Bachelor’s degree student dropouts: Who tend to stay and who tend to leave?

Highlights

• The percentage of lost credit vouchers is the most important variable for the classification.

• Starting from the second semester, pre-entry attributes become irrelevant for classification.

• The main pre-entry attribute related to leaving after the first semester is the time gap between high school and university.

• Decision trees provide accurate and interpretable models.

Abstract

Factors of students’ dropout can be studied either by surveys among students or by analyzing data the university collects. In the work reported in this paper, we analyzed data known about students at the time of admission as well as data about the students' study achievements collected on a semester basis. Using data about students who enrolled in the academic year 2013/14, we created several data mining models to predict who will finish their studies successfully and who will not. Our results show that the key factor is the percentage of lost credit vouchers in the most recent semester. The pre-entry attributes have only a very small impact. We also created association rules of different types to find characteristics of students who did not successfully complete the first semester of study. Here, the factor that mainly increases the probability of a failure is the time gap between secondary and tertiary education.

Introduction

Large dropout rates of university students can be observed all over the world. According to Aulck, Nambi, Velagapudi, Blumenstock, and West (2019), "Each year, roughly 30 % of first-year students at US baccalaureate institutions do not return for their second year". According to OECD (2013), the European dropout rate is 30 %. Norton and Cherastidtham (2018) report that "Nearly a quarter of a million students will start a bachelor degree in Australia in 2018, but more than 50,000 of them will leave university without getting a degree". Analytical materials of the Ministry of Education, Youth and Sport of the Czech Republic show that about 25–35 % of students who enrolled for bachelor's degree studies at Czech universities between the years 2003 and 2015 ended their studies during the first year (MSMT, 2016). So regardless of the country, the numbers are very similar: a significant percentage of students who enroll in study toward the bachelor degree leave the university before obtaining the degree.

As pointed out by Larsen, Kornbeck, Kristensen, Larsen, and Sommersel (2013), student dropout has consequences at the student, university, and society levels. The consequences at the student level are psychological (feeling of individual failure and loss of self-confidence) and economic (waste of personal resources). The consequences at the university level are economic, organizational, and academic. As many universities are funded according to the number of students, losing students means losing money. Organizational consequences are related to scheduling teachers, classrooms, and laboratories for students who might not enroll for the forthcoming semester. Academic consequences are related to the fact that, in the case that students are leaving by their own free will, the university can lose potentially successful students and graduates. At the society level, a high dropout rate affects both the educational system and labor market in a given country.

Complex dropout and retention models have been proposed to explain this phenomenon. Tinto's model of institutional departure, grounded in psychology and sociology, understands university as a composite of two systems: academic and social (Tinto, 1993). A student needs to be integrated into both of these systems to successfully finish his/her study. Tinto identifies family background, skills and abilities, prior schooling (considered as pre-entry attributes), institutional experience with both academic and social systems, integration into both academic and social systems, and external commitments as key factors affecting the goals and commitments of a student during his/her study, thus affecting the dropout decision. Bean's student attrition model is based on an analogy of student dropout with employee turnover in an organization (Bean, 1983). Like that of employees, students' satisfaction is also affected by organizational determinants. Bean models the dropout decision using four groups of variables: background, organizational determinants, satisfaction, and institutional commitment (Aljohani, 2016).

It is important for universities to understand the factors that lead to student success and student dismissal, and researchers can help with this understanding. Larsen et al. (2013) distinguish between two research approaches. The first approach, called "German" relies heavily on the dropout and retention theories. Thoroughly organized questionnaire surveys cover questions related to all aspects of dropout as described in the theory. The second approach, called "British" is more data-driven; university register data are analyzed with the aim to create classification or prediction models. While the first approach goes more in-depth, the response rate of the survey is usually low. Besides this, the information about the eventual dropout of a student is not known when he/she answers the questionnaire. In the second approach, on the contrary, large samples with the information about the dropout are available but cover only a limited range of dropout-related factors.

The "German" approach is represented by detailed studies that report various factors related to the theoretical findings that influence the likelihood of a student dropping out. Sagenmuller (2018) identifies financial problems, poor secondary school preparation, poor identification with the study field, conflict with work and family commitment, increasingly failing courses, lack of quality time with teachers, de-motivating school environment, and lack of student support as key factors. Norton and Cherastidtham (2018) discussed factors related to personal and family background, academic performance, the field of education, and engagement with the study. The EU report on dropout and completion in higher education in Europe presents a comparative study based on an extensive review of literature and policy documents on study success, a survey of national higher education experts, and eight in-depth country case studies including the Czech Republic (Vossensteyn et al., 2015). A survey of attitudes and motivation of students at Czech universities was carried out as a part of a broad, cross-country European comparison (the Eurostudent VI project) in 2016. The studied factors related to dropout were age, gender, subjective assessment of study achievements, the form of study, type of study program, and type of school (Fischer et al., 2016).

The “British” approach is represented by the area of educational data mining, a sub-field of data mining that aims "to detect patterns in large collections of educational data that would otherwise be hard or impossible to analyze" (Romero & Ventura, 2013). Data mining in general can be used for segmentation, concept description, classification, prediction, dependency analysis, or deviation detection in many areas (Chapman et al., 2000). Statistical methods (e.g. logistic and linear regression, or k-nearest neighbor clustering), as well as artificial intelligence and machine learning methods (e.g. decision trees, decision rules, association rules, neural networks, Bayesian methods or SVM), are applied to find useful knowledge in the data. These methods can be grouped into two categories. Symbolic methods, e.g. decision trees or rules, represent the found models in a human-understandable way, so both accuracy and interpretability are important. Sub-symbolic methods, e.g. neural networks or SVM, are not readily intelligible by humans, so only the accuracy of the models matters. In educational data mining, the data collected by the university about their students are used to create models to predict students' success (here both symbolic and sub-symbolic methods can be used) or to look for important characteristics related to students dropping out (here only the symbolic methods make sense).

Dekker, Pechenizkiy, and Vleeshouwers (2009) used decision trees, Bayesian networks, logistic regression, and rule-based methods to predict dropout of freshmen at the Electrical Engineering department at Eindhoven Technical University. They considered both pre-entry data and data from the first year of study and created models (1) from pre-entry data only, (2) from study data only, and (3) from all data. Nicholes and Reimer used logistic regression to analyze the impact of the first-year composition on student persistence. They explored the pathway through composition-1 and composition-2 courses to graduation with the aim of early recognition of students needing some support. The data were collected at one university in the US (Nicholes & Reimer, 2020). Rovira, Puertas, and Igual (2017) used different machine learning techniques for the early prediction of student dropout as well as for the prediction of grades. For the first, classification task the authors used logistic regression, a naive Bayes classifier, SVM, random forest, and AdaBoost. For the second, prediction task, linear regression and support vector regression were used. The data collected at the University of Barcelona consists of grades in all courses taken by a student. Abu-Oda and El-Halees analyzed data about grades in particular courses of bachelor-degree computer science students at Alaqsa University. They used these data to classify dropouts using decision trees and a naive Bayes classifier and to find relationships between grades using association rules mining (Abu-Oda & El-Halees, 2015). Pal and Pal used decision trees to analyze data about socio-demographic characteristics of students at Veer Bahadur Singh Purvanchal University in Jaunpur, India. They tried to predict the students' performance with the aim to identify students who need special advising or counseling. (Pal & Pal, 2013). Montmarquette, Mahseredjian, and Houle (2001) describe a bivariate probit model created to differentiate between persistence and dropout using longitudinal data about enrolment of students at the University of Montreal. Aulck et al. (2019) applied logistic regression, k-nearest neighbors, random forests, SVM, and gradient boosted trees to predict graduation or non-completion of students using data from the University of Washington.

Our research fits into the data-driven (i.e. “British”) approach. The data used in our work come from one Czech university (University). In our previous work, we focused on data known about the students at the time of their admission (Berka, Marek, & Vrabec, 2019). In the work reported in this paper, we also analyze the data about the students' study achievements collected on a semester basis. We aim at differentiating between bachelor's students who successfully completed their study at the University and those who did not. We used decision trees, random forest and logistic regression to be able to predict who will finish their studies successfully and who will not (either being dismissed or leaving by his/her own choice). We also perform association rules mining to find interesting characteristics of students who did not complete the first semester/year.

In the rest of the paper, we describe the data used, define the data mining tasks, describe the experiments we carried out, discuss the results, and show future research directions. As we follow the CRISP-DM methodology (Chapman et al., 2000) in our work, we organize the main part of the paper according to its basic steps: business understanding, data understanding, data preparation, and modeling.