1 Introduction

In a world in constant evolution, the needs of the labor market are also frequently changing (Fareri et al., 2020). Technological innovations lead to the creation of jobs with new requirements. With these innovations, the established tasks evolve, and the skills associated with them change (Fareri et al., 2020). It is then imperative that education is adapted to better prepare each citizen for these new market requirements (Skryabin et al., 2015). Indeed, education is seen as a factor contributing to the growth and development of economies (Hanushek & Woessmann, 2008). It is through education that citizens acquire the necessary skills to enter the labor market and become creators of value for society (Srijamdee & Pholphirul, 2020). As such, education must keep up with the digitalization process of the most diverse sectors of the economy.

The relevance of digitization became even clearer in 2020 with the spread of the COVID-19 pandemic. Educational institutions were some of the most affected by the pandemic. The need to move teaching to a remote setting brought several obstacles and caught many education institutions unprepared, as they were not yet able or accustomed to working with access to information and communication technologies (Gillis & Krull, 2020). Society in general was not prepared for such a sudden change, as were neither teachers nor students (Shim & Lee, 2020).

Aligned with this digitalization phenomenon, many tasks are currently performed with the aid of information and communication technologies (Srijamdee & Pholphirul, 2020). Therefore, it is worth understanding which teaching method best provides the skills required to work with information and communication technologies. The answer seems to lie in education using information and communication technologies, contrary to what has been the norm until now (traditional classrooms) (Fullan & Langworthy, 2013). This means not only learning how to use information and communication technologies well but also using these technologies for learning purposes in general. An example of this trend can be found in the Portuguese education system, where the use of information and communication technologies is seen as a necessity for education purposes (Cnedu, 2019).

To the best of our knowledge, no study has yet determined the factors that promote students’ use of information and communication technology (ICT) for educational purposes, even though several studies specify the use of information and communication technologies for educational purposes as a good predictor of student performance (Bielefeldt, 2005; Kubiatko & Vlckova, 2010; Skryabin et al., 2015; Srijamdee & Pholphirul, 2020). Therefore, it is very relevant to find out what factors influence this type of use so that better incentive measures can be designed to promote it. Regarding student performance, several studies have analyzed how it is impacted by information and communication technologies, however, no consensus has yet been reached.

Within this context, our aim in the present study is to determine which factors promote students’ use of ICT for educational purposes so that ICT implementation in schools can be carried out with optimal effectiveness, easiness, and speed. Accordingly, we seek to answer two main questions:

How do socio-demographic factors impact students' use of information and communication technologies for educational purposes? To carry out this analysis, we used the 2018 database from the Program for International Student Assessment (PISA).

Which factors impact students' use of information and communication technologies for educational purposes the most?

The paper is organized in five sections. After this introduction to the research topic, we review the literature on the implementation of information and communication technologies in education. In Section 3, we describe the methodology used, based on the construction of a classification machine learning model, using the CRoss-Industry Standard Process for Data Mining framework. Section 4, presents and discusses the results obtained from testing the model with data from PISA 2018. Finally, in Section 6, we outline our conclusions and provide practical recommendations to assist in the adoption and implementation of ICT for educational purposes.

2 Literature review

The literature on the implementation of information and communication technologies in education is vast. However, there is no consensus regarding how the use of information and communication technologies impacts student performance. It is believed that the skills developed through the frequent use of ICT have a positive influence on student performance. The improvements confirmed at various student levels include increased autonomy related to school tasks (Clark & Lee, 2019), higher awareness of news about their curriculum (Chou & Block, 2018), increased creativity and critical thinking (Chou & Block, 2018), better insight into their performance and a higher level of independence and ownership (Clark & Lee, 2019), a more personalized learning experience (Chou & Block, 2018; Harper & Milman, 2016), the development of the skills needed to solve real problems (Chou & Block, 2018); easier interaction with peers and teachers, e.g., to ask questions (Clark & Lee, 2019), and greater enjoyment and motivation to learn (Clark & Lee, 2019; Harper & Milman, 2016).

Most studies show that the relationship between the use of information and communication technologies in education and student performance is mostly positive. For example, the use of ICT in education seems to improve student performance in various subjects such as science, reading, and mathematics (Areepattamannil & Santos, 2019; Bielefeldt, 2005; Ferraro, 2018; Kubiatko & Vlckova, 2010; Skryabin et al., 2015; Srijamdee & Pholphirul, 2020; Wei et al., 2020; Xiao & Hu, 2019a). However, it is important to note that this improvement appears to be higher in students who are average at mathematics than in those who are very good or very bad, i.e., very good students would still be good regardless of access to ICT and very bad students do not become good in a short period of time just because of access to ICT (Wei et al., 2020). The literature further indicates that student performance in mathematics is not related to having access to information and communication technologies at school or at home, but rather to the actual use of these technologies (Bielefeldt, 2005; Srijamdee & Pholphirul, 2020). On the other hand, the use of ICT for recreational purposes such as games and programming seem to negatively influence performance (Kubiatko & Vlckova, 2010; Srijamdee & Pholphirul, 2020). Students’ perceptions of ICT, namely regarding perceived autonomy and perceived competence also appear to have a positive influence on their performance in science subjects, with the effect of autonomy being higher (Areepattamannil & Santos, 2019). Finally, Skryabin et al. (2015) found that the level of ICT use in a country is positively related to students' achievement in mathematics, sciences, and reading. The effects of individual usage of ICT proved to be good for 4th-grade students’ performance, both at home and at school. For 8th-grade students, the use of ICT at home was positive regardless of whether it was for entertainment or educational purposes, while the use of ICT at school was negative.

On the other hand, some studies found that the impact of ICT use on students’ performance is not so positive. Biagi and Loi (2013) found that gaming is the only activity that positively relates the intensity of ICT use with students' achievement. Students' perceived ICT competence does not seem to have a positive influence on student performance in science, which can be explained by students' overconfidence in their ability to master ICT and by cultural factors,Footnote 1 although autonomy and interest in ICT appear to have a positive influence (Li et al., 2020; Meng et al., 2019). Petko et al. (2016) argued that ICT use at school and ICT use for entertainment have a negative relationship with student performance, even though ICT use at home is positively associated with student performance. Despite believing that ICT use is good for education, Petko et al. (2016) pointed out that this type of analysis should focus on the quality of ICT use rather than just its intensity and concluded that there is a positive relationship between a positive attitude toward ICT and performance. Lastly, Zhai et al. (2019) concluded that the relationship between ICT-related activities and students' achievements in physics depends on the intensity of ICT use in those activities.

Corroborating the above-mentioned results, the meta-analysis of Odell et al. (2020) concluded that the impact of ICT use on student performance is not consensual, with results existing in all directions.Footnote 2 However, these authors mentioned some trends in the results, such as perceived autonomy in ICT being a good influence for students. Competence and interest also appear to be mostly positive for students, although this is not always the case.

Regarding the factors that promote the use of ICT by students, the literature is scarce. The existing studies mostly take into consideration some socio-demographic factors to explain the use of technology in education. However, few other factors are considered, and these socio-demographic factors are only used as controls. Examples of these factors are gender, family structure, nationality, and repetition of school level (Coovadia & Ackermann, 2020; Petko et al., 2016; Skryabin et al., 2015; Xiao & Hu, 2019a, 2019b).

A small number of studies have explored how economic, social, demographic, and technological factors (namely regarding the use of technologies) influence ICT adoption, approaching the topic from different perspectives. The most studied issue relates to the identification of the factors that influence the adoption and use of ICT by teachers in classrooms (Al-Mamary, 2020; Basak, 2014; Juggernath & Govender, 2020; Mirzajani et al., 2016; Prieto et al., 2014; Salinas et al., 2016; Spiteri & Chang Rundgren, 2018; Wang & Han, 2018). Basak (2014) concluded that some of the barriers to ICT use by teachers are a lack of competence, lack of confidence, and lack of time. In terms of the factors enabling ICT use by teachers, Basak (2014) gave the examples of teacher familiarity with ICT, availability of ICT, reliability in ICT, and classes being more enjoyable for students, while Salinas et al. (2016) pointed out ICT training and the perceived contribution of ICT to student learning. Juggernath and Govender (2020) added that most teachers believe that ICT can bring improvements to teaching, that a lack of ICT mastery can make it difficult to implement ICT in teaching, and that most teachers are willing to learn how to work with ICT.

Some other research topics have been less frequently explored. For example, Chen and Hu (2020) discovered a strong relationship between interest in ICT and ICT self-efficacy that was in part mediated by ICT use. Additionally, Xiao and Hu (2019b) found a relationship between socio-economic factors and students' performance in the form of reading test scores that was mediated by ICT use. Finally, Juhaňák et al. (2019) studied the relationship between the age of first contact with ICT and perceived ICT competence and autonomy and concluded that the earlier the first contact with ICT, the better the perceived autonomy and competence. Furthermore, the authors found that the moderating factorsFootnote 3 were positively related to perceived ICT competence and autonomy (apart from the use of ICT outside of school for schoolwork).

In the present study, we used the most recent data (PISA 2018 database) and approached ICT use very granularly (at the student level) to identify the factors that influence students' use of ICT for educational purposes. We hope to contribute to a consensus regarding the use of ICT in education and introduce a novel approach to studying the factors that promote students’ use of ICT. This aspect is just beginning to be addressed in the literature and several studies refer to it not only as an opportunity for future research but also as a contribution to improve future discussions surrounding the impact of ICT use on student performance (Buabeng-Andoh et al., 2018; Eickelmann et al., 2016; Li et al., 2020; Odell et al., 2020).

3 Methodology

We sought to construct a classification machine learning model, more specifically a decision tree model. We chose a decision tree model since it is easy to interpret, handles numerical and categorical variables well, and, unlike traditional statistical methods, is non-parametric and non-linear (Önder & Uyar, 2017).

We have adopted the CRoss-Industry Standard Process for Data Mining (CRISP-DM) framework to guide the present study from the definition of the problem to the evaluation of the results. This is one of the most popular frameworks to support a data mining project. The CRISP-DM framework was developed by an association of companies involved in data mining and has since been adopted by the research community working in the field.

CRISP-DM is used to provide solutions to problems that are solved using data mining and considers different stages for the data mining process, going through the understanding of the data, the preparation of the data, and the modeling. In the following sections, we describe the work developed within the scope of each CRISP-DM stage.

3.1 Business understanding

It is of the utmost importance to identify the factors that may influence students’ use of ICT for educational purposes. This identification may help educational institutions to effectively implement ICT into their teaching and learning practices and to promote better results for their students.

Therefore, in the present study, we adopted and adapted the conceptual framework proposed by Biagi and Loi (2013). The main change to the original framework consisted of not considering “ICT use” but instead “ICT use for educational purposes”, as it potentially impacts more directly the student’s school performance. Other changes included the separation of Student-level factors into two categories to distinguish between factual variables and interpretative variables, allowing a more comprehensive understanding of the influence of each category on the dependent variables, and the creation of a dedicated group of variables for teachers, as they are potentially the most promising influencer regarding the use of ICT for educational purposes at school. This framework assumes that factors related to school, teachers, family, and students affect students' use of ICT for educational purposes, which, in turn, is expected to influence students' school performance.

In the present study, we focused on whether certain factors have an influence on the use of ICT for educational purposes, considering most of the factors proposed by Biagi and Loi (2013), as seen in Fig. 1. Each group of variables considered included both variables related to ICT and others that were not directly related to ICT.

Fig. 1
figure 1

Conceptual framework systematizing factors that influence the use of ICT for education purposes (adapted from Biagi & Loi, 2013). Note. Own elaboration based on Biagi and Loi (2013). a “Ratio ‘no. students/no. teachers’” refers to the ratio between the number of students and the number of teachers, to provide an idea of how many students per teacher there are in the respective school

Full size image

3.2 Data understanding

The data used in the present study is from the 2018 PISA databaseFootnote 4 of the Organization for Economic Co-operation and Development (OECD). This is an international, standardized educational research study of 15-year-olds. PISA 2018 is the seventh three-year cycle of this OECD research program, which began in 2000. PISA 2018 involved 79 countries. Around 600,000 students participated, representing a population of approximately 32 million 15-year-old students. The sample consists of approximately 50% female students and 50% male students.

PISA databases are constructed based on questionnaires and tests. The variables used for the present study come from the student survey, the teacher questionnaire, the student computer familiarity questionnaire, and the parent questionnaire.

The database used in the present study originally included 612,004 students, 40 variables related to the characteristics of the students, and two variables on the use of ICT by students for educational purposes.

3.3 Data preparation

To accomplish the goal of the present study, and in line with previous studies (Salloum et al., 2017; Jamil et al., 2018), the data preparation step was very important. Data preprocessing allowed the databases to be error-free and ready to use and ultimately enabled us to draw valid conclusions from them.

We selected and downloaded three PISA 2018 databases from https://www.oecd.org/pisa/data/2018database/: SCHQQQQ, which refers to the survey filled out by school officials, TCHQQQ, which refers to the survey filled out by teachers, and STUQQQ, which refers to the survey filled out by students.

After a detailed analysis of the variables in each of the transferred databases, and in line with the framework of Biagi and Loi (2013), we selected the variables with the highest potential of explaining/predicting the evolution of the behavior of the variables that characterize the use of ICT: the use of ICT for educational purposes in the classroom and the use of ICT for educational purposes outside the classroom. To build the databases used in the present study, we used three variables from the SCHQQQ database, 10 variables from the TCHQQQ database, and 29 variables from the STUQQQ database.Footnote 5 When preparing the database, we replaced the values of some of the attributes. Values with the following contents were replaced with a "missing values" category: "95—valid skip", "97—not applicable", "98—invalid", "99—no response", "9995—valid skip", "9997—not applicable", "9998—invalid", "9999—no response", "5—valid skip", "7—not applicable", "8—invalid", and "9—no response". Once this procedure was concluded, we performed a descriptive analysis of the database to derive statistics such as the number of missing values, the mean, the standard deviation, the minimum, the maximum, and the mode and thus get a better view of the distribution of variables and their usefulness for the present study.

As a result of this analysis, we removed the variables "How many languages […] do you and your parents speak well enough to converse with others? Your mother" and "How many languages […] do you and your parents speak well enough to converse with others? Your father" due to the excessive lack of data, i.e., missing values. Moreover, we included the variables "School size (sum)" and "School ownership",Footnote 6 which were extracted from the SCHQQQ database and added to the database via the ID of the schools, i.e., each student got the values of these variables according to their school.

We had to transform the variables "Teacher's use of specific ICT applications (WLEFootnote 7)", "Teacher's satisfaction with the current job environment (WLE)", "Teacher's satisfaction with teaching profession (WLE)", "Teacher's self-efficacy in classroom management (WLE)", and "Teacher's self-efficacy in maintaining positive relationships with students (WLE)" in order to link them to the students. These variables refer to each teacher, and since each student has several teachers, there are several values of these variables for each student. A simple way to solve this problem would be to use the average of the teachers in the school for each student. However, the average can be the same between schools with very different teacher profiles and these are variables that represent differences from an OECD average value, which led us to another transformation to better capture those differences: we decided to create new variables from the old ones. The new variables represent the "% of teachers in the X range of variable Y". We achieved these transformations by discretizing the variables into three intervals where each variable took only three values (interval1, interval2, or interval3, where the 1st represents the worst-case scenario, the 2nd represents the middle-case scenario, and the 3rd represents the best-case scenario). Next, we created binary variables for each interval, with ‘1’ meaning a teacher is assigned to interval X and ‘0’ meaning the opposite. Finally, we created variables corresponding to the sum of the teachers in each school to obtain the denominator for the ratios. To finish this transformation, we merged the two databases using the ID of the schools.Footnote 8 In the end, we obtained the variables "% of teachers in the X range of variable Y" for the five transformed variables, which resulted in 15 variables. An example is shown in Fig. 2 and explained in the following paragraph.

Fig. 2
figure 2

Source: Own elaboration

Example of variable transformations to link it to the students. Note.

Full size image

As an example, the variable “Teacher’s use of specific ICT applications” in one school initially contained 9 values, i.e., there were 9 teachers for that school, ranging from -1.635 to 1.685, with a mean of 0.022. After discretization, the variable included the following bins: range 1]-∞; -1.339], range 2]-1.339; 0.137], and range 3]0.137; + ∞ [. Once this division was defined, we created the binary variables for each interval and estimated a variable corresponding to the sum of all "1" values in each of the intervals. Afterward, we aggregated the teachers' information for each school. For this, we estimated the proportion of teachers in each school belonging to each of the previously mentioned intervals. Finally, we merged this database with the one containing the students' variables through the school ID so that all the students in the same school would obtain these values.

Still on the topic of the independent variables, there were other variables we had to modify.Footnote 9 In these cases, we only considered the average values of the teachers for each student in the same school, except for the variable “Originally trained teacher (wide definition): standard, in-service, or work-based teacher training (composite)”, for which we considered the mode of the teachers' answers for each student, given that this is a categorical variable. We decided what type of transformation to use based on the variance of the values. In case the variance observed was limited, we considered that the mean or the mode of the teachers' values would be an acceptable representation of how these characteristics impact students.

Data preparation also involved eliminating all students for whom there were no values for both dependent variables in the database, which reduced the data by about 50% (from 612,004 to 324,956). Given that the variables from the teacher survey (TCHQQQ) also had a very high number of missing data, we decided to perform two separate analyses. In the first analysis, we used the database excluding the variables from the teachers’ survey (TCHQQQ), hereafter called database 1. In the second analysis, we used database 2, which includes all variables. After creating the two databases, we eliminated all the students who had missing values in any of the variables, which made the databases smaller. Database 1 was left with 197,267 students and database 2 was left with 44,899 students.

3.4 Modeling and evaluation

The modeling phase consisted in constructing two decision tree models (Tables 7, 8, 9, and 10 in the Appendix) using the RapidMiner software. The independent and dependent variables included in the models are the ones shown in Tables 1 and 2, respectively. We considered these variables because they were available in the PISA 2018 database, a reliable and reputable database, and because they were able to represent each of the factors that influence students' use of ICT for educational purposes according to Biagi and Loi (2013).

Table 1 Independent variables used
Full size table
Table 2 Dependent variables used
Full size table

Table 2 presents the variables used to support the creation of the dependent variables. Based on these variables, we assumed that each student could belong to one of four possible clusters: cluster 0 (ICT use during lessons < 0 and ICT use outside of lessons < 0), cluster 1 (ICT use during lessons > 0 and ICT use outside of lessons < 0), cluster 2 (ICT use during lessons < 0 and ICT use outside of lessons > 0), and cluster 3 (ICT use during lessons > 0 and ICT use outside of lessons > 0). This is shown in Fig. 3. Since these variables originally represented the difference from the OECD student average, zero acted as the mean value.

Fig. 3
figure 3

Source: Own elaboration based on the PISA 2018 database

Student clusters for dependent variables (ICT use during lessons and ICT use outside lessons). Note.

Full size image

In cluster 0, students’ use of ICT during lessons and students’ use of ICT outside of lessons were below average. In cluster 1, students’ use of ICT during lessons was above average but the use of ICT outside of lessons was below average. In cluster 2, students’ use of ICT during lessons was below average and the ICT use outside of lessons was above average. Lastly, in cluster 3, students’ use of ICT during lessons and students’ use of ICT outside of lessons were above average.

After observing the distribution of students within these clusters, we decided it would be better for the model to predict only two possible extreme situations: above-average use of ICT for educational purposes (cluster 3) (regardless of location, i.e., at school or outside school) and below-average use of ICT for education purposes (cluster 0) (regardless of location), as most of the students fall within one of these two clusters. In database 1, cluster 0 had 61,383 students, cluster 1 had 26,904, cluster 2 had 37,262, and cluster 3 had 71,718. In database 2, cluster 0 had 15,907 students, cluster 1 had 8222, cluster 2 had 7932, and cluster 3 had 12,838. After disregarding all students belonging to the intermediate clusters, database 1 was left with 133,101 students while database 2 was left with 28,745 students.

To obtain reliable classification models, we divided the data into training, testing, and validation groups, which we then used to train the model. This division resulted in 80% training data and 20% test data. Of the 80% training data, 20% was validation data. To determine the optimal parameters, we used a grid search procedure. We used the validation data to determine the optimal decision tree parameters (minimum leaf size, maximum tree depth, distribution criteria, and minimum leaf size for distribution). To evaluate the quality of the models obtained, we used the "precision", "recall", and "accuracy" metrics.

4 Results and discussion

In this section, we present and discuss the results obtained. First, we present and discuss the model that does not consider the data from the teacher survey (decision tree 1), and then the model that considers all the data (decision tree 2).

The distribution of the students per cluster for the dataset that considers all the data is presented in Fig. 4. Cluster 0 consists of 61,383 students who had below-average ICT use during lessons and outside of lessons. Cluster 3 consists of 71,718 students who had above-average ICT use during lessons and outside lessons. In database 2 (Fig. 5), cluster 0 consists of 15,907 students who had below-average ICT use during lessons and outside of lessons. Finally, cluster 3 consists of 12,838 students who had above-average ICT use during lessons and outside lessons.

Fig. 4
figure 4

Source: Own elaboration based on the PISA 2018 database

Division of database 1 (not considering data from teacher survey) into quadrants (according to the clusters in Fig. 3). Note.

Full size image
Fig. 5
figure 5

Source: Own elaboration based on the PISA 2018 database

Division of database 2 (considering all the data) into quadrants (according to the clusters in Fig. 3). Note.

Full size image

For the construction of decision tree models 1 and 2, in the first parameter (minimum leaf size), we considered a minimum of 2 and a maximum of 100 (considering 20 possible values in this intervalFootnote 10), in the second parameter (maximum tree depth), we considered a minimum of 1 and a maximum of 20 (considering 15 possible values in this intervalFootnote 11), in the third parameter (distribution criteria), we considered the possibilities of the Gini index, the information gain, and the gain ratio, and finally, in the last parameter (minimum leaf size for distribution), we considered a minimum of 1 and a maximum of 100 (considering 10 possible values in this intervalFootnote 12). Once we determined the optimal parameters, we used them to train the final decision trees, which we then tested using the test data.

4.1 Summary of results and discussion—decision tree 1

Tables 7 and 8 in Appendix present the classification rules inferred from the decision tree model 1, which disregards data from the teacher survey. Having tuned the parameters, we ended up using the Gini index as the distribution criterion, a minimum leaf size of 22, a maximum decision tree depth of 5, and a minimum leaf size for distribution of 60. This model achieved an accuracy of 68.04% (Fig. 6), which demonstrates it has a good ability to distinguish between cluster0 and cluster3 students.

Fig. 6
figure 6

Source: Own elaboration based on the PISA 2018 database

Performance of decision tree 1 (not considering data from teacher survey). Note.

Full size image

Regarding the recall metric, the model correctly predicted 72.86% of the true values (cluster = cluster 3) and 62.42% of the false values (cluster = cluster 0). In terms of the precision metric, the model correctly predicted 69.37% of the true values and 66.31% of the false values. Given the high scores achieved in all three metrics, we can assume that this model can distinguish different patterns in ICT use for educational purposes.

Table 3 lists the obtained weight of the attributes in descending order of magnitude.

Table 3 Weight of the attributes (database 1)
Full size table

The weight of an attribute indicates the improvement in the performance of the tree when the attribute in question is selected for a given node.

When analyzing the results, a given level of a variable may be considered high or low depending on the country. Therefore, we analyzed the results per country, given the decision trees obtained. For example, a school size of 1000 students can be considered high in a country where the average is 500 and low in a country where the average is 1200. It is important to have this in mind in order to better understand the discussion of the results. However, we did not follow a per-country analysis for all variables, because, for most of them, the averages were very close.

The decision tree obtained allowed us to conclude that ICT access at school had a positive influence on the use of ICT for education purposes in almost all countries. In fact, access to ICT at school is crucial, as the school is where students are expected to more easily find support on how to use ICT for educational purposes. Access to ICT at home and the existence of ICT resources at home also had a positive influence on the use of ICT for educational purposes. This is entirely reasonable since the use of ICT for educational purposes outside of school is mainly expected to happen at home. In this sense, both types of access are central to surpassing the OECD average for ICT use for educational purposes in both locations, i.e., in and out of school. These results seem to be in line with the findings of Basak (2014). This author showed that the availability of ICT resources is one of the factors that promotes ICT adoption by teachers and our results suggest that this pattern may also apply to students.

The use of ICT for entertainment had a positive effect on the use of ICT for education purposes in almost all countries. The use of ICT for entertainment contributes to improving ICT skills, which in turn may favor the use of ICT for educational purposes. Nevertheless, it is curious to note that this type of use, which could have a negative influence because of the distraction it can cause, turned out to be generally positive. This conclusion may help to understand why Biagi and Loi (2013) reported that the only activity with a positive relationship between frequency of use and performance improvement was online gaming, i.e., the use of ICT for entertainment led to an increase in the use of ICT for educational purposes. Despite this, according to Srijamdee and Pholphirul (2020), the use of ICT for entertainment should not exceed a certain limit, otherwise, it consumes all the students' free time (Kubiatko & Vlckova, 2010) and compromises their availability to learn.

The interest of teachers (as perceived by students) had a positive influence on the use of ICT for educational purposes in most countries. A more interested teacher should be able to make the lessons more dynamic and try to be up-to-date with the state of the art in their teaching, which includes promoting the use of ICT, especially for educational purposes.

Parental emotional support (as perceived by students) had a negative influence on the use of ICT for educational purposes in general, contrary to what we expected. One explanation for this result may be that students with too much parental support do not feel the need to adopt new ways of studying or get out of their comfort zone, restricting themselves to "normal" studying, i.e., without the use of ICT.

Autonomy in using ICT had a positive influence on the use of ICT for educational purposes in most countries. A student must be able to use ICT autonomously in order to feel ready to use ICT for educational purposes. Likewise, ICT competence had a positive effect on good ICT use for educational purposes. We already expected this result, since a good use of ICT in any kind of activity requires the user to have the necessary skills. These results may help to explain why Areepattamannil and Santos (2019) observed a link between autonomy and competence in using ICT and a greater interest in topics such as sciences. Finally, these results may help to explain why Xiao and Hu (2019a) indicated that these variables can improve literacy.

Financial and social well-being also had a positive effect on the use of ICT for educational purposes. In terms of financial well-being, ICT involves some investment. This is reinforced by the fact that the use of ICT for educational purposes is highest among students enrolled in private schools. As for social well-being, a calm and positive environment should be beneficial for almost all types of student activities, including the use of ICT for educational purposes. This result (as well as those of similar types of influence such as parents' level of education and type of school attended) is consistent with those of Skryabin et al. (2015) and Xiao and Hu (2019a), who showed that the higher the socio-economic status of the student, the better the results. As is the case with other variables, the better performance associated with these variables in the previous studies might be explained by their positive effect on the use of ICT for educational purposes.

The level of education of parents also had a positive effect on ICT use for educational purposes. A high level of education of parents is usually linked with a culture that fosters study, interest in learning, and goal orientation. Thus, these students are more culturally inclined to follow their parents’ example and dedicate themselves to studying and achieving good results, which might mean using ICT for educational purposes.

Attitudes toward learning, as well as interest in ICT, proved positive in most of the results. These results are not surprising. Students with a positive and constructive attitude toward the role of school and learning are more willing to engage in study-related activities, namely the use of ICT for educational purposes. As for interest in ICT, students interested in ICT are expected to carry out more ICT-related activities (namely for educational purposes), which may help to explain why Li et al. (2020) found a positive relationship between interest in ICT and performance in science.

Learning about ICT-related topics was positive for the use of ICT for educational purposes. When students master ICT and recognize its dangers and possible uses, they are more apt to use it for educational purposes.

Finally, the results of the no. of students/no. of teachers ratio and school size variables point in both directions. Starting with the ratio, the results show that a lower ratio was beneficial for a better use of ICT for educational purposes, except in the UK, where the opposite was observed (a higher ratio can indicate more autonomy and thus better results, however, this is only a possible explanation, and this result should be explored further), and Hungary and Albania, where an intermediate ratio was more beneficial. A low ratio may be beneficial as it means that teachers can pay more attention to each student and manage the classroom more easily, which are factors that may lead to more dynamic lessons, in which ICT might be used for educational purposes. As for school size, the generality of the results indicates that larger schools were better when it came to ICT use for educational purposes. However, some results point in the opposite direction. Larger schools are typically more modern, located in more developed locations, have good ICT resources, and have more modern teaching methods, which might foster the use of ICT for educational purposes. On the other hand, in locations with more rudimentary schools that have poor ICT resources and worse infrastructure, a lower ratio might be important because it means more ICT resources per student, fostering ICT use for educational purposes. This explanation makes sense in the context of the present study since the countries where smaller school sizes were beneficial for the use of ICT for educational purposes (Bulgaria, Morocco, Switzerland, Turkey, and the Tatarstan region in Russia) have a lower Human Development Index (HDI) than those where the opposite happened (Estonia, France, Hong Kong, Singapore, Slovakia, the Moscow region in Russia, and Uruguay), apart from Switzerland and Uruguay, which did not fit this pattern.Footnote 13

4.2 Summary of results and discussion—decision tree 2

Tables 9 and 10 in Appendix present the classification rules inferred from the decision tree model 2, which considers all variables, namely those collected from the teachers’ survey. We based the final tree on the Gini index and considered a minimum leaf size of 27, a maximum decision tree depth of 7, and a minimum distribution size of 4.

This model achieved an accuracy of 64.78% (Fig. 7). This decrease in accuracy when compared to decision tree model 1 may be related to the significant reduction in the number of observations when considering all the predictors.

Fig. 7
figure 7

Source: Own elaboration based on the PISA 2018 database

Performance of decision tree 2 (considering all the data).

Full size image

As far as the recall is concerned, the model correctly predicted 62.54% of the true values (cluster = cluster 3) and 66.58% of the false values (cluster = cluster 0). Regarding precision, the model correctly predicted 60.17% of the true values and 68.77% of the false values.

Table 4 lists the obtained weight of the attributes in descending order of magnitude.

Table 4 Weight of the attributes (database 2)
Full size table

The relationships that emerged in this decision tree are similar to those found in decision tree 1, whose model excluded the teacher-related variables. Access to ICT both at school and at home was important for a good ICT use for educational purposes. ICT use for entertainment was once more positive for the use of ICT for educational purposes in most countries. Autonomy in ICT use and resilience were also important for students’ use of ICT for educational purposes. Learning about ICT-related topics such as search engines was also a determining factor in dictating the use of ICT for educational purposes.

Teacher-related variables such as satisfaction with the profession, level of ICT use in the classroom, satisfaction with the work environment, effort in maintaining good relationships with students, and need for ICT training also appeared in this decision tree. The majority of these variables had a positive influence on students’ ICT use for educational purposes. The only exception was the need for ICT training, which, as expected, had a negative influence, i.e., a greater need for ICT training by teachers was detrimental to students’ ICT use for educational purposes. This result is in line with that found by Juggernath and Govender (2020), who showed that teachers’ need for ICT training is a barrier to ICT implementation in classrooms.

Unlike in decision tree model 1, the emotional support perceived by students promoted ICT use for educational purposes in most countries. However, it rarely appeared in the decision tree, so it would be interesting to conduct further studies to clarify this relationship.

The no. of students/no. of teachers ratio and the school size were less relevant in this second model. The results show that a larger school size was beneficial for a good ICT use for educational purposes (again, countries confirming this rule have a high HDI, namely Hong Kong and the US). As for the ratio, it appeared only once in this decision tree, being associated with the UK. The UK stood out in decision tree 1 in this regard, and in decision tree 2, the results went in the same direction. Again, the need to promote autonomy might be higher in this country for some reason that should be further explored.

Finally, the attributes that influenced ICT use for educational purposes differently from what was expected were the ICT training of the teachers, the teachers' background in education, and the age of the teachers. We expected teachers’ ICT training and background in education to be beneficial for the dependent variables and expected teachers' age to have an inverted U-shaped relationship with the dependent variables. However, the results indicate that this was not the case. As for ICT training, one possible explanation is that teachers with no ICT training see ICT as a great benefit and innovation, and even without the necessary skills, they try to implement it into their teaching, making a greater effort than teachers for whom ICT is something more common. However, these results contradict those of Salinas et al. (2016), who found teachers' ICT training to have a positive influence on their adoption of technology. As for teachers' age, older ages were found to be more beneficial. This relationship may be related to younger teachers being less confident in their ability to innovate in teaching and to promote a teaching style that is different from the traditional one, despite their greater ability with ICT. Finally, teachers having a background in education had a negative influence on ICT use for educational purposes. One possible explanation is that teachers coming from other fields have different views about school subjects and more innovative teaching methodologies, which may include promoting ICT use.

5 Conclusions

In the present study, we researched which factors have the greatest impact on students' ICT use for educational purposes with the aim of assisting the implementation and use of ICT in schools. Given that ICT are massively adopted in society it is of crucial importance that schools have identified those factors and consequently adapt their ecosystem to meet the context in which students learn and grow up in today’s society. We used two decision tree models to answer the research questions. The obtained decision trees performed well, allowing us to derive valid and valuable conclusions about which factors influence students' use of ICT for educational purposes. We observed strong trends in the results among countries, however, some results turned out to be country-specific, so we recommend that each country analyzes its specificities and adopts the best measures to improve students’ use of ICT for educational purposes within its own context.

The database used for one of the decision tree models included variables used to characterize the teachers at the schools included in the study, while the database used for the other decision tree model did not include this data. The results obtained allowed us to conclude that the teacher-related variables were important to correctly predict the use of ICT for educational purposes, since the results of the two models were slightly different, and teacher-related variables were relevant in influencing the dependent variables in the second model. However, it is important to highlight that the model with the teacher-related variables performed worse than the other model. One possible explanation is the smaller number of observations included in the model with more variables, i.e., the model with the teacher-related variables.

Summarizing the observed trends, we found that the three most important variables for a good use of ICT for educational purposes were ICT use for entertainment, access to ICT at home, and access to ICT at school, which positively impacted ICT use for education purposes in almost all countries. These were followed by school size and the no. of students/no. of teachers ratio. However, the impact of these two variables varied more among countries than those of the previous three variables, so it would be interesting for future studies to focus on these two variables in order to better clarify their impact. There seem to be some trends related to the HDI of the countries that explain these variations, which could be a starting point to further explore this topic. A very interesting observation is that ICT use for entertainment seems to be able to stimulate ICT use for educational purposes as long as it is not excessive. Apart from these variables, there were several other important ones, such as learning ICT-related topics at school, teacher interest, autonomy with ICT, resilience, financial and social well-being, parental education level, attitudes toward school, ICT skills, interest in ICT, the type of school, and emotional support.

As expected, most of the teacher-related variables had a positive impact on promoting the use of ICT for educational purposes. Nonetheless, there were some surprising results, such as higher teacher ages being favorable for the use of ICT for educational purposes. Teachers having training in education or ICT was detrimental to students’ use of ICT for educational purposes. Finally, and as expected, teachers' need for ICT training was also detrimental to students’ use of ICT for educational purposes.

In the present study, we identified the factors with the highest impact on ICT use for educational purposes, which in turn is a very promising factor influencing student performance in various subjects, as stated in the literature. The present study can be used as a basis for future research seeking to build a model capable of effectively predicting between more than two clusters of ICT use for educational purposes and to stratify the various levels of such use. It would also be interesting for future studies to include more relevant variables in order to improve predictions and to gather more data on teacher-related variables in order to obtain a larger database. Finally, we believe that it would be important to focus on the variables that had non-consensual or unpredictable results and try to obtain new results, as these could lead to more robust conclusions.