 0.34 papers per author and 2.93 authors per paper. The average number of co-authors per document is 3.45. These data again highlight that authors show a strong interest in collaboration, likely driven by the need to comprehensively analyze the factors influencing MOOC dropout.

3.2. Sources

So far, we have discussed the authors, but it is also crucial to note their publication preferences. Figure 4 displays the top 15 preferred journals for publishing research. Leading the list is the journal “IEEE Access”, where 14 papers on MOOC abandonment were published. This journal is interdisciplinary and application-oriented, operating exclusively online, which aligns with the high volume of papers published in this field.

The next journal, “Education and Information Technologies”, has published 10 articles, combining education with information technology to provide insights across various levels. Following this is “Computers & Education” with nine articles and “International Review of Research in Open and Distributed Learning” with eight articles, both of which focus on how technology supports the open learning process in online environments. Additionally, five journals each have four publications in the top list: “Computers & Electrical Engineering”, “Distance Education”, “Education Sciences”, “International Journal of Engineering Pedagogy” and “ Sustainability”.

To examine the most cited publications, Figure 5 applies Bradford’s law. Essentially, Bradford’s law identifies a core group of journals with the highest concentration of articles on a specific topic, surrounded by additional clusters with varying publication frequencies. These peripheral clusters, being further from the central topic, contain fewer articles in the analyzed domain [86,87].

As anticipated, Figure 5 shows that the core or Zone 1 contains the top nine journals with the highest number of publications on MOOC dropout. These journals are “IEEE Access”, “Education and Information Technologies”, “Computers & Education”, “International Review of Research in Open and Distributed Learning”, “Turkish Online Journal of Distance Education”, “Computer Applications in Engineering Education”, “International Journal of Emerging Technologies in Learning”, “Computers & Electrical Engineering”, and “Distance Education”.

Figure 6 shows that the top two ranked journals, “Computers & Education” and “IEEE Access”, both have an H-index of 9. Most of the journals depicted in the figure also appear in both the core field identified by Bradford’s law (Figure 5) and the list of most relevant journals (Figure 4).

To provide a clearer picture of the most frequently cited publications by authors studying MOOC dropout, Figure 7 illustrates the scientific output over the years for the top 5 journals identified as favorites (Figure 4).

In 2013, none of these journals published articles in the analyzed field. It was not until 2014 that “International Review of Research in Open and Distributed Learning” published two articles. As anticipated, despite its first publication appearing in 2018, “IEEE Access” has the highest number of publications throughout the analyzed period.

3.3. Authors

Figure 8 identifies the top 15 authors based on the number of articles published on MOOC abandonment. Among them, four authors (El Kabtane H, Mourdi Y, Sadgal M, and Xing WL) have each published five papers, while two authors (Feng J and Sun X) have published four papers each, and the remaining nine authors have each published three papers. With a total of 193 articles in the dataset, the combined scientific output of the top four authors accounts for 10.36% of the total number of selected articles.

Figure 9 illustrates the scientific output of the top authors from 2016 to 2023. The graph shows that 2019 was a particularly fruitful year for dropout research, with the top five authors (El Kabtane H, Mourdi Y, Sadgal M, Feng J, and Sun X) each publishing two papers, while Xing WL published three papers. This increase aligns with the observed rise in annual scientific output seen in Figure 2, which began to accelerate around 2016 and peaked in 2019.

Table 7. Top 15 most relevant affiliations.

Affiliations	Articles	Percentage
Beijing Normal University	11	5.70%
Cadi Ayyad University of Marrakech	6	3.11%
Harvard University	5	2.59%
IBN Tofail University of Kenitra	5	2.59%
Texas Tech University	5	2.59%
Mohammed V University in Rabatg	4	2.07%
Cadi Ayyad University of Marrakech	4	2.07%
Guilin University of Electronic Technology	4	2.07%
Central China Normal University	4	2.07%
Northwest University Xi’an	4	2.07%
Texas Tech University System	4	2.07%
Universidad Carlos III de Madrid	3	1.55%
Abdelmalek Essaadi University of Tetouan	3	1.55%
Chulalongkorn University	3	1.55%
Nanjing Agricultural University	3	1.55%

presents the top affiliations of authors based on the number of articles published. “Beijing Normal University” from Beijing and “Cadi Ayyad University of Marrakech” from Morocco occupy the first two positions, with 11 and 6 published articles, respectively, followed by other prestigious institutions. Although only 3 universities appear in the top 15 affiliations—“Abdelmalek Essaadi University of Tetouan”, “Chulalongkorn University”, and “Nanjing Agricultural University”—a total of 14 universities have each published three articles. These include “National Taiwan Normal University”, “Smithsonian Astrophysical Observatory”, “Smithsonian Institution”, “University of California, Berkeley”, “Yeungnam University”, “Amity University Uttar Pradesh”, “Beijing Information Science and Technology University”, “Egyptian Knowledge Bank (EKB)”, “Open University Netherlands”, “Tecnologico de Monterrey”, and “University of Tartu”. Overall, 37.82% of the total number of papers included in the dataset belong to the top 15 universities listed in Table 7.

Figure 10 illustrates the country of origin for each author, along with the SCP (Single-Country Publication) and MCP (Multiple-Country Publication) indexes. According to the graph, China leads with the highest number of published articles, totaling 55, which represents 28.5% of the dataset. Of these, 42 articles are indicated by the SCP index, while 13 are noted by the MCP index, reflecting a tendency for more internal collaboration within China. Following China, the USA has 19 published articles, all classified under SCPs, with an MCP value of 0, indicating that these articles were published within the country only. The same pattern is observed for Italy, Korea, Egypt, and Estonia.

Figure 11 depicts the scientific output by country, with color intensity increasing from light blue to dark blue to represent higher contributions, and with gray areas representing no contribution.

As anticipated from previous analysis, China leads with the highest number of papers, totaling 113. Tan and Tasir [88] observed a substantial increase in MOOCs in China over the past two decades, attributed to the large population and a strong emphasis on education. Ma and Mendez [89] noted that China launched its first comprehensive MOOC platform in 2013, with many others following. The COVID-19 pandemic further boosted MOOC platforms in China, resulting in increased courses and learners. In 2022, China’s Ministry of Education initiated “China’s Smart Education”, consolidating all MOOC platforms under a single website.

Following China, the USA has a notable output with 40 papers, while India and Morocco each have 30. Spain has 28 papers, the UK has 21, and Italy and South Korea each have 10. Australia, Portugal, and Turkey each have eight papers.

Figure 12 illustrates the number of citations by country. China leads with 1155 total citations (TCs), followed by the US with 798, the UK with 298, and Spain with 292. The remaining countries in the top 15 have a TC value below 200.

The Netherlands tops the list with an average of 64 citations per article, followed by the UK with 42.60, the US with 42, Sweden with 36.7, Australia with 33.7, and Mexico with 31.7. Other countries in the top 15 have average citations per article below 30.

Figure 13 displays a map of international collaborations between countries. It should be noted that the countries with contributions in this area, measured by the number of publications, are marked using various shades of blue, with dark blue representing the highest number of contributions, while the countries with no contributions are marked in grey. Furthermore, the lines in red depict the collaborations among countries, while the size of the line characterize the number of collaborations. China leads with the highest number of collaborations, engaging with 12 countries: Ecuador, Australia, Bangladesh, Indonesia, Korea, Saudi Arabia, Singapore, Spain, Tanzania, the United Kingdom, the USA, and Yemen. The most frequent collaborations were with Australia and the USA (five times each) and with Spain (three times). The UK collaborated with seven countries: Azerbaijan, Hungary, Ireland, Malaysia, the Netherlands, Saudi Arabia, and Sweden, with the highest frequency being two collaborations, which were with Malaysia. The USA engaged with five countries: Australia, Canada, Indonesia, Mexico, and Tanzania, with each collaboration occurring once.

To be able to even better observe the existing collaborations between authors, Figure 14 is attached, representing the collaboration network of the top 40 researchers (being divided from Cluster #1 to Cluster #11). Even though the top 40 were chosen to be displayed, there are authors whose research has no correspondences with other researchers, as well as a setting that does not illustrate isolated nodes. This is why only 28 authors are shown in the figure below.

The obtained clusters consist of two to five authors each. Each cluster includes the following:

• Cluster #1 (in red): This consists of Alario-Hoyos C and Kloos CD. These two researchers have focused on analyzing how self-regulated learning strategies (SRLs), particularly event-driven and self-reported SRLs, can be integrated into predictive models for self-paced MOOCs [90]. The authors have also investigated the relationships between SRLs and the information obtained from MOOC learners [91].
• Cluster #2 (in blue): This consists of Lepp M and Luik P. They investigated performance metrics recorded before dropout and identified periods with the highest dropout rates in MOOCs dedicated to computer programming [92]. Using non-parametric tests and descriptive statistics, they analyzed performance in assessments of those who did not complete the courses, those who completed, and those who managed to complete based on involvement or difficulty [93].
• Cluster #3 (in green): This contains Distante D and Faralli S. This cluster of authors focused on both literature reviews and research related to predicting student dropout using machine learning algorithms [94,95].
• Cluster #4 (in purple): This includes Burgos D and Chen L. These researchers described the emergence and development of MOOC platforms over time [96], and analyzed the types and amount of support participants receive during courses, based on 14 MOOC platforms and 621 courses in China [97].
• Cluster #5 (in yellow): This cluster consists of Azzouzi S and Charaf ME. They used optimization algorithms such as ant colony optimization, recurrent neural networks, and classification algorithms to create predictive models aimed at optimizing courses to prevent dropout [98,99].
• Cluster #6 (in brown): This cluster includes Bachelet R and Chaker R. They employed structural equation modeling and path analysis to investigate causal links between theoretical self-experience, MOOC learning outcomes, and social intentions [100].
• Cluster #7 (in pink): This consists of Asensio-Pérez JI and Bote-Lorenzo ML. They examined learners’ motivation to participate in MOOCs, focusing on the effects of redeemable tokens or rewards on dropout behavior [101,102].
• Cluster #8 (in gray): This contains Chen C and Fredericks C. This cluster of authors analyzed how course completion is impacted by students’ misconceptions about the course material [103] and dropout rates related to chapter transitions, using disjoint survival analysis [16].
• Cluster #9 (in turquoise): This includes Gao ZH, Zheng YF, and Fu Q. They addressed topics such as MOOC dropout prediction using convolutional neural networks, bidirectional short-term memory networks, and deep fusion models [104,105,106];
• Cluster #10 (in orange): This is the largest cluster, consisting of Feng J, Sun X, Liu Y, Chen J, and Gao Y. The authors focused on developing a new algorithm combining extreme learning machines and decision trees for more accurate dropout predictions [107], creating a hybrid neural network for selecting posts needing immediate teacher attention [108] and developing a parallel neural network for grouping MOOC forum sentiments [109].
• Cluster #11 (in navy blue): This is the second largest cluster, made up of El Kabtane H, Mourdi Y, Sadgal M, and El Adnani M. The research conducted by the authors included predicting learner behavior leading to dropout [27,110], creating individual learner behavior profiles throughout courses [25] and incorporating online manipulations into MOOCs [28].

3.4. Analysis of the Literature

This section first presents the top 10 most cited articles, detailing the number of authors, the data used, the objectives, the total number of citations, and the methods employed. Additionally, a summary of each article is provided to illustrate how MOOC dropout rates were analyzed. In the latter part of the section, a comprehensive word analysis is conducted.

3.4.1. Top 10 Most Cited Papers—Overview

The most globally cited paper, authored by Xing et al. [111], as shown in

Table 8. Top 10 most global cited documents.

No.	Paper (First Author, Year, Journal, Reference)	Number of Authors	Region	Total Citations (TC)	TC per Year (TCY)	Normalized TC (NTC)
1	Xing WL, 2016, Computers in Human Behavior [111]	4	USA	157	17.44	2.93
2	Dai HM, 2020, Computers & Education [113]	4	China	147	29.40	6.10
3	Xing WL, 2019, Journal of Educational Computing Research [114]	2	USA	133	22.17	5.53
4	Tsai YH, 2018, Computers & Education [115]	4	Taiwan	124	17.71	4.27
5	Henderikx MA, 2017, Distance Education [116]	3	Netherlands	97	12.13	1.88
6	Eriksson T, 2017, Journal of Computing in Higher Education [117]	3	Sweden	97	12.13	1.88
7	Aldowah H, 2020, Journal of Computing in Higher Education [24]	4	United Kingdom	87	17.40	3.61
8	Almatrafi O, 2018, Computers & Education [19]	3	USA	87	12.43	2.99
9	Sunar AS, 2017, IEEE Transactions on Learning Technologies [118]	4	United Kingdom, Asia	85	10.63	1.64
10	Moreno-Marcos PM, 2020, Computers & Education [90]	6	Spain, Chile	78	15.60	3.24

, was published in Computers in Human Behavior. The paper has four authors and has garnered 157 total citations (TCs), with an annual average of 17.44 citations per year (TCY), and a normalized total citation (NTC) score of 2.93. The NTC is calculated by dividing the total number of citations (TCs) by the average number of citations per article published in 2016, which was 53.58. The NTC indicator reflects the relative impact of a paper in comparison to others published in the same year [112]. Based on this metric, the work of Xing et al. [111] was cited 2.93 times more frequently than the average for papers in the dataset.

The second most globally cited paper, authored by Dai et al. [113], and published in Computers & Education, has accumulated 147 total citations (TCs). Notably, it surpasses the first-ranked paper in both annual citation rate, with an average of 29.40 citations per year (TCY), and normalized total citations (NTCs), achieving a score of 6.10, calculated by dividing the 147 total citations by the 24.11 average citations per article in 2020. In third place is a paper published in the Journal of Educational Computing Research [114], with 133 total citations, while the fourth-ranked paper, published in Computers & Education [115], has registered 124 total citations.

Compared to the top four papers, the remaining articles among the most globally cited have fewer than 100 total citations (TCs). Nevertheless, all of these papers have accumulated more than 78 citations, with an annual average of over 10 citations per year (TCY) and a normalized total citation score (NTC) exceeding 1.5. This suggests that these works have provided a foundational basis for significant subsequent analyses in the field of dropout research.

3.4.2. Top 10 Most Cited Papers—Review

This section presents the abstracts of the top 10 most globally cited papers to examine the authors’ approaches to MOOC dropout, the techniques employed, and the results obtained.

The paper by Xing et al. [111] analyzes high dropout in MOOCs and proposes a predictive model for the early identification of students at risk of dropping out. It focuses on learners who are active on forums, considering that they are more likely to complete the course if they receive adequate support. This study uses temporal modeling, combining historical data with current week data, considering that a student’s inactivity actually represents their dropout by taking social interactions into account as well, and applies principal component analysis (PCA) to determine the optimal time to exclude prior information. It also uses a stacking technique to improve the robustness and accuracy of the predictions. Through the results obtained, it was deduced that this method is more efficient than traditional models and provides instructors with a useful technique for targeted interventions, thus contributing to an increase in the retention rate in MOOCs.

In contrast to other studies, Dai et al. [113] focused on identifying the factors that influence learners to persist with MOOCs. Their paper aims to explore the roles of learning, technology, and teaching in this complex process. As an analysis technique, they modified the Expectancy Confirmation Model, incorporating the affective and cognitive variables of ‘attitude’ and ‘curiosity’ to capture both past reflections and future expectations. Two-step structural equation modeling was employed to analyze the interconnectedness and multiple dependency relationships, while confirmatory factor analysis was used to assess validity and reliability. Based on data from 192 Chinese students who completed a survey, the model was found to explain 48% of participants’ intentions to continue with MOOCs, with both newly introduced variables proving significant. Based on the literature review, the researchers start from several hypotheses in which they use the following as influencing variables: confirmation, satisfaction, attitude, curiosity and perceived usefulness. The study shows that although curiosity influences learners’ intention to continue, attitude plays a more important role. To improve retention, interventions should focus on shaping learners’ attitudes. Additionally, instructors should avoid overemphasizing the benefits of MOOCs to prevent negative impacts on learner satisfaction and attitudes.

The study by Xing and Du [114] focuses on a MOOC course hosted on the Canvas platform, which began in August 2014 and lasted eight weeks. The course comprised 11 modules, 14 discussion forums, and 12 multiple-choice quizzes, with a total enrollment of 3617 students. The dataset used in this research was sourced from two main areas: clickstream data from Canvas, which tracked student activity on the platform, and data on discussion forums and quiz scores, extracted via the Canvas API. Utilizing deep learning techniques, along with K-nearest neighbors, support vector machines, and decision trees, the study aimed to optimize a model for predicting MOOC dropout, with a focus on personalizing interventions for at-risk students. The model developed in the study estimates the dropout probability for each individual student. Based on the number of views of announcements, calendar, gradebook, homework, module pages, quizzes, etc., by students not enrolled in the MOOC, the researchers propose as a strategy the use of the prediction model developed to provide teachers with a technique to intervene in time to stop dropout.

The study by Tsai et al. [115] examines the challenges of MOOCs in online teacher training, specifically focusing on the high dropout rates and the low intention of learners to continue their studies. The researchers proposed a model that integrates metacognition and learning interest, drawing on data from 126 respondents. To analyze the data, confirmatory factor analysis and structural equation modeling were employed to test the relationships between variables, starting from different claims that metacognition is influenced by liking, enjoyment and engagement. The results indicated that metacognition positively correlates with the three levels of interest which in turn significantly influence learners’ intention to continue using MOOCs. The study suggests that improving metacognitive skills may foster greater engagement and persistence in online learning, thus supporting the development of effective teacher training programs through MOOCs.

The paper by Henderikx et al. [116] proposes an alternative typology for evaluating MOOC success and dropout, incorporating participants’ intentions and subsequent behaviors. Data were gathered through two questionnaires, administered before and after the course, with participants from two MOOCs. The first questionnaire initially had 689 respondents, later reduced to 163, while the second started with 821 respondents, narrowing down to 126. This exploratory study revealed that although the success rates based on course completion were only 6.5% and 5.6%, the success rates from the participants’ perspectives were considerably higher, reaching 59% and 70%. These findings suggest that course completion alone is an insufficient measure of success and highlight the importance of considering participants’ perspectives in evaluating MOOC outcomes. In the paper, the researchers analyze the following as factors related to dropout, the difference between intention and behavior, the definition of success, the level of commitment, the typologies of participants. Thus, they came to the conclusion that for the fight against dropout, it is very important that the perspective of institutional goals be complemented by that of personal goals.

Drawing on data from semi-structured interviews with 34 learners who participated in two MOOCs (18 from ChM001x and 16 from ChM002x), Eriksson et al. [117] employed a qualitative case study approach to analyze learners’ experiences and identify the factors influencing their decision to drop out or continue. The study identified four hypotheses that were found to impact dropout decisions: (1) the learner’s perception of course content, (2) the learner’s perception of course design, (3) the learner’s social situation and characteristics, and (4) the learner’s ability to manage time effectively. Under the first hypothesis, factors such as motivation and course difficulty were highlighted. The second hypothesis revealed discouragement stemming from poorly designed tasks. The third hypothesis identified external and socio-economic factors, as well as personality traits, while the fourth hypothesis focused on time management challenges and study techniques. Moreover, through in-depth interviews, high workload, lack of time, lack of awareness or pressure, social influence and course content were also identified as factors leading to dropout. As strategies, the authors proposed identifying the best times for MOOCs or their duration, and to adapt them according to the learners (in terms of their language knowledge, expectations, etc.).

The study by Aldowah et al. [24] employed a multiple-criterion decision-making method to identify and analyze factors contributing to the high dropout rates in MOOCs. Through a literature review, twelve primary factors, categorized into four dimensions, were identified. These factors were subsequently evaluated by 17 experienced MOOC instructors. The analysis revealed six core factors that directly influence dropout rates: prior experience, social presence, social support, academic skills, feedback, and course design. Additionally, factors such as engagement, interaction, motivation, time management, course difficulty, and personal circumstances were found to have a secondary impact. The study also described the causal relationships among these factors, providing valuable insights for developing interventions aimed at reducing dropout rates in MOOCs. As strategies, the paper offers an important analysis of the identified factors, deducing which are those with a significant influence, thus being able to analyze the respective factors and subsequently to combat dropout. Moreover, the relationships between the primary and secondary factors have a reciprocal effect and are also interdependent.

Almatrafi et al. [19] have identified the challenge of retaining MOOC learners until course completion, particularly due to the high volume of messages on discussion forums that often go unanswered. They considered as an abandonment factor situations in which urgent messages, crucial for learners’ progress, do not receive timely attention, thus having a negative impact on collaborative learning. The study proposes a model designed to identify these urgent messages that require immediate support from coordinators. By using different data mining techniques and feature sets, the research developed a reliable classification model capable of recognizing urgent messages in different MOOCs. The work helps reduce dropout by using the created model, which helps instructors prioritize responses, thus improving learning support.

Sunar et al. [118] investigated the social behaviors of MOOC participants and the impact of their engagement on course completion. The researchers identified the social interactions that exist between learners as a factor increasing course completion rates. An analysis of discussions from a FutureLearn MOOC, which ran for eight weeks and included 9855 learners, showed that following and interacting with other learners significantly increased the likelihood of course completion. These interactions have been identified as key predictors of success in MOOCs. The present work is of particular importance, providing one of the key factors that can reduce MOOC dropout.

Moreno-Marcos et al. [90] observed that many existing predictive models for identifying learners at risk of dropout in MOOCs fail to incorporate complex high-level variables, such as self-regulated learning (SRL) strategies. This paper addresses the limitations of current predictive models, particularly for self-paced MOOCs, where learners can start the course at any time. The variability in start dates and learner engagement makes early predictions challenging. The study investigates how SRLs can be integrated into predictive models by comparing event-based SRLs with self-reported SRLs to assess their impact on dropout prediction. The paper introduces a novel methodology for analyzing self-paced MOOCs, employing a temporal approach to enhance early prediction capabilities. The results indicate that event-driven SRL techniques possess strong predictive power and can effectively forecast dropout even when other data are sparse. The proposed methodology enables reliable predictions from as early as the first 25–33% of the course duration, offering a new approach to identifying at-risk learners in self-paced MOOCs. Researchers bring a new perspective that the best predictors to be analyzed as a strategy to reduce dropout are interactions between participants and exercises within the platform, as well as openness, skimming and completion of the content presented. By constantly analyzing and intervening on low-activity learners, dropout can also be reduced.

A brief summary of the content of the papers discussed above is provided in

Table 9. Brief summary of the content of top 10 most globally cited documents.

No.	Paper (First Author, Year, Journal, Reference)	Title	Methods Used	Data	Purpose
1	Xing WL, 2016, Computers in Human Behavior [111]	Temporal predication of dropouts in MOOCs: Reaching the low hanging fruit through stacking generalization	Machine learning methods, principal components analysis	14 discussion forums and 12 multiple-choice quizzes, gathered from a course that had 3617 students enrolled	To create a mechanism to identify students at risk of dropping out as accurately as possible
2	Dai HM, 2020, Computers & Education [113]	Explaining Chinese university students’ continuance learning intention in the MOOC setting: A modified expectation confirmation model perspective	Expectation confirmation model, structural equation modeling, confirmatory factor analysis	192 Chinese students were recruited as participants to complete a questionnaire	To identify and explore the factors that influence students to continue MOOC studies
3	Xing WL, 2019, Journal of Educational Computing Research [114]	Dropout Prediction in MOOCs: Using Deep Learning for Personalized Intervention	Techniques of deep learning, K-nearest neighbors, support vector machines, anddecision tree, baseline algorithms	3617 students participants in a course organized by Canvas. JSON data for test scores or discussion form data, as well as trace or click-stream data	To optimize a MOOC dropout prediction model customized to intervention
4	Tsai YH, 2018, Computers & Education [115]	The effects of metacognition on online learning interest and continuance to learn with MOOCs	First order confirmatory factor analysis, structural equation modeling	Data were collected from a total of 126 respondents	To create a unified model that combines both learning interest and metacognition to investigate MOOCs’ continuance intention
5	Henderikx MA, 2017, Distance Education [116]	Refining success and dropout in massive open online courses based on the intention–behavior gap	The traditional approach tracking course success rates	Data were collected using two questionnaires (before and after the course) completed by the participants of two MOOCs. The first questionnaire had a total of 689 respondents, subsequently completed by 163 respondents, and the second questionnaireinitially had 821 respondents and subsequently had 126 respondents	The aim was to test the applicability of the typology by conducting an exploratory study
6	Eriksson T, 2017, Journal of Computing in Higher Education [117]	“Time is the bottleneck”: a qualitative study exploring why learners drop out of MOOCs	Qualitative case study approach	Application of semi-structured interviews, on a total of 34 learners who recorded different degrees of course completion for two MOOCs	To identify the reasons that influence participants to both complete and abandon the MOOC
7	Aldowah H, 2020, Journal of Computing in Higher Education [24]	Factors affecting student dropout in MOOCs: a cause and effect decision-making model	Multi-criteria decision-making	Identification of 12 factors from the literature	To find the underlying factors and possible causal relationships that are responsible for the rather high dropout rate in MOOCs
8	Almatrafi O, 2018, Computers & Education [19]	Needle in a haystack: Identifying learner posts that require urgent response in MOOC discussion forums	Linguistic inquiry and word count, metadata, term frequency, classification methods and sampling groups	The Stanford MOOCPosts dataset (with a large number of posts—29,604. 29,584, after excluding posts with insignificant information	To create a model that is able to identify posts of an urgent nature that need the immediate attention of the coordinator
9	Sunar AS, 2017, IEEE Transactions on Learning Technologies [118]	How Learners’ Interactions Sustain Engagement: A MOOC Case Study	Social network analysis techniques, prediction model	Discussions in a FutureLearn MOOC, which had a total of 9855 registered learners	To investigatet the social behaviors learners exhibit in MOOCs and what the impact of engagement is in terms of course completion
10	Moreno-Marcos PM, 2020, Computers & Education [90]	Temporal analysis for dropout prediction using self-regulated learning strategies in self-paced MOOCs	Predictive models, self-regulated learning	Questionnaire for MOOC participants on Electronics, named “Electrons in Action”, Open edX platform	To explore how self-regulated learning (SRL) strategies can be integrated into predictive models for self-paced MOOCs; it also introduces a new temporal analysis methodology for self-paced MOOCs aimed at early detection of learners at risk of dropout

.

3.4.3. Words Analysis

This section examines the most frequently used keywords, including both ‘keywords plus’ and authors’ keywords, as well as keyword groups (bigrams or trigrams) in titles and abstracts, to provide a clearer understanding of the research focus.

Table 10. Top 10 most frequent words in keywords plus and authors’ keywords.

Keywords Plus	Occurrences Keywords Plus	Authors Keywords	Occurrences Authors Keywords
students	31	mooc/moocs	103
engagement	19	dropout prediction	21
motivation	18	massive open online courses	21
education	15	machine learning	20
performance	15	dropout	19
online	13	learning analytics	11
model	12	online learning	11
motivations	10	deep learning	10
open online courses	10	distance education	9
participation	10	e-learning	9

presents the most commonly used ‘keywords plus’ and authors’ keywords. Among the ‘keywords plus’, ‘students’ appears most frequently, with 31 occurrences, reflecting MOOC participants. Other notable terms include ‘engagement’ (19 occurrences), ‘motivation/motivations’ (18/10 occurrences), and ‘performance’ (15 occurrences), which relate to the characteristics and objectives of MOOC enrollees. Additional frequently occurring terms are ‘education’ (15 occurrences), ‘online’ (13 occurrences), ‘model’ (12 occurrences), ‘open online courses’, and ‘participation’ (both with 10 occurrences).

For authors’ keywords, ‘MOOC/MOOCs’ is the most common term with 103 occurrences, followed by ‘dropout’ (19 occurrences) and ‘e-learning’ (9 occurrences), which are central to the dataset. The most frequent keyword groups include ‘massive open online courses’ (21 occurrences), ‘online learning’ (11 occurrences), and ‘distance education’ (9 occurrences). Keywords related to analytical techniques, such as ‘dropout prediction’ (21 occurrences), ‘machine learning’ (20 occurrences), ‘learning analytics’ (11 occurrences), and ‘deep learning’ (10 occurrences), highlight the methods used to investigate dropout.

This research focused on retaining only those word groups related to analytical techniques and factors influencing MOOC dropout, while excluding groups associated with search terms used in paper selection.

Table 11. Top 10 most frequent bigrams in abstracts and titles.

Bigrams in Abstracts	Occurrences	Bigrams in Titles	Occurrences
dropout rate/rates	160	dropout prediction	29
online courses	137	online courses	20
moocs	92	mooc dropout	14
dropout prediction	54	student dropout	11
online learning	54	machine learning	9
machine learning	43	mooc learners	8
student dropout	30	deep learning	7
neural network	29	data mining	5
completion rates	24	discussion forums	5
continuance intention	24	continuance intention	5

displays bigrams at both the title and abstract levels. It is noted that some bigrams appear in both categories. For instance, ‘dropout prediction’ appears 29 times in titles and 54 times in abstracts. Similarly, ‘online courses’ occurs 137 times in abstracts and 20 times in titles; ‘machine learning’ appears 43 times in abstracts and 9 times in titles; ‘student dropout’ is found 30 times in abstracts and 11 times in titles; and ‘continuance intention’ appears 24 times in abstracts and 5 times in titles.

Among abstract-level bigrams, ‘dropout rate/rates’ has the highest frequency, with 160 occurrences, highlighting its significance as a central research interest in this field. Other notable bigrams include ‘continuance intention’. Techniques such as ‘dropout prediction’, ‘machine learning’, ‘neural network’ (29 occurrences in abstracts), and ‘data mining’ (5 occurrences in titles) are prominent at both levels, reflecting methods used to analyze MOOC dropout. Additionally, ‘discussion forums’, found five times in titles, frequently appears in the literature as a factor influencing MOOC dropout both positively and negatively.

In realizing a paper related to the field of education, it could have been expected that the main word clusters identified would be related to both ‘dropout rates’ and ‘dropout predictions’. According to the literature, it has been observed that dropout is one of the most adverse and complex events, whether we are talking about its influence on the student or on the institution [119]. As the identification of predictive factors or patterns, involving consistent datasets and the need to create the most appropriate models, groups of words such as ‘machine learning’, ‘neural network’ or ‘deep learning’ or ‘data mining’ are normally among the most common ones, and researchers often use them in their analysis.

Over the years, there have been numerous researchers who have tried to realize these predictions using different techniques or combinations. Lykourentzou et al. [120] have combined, for the prediction of e-learning course dropout, three different machine learning techniques: support vector machines, feed-forward neural networks and the simplified fuzzy probabilistic ensemble ARTMAP; another example is represented by Lee and Chung [121], who used synthetic minority oversampling and ensemble methods in machine learning, as well as precision–recall curves and ROC (receiver operating characteristic) curve evaluation in order to predict dropout.

Table 12. Top 10 most frequent trigrams in abstracts and titles.

Trigrams in Abstracts	Occurrences	Trigrams in Titles	Occurrences
online courses mooc/moocs	97	mooc dropout prediction	10
machine learning algorithms	14	student dropout prediction	6
dropout prediction model	12	mooc discussion forums	3
convolutional neural network/networks	16	online courses moocs	3
structural equation modeling	8	self-regulated learning strategies	3
low completion rates	6	Chinese university students	2
neural network model	6	convolutional neural networks	2
self-regulated learning srl	6	deep learning model	2
student dropout prediction	6	dropout prediction model	2
accuracy precision recall	5	educational data mining	2

provides an overview of the 10 most common trigrams at both the title and abstract levels. At the abstract level, ‘online courses MOOC/MOOCs’ is the most prevalent trigram, appearing 97 times. In the category of analytical techniques, ‘dropout prediction model’ appears 12 times in abstracts and 2 times in titles, while ‘student dropout prediction’ is present 6 times in both categories, indicating its significance as a common method for analyzing MOOC dropout.

Other notable trigrams include ‘convolutional neural network/networks’, which appears 16 times in abstracts and 2 times in titles. Additionally, ‘neural network model’ is found 6 times in titles. The trigram ‘Chinese university students’ is noted with 2 occurrences. This observation aligns with the high citation rate from China (Figure 12) and the significant scientific output from Chinese researchers (Figure 10).

3.5. Mixed Analysis

This section presents a mixed analysis using various techniques to examine the correlations between different categories, track their evolution over time, and identify emerging areas of focus for future research.

The thematic map was chosen for this analysis in order to be able to infer the relevance of the theme and its development, each bubble in the figure helping to identify those words or groups of words at the level of the authors that have a high frequency of occurrence, their position being also representative [122]. Factor analysis is relevant because it helps to observe both the areas of interest and the research topics; this representation facilitates the way in which keywords are related to different concepts in the literature [123]. Moreover, both the thematic evolution and the three-fields plot have a special significance, making the reader able to understand the evolution over time as well as the way in which certain words, authors, countries or affiliations have collaborated or been used [122,123]. Each of the analyses mentioned has particular significance, and the information within them is diverse, even though some of the most relevant words or groups of words may be repeated. Considering other papers from the field, the prevalence of these analyses in better shaping and understanding the selected research domain has been observed [124,125].

3.5.1. Thematic Map

Figure 15 illustrates the thematic map of authors’ keywords, generated using specific filters. A total of 80 keywords were selected, with a minimum cluster frequency of 15 per thousand documents, a tag size of 0.35, and three labels. According to Nasir et al. [126], the thematic map classifies research topics into two dimensions: density and centrality. Density indicates the significance of a theme, while centrality reflects the volume of research effort devoted to it.

The top right quadrant, labeled ‘Motor Themes’, includes well-developed, frequently addressed topics that are highly relevant to the field. Conversely, the top left quadrant, labeled ‘Niche Themes’, features themes that are less prominent. In between these quadrants, the second largest cluster, shown in pink, comprises 15 significant terms. The most frequently occurring terms in this cluster include ‘feature extraction’ (nine occurrences), ‘massive open online courses’ (six occurrences), ‘e-learning’ (five occurrences), and ‘predictive models’ (five occurrences). This cluster encompasses MOOC dropout analysis techniques such as ‘machine learning algorithms’, ‘predictive models’, ‘random forest’, and ‘convolutional neural networks’, as well as domain characteristics like ‘education’, ‘collaboration’, and ‘electronic learning’.

Although small, the ‘Niche Themes’ quadrant contains an orange cluster with a single term: ‘course completion’. This represents a focal point of analyses aimed at identifying factors influencing MOOC completion or dropout, and addressing these factors to improve retention.

The bottom right quadrant, labeled ‘Basic Themes’, includes themes that are still in the developmental stage and require further research. The largest cluster in this quadrant, shown in blue, consists of 35 terms. Key terms with the highest occurrences include ‘MOOC/MOOCs’ (103 occurrences as shown in Table 9), ‘dropout’ (19 occurrences), ‘learning analytics’ (11 occurrences), and ‘online learning’ (11 occurrences). This cluster also includes essential MOOC characteristics such as ‘continuance intention’, ‘learner motivation’, ‘higher education’, ‘completion rate’, ‘connectivism’, ‘distance education’, ‘e-learning’, and ‘engagement’.

On the boundary between ‘Motor Themes’ and ‘Basic Themes’, a green cluster containing 14 terms is observed. The most common terms in this cluster include ‘dropout prediction’ (21 occurrences), ‘massive open online courses’ (21 occurrences), ‘machine learning’ (20 occurrences), and ‘deep learning’ (10 occurrences). These represent research themes and techniques frequently used in dropout studies, hence their positioning at the intersection of the two quadrants.

In the lower-left quadrant, labeled ‘Emerging or Declining Themes’, are topics with low density and centrality, indicating that they are underdeveloped. This quadrant features a purple cluster with three terms, ‘distance education and online learning’, ‘data science applications in education’, and ‘post-secondary education’, which are more general themes.

3.5.2. Thematic Map Evolution

Considering the significant increase in scientific output related to MOOC dropout research in 2020 (Figure 2), analyzing thematic evolution has been deemed relevant, as illustrated in Figure 16. Domain-specific terms such as ‘MOOC’, ‘MOOCs’, ‘massive open online course’, ‘massive open online courses’, ‘massive open online course (MOOC)’, and ‘distance education and online learning’ have been excluded from this analysis to focus on thematic evolution.

Figure 16 reveals the most frequently encountered terms in authors’ keywords between 2013–2020 and 2021–2023, highlighting how these terms have evolved to define new concepts.

Notably, ‘blended learning’ has led to the emergence of terms such as ‘learner autonomy’ and ‘motivation’, which also encompasses ‘self-regulated learning’. Additionally, ‘dropout’ and ‘dropout rate’ have been associated with ‘online learning’, reflecting ongoing challenges in online education. The integration of ‘machine learning’ with ‘dropout prediction’ illustrates the development of advanced techniques for analyzing dropout, a trend evident in the papers selected for our database.

Based on the thematic evolution and considering the papers included in the database, a series of papers related to factors in dropout have been identified in various studies, such as but not limited to the papers discussed in the following. Azhar et al. [127] used semi-structured interviews in order to observe what the factors leading to dropout were, grouping the responses into three main themes: inability to manage time, perception of course content and lack of motivation.

Eriksson et al. [117] interviewed 34 learners and were able to identify a total of four factors that influence the decision to drop out: learners’ ability to manage their time effectively, the learner’s characteristics and social situation, the learner’s perception of the course design, and the learner’s perception of the course. Bozkurt and Akbulut [128] focused their attention on dropout in networked learning, analyzing this phenomenon in terms of cultural contexts. Using a mixed approach (group comparison and social network analysis), the two researchers analyzed a total of 179 MOOC participants. Using the results obtained, the researchers concluded that those students who do not drop out occupy central positions in the network, and that students from high cultural backgrounds tend to drop out more often, while those from low cultural backgrounds are more likely to complete the course.

In terms of dropout, dropout intention and continuance intention, a series of works have been identified in the extracted dataset. For example, Guo et al. [108] identified the problem of identifying those posts in MOOC forums that are considered to be “urgent”, but in most cases end up being influenced by a large volume of information and confusion. As a solution to this problem, the researchers propose a hybrid neural network that identifies urgent posts. This approach helps teachers to prioritize responses and better manage postings, helping to reduce dropout rates and improve course completion rates.

Xing and Du [114] set out to customize interventions for each student on the verge of dropping out. Starting from a temporal prediction mechanism based on deep learning algorithms, the researchers build a dropout prediction model that estimates the probability of dropout for each individual student. By harnessing the power of deep learning, the algorithm allows for both the prioritization and customization of interventions for students at risk of dropping out based on these individual probabilities.

Furthermore, the use of machine learning has been highlighted in various works included in the extracted dataset. It has been observed that due to the increasing developments in recent years in the use of machine learning techniques, there has been a massive use of machine learning in the research found in the database used to create this analysis. The use of machine learning in dropout analysis is extremely valuable due to its ability to identify complex patterns and predict dropout risks based on detailed data analysis.

Based on the high MOOC dropout rate, Chi et al. [129] built prediction models using different machine learning techniques, such as K-nearest neighbor, logistic regression or Random Forest. Subsequently, by evaluating the models with five metrics (accuracy, precision, recall, F1 score, and AUC), they concluded that the random forest-based model performed the best compared to the logistic regression and K-nearest neighbor-based models. Thus, the study suggests that random forest represents the most effective technique in predicting dropout in MOOCs, having the ability to assist educational personnel in making decisions to reduce dropout.

Hong et al. [130] describe the benefits of using a two-layer cascade classifier for dropout prediction, combining three different machine learning techniques: support vector machine, multinomial logistic regression and random forest. The authors choose to use this approach to improve the accuracy of dropout prediction. The experimental results suggest that the technique has an accuracy of 97%, which indicates a very good performance in predicting dropout risk.

Considering the behavior of the students enrolled in the MOOCs, Sr and Saravanan [131] used the frequent pattern growth technique to be able to identify frequent items and create useful features for the prediction model and used artificial neural networks for feature selection, implemented using frequent itemset-3. The obtained prediction models were evaluated using multilayer perceptron and compared with machine learning methods such as naive Bayes, random forest or K-nearest neighbor. The results showed that the feature importance association rule for the artificial neural network model had an accuracy of 92.42%, being about 18% more accurate than the multilayer perceptron–neural network model. The ultimate goal of the research was to reduce dropout rates and increase student retention.

In the study conducted by researcher Şahin [132], the adaptive neuro-fuzzy inference system technique was used to predict dropout rates in MOOCs. This method has the ability to combine two machine learning approaches: fuzzy inference systems and neural networks. The neural network helps with learning from the available data, and the fuzzy system helps with handling uncertainty and variations in the input data. The study compares the performance of this model with that of others developed based on traditional machine learning methods (support vector machine, decision tree, ensemble learning, logistic regression and K-nearest neighbor). The researchers concluded that the adaptive neuro-fuzzy inference system technique resulted in higher statistical accuracy compared to the reference models.

Mourdi et al. [27] conducted an exploratory study as well as a multivariate analysis in order to reduce the dimensionality and obtain the most relevant features. During the analysis, the researchers compared five machine learning algorithms, using association rules to extract similarities between the behaviors of those learners who dropped out of MOOCs. The results of the study showed that deep learning technology provides the most accurate predictions in terms of accuracy, recording an average of 95.8%. What this study contributes, in addition, is the provision of a model capable of not only predicting who the learners at risk of dropping out are, but also who those who will fail or succeed are.

3.5.3. Factorial Analysis

Figure 17 presents a Multiple Correspondence Analysis (MCA) for KeyWords Plus, which extends Correspondence Analysis (CA) by examining relationships between multiple categorical variables [90]. The analysis reduces the initial data dimensions to two primary dimensions, referred to as Dim. 1 and Dim. 2. Dim. 1 accounts for 25.53% of the variation, while Dim. 2 accounts for 19.15% of the variation. These dimensions reveal the main trends and structures within the data, with terms in each dimension clustered based on similarity.

Cluster 1, shown in blue, is larger than Cluster 2, shown in red. Cluster 1 comprises 47 words, whereas Cluster 2 includes only 7 words. Cluster 1 exhibits both positive and negative loadings, with a predominant positive loading in both dimensions. Cluster 2, on the other hand, shows a significantly positive loading on Dim. 1 and mixed positive and negative loadings on Dim. 2, with a predominant negative loading.

Cluster 1 includes terms related to MOOC participant traits such as ‘students’, ‘engagement’, ‘motivation’, ‘performance’, ‘behavior’, ‘patterns’, ‘retention’, ‘satisfaction’, ‘self-determination’, ‘strategies’, ‘classification’, ‘persistence’, and ‘intention’. These terms have been identified as significant determinants of dropout in the selected papers. In contrast, Cluster 2 contains terms such as ‘achievement’, ‘continuance intention’, ‘experience’, ‘perspective’, ‘technology’, ‘responses’, and ‘self-efficacy’, which are central to describing the research field.

3.5.4. Three-Fields Plot

Figure 18 helps us to identify links between countries, authors and publications. As expected, the country with the most authors is China, advancing Gao ZH, Zheng YF, Liu Y, Feng J, Sun X and Zhou YH, which, as seen in Figure 8, are among the top most relevant authors. In contrast to China, the country with the top three authors, among the most relevant ones, is Morocco (El Kabtane H, Mourdi Y, Sadgal M). Most of the authors illustrated in the figure were rather reluctant to collaborate with other countries, with only 4 out of the 20 authors opting to work with another country. This can be considered to be due to the existing educational legislation in each country. A similar situation was found within journals, with most authors choosing to publish in more than one journal, while others remained loyal to the same journal. The most preferred journal was “IEEE Access”, as shown in Figure 4, in top most relevant journals.

Figure 19 contains the second mixed analysis, this time between publications, authors and keywords. The keywords identified in the figure below were later analyzed in our research, and can be consulted in Table 10. The same situation can be observed with the affiliations (Table 7), which are part of the top of the authors’ most chosen publications.

“Cadi Ayyad University of Marrakech” from Morocco is connected with several authors, namely El Kabtane H, Mourdi Y, and Sadgal M, who have distinguished themselves by consistent contributions to the research on the factors leading to MOOC dropout. It is worth mentioning that there were five authors who were not affiliated to one of the universities included in the figure, even though they are among the most significant writers.

Analyzing the annual scientific output since 2013, a significant increase was observed from 2016 to 2019. Later increases were also identified in years such as 2021 and 2022. The flourishing period in 2019 may have been due to the COVID-19 pandemic; according to other studies conducted by other researchers, the COVID-19 period was a beneficial period for MOOCs, with dedicated courses for different fields of activity [133,134]. The bibliometric analysis by Liu et al. [135] also showed a significant increase in 2019, even though their number of articles entered in the research was higher (1078 articles). Moreover, a steady increase in the number of articles for the period 2013–2019 was also observed in our study. The work of Wahid et al. [136], which contains a total number of 3118 articles in the MOOC field, observed the highest increase in 2018, with a total of 678 articles published compared to the total, 2019 for which they identified only 298.

Regarding the most relevant sources identified, the journals “IEEE Access”, “Education and Information Technologies” and “International Review of Research in Open and Distributed Learning” represented the top three journals that stood out for their high number of articles published on MOOC dropout, due to their high interest in education. “IEEE Access” even has a dedicated section for educational technology and theory entitled “IEEE Access: IEEE Education Society Section”. Compared to the work of Wahid et al. [136], “International Review of Research in Open and Distributed Learning” is ranked fourth in terms of number of sources, which means that this journal is a significant one, having more publications related to our research area. Moreover, it denotes that the publication has a solid reputation, being able to give visibility and credibility to the published work. Making another comparison between the article cited above and our paper, we can see that Alario-Hoyos C., who appears in the top authors identified in terms of scientific creation, is also found in the cited article, but with a higher number of publications.

Regarding regional contributions to the field, China stands out with the highest number of publications, totaling 133 papers. These papers have accumulated a total of 1155 citations, and China has demonstrated the most extensive collaboration with other countries in the region. Following China, the USA has produced 40 papers with a total of 798 citations. In addition to these two countries, the United Kingdom, Spain, India, Morocco and India were also considered representative. In the paper by Wahid et al. [136], the top five significant countries are the USA, China, Spain, UK and Australia, a result that is very similar to ours. Another paper by Liu et al. [135] identifies USA, UK and Spain as the most significant countries.

Based on the papers included in the dataset, the following sections highlights some research directions, as well as a sample of papers from the identified research areas. Some of the research directions focus on the methods used (e.g., machine learning algorithms), while other focus more on the practical aspects related to MOOCs, such as motivational and psychological factors.

4. Conclusions and Limitations

The aim of this paper was to highlight the most significant papers indexed in the ISI Web of Science database related to school dropout and the factors contributing to it. The study covered the period from 2013 to 2023, with papers selected using specific filters relevant to our research. These papers were then subjected to a comprehensive bibliometric analysis to extract the most pertinent information.

MOOC analysis is particularly important as it is an essential technique for democratizing education, providing access to quality learning for a global and diverse audience. By analyzing the most globally significant MOOCs, one can better understand how learners interact with learning materials, what factors influence course success or dropout, and how pedagogical and technological strategies can be improved to enhance their effectiveness. It also helps to identify barriers to access and participation, enabling the development of solutions to make online education more inclusive and tailored to learners’ individual needs. With this work, the importance of education has been highlighted, regardless of the channel (online/offline) on which it is realized, as well as the importance of stopping dropout.

Based on the analysis conducted in the paper, aside from the identification of the top researchers in terms of citations and number of papers, top universities and journals, the description of the most cited papers, the characteristics of the collaboration networks, and the most used keywords, the research has also pointed out a series of research directions that can be followed in future research in the area. As a result, based on the thematic map, the prevalence of the transition from “self-regulated learning” to “motivation” and “learner autonomy”, which could be further supported by building personalized learning strategies that can boost the learners’ motivation, is underlined in this study.

Also, a more proactive approach that shifts the focus from an analysis of the “dropout rate” to an analysis of a “completion rate” or “continuance intention” could be another avenue for research in which the focus could be on developing models tailored for exploring predictive models and engagement strategies in order to boost student retention in the learning process.

Nevertheless, the analysis of the psychological and social factors that might affect the success of MOOCs could be another research direction, as can the analysis of other learning options one might find freely available in the online environment which could influence MOOC adoption and retention rates.

For a more comprehensive future study in this field to be conducted, we aim to expand the research by including indexed papers from additional databases to capture a broader spectrum of published work. Furthermore, an interesting area of research is represented by the integration of sustainable education frameworks into the MOOCs in order to provide a higher retention rate, which could bring long-term benefits to various learning communities.

Although the analysis provides significant information on MOOC dropout, it is essential to acknowledge and describe its limitations. Based on the database used, the selection of articles was restricted to those available in the WoS database. While this database includes many relevant papers, relying solely on WoS may be seen as a limitation. Incorporating other databases, such as Scopus, could have provided a more comprehensive view of the research domain. Limiting the analysis to English-language papers or those classified as “article” may also be seen as a constraint, as relevant studies in other languages or with different classifications might have been excluded, potentially affecting the findings. By focusing solely on articles indexed as “article” in the WoS database, the analysis includes both conference papers and journal articles. Additionally, the time frame restriction to 2013–2023 could mean that some research was omitted, as papers published after this date are not included in the analysis.