Sage Journals: Discover world-class research

Abstract

The purpose of this study is to discover which noncognitive variables provide more information about reading performance. To answer this question, data mining based on information gain, decision tree and random forest methods were utilized in the study. The participants of the study consisted of 606,627 15-year-old students (49.8% female) in a total of 78 countries or economies, 37 of which are OECD members. Reading performance and plausible values of reading, the Student, ICT Familiarity, Financial Literacy, Educational Career, Well-Being and Parent Questionnaire data in PISA 2018 were analyzed to answer the research questions. When 108 features were analyzed as independent variables, it was found that SES (home possessions, cultural possessions, and ICT resources at home), metacognitive skills (assessing credibility and summarizing), and liking/enjoying reading were major variables predicting reading performance. The path analysis revealed that these variables explain 53.3% of the variability in reading performance. It is also remarkable that the decision tree model has a 74.61% accuracy value in estimating the reading performance.

Keywords

reading performance SES metacognition data mining

Introduction

Education is one of the most important factors in the scientific, technological, economic, and social development of a country (Hanushek & Wößmann, 2010; Kirikkaleli et al., 2021; McClelland, 1966; Uljens, 2007). Education has always been seen as the main source of power in winning political wars, and continuing revolutions and reforms in history. For example, in the restoration movement the Japanese Emperor Meiji started in 1868, education reform was one of the main sources of motivation for the development of the country (Morito, 1955). Defeated in World War II, Germany invested in education as the most important driver of postwar development (Füssl & Kubina, 1985). During the Cold War era, the Soviet Union sent the first satellite into orbit around the Earth in 1957, which led to comprehensive educational reform in the USA (Powell, 2007). Consequently, developed countries have realized that the real capital of a country is well-trained human resources rather than underground and surface resources (Hanushek & Wößmann, 2010; Sanders & Barth, 1968; Sauer & Zagler, 2014). Indeed, many countries rich in natural resources from South America, the Middle East, and North Africa (MENA) countries have still not reached the desired level of development today (Davoodi & Abed, 2003; Gylfason, 2001; Ozkaya et al., 2021).

Understanding that education has a key role in development has pushed countries into a race in this field. A competition based on comparisons in K-12 as well as higher education draws attention. For example, “Trends in International Mathematics and Science Study” (TIMSS), which tests the science and mathematics achievements of fourth- and eighth-grade students, “Progress in International Reading Literacy Study” (PIRLS), which tests the reading success of fourth-grade students (IEA, n.d.), and “Programme for International Student Assessment” (PISA), which tests the reading, mathematics, and science achievements of 15-year-old high school students (OECD, n.d.), are the most obvious indicators of this competition. The ranking of universities in higher education according to different criteria is also one of the indicators of this race (QS, n.d.; The World University Rankings, n.d.).

These examinations, which have been going on for several decades, have become the focus of the attention of researchers. Economists, statisticians, and engineers, as well as educators, are trying to discover the main factors that affect educational achievement. In particular, assessments such as TIMSS and PISA have special importance as they collect data on students’ noncognitive socioeconomic and psychosocial characteristics as well as academic skills. Hundreds of thousands of data collected from many different countries, in fact, have special importance because they help us see the big picture regarding education and the factors affecting it.

Considering the publications on factors affecting educational achievement, it is clear that the vast majority of researchers conduct studies in which certain countries are evaluated within themselves (Berger et al., 2020; Rolfe, 2021) or more than one country is compared in terms of performance (Kılıç-Depren & Depren, 2021; Wu et al., 2021). In addition, it is noteworthy that most researchers analyze a few theoretically selected variables (Wang & Liou, 2018; Wu et al., 2020). The main reason we focused on data from all countries was to see the big picture and explore the noncognitive variables affecting reading performance in the same context. When the Web of Science, Scopus and ERIC databases were searched, few studies were found in which PISA data obtained from all countries based on student perception was analyzed (e. g. Gamazo & Martínez-Abad, 2020; Lee & Stankov, 2018).

Background of the study

This study focused on noncognitive features defined by Duckworth and Yeager (2015), and Stankov and Lee (2014). Noncognitive features were defined as “a wide variety of individual attributes, skills, and characteristics representing one's attitudinal, behavioral, emotional, motivational, and other psychosocial dispositions” by Lee and Stankov (2018, p. 50). Noncognitive structures (a) are conceptually independent of cognitive abilities; (b) benefit the individual and others; (c) remain constant when external factors remain constant; (d) can be changed by training; and (e) depend on other situation-specific factors (Duckworth & Yeager, 2015). According to Lee and Stankov (2018), the underlying assumption of this concept is that the most important factor determining the academic achievement of students is the students themselves. Despite such importance being attributed, there is no definite opinion as to which of the noncognitive features are directly or indirectly related to student achievement.

To answer this question, Lee and Stankov (2018) examined the relationship of noncognitive variables in TIMSS (2003, 2007, 2011) and PISA (2003, 2012) with mathematics achievement. According to the results of this study, “self-efficacy” in PISA, “confidence” in TIMSS and “educational aspiration” in both assessments were the best predictors of mathematics achievement. Likewise, examining the results of five different previous studies, Stankov and Lee (2014) stated that “confidence” is the best noncognitive predictor of academic achievement. In addition, they stated that the predictive power of self-belief variables varies: for self-concept they are low, for academic anxiety medium, and for self-efficacy high.

Säälik et al. (2015) analyzed PISA 2009 data from the Nordic and Baltic countries and reported that student awareness of metacognitive learning strategies was the best predictor of reading proficiency both at student and school levels.

Comparing the PISA 2018 data from China and Turkey, Kılıç-Depren and Depren (2021) found that metacognitive competencies and the socioeconomic and cultural status of students were the significant factors influencing high reading achievement.

Gamazo and Martínez-Abad (2020) examined the predictive power of noncognitive features in the PISA 2018 on overall school achievement. According to the results of the analysis, performance in high socioeconomic status (SES) schools is mostly affected by metacognitive strategies and academic motivation, while performance in low SES schools is influenced by socioeconomic indicators across the country, such as Gross Domestic Product (GDP).

In conclusion, the initiation of our study is in line with the starting point of Lee and Stankov (2018) and Gamazo and Martínez-Abad (2020). The main purpose of this study is to discover which noncognitive variables provide more information about reading performance, with an exploratory approach, instead of modeling certain variables theoretically as a starting point. It would be more meaningful to undertake detailed statistical analysis after determining the variables that provide the most information. This study is intended to contribute to the literature in terms of defining the factors to be utilized in future studies in this regard.

Why is reading performance taken as the dependent variable?

Reading is one of the most important and complex functions of the human brain. It plays a wide range of roles in daily life, from simple written communication activities to high-level cognitive activities such as critical thinking (Hidayati et al., 2020; OECD, 2019a). As is known, reading performance is the major pillar for understanding and solving problems in all other academic domains. In other words, achievement in other areas depends on reading performance. PISA (2019b) addressed this issue as follows:

We live in a rapidly changing world, in which both the quantity and variety of written materials are increasing and where more and more people are expected to use these materials in new and increasingly complex ways. It is now generally accepted that our understanding of reading literacy evolves along with changes in society and culture. The reading literacy skills needed for individual growth, educational success, economic participation and citizenship 20 years ago were different from those of today; and it is likely that in 20 years’ time they will change further still. (p. 22)

Consistent with this explanation, many studies have shown that reading performance has a significant impact on achievement in other academic fields, such as mathematics and science (Grimm, 2008; Larwin, 2010; Schlatter et al., 2020). In addition, reading skills are an important determinant on high-level standard tests such as the Scholastic Aptitude Test (SAT) and the American College Test (ACT) (Hendricks, 2013).

Every three years, a randomly selected sample of 15-year-olds take exams in major areas such as reading, mathematics, and science, with each year of assessment focusing on a different topic. Reading was the focus in 2000, math in 2003, science in 2006, and reading again in 2009 (OECD, n.d.). The main subject of PISA 2018 was also reading (OECD, 2019a, s.15). Therefore, the dependent variable of this study was chosen as reading because of the main importance of reading performance explained above.

It is also important to clarify the terms used in this study, such as reading, reading achievement, and reading performance, to avoid ambiguity. “Reading” was used as the name of a general domain (e.g., “the main subject of PISA 2018 was also reading”) in this paper. Although reading performance covers reading and reading achievement conceptually, they were used interchangeably according to the cited articles (e.g., see Cheema, 2018; Kılıç-Depren & Depren, 2021; OECD, 2019a).

Why was PISA data chosen instead of TIMSS?

TIMSS and PISA measure different attributes. As seen from its title (Trends in International Mathematics and Science Study), TIMSS focuses on math and science; PISA focuses on reading additionally. While TIMSS consists of more curriculum-based questions, PISA measures the practical application of skills to real-life problems. While TIMSS is applied to fourth- and eighth-grade students, PISA is administered to 15-year-old high school students (Swensson, 2017). In this respect, PISA can also provide clues as to the achievements of previous years. Finally, while 64 countries were included in the last TIMSS in 2019, 79 countries participated in the last PISA in 2018.

Research questions

This study was conducted to answer three basic questions:

Of all the variables in the PISA student and parent surveys, which variables provide the most information gain (>2 SD) about reading performance?

When the PISA variables that provide the most information gain (>2 SD) about reading performance are tested with the decision tree model, with what accuracy can individuals with low and high reading levels be classified?

When the PISA variables that provide the most information gain (>2 SD) about reading performance are taken as independent variables, to what extent do they predict reading performance?

Method

The current study was designed based on secondary data analysis. We used the Student, ICT Familiarity, Financial Literacy, Educational Career, Well-Being and Parent Questionnaire data in PISA 2018, which can be downloaded from https://www.oecd.org/pisa/data/2018database/. Ethical approval from the institutional review board (IRB) for studies involving human subjects was not required for secondary data analysis. To answer our research questions, data-mining techniques and structural equation modeling (SEM) were used.

Participants

The participants of the study consisted of 612,004 students studying in the seventh grade and above in a total of 79 countries or economies, 37 of which are OECD members. These students were selected from those who were enrolled in any educational institution full-time or part-time, participating in academic or vocational programs, or attending public or private schools in the country or foreign schools. The ages of the participating students ranged from 15 years 3 months to 16 years 2 months at the time of the assessment, and they had at least 6 years of formal education (OECD, 2019a). 50.2% of the participants were male and 49.8% were female.

Since Reading Adaptive Performance (Core + First Stage) scores and plausible values (PV) for reading were not available in the dataset, the Vietnam sample could not be taken into analysis. In addition, because the Reading Adaptive Performance (Core + First Stage) scores of Argentina, Jordan, Lebanon, Moldova, Romania, Saudi Arabia, Ukraine, and North Macedonia were not used in the assessment, these countries were not taken for data-mining analysis. Consequently, we performed data-mining analyses using the data of 70 countries (N = 301,946; 48.2% female), and path analysis using the data of 78 countries (N = 606,627; 49.8% female).

Variables/features

Variables in PISA 2018 were measured via two group questions according to their content: noncognitive and cognitive questions. Noncognitive questions consist of the background of students and families, including their economic, social, and cultural capital, students’ attitudes towards learning, students’ metacognitive strategies, and their habits and lives in and out of school. Cognitive questions are performance tests prepared to evaluate reading, mathematics, and science achievement (OECD, 2019a).

The Student Questionnaire data file (CY07_MSU_STU_QQQ) provided by the OECD (2018b) included all the data measured by the ICT Familiarity Questionnaire, the Financial Literacy Questionnaire, the Educational Career Questionnaire, the Well-Being Questionnaire, and the Parent Questionnaire. In the current study, we merged the Student Questionnaire (SQ) data file and the Cognitive Item (CI) data file (CY07_MSU_STU_COG) according to students’ ID (CNTSTUID), school ID (CNTSCHID), and country ID (CNTRYID). We used 108 features/variables from the SQ dataset. Nine of 108 features are questions responded to by parents. Some meaningless variables for the study, such as unique national study program code (PROGN), Country of Birth National Categories—Father and Mother (COBN_F/M) were not taken into consideration. All noncognitive features and their codes in the dataset can be seen in Table 1.

Table 1.

Variables/features and codes in PISA 2018.

Questionnaire	Variable/Feature Code	Count
Student	ADAPTIVITY, AGE, ATTIMM, ATTLNACT, AWACOM, BEINGBULLIED, BELONG, BFMJ2, BMMJ1, BSMJ, COGFLEX, COMPETE, CULTPOSS, DIRINS, DISCLIMA, DISCRIM, DURECEC, EFFORT1, EFFORT2, EMOSUPS, ESCS, EUDMO, FISCED, FISCED_D, GCAWARE, GCSELFEFF, GFOFAIL, GLOBMIND, GRADE, HEDRES, HISCED, HISCED_D, HISEI, HOMEPOS, ICTRES, IMMIG, INFOJOB1, INTCULT, ISCEDL, ISCEDO, JOYREAD, LANGFATHER, LANGFRIEND, LANGMOTHER, LANGSCHMATES, LANGSIBLINGS, LMINS, MASTGOAL, METASPAM, METASUM, MISCED, MISCED_D, MMINS, OCOD3, PARED, PAREDINT, PERCOMP, PERCOOP, PERFEED, PERSPECT, REPEAT, RESILIENCE, RESPECT, SCREADCOMP, SCREADDIFF, SMINS, ST004D01 T, STIMREAD, SWBP, TEACHINT, TEACHSUP, TMINS, UNDREM, WEALTH, WORKMAST	75
ICT Familiarity	AUTICT, COMPICT, ENTUSE, HOMESCH, ICTCLASS, ICTHOME, ICTOUTSIDE, ICTSCH, INTICT, SOIAICT, USESCH	11
Financial Literacy	FCFMLRTY, FLCONFIN, FLCONICT, FLFAMILY, FLSCHOOL	5
Educational Career	CHANGE, INFOCAR, INFOJOB2, ISCEDD, SCCHANGE	5
Well-being	BODYIMA, SOCONPA, STUBMI	3
Parent	ATTIMMP, CURSUPP, EMOSUPP, GCAWAREP, INTCULTP, JOYREADP, PASCHPOL, PQSCHOOL, PRESUPP	9
Total		108

Note: Whole names of the variables can be found in Appendix A.

In the SQ dataset, many items were designed to measure latent constructs that cannot be directly observed. Significant indices were created by applying various transformations or scaling procedures to these items. These indices are called “derived variables.” Derived variables were generated using methods such as transforming or re-coding one or more questionnaire items with simple arithmetic operations, and Item Response Theory (IRT) scaling (OECD, 2009).

According to IRT scaling, for each item, person j’s responses to item i are modeled as a function of the latent construct (Q_j). With one-parameter model proposed by Rasch (1960) for dichotomous items (True-False or 1-0), the probability of person j selecting category 1 (or True) is calculated according to the equation 1 (Rasch, 1960).

P (X_{j i} = 1 | Q_{j}, β_{i}) = \frac{e^{(Q_{j -} β_{i})}}{1 + e^{(Q_{j -} β_{i})}}

(1)

where $P (X_{j i} = 1)$ is the probability of selecting category 1 (or True) on item i of person j. Q _j is the estimated latent feature of person j and β _i is the estimated difficulty of item i. For items with more than two categories, such as Likert-type items, this model has been generalized to other models, such as the “Partial Credit Model” (Masters & Wright, 1997), and the “Generalized Partial Credit Model” (Muraki, 1992). In PISA 2018, the derived variables scaled to the IRT were calculated using the Generalized Partial Credit Model (OECD, 2009).

As mentioned in the “Introduction,” we took reading performance as a dependent variable. Therefore, we used the “Reading Adaptive Performance: Core + First Stage” (RCO1S_PERF) variable when calculating information gain, the decision tree, and the random forest since these techniques required the data to be categorical. The Multistage Adaptive Design (MSAT) method for the reading performance consists of three stages: Core Stage, Stage 1, and Stage 2. In the reading literacy test, there are a total of 245 items in total, with 45 different units (5 units in the core stage, 24 units in Stage 1, and 16 units in Stage 2). Students were classified as “Low Level,” “Medium Level,” and “High Level” according to their scores at the Core plus first stage (OECD, 2012).

We also used ten plausible values (PV) while performing path analysis since the data must be continuous. PISA is a large-scale international testing application focusing on the population rather than the individual performance of students. Therefore, student proficiency levels are reported using plausible values (PV). In PISA 2018, students’ reading performance/proficiency levels were reported with 10 PVs. A common mistake when analyzing PVs is calculating the average of the PVs before analysis. Although PVs are useful in large-scale applications with large populations, they are not recommended for individual-level scores (OECD, 2009; Wu, 2005).

Data analysis procedure

Data analysis in the study can be grouped under two headings. These are data-mining analysis and structural equation modeling (SEM). There are different techniques used to analyze big data. Information gain, decision tree, and random forest methods are the most commonly used techniques in the context of machine learning. Although these are different techniques, each one is used to support and explain others’ findings respectively. For example, information gain was used to determine the most significant features, providing more information among all features in the current study. Later, decision tree classification analysis was performed using the significant features selected in the previous step to build a hierarchical tree structure as an output. Thus, this algorithm could be used to predict reading performance using the most significant features in a hierarchical model. Actually, this model becomes the basis for artificial intelligence to make decisions in future applications. After the decision tree model was built, the random forest method was used to determine the order of importance of the top features used in the analysis and to consolidate the findings. As explained in the Introduction, the main purpose of this study was to discover which noncognitive variables provided more information about reading performance with an exploratory approach. These techniques facilitated answering the main research question. In the final section, path analysis in the context of SEM was used to generate a meaningful model based on earlier exploratory findings. Thus, after determining the top features in the big dataset (without premeditating) via data-mining methods, we finally reached a statistically more robust model.

The R software (R Core Team, 2016) was used for dataset preparation. After the dataset preparation process, the information gain values of each feature (independent variables) were calculated according to the target (dependent) variable. Then we created a decision tree model using the features selected. We also calculated feature importance values via a random forest technique. Finally, we computed the predictive power of these selected variables on reading performance. The details of the data analysis procedure are shown in Figure 1.

Figure 1.

Data analysis procedure.

Information gain

The purpose of the information gain analysis was to calculate which features provide the most information gain scores according to students’ reading performance. We selected only low and high-level students to reach clearer findings. Therefore, 301,946 students from 70 countries and economies were taken for analysis. We calculated the information gain scores for 108 features. Information gain scores of 108 variables are given in Appendix A.

The merged data file was converted to CSV format and was used in the analysis. The information gain is based on the entropy formula introduced to computer science by Shannon (1948). Entropy is a numerical measure of uncertainty in data and is calculated according to equation 2:

H = - \sum_{k = 1}^{n} p_{k} \log_{2} p_{k}

(2)

where k = 1, 2, 3, . . ., m represents the m classes of the target variable. $p_{k}$ represents the proportion of samples that belong to class k. Entropy can be a value between 0 and 1, and data with higher complexity have a higher entropy value. The information gain value of a feature in a dataset is the amount of reduction in the entropy value of the target variable and is calculated according to equation 3:

I G (C, A) = H (C) - \sum_{t \in T} p (t) H (t) = H (C) - H (C | A)

(3)

where A represents a feature in the dataset, C represents the target variable, H(C) represents the entropy of the target variable, T represents the subset of the class C partitioned according to the A feature, H(C|A) represents the entropy of the target variable according to the A feature. IG(C, A) is the information gain value of the A feature according to the target variable C.

Information gain is also used as a splitting criterion in some decision tree algorithms (Quinlan, 1986, 1993). In addition, information gain is one of the methods used to determine the importance of features in filtering-based feature selection methods (Kotu & Deshpande, 2019).

The purpose of the information gain analysis in the current study is to filter the features in the dataset and continue the next analysis with features having a two standard deviation or higher information gain score. According to Prasetiyowati et al. (2021), threshold values for information gain can be found via different techniques. They proposed “threshold rate determination using the information gain value's standard deviation generated by each feature in the dataset” (p. 1). Thus, we calculated the mean (M = 0.038141173686) and standard deviation (SD = 0.03091608722) of all information gain scores for 108 variables. To be more conservative, we accepted two standard deviations [0.04 + (0.03 × 2) = 0.10] as the cut-off point, which is higher than 97.5% of all features. KNIME software (Berthold et al., 2009) was used for information gain analysis.

Decision tree

In the next step of the data-mining analysis, decision tree classification analysis was performed using the six features selected in the previous step. Many decision tree algorithms, such as ID3 (Quinlan, 1986), C4.5 (Quinlan, 1993), CART (Breiman et al., 1984) have been developed. In decision tree analysis, the dataset is divided into partitions according to the splitting criterion of the algorithm and a hierarchical tree structure is built as an output. The hierarchically top element of the tree is called the root. The root is followed by a top-down structure consisting of nodes and leaves. The bottom leaves contain the class labels. When the decision tree algorithm predicts the class of an instance, it starts at the root of the tree and moves from top to bottom according to the features of the instance, ending at the last leaf. The class label on this leaf is considered the class label for instance.

One of the important issues while building the tree structure in decision tree analysis is to determine the splitting criterion in the root and child nodes. The aim is to identify the most distinctive features in the dataset and to choose them as splitting criteria. Metrics such as information gain, entropy-based gain ratio, Gini Index, and chi-square are used as splitting criteria. Algorithms such as ID3 and C4.5 use information gain as the splitting criteria, while the CART algorithm uses the Gini Index as the splitting criteria. In the current study, the entropy-based gain ratio was used as the splitting criterion.

In decision tree analysis, a process called “pruning” is applied to reduce the complexity of the tree structure. There are some methods of performing pruning before (Takamitsu et al., 2004) and after (Fürnkranz, 1997; Mehta et al., 1995) the tree is built. These methods are called “pre-pruning” and “post-pruning.” The Minimum Description Length (MDL) method, which is a post-pruning method, was used in the decision tree analysis in this study (Mehta et al., 1995). The KNIME software (Berthold et al., 2009) was used for the analysis.

Random forest

After the decision tree model was built, the random forest (Breiman, 2001) method was also used to determine the order of importance of the six features used in the analysis. The random forest method is an ensemble learning method that works by creating a decision tree model collection from a dataset. In the random forest method, the decision tree collection is created by random feature selection called “bagging.” When predicting the class of an instance in the random forest method, each built tree model predicts the class for the instance, and the class with the most votes is returned as the prediction class. The random forest method is also recommended to rank the importance of features in a regression or classification analysis (Breiman, 2001). Scikit-learn library for Python (Pedregosa et al., 2011) was used for random forest analysis.

Model evaluation measures

The best way to evaluate the success of a classification model in data-mining analysis is to use a truth table called the confusion matrix (Kotu & Deshpande, 2019). There are four possible situations when a binary classification model tries to predict the class of an instance:

The instance where actual class is positive is predicted as positive by the model (called true positive or TP)

The instance where actual class is positive is predicted as negative by the model (called false negative or FN)

The instance where actual class is negative is predicted as negative by the model (called true negative or TN)

The instance where actual class is negative is predicted as positive by the model (called false positive or FP)

In the data-mining analysis in this study, students with “high level” reading performance are positive instances, while students with “low level” reading performance are negative instances. In a model with excellent prediction accuracy, the FP and FN numbers on the confusion matrix are equal to zero. The confusion matrix format used when evaluating a classification model in data-mining analysis is presented in Figure 2.

Figure 2.

Confusion matrix.

There are measures such as Sensitivity (TPV), Specificity (TNR), Precision (PPV), Accuracy (ACC), Matthews Correlation Coefficient (MCC) are calculated using the values on the confusion matrix to evaluate the model's success. The meaning of these measures and how they are calculated are presented in Table 2.

Table 2.

Measures used to evaluate the model.

Measure	Calculation	Definition
Accuracy (ACC)	$\frac{T P + T N}{T P + T N + F P + F N}$	Aggregate measure of classifier performance
True positive rate or sensitivity (TPR)	$\frac{T P}{T P + F N}$	Ability to select what needs to be selected
True negative rate or specificity (TNR)	$\frac{T N}{T N + F P}$	Ability to reject what needs to be rejected
Positive predictive value or precision (PPV)	$\frac{T P}{T P + F P}$	Proportion of cases found that were relevant
Matthews correlation coefficient (MCC)	$\frac{T P x T N - F P x F N}{\sqrt{(T P + F P) (T P + F N) (T N + F P) (T N + F N)}}$	The correlation of the true classes with the predicted labels

Source: Chicco et al. (2021); Kotu and Deshpande (2019).

In the decision tree and random forest analysis, 80% of the instances in the dataset (241,556 instances) were used to build the model and 20% (60,390 instances) were used to test the models.

Structural equation modeling

To examine the predictive power of independent variables on reading performance, a path analysis using structural equation modeling was performed in AMOS 16 (Arbuckle, 2007). Path analysis is a technique for evaluating a theoretical model that includes direct and possibly indirect effects between independent (exogenous) and dependent (endogenous) variables. Although path analysis and multiple regression analysis are comparable approaches, path analysis gives a stronger framework for establishing a specific theoretical model regarding the relationship between a collection of predictor and dependent variables than multiple regression. Maximum likelihood (ML) estimation was used because multivariate normality was assumed for the population distributions of the dependent variables (reading plausible values) (Kline, 2011).

In the path model, home possessions (HOMEPOS), cultural possessions at home (CULTPOSS), and ICT resources at home (ICTRES) were considered in the same category, which means SES is a latent variable. In addition, metacognition: assessing credibility (METASPAM) and metacognition: summarizing (METASUM) were taken as another latent variable named “metacognition.” We also took 10 plausible values of reading performance as a latent variable because they are not recommended for individual estimation as mentioned before. OECD experts (2009, p. 98) suggested that plausible values can be used as a latent variable in estimation of population parameters.

Results

RQ 1: Of all the variables in the PISA student and parent surveys, which variables provide the most information gain (>2 SD) about reading performance?

Information gain analysis

The information gain values of all the features in the dataset were calculated to answer the first research question. The values of 108 features vary between 0.0004 and 0.1554. The features with information gain values of 0.10 (2 SD) and higher are “cultural possessions” (0.1554), “metacognition: assessing credibility” (0.1239), “ICT resources” (0.1171), “metacognition: summarizing” (0.1131), “home possessions” (0.1118), and “liking/enjoying reading” (0.1080). The six features with the highest information gain value are presented in Table 3. Information gain values for all features can be seen in Appendix A.

RQ 2: When the PISA variables that provide the most information gain (>2 SD) about reading performance are tested with the decision tree model, with what accuracy can individuals with low and high reading levels be classified?

Table 3.

Top six features according to information gain values.

Rank	Feature Code	Description	Information Gain Value
1	CULTPOSS	Cultural possessions at home	0.1554
2	METASPAM	Metacognition: Assessing credibility	0.1239
3	ICTRES	ICT resources at home	0.1171
4	METASUM	Metacognition: Summarizing	0.1131
5	HOMEPOS	Home possessions	0.1118
6	JOYREAD	Liking/enjoying reading	0.1080

Decision tree analysis

In the decision tree analysis, a prediction model with six variables (CULTPOSS, METASPAM, ICTRES, METASUM, HOMEPOS, and JOYREAD) was created and the performance of the model was evaluated. The data revealed that the accuracy value of the model was 0.7461. The performance evaluation measures of the decision tree model are presented in Table 4.

Table 4.

Decision tree model evaluation measures.

Measure	Value
Accuracy (ACC)	0.7461
True positive rate or sensitivity (TPR)	0.7855
True negative rate or specificity (TNR)	0.6853
Positive predictive value or precision (PPV)	0.7936
Matthews correlation coefficient (MCC)	0.4696

The root of the tree structure built with the decision tree model is the HOMEPOS. After the root level, the METASPAM and METASUM features are located at the second level hierarchically. The tree structure is presented with Figure 3. The whole tree structure can be downloaded as a PDF from Appendix B.

Figure 3.

Tree structure of the model.

Random forest analysis

The random forest method was also used to determine the order of importance of six features used in the decision tree model and to verify the hierarchy in the constructed tree structure. According to the analysis, HOMEPOS is located in the first place, METASPAM is located in the second place, and METASUM is located in the third place in order of importance. It is noteworthy that the METASPAM is also located in the second place in the decision tree model and information gain analysis. Another remarkable result is that HOMEPOS, METASPAM, and METASUM are among the first five features in information gain analysis and random forest analysis, and they are located at the root and second levels of the decision tree. The importance scores of six features calculated by random forest analysis are presented in Table 5.

RQ 3: When the PISA variables that provide the most information gain (>2 SD) about reading performance are taken as independent variables, to what extent do they predict reading performance?

Table 5.

Importance scores of the six features.

Feature	Random forest feature importance rank	Information gain rank	The level in the decision tree	Mean accuracy decrease
HOMEPOS	1	5	1	0.06762052
METASPAM	2	2	2	0.04468196
METASUM	3	4	2	0.03434567
JOYREAD	4	6	less than 2	0.01300132
ICTRES	5	3	less than 2	0.00703708
CULTPOSS	6	1	less than 2	0.00642647

Path analysis

Path analysis was utilized in the context of structural equation modeling to answer the last research question. As seen in Figure 4, SES (CULTPOSS, ICTRES, HOMEPOS), metacognition (METASPAM and METASUM), and liking/enjoying reading (JOYREAD) were considered independent (predictor/exogenous) variables, and reading performance was taken as a dependent (outcome/endogenous) variable. SES, metacognition, and reading performance are latent variables, while liking/enjoying reading is an observed variable.

Figure 4.

Path analysis: Predictors of reading performance.

Single-headed arrows in the model represent the direction of the effect of the predictor variable on the outcome variable. The number related to each of the single-headed arrows is the path (standardized) coefficient. The curved two-headed arrows connecting two variables are correlation coefficients between predictor variables. Small circles illustrate errors in the prediction of endogenous variables (Kline, 2011).

RMSEA, NFI, TLI, and CFI were selected to test the model's fit (Hu & Bentler, 1999; Tabachnick & Fidell, 2007). The model fit indices showed a good fit, indicating that the sample size had sufficient power (RMSEA = .029; NFI = 0.997; TLI = 0.995; CFI = 0.997).

The path model revealed that SES, metacognition, and enjoying reading predict reading performance significantly (β = 0.16, p = .001, β = 0.67, p = .001, and β = 0.06, p = .001, respectively). In other words, when SES increases by one point, reading performance correspondingly increases by 12.30 points. When metacognition increases by one point, reading performance increases by 107.87 points. When enjoying reading increases by one point, reading performance increases by 6.46 points. In addition, these three variables explain 53.3% of the variability in reading performance. Finally, these three variables have significant correlations with each other (p < .001).

Discussion

As aforementioned, the main purpose of this study was to discover the noncognitive variables that provided more information about reading performance with an exploratory approach. Data-mining methods provided answers to the first two research questions. Thus, we could create a more comprehensive model and test it via path analysis to answer the third research question. After determining the top features in the big dataset via data-mining methods, we finally reached a statistically more robust model.

Reading performance was defined as “understanding, using, evaluating, reflecting on, and engaging with texts in order to achieve one's goals, to develop one's knowledge and potential, and to participate in society” in PISA 2018 (OECD, 2019b, p. 28).

For 99 variables in the student questionnaires and 9 variables (a total of 108 features) in the parent questionnaire, the information gain values were calculated and ranked according to the reading performance levels of the students. As a result of this ranking, it was seen that the number of variables with an information gain value of 0.10 (2 SD) and above was six. These six variables are “cultural possessions” (CULTPOSS), “Metacognition: assessing credibility” (METASPAM), “ICT Resources” (ICTRES), “Metacognition: summarizing” (METASUM), “home possessions” (HOMEPOS), and “liking/enjoying reading” (JOYREAD), respectively.

When these variables are considered according to their characteristics, it is seen that cultural possessions and ICT resources at home are in the same category as home possessions in the student questionnaire. Indeed, five of the 25 questions about home possessions belong to questions about cultural possessions (ST011Q07TA, ST011Q08TA, ST011Q09TA, ST011Q16NA, and ST012Q09NA), and six of them are belong to questions about the ICT resources (ST011Q05TA, ST011Q06TA, ST012Q05NA, ST012Q06NA, ST012Q07NA, and ST012Q08NA) (OECD, 2018a, p. 14). Therefore, it is possible to evaluate these three features as a latent socioeconomic (SES) variable. Considering the six variables, the three features providing the highest information gain about reading performance are socioeconomic and cultural variables. In addition, the decision tree model revealed that home possessions are located at the root level, which is the top level of the tree hierarchically, so it is the most distinctive feature in the model. The path model also showed that when SES increased by one point, the reading performance increased by 12.30 points. These findings are consistent with those of Güven (2019), Gamazo and Martínez-Abad (2020), and Kılıç-Depren and Depren (2021). The meta-analysis study conducted by Çiftçi and Cin (2017) also showed that socioeconomic status has a high level of influence on predicting student achievement.

Socioeconomic status is not only the family's property or cultural assets, but also a reflection of the family's cultural heritage and worldview. This worldview directly affects parental attitudes and the child's view of education (Kotchick & Forehand, 2002). In the study of Sarsour et al. (2011) examining the relationship between the SES of the family and the child's neurocognitive development, they found that SES has an effect on the child's higher-order executive functions. In particular, these executive functions are related to inhibitory control such as blocking distracting stimuli and cognitive flexibility, which is the ability to adjust cognitively to unexpected situations. According to Sarsour et al. (2011), these abilities are also important for daily functioning at home, in school, and in critical social settings. Additionally, they reported that the home environment mediated the relationship between parents’ SES and children's executive functions. This result is also important for understanding the association between metacognitive skills and reading performance, which is another significant finding of our study and discussed below. Indeed, our theoretical model (SEM) indicated that there is a significant relationship between metacognitive skills and SES.

Although SES is not the only variable affecting academic achievement (Caldwell & Ginther, 1996; Milne & Plourde, 2006), it is one of the necessary prerequisites for success (Koyuncu & Fırat, 2020; Reardon, 2016). As a matter of fact, the availability of a child's own room, work desk, supporting materials (e.g., books, dictionaries, educational software), and technical support (e.g., ruler, miter, computer) are the basic elements for effective study (Filiz & Öz, 2019).

Two other remarkable features are metacognitive skills: assessing credibility (2nd rank) and summarizing (4th rank). Similarly, the decision tree model showed that both metacognitive skills are located at the second level of the tree hierarchically. At the same time, these features are positioned in the second and third order according to the random forest analysis. Path analysis also indicated that metacognitive skills are the best variable contributing to the model. When metacognitive skills increased by one point, the reading performance increased by 107.87 points.

Metacognition is defined as “an individual's ability to think about and control his or her reading and comprehension strategies” in PISA 2018 (OECD, 2019b, p. 52). Artelt and Schneider (2015) acknowledged that there is a limit to evaluating metacognitive strategies based on self-report in PISA, but they stated that self-report is a valid indicator of habitual strategic behavior.

Assessing quality and credibility is “evaluating the information in a piece of text: whether the information is valid, up-to-date, accurate, and/or unbiased” (OECD, 2019b, p. 35). Assessing credibility can be considered a metacognitive skill related to critical thinking as well (Hartman, 2001; OECD, 2020). The term “assessing credibility” fits the metacognitive function described by Pressley and Afflerbach (1995) as “constructively responsive reading.” According to Van Kraayenoord (2010), constructively responsive readers read with a purpose and actively create meaning from text. Indeed, the question asked in PISA 2018 (ST166) aimed to measure this skill through a vignette in the online environment (OECD, 2018a, p. 13). Since the question in PISA 2018 includes evaluating the credibility of information in the online environment, it is similar to the study of Van Zyl et al. (2020). Van Zyl et al. (2020) emphasized that assessing the credibility of information in the online environment can only be realized with critical thinking skills.

The skill of summarizing includes not only understanding what you have read, but also being able to briefly and concisely explain the information obtained in one's own words. The competence of the individual in this area will naturally affect reading performance positively. Good readers utilize several methods before, during, and after reading, and typically use them in a coordinated manner. Predicting forthcoming text content before reading, utilizing questions, forming mental images while reading, and summarizing after reading are some of the methods used (Pressley & Afflerbach, 1995; Van Kraayenoord, 2010).

There are many studies showing the direct relationship between metacognitive and reading skills (Hendricks, 2013; Johnson et al., 2010; Van Kraayenoord, 2010). In a recent study conducted by Kılıç-Depren and Depren (2021), it was determined that metacognitive skills are the most important variable affecting reading performance. Artelt and Schneider (2015) and Zhou et al. (2020) showed that metacognitive skills are significantly related to reading competence in PISA 2009. Artelt et al. (2001) also stated that metacognitive skills are the best variable for predicting reading performance when the SES is controlled.

Information gain analysis indicated that liking/enjoying reading is the sixth variable providing information about reading performance. In the decision tree, this feature is located at the lower levels of the tree structure. In the random forest analysis, it is in the fourth place in importance. Similarly, in the path analysis, enjoying reading contributes the least to the model. When enjoying reading increased by one point, the reading performance increased by 6.46 points.

Liking/enjoying reading (JOYREAD) is a question (ST160) asked with five items in PISA 2018 (OECD, 2018a, p. 17). It is also noteworthy that, among more than 100 variables, enjoying reading is one of the variables providing the most information gain on reading performance. The presence of this variable after SES and metacognitive skills is important in terms of showing the most important affective factor in reading performance.

There are many internal and external motivation sources for reading behavior. As there are external motivation sources such as grades, competitions, and social context, there are also intrinsic motivation sources such as curiosity, efficacy, and enjoyment (Hebbecker et al., 2019; Schiefele et al., 2012). As a result of the systematic review study by Schiefele et al. (2012), in which they examined the findings of 20 years of studies on reading motivation and reading competence, it was seen that reading for enjoyment is the most important variable having a relationship with reading competence. Similar to our study, Park (2011), Schwabe et al. (2015), Cheema (2018), and Kavanagh (2019) also found that enjoyment of reading (intrinsic reading motivation) significantly predicted reading achievement.

When 99 features based on students’ perceptions and 9 features obtained from the parent questionnaire were analyzed as independent variables, it was seen that SES, metacognitive skills, and liking/enjoying reading were major variables predicting reading performance. It is noteworthy that the features measured by the parent questionnaire were in the last place (91 and later) according to the calculated information gain values.

Finally, SES, metacognitive skills, and enjoying reading explain 53.3% of the variability in reading performance. This rate can be considered high in the context of social sciences. In this respect, our study also sheds light on future research. It is also remarkable that the decision tree model created with only these six variables has a 74.61% accuracy value in estimating the reading performance. This means that the decision tree model with the six features can be used to predict the reading performance of students in artificial intelligence algorithms.

Limitations of the study

This study has some limitations for methodological reasons. First of all, the dataset is provided by the OECD (2018b) as open access. This dataset, which is used by many researchers, also constitutes the dataset for our study. In this respect, the study is considered a secondary data analysis. However, the research questions we have addressed and the analysis techniques we have used make this study different from other studies.

In our study, only the noncognitive features responded to by the students and the parents were included in the analysis as independent variables. The features obtained from the teacher questionnaire were not included in the analysis. The main reason for this was to evaluate the students’ self-perceived characteristics and their parents’ attitudes (in the microsystem).

Finally, the dependent variable of this study is only general reading performance. In PISA 2018, besides different reading skills at a detailed level, competencies in mathematics and science were also measured. However, such a path was followed since reading competence was the main subject in PISA 2018 and to narrow the scope of the study.

Suggestions for future studies

We propose that research on large educational datasets consider all variables in the dataset rather than just a few. Although some theoretical models and pre-designed variables facilitate the research process, they also cause biased approaches and results. As a relatively new approach to educational data, data-mining strategies can provide more opportunities to see the big picture. In future studies, it is thought that determining the variables providing the most information gain in mathematics and science competence with the same methods is important to see the big picture. In addition, it is thought that examining the relations between the features we found in this study and different variables will provide more information to educational scientists and politicians.

Footnotes

Acknowledgements

We would like to thank Dr. Jack Cummings for proofreading of the manuscript.

Authors' Note

Osman Tolga Aricak is also affiliated at Department of Primary Education at Bogazici University, Istanbul, Turkey.

Declaration of conflicting interests

The authors declared no potential conflicts of interest with respect to the research, authorship, and/or publication of this article.

Ethics statement

Ethical review and IRB approval were not required because the study is based on the public databases of the PISA 2018 assessment. Data collection for PISA studies is the responsibility of the governments of the participating countries.

Funding

The authors received no financial support for the research, authorship, and/or publication of this article.

ORCID iD

Osman Tolga Aricak

Appendix

Appendix A.

Information gain scores of 108 variables

Rank	Feature code	Description	Information gain value
1	CULTPOSS	Cultural possessions at home (WLE)	0.155399649188
2	METASPAM	Meta-cognition: Assess credibility	0.123902460153
3	ICTRES	ICT resources (WLE)	0.117125626705
4	METASUM	Meta-cognition: Summarizing	0.113057507539
5	HOMEPOS	Home possessions (WLE)	0.111763858037
6	JOYREAD	Joy/Like reading (WLE)	0.108001036324
7	ESCS	Index of economic, social and cultural status	0.093091451295
8	TMINS	Learning time (minutes per week)—In total	0.089851125524
9	BMMJ1	ISEI of mother	0.088895263481
10	HISEI	Index highest parental occupational status	0.088508331643
11	UNDREM	Meta-cognition: Understanding and remembering	0.086823021257
12	WEALTH	Family wealth (WLE)	0.083049803432
13	SCREADDIFF	Self-concept of reading: Perception of difficulty (WLE)	0.079855398884
14	HEDRES	Home educational resources (WLE)	0.076968710450
15	GCAWARE	Student's awareness of global issues (WLE)	0.069194130227
16	OCOD3	ISCO-08 Occupation code—Self	0.067989645645
17	INTICT	Interest in ICT (WLE)	0.064279145259
18	BFMJ2	ISEI of father	0.063950724175
19	MMINS	Learning time (minutes per week)—	0.063300355981
20	RESILIENCE	Resilience (WLE)	0.059829061347
21	WORKMAST	Work mastery (WLE)	0.058291749537
22	LMINS	Learning time (minutes per week)—<test language>	0.057695494684
23	BELONG	Subjective well-being: Sense of belonging to school	0.057625505905
24	SMINS	Learning time (minutes per week)—	0.057371857553
25	GCSELFEFF	Self-efficacy regarding global issues (WLE)	0.055219615194
26	BSMJ	Student's expected occupational status (SEI)	0.052911889526
27	COMPICT	Perceived ICT competence (WLE)	0.052740315511
28	BEINGBULLIED	Student's experience of being bullied (WLE)	0.051444810823
29	ENTUSE	ICT use outside of school (leisure) (WLE)	0.050162215683
30	HOMESCH	Use of ICT outside of school (for school work activities) (WLE)	0.049729083027
31	DIRINS	Teacher-directed instruction (WLE)	0.049255263999
32	MASTGOAL	Mastery goal orientation (WLE)	0.046605196575
33	EFFORT2	How much effort would you have invested?	0.045475854056
34	DISCRIM	Discriminating school climate (WLE)	0.045303606335
35	ICTHOME	ICT available at home	0.044751724644
36	SCREADCOMP	Self-concept of reading: Perception of competence (WLE)	0.043727867098
37	ICTSCH	ICT available at school	0.043553649520
38	USESCH	Use of ICT at school in general (WLE)	0.042539185725
39	ATTLNACT	Attitude towards school: Learning activities (WLE)	0.042046686210
40	ATTIMM	Student's attitudes towards immigrants (WLE)	0.041943154524
41	AUTICT	Perceived autonomy related to ICT use (WLE)	0.041386967845
42	AWACOM	Awareness of intercultural communication (WLE)	0.041367900404
43	STIMREAD	Teacher's stimulation of reading engagement perceived by student (WLE)	0.038109831265
44	RESPECT	Respect for people from other cultures (WLE)	0.037305002615
45	GRADE	Grade compared to modal grade in country	0.036241075337
46	SOIAICT	ICT as a topic in social interaction (WLE)	0.036130216452
47	ICTOUTSIDE	Subject-related ICT use outside of lessons (WLE)	0.034710226041
48	EFFORT1	How much effort did you put into this test?	0.033904814714
49	PARED	Index highest parental education in years of schooling	0.033390519681
50	GLOBMIND	Global-mindedness (WLE)	0.033142677476
51	COMPETE	Competitiveness (WLE)	0.032418503698
52	ICTCLASS	Subject-related ICT use during lessons (WLE)	0.032315217127
53	REPEAT	Grade repetition	0.031664088301
54	MISCED_D	Mother's education—Alternate definition (ISCED)	0.031087258262
55	PERSPECT	Perspective-taking (WLE)	0.030139209579
56	HISCED_D	Highest education of parents (ISCED)	0.029228546305
57	PAREDINT	Index highest parental education (international years of schooling scale)	0.029174175658
58	ADAPTIVITY	Adaptation of instruction (WLE)	0.028043667953
59	INTCULT	Student's interest in learning about other cultures (WLE)	0.027883918596
60	FISCED_D	Father's education—alternate definition (ISCED)	0.027065242520
61	PERCOMP	Perception of competitiveness at school (WLE)	0.026322638591
62	TEACHSUP	Teacher support in test language lessons (WLE)	0.025875187693
63	DISCLIMA	Disciplinary climate in test-language lessons (WLE)	0.025760908790
64	MISCED	Mother's education (ISCED)	0.025465822277
65	DURECEC	Duration in early childhood education and care	0.024900483638
66	EMOSUPS	Parents’ emotional support (WLE)	0.024620602155
67	HISCED	Highest education of parents (ISCED)	0.024620602155
68	COGFLEX	Cognitive flexibility/adaptability (WLE)	0.024553142517
69	EUDMO	Eudaemonia: Meaning in life (WLE)	0.024506295691
70	FCFMLRTY	Familiarity with concepts of finance (Sum)	0.024135260275
71	INFOCAR	Information about careers (WLE)	0.024115171843
72	TEACHINT	Perceived teacher's interest (WLE)	0.023118951014
73	GFOFAIL	General fear of failure (WLE)	0.022232099820
74	ISCEDO	ISCED orientation	0.022137720893
75	FISCED	Father's education (ISCED)	0.021511013845
76	FLCONFIN	Confidence about financial matters (WLE)	0.021370613160
77	PERCOOP	Perception of cooperation at school (WLE)	0.021293239750
78	FLFAMILY	Parental involvement in matters of financial literacy (WLE)	0.018708469291
79	SWBP	Subjective well-being: Positive affect (WLE)	0.017288920967
80	INFOJOB2	Information about the labour market provided outside of school (WLE)	0.017213962729
81	FLCONICT	Confidence about financial matters using digital devices (WLE)	0.016952390610
82	FLSCHOOL	Financial education in school lessons (WLE)	0.015377520091
83	LANGSCHMATES	Language spoken with their school mates for students who do not speak the test language at home	0.014828724403
84	LANGFRIEND	Language spoken with their best friend for students who do not speak the test language at home	0.014138859678
85	INFOJOB1	Information about the labour market provided by the school (WLE)	0.013960765986
86	LANGSIBLINGS	Language spoken with their brother(s) and/or sister(s) for students who do not speak the test language at home	0.012798065229
87	LANGMOTHER	Language spoken with their mother for students who do not speak the test language at home	0.011998892179
88	LANGFATHER	Language spoken with their father for students who do not speak the test language at home	0.010693657576
89	PERFEED	Perceived feedback (WLE)	0.010309013928
90	CHANGE	Number of changes in educational biography (Sum)	0.009090197733
91	JOYREADP	Parents’ enjoyment of reading (WLE)	0.008689155399
92	GCAWAREP	Parents’ awareness of global issues (WLE)	0.008444286892
93	ISCEDL	ISCED level	0.008405156100
94	SCCHANGE	Number of school changes	0.007972945309
95	ST004D01T	Gender	0.006389637760
96	ISCEDD	ISCED designation	0.005486783878
97	PASCHPOL	School policies for parental involvement (WLE)	0.005110396480
98	CURSUPP	Current parental support for learning at home (WLE)	0.004152939545
99	ATTIMMP	Parents’ attitudes towards immigrants (WLE)	0.003513754487
100	PQSCHOOL	Parents’ perceived school quality (WLE)	0.003240439638
101	EMOSUPP	Parents’ emotional support (WLE)	0.002278687828
102	IMMIG	Index immigration status	0.002274806757
103	INTCULTP	Parents’ interest in learning about other cultures (WLE)	0.001835338586
104	SOCONPA	Social connections: Parents (WLE)	0.000935000188
105	AGE	Age	0.000781715641
106	STUBMI	Body mass index of student	0.000762313214
107	PRESUPP	Previous parental support for learning at home (WLE)	0.000719721870
108	BODYIMA	Body image (WLE)	0.000414391539

Appendix B. Link of whole decision tree structure

Figure_3_SuppInfo.pdf

References

Arbuckle

J. L.

(2007). Amos™ 16.0 user’s guide. SPSS.

Artelt

Schiefele

Schneider

(2001). Predictors of reading literacy. European Journal of Psychology of Education, 16(3), 363–383. https://doi.org/10.1007/BF03173188

Artelt

Schneider

(2015). Cross-country generalizability of the role of metacognitive knowledge in students’ strategy use and reading competence. Teachers College Record, 117(1), 1–32. https://www.tcrecord.org/books/Content.asp?ContentID=17695 https://doi.org/10.1177/016146811511700104

Berger

Mackenzie

Holmes

(2020). Positive attitudes towards mathematics and science are mutually beneficial for student achievement: A latent profile analysis of TIMSS 2015. The Australian Educational Researcher, 47, 409–444. https://doi.org/10.1007/s13384-020-00379-8

Berthold

M. R.

Cebron

Dill

Gabriel

T. R.

Kötter

Meinl

Wiswedel

(2009). KNIME—the Konstanz information miner: Version 2.0 and beyond. AcM SIGKDD Explorations Newsletter, 11(1), 26–31. https://doi.org/10.1145/1656274.1656280

Breiman

(2001). Random forests. Machine Learning, 45(1), 5–32. https://doi.org/10.1023/A:1010933404324

Breiman

Friedman

J. H.

Olshen

R. A.

Stone

C. J.

(1984). Classification and regression trees (1st ed.). Routledge. https://doi.org/10.1201/9781315139470

Caldwell

G. P.

Ginther

D. W.

(1996). Differences in learning styles of low socioeconomic status for low and high achievers. Education, 117(1), 141–148. https://link.gale.com/apps/doc/A18960236/AONE?u=anon∼892a03b1&sid=googleScholar&xid=55e878cf

Cheema

J. R.

(2018). Adolescents’ enjoyment of reading as a predictor of reading achievement: New evidence from a cross-country survey. Journal of Research in Reading, 41(S1), S149–S162. https://doi.org/10.1111/1467-9817.12257

10.

Chicco

Tötsch

Jurman

(2021). The Matthews correlation coefficient (MCC) is more reliable than balanced accuracy, bookmaker informedness, and markedness in two-class confusion matrix evaluation. BioData Mining, 14(1), 1–22. https://doi.org/10.1186/s13040-021-00244-z

11.

Çiftçi

ŞK

Cin

F. M.

(2017). The effect of socioeconomic status on students’ achievement. In Karadag

(Ed.), The factors effecting student achievement (pp. 171–181). Springer. https://doi.org/10.1007/978-3-319-56083-0_10

12.

Davoodi

H. R.

Abed

G. T.

(2003). Challenges of growth and globalization in the Middle East and North Africa. International Monetary Fund. https://www.imf.org/external/pubs/ft/med/2003/eng/abed.htm

13.

Duckworth

A. L.

Yeager

D. S.

(2015). Measurement matters: Assessing personal qualities other than cognitive ability for educational purposes. Educational Researcher, 44(4), 237–251. https://doi.org/10.3102/0013189X15584327

14.

Filiz

Öz

(2019). Finding the best algorithms and effective factors in classification of Turkish science student success. Journal of Baltic Science Education, 18(2), 239–253. https://www.ceeol.com/search/article-detail?id=950376 https://doi.org/10.33225/jbse/19.18.239

15.

Fürnkranz

(1997). Pruning algorithms for rule learning. Machine Learning, 27(2), 139–172. https://doi.org/10.1023/A:1007329424533

16.

Füssl

K. H.

Kubina

(1985). Educational reform between politics and pedagogics—The development of education in Berlin after World War II. History of Education Quarterly, 25(1-2), 133–153. https://doi.org/10.2307/368894

17.

Gamazo

Martínez-Abad

(2020). An exploration of factors linked to academic performance in PISA 2018 through data mining techniques. Frontiers in Psychology, 11, 1–17. https://doi.org/10.3389/fpsyg.2020.575167

18.

Grimm

K. J.

(2008). Longitudinal associations between reading and mathematics achievement. Developmental Neuropsychology, 33(3), 410–426. https://doi.org/10.1080/87565640801982486

19.

Güven

(2019). The effect of socioeconomic status on student achievement: A TIMSS study. In Tokcan

Altunçekiç

(Eds.), Contemporary approaches in education and social science (pp. 89–99). Old Publishing.

20.

Gylfason

(2001). Natural resources, education, and economic development. European Economic Review, 45(4-6), 847–859. https://doi.org/10.1016/S0014-2921(01)00127-1

21.

Hanushek

E. A.

Wößmann

(2010). Education and economic growth. In Peterson

Baker

McGaw

(Eds.), International encyclopedia of education (pp. 245–252). Elsevier.

22.

Hartman

H. J.

(Ed.) (2001). Metacognition in learning and instruction: Theory, research and practice (Vol. 19). Springer Science & Business Media. https://doi.org/10.1007/978-94-017-2243-8

23.

Hebbecker

Förster

Souvignier

(2019). Reciprocal effects between reading achievement and intrinsic and extrinsic reading motivation. Scientific Studies of Reading, 23(5), 419–436. https://doi.org/10.1080/10888438.2019.1598413

24.

Hendricks

(2013). Reading and test taking in college English as a second language students [Doctoral dissertation]. Syracuse University. https://surface.syr.edu/psy_etd/181

25.

Hidayati

Inderawati

Loeneto

(2020). The correlations among critical thinking skills, critical reading skills and reading comprehension. English Review: Journal of English Education, 9(1), 69–80. https://doi.org/10.25134/erjee.v9i1.3780

26.

L. T.

Bentler

P. M.

(1999). Cutoff criteria for fit indexes in covariance structure analysis: Conventional criteria versus new alternatives. Structural Equation Modeling: A Multidisciplinary Journal, 6(1), 1–55. https://doi.org/10.1080/10705519909540118

27.

IEA (n.d.). Retrieved June 18, 2021, from https://timssandpirls.bc.edu/

28.

Johnson

T. E.

Archibald

T. N.

Tenenbaum

(2010). Individual and team annotation effects on students’ reading comprehension, critical thinking, and meta-cognitive skills. Computers in Human Behavior, 26(6), 1496–1507. https://doi.org/10.1016/j.chb.2010.05.014

29.

Kavanagh

(2019). Relations between children’s reading motivation, activity and performance at the end of primary school. Journal of Research in Reading, 42(3-4), 562–582. https://doi.org/10.1111/1467-9817.12284

30.

Kirikkaleli

Ertugrul

H. M.

Sari

Ozun

Kiral

(2021). Quality of education and technological readiness: Bootstrap panel causality analysis for northern European countries. Scandinavian Journal of Educational Research, 65(2), 276–287. https://doi.org/10.1080/00313831.2019.1705892

31.

Kılıç-Depren

Depren

(2021). Cross-cultural comparisons of the factors influencing the high reading achievement in Turkey and China: Evidence from PISA 2018. The Asia-Pacific Education Researcher, 31, 427–437. https://doi.org/10.1007/s40299-021-00584-8

32.

Kline

R. B.

(2011). Principles and practice of structural equation modeling. Guilford publications.

33.

Kotchick

B. A.

Forehand

(2002). Putting parenting in perspective: A discussion of the contextual factors that shape parenting practices. Journal of Child and Family Studies, 11(3), 255–269. https://doi.org/10.1023/A:1016863921662

34.

Kotu

Deshpande

(2019). Data science: Concepts and practice (2nd ed.). Elsevier. https://doi.org/10.1016/B978-0-12-814761-0.00014-9

35.

Koyuncu

Fırat

(2020). Investigating reading literacy in PISA 2018 assessment. International Electronic Journal of Elementary Education, 13(2), 263–275. https://10.26822/iejee.2021.189

36.

Larwin

K. H.

(2010). Reading is fundamental in predicting math achievement in 10th graders? International Electronic Journal of Mathematics Education, 5(3), 131–145. https://doi.org/10.29333/iejme/254

37.

Lee

Stankov

(2018). Non-cognitive predictors of academic achievement: Evidence from TIMSS and PISA. Learning and Individual Differences, 65, 50–64. https://doi.org/10.1016/j.lindif.2018.05.009

38.

Masters

G. N.

Wright

B. D.

(1997). The partial credit model. In van der Linden

W. J.

Hambleton

R. K.

(Eds.), Handbook of modern item response theory (pp. 101–121). Springer. https://doi.org/10.1007/978-1-4757-2691-6_6

39.

McClelland

D. C.

(1966). Does education accelerate economic growth? Economic Development and Cultural Change, 14(3), 257–278. https://doi.org/10.1086/450163

40.

Mehta

Rissanen

Agrawal

(1995). MDL-based decision tree pruning. In Knowledge Discovery in Databases 1995 (KDD-95) Proceedings (Vol. 21, No. 2, pp. 216–221). https://www.aaai.org/Papers/KDD/1995/KDD95-025.pdf

41.

Milne

Plourde

L. A.

(2006). Factors of a low-SES household: What aids academic achievement? Journal of Instructional Psychology, 33(3), 183–193. https://web.b.ebscohost.com/ehost/pdfviewer/pdfviewer?vid=0&sid=1ce733a0-c3dc-4ebb-bd41-81b53d9700f5%40pdc-v-sessmgr01

42.

Morito

(1955). Educational reform and its problems in post-war Japan. International Review of Education, 1(3), 338–351. https://doi.org/10.1007/BF01421722

43.

Muraki

(1992). A generalized partial credit model: Application of an EM algorithm. ETS Research Report Series, 1992(1), 1–33. https://onlinelibrary.wiley.com/doi/pdfdirect/10.1002/j.2333-8504.1992.tb01436.x https://doi.org/10.1002/j.2333-8504.1992.tb01442.x

44.

OECD. (2018a). Technical report 2018. Retrieved June 2, 2021, from https://www.oecd.org/pisa/data/pisa2018technicalreport/PISA2018_Technical-Report-Chapter-16-Background-Questionnaires.pdf

45.

OECD. (2018b). Retrieved July 6, 2021, from https://www.oecd.org/pisa/data/2018database/

46.

OECD. (n.d.). Retrieved June 21, 2021, from https://www.oecd.org/pisa

47.

OECD. (2009). PISA data analysis manual: SPSS, second edition. OECD Publishing. https://doi.org/10.1787/9789264056275-9-en

48.

OECD. (2012). PISA 2009 technical report. OECD Publishing. https://doi.org/10.1787/9789264167872-3-en

49.

OECD. (2020). PISA 2018 results (volume VI): Are students ready to thrive in an interconnected world? PISA, OECD Publishing. https://doi.org/10.1787/d5f68679-en

50.

OECD. (2019b). PISA 2018 assessment and analytical framework. PISA, OECD Publishing. https://doi.org/10.1787/b25efab8-en

51.

OECD. (2019a). PISA 2018 results (volume I): What students know and can do. PISA, OECD Publishing. https://doi.org/10.1787/5f07c754-en

52.

Ozkaya

Timor

Erdin

(2021). Science, technology and innovation policy indicators and comparisons of countries through a hybrid model of data mining and MCDM methods. Sustainability, 13(2), 1–49. https://doi.org/10.3390/su13020694

53.

Park

(2011). How motivational constructs interact to predict elementary students’ reading performance: Examples from attitudes and self-concept in reading. Learning and Individual Differences, 21(4), 347–358. https://doi.org/10.1016/j.lindif.2011.02.009

54.

Pedregosa

Varoquaux

Gramfort

Michel

Thirion

Grisel

Blondel

Müller

Nothman

Louppe

Prettenhofer

Weiss

Dubourg

Vanderplas

Passos

Cournapeau

Brucher

Perrot

Duchesnay

(2011). Scikit-learn: Machine learning in Python. The Journal of Machine Learning Research, 12, 2825–2830. https://www.jmlr.org/papers/volume12/pedregosa11a/pedregosa11a.pdf?source=post_page

55.

Powell

(2007). How Sputnik changed US education. Harvard Gazette, 11 . Retrieved June 17, 2021, from https://news.harvard.edu/gazette/story/2007/10/how-sputnik-changed-u-s-education/

56.

Prasetiyowati

M. I.

Maulidevi

N. U.

Surendro

(2021). Determining threshold value on information gain feature selection to increase speed and prediction accuracy of random forest. Journal of Big Data, 8(1), 1–22. https://doi.org/10.1186/s40537-021-00472-4

57.

Pressley

Afflerbach

(1995). Verbal protocols of reading: The nature of constructively responsive reading. Erlbaum. https://doi.org/10.4324/9780203052938

58.

QS. (n.d.). Retrieved June 18, 2021, from https://www.qs.com

59.

Quinlan

J. R.

(1986). Induction of decision trees. Machine Learning, 1(1), 81–106. https://doi.org/10.1007/bf00116251

60.

Quinlan

J. R.

(1993). C4.5: Programs for machine learning. Morgan Kaufmann Publishers.

61.

Rasch

(1960). Probabilistic model for some intelligence and achievement tests. Danish Institute for Educational Research.

62.

R Core Team. (2016). R: A language and environment for statistical computing. R Foundation for Statistical Computing, Vienna, Austria. https://www.R-project.org/

63.

Reardon

S. F.

(2016). School district socioeconomic status, race, and academic achievement. Stanford Center for Educational Policy Analysis. Retrieved June 26, 2021, from https://cepa.stanford.edu/sites/default/files/reardon%20district%20ses%20and%20achievement%20discussion%20draft%20april2016.pdf

64.

Rolfe

(2021). Tailoring a measurement model of socioeconomic status: Applying the alignment optimization method to 15 years of PISA. International Journal of Educational Research, 106, 1–12. https://doi.org/10.1016/j.ijer.2020.101723

65.

Säälik

Nissinen

Malin

(2015). Learning strategies explaining differences in reading proficiency: Findings of Nordic and Baltic countries in PISA 2009. Learning and Individual Differences, 42, 36–43. https://doi.org/10.1016/j.lindif.2015.08.025

66.

Sanders

D. P.

Barth

P. S.

(1968). Chapter II: Education and economic development. Review of Educational Research, 38(3), 213–230. https://doi.org/10.3102/00346543038003213

67.

Sarsour

Sheridan

Jutte

Nuru-Jeter

Hinshaw

Boyce

W. T.

(2011). Family socioeconomic status and child executive functions: The roles of language, home environment, and single parenthood. Journal of the International Neuropsychological Society, 17(1), 120–132. https://doi.org/10.1017/S1355617710001335

68.

Sauer

Zagler

(2014). (In) equality in education and economic development. Review of Income and Wealth, 60, S353–S379. https://doi.org/10.1111/roiw.12142

69.

Schiefele

Schaffner

Möller

Wigfield

(2012). Dimensions of reading motivation and their relation to reading behavior and competence. Reading Research Quarterly, 47(4), 427–463. https://doi.org/10.1002/RRQ.030

70.

Schlatter

Molenaar

Lazonder

A. W.

(2020). Individual differences in children’s development of scientific reasoning through inquiry-based instruction: Who needs additional guidance? Frontiers in Psychology, 11, 1–14. https://doi.org/10.3389/fpsyg.2020.00904

71.

Schwabe

McElvany

Trendtel

(2015). The school age gender gap in reading achievement: Examining the influences of item format and intrinsic reading motivation. Reading Research Quarterly, 50(2), 219–232. https://doi.org/10.1002/rrq.92

72.

Shannon

C. E.

(1948). A mathematical theory of communication. Bell System Technical Journal, 27(3), 379–423. https://doi.org/10.1002/j.1538-7305.1948.tb01338.x

73.

Stankov

Lee

(2014). Quest for the best non-cognitive predictor of academic achievement. Educational Psychology, 34(1), 1–8. https://doi.org/10.1080/01443410.2013.858908

74.

Swensson

(2017). How similar are the PISA and TIMSS studies? Retrieved June 18, 2021, from https://ioelondonblog.wordpress.com/2017/12/04/how-similar-are-the-pisa-and-timss-studies/

75.

Tabachnick

B. G.

Fidell

L. S.

(2007). Using multivariate statistics (5th ed.). Allyn and Bacon.

76.

Takamitsu

Miura

Shioya

(2004, August). Pre-pruning decision trees by local association rules. In Yang

Z. R.

Everson

Yin

(Eds.), International conference on intelligent data engineering and automated learning (pp. 148–151). Springer. https://link.springer.com/content/pdf/10.1007%2Fb99975.pdf

77.

The World University Rankings. (n.d.). Retrieved June 18, 2021, from https://www.timeshighereducation.com

78.

Uljens

(2007). Education and societal change in the global age. In Jakku-Sihvonen

Niemi

(Eds.), Education as a societal contributor (pp. 23–51). Peter Lang. http://www.vasa.abo.fi/users/muljens/pdf/Education.pdf

79.

Van Kraayenoord

C. E.

(2010). The role of metacognition in reading comprehension. In Focal points of the research and development of pedagogically-psychological perspectives (Eds Hans-Peter Trolldenier, Wolfgang Lenhard, and Peter Marx), Gottingen: Hogrefe (pp. 277–302). https://www.researchgate.net/profile/Christina-Kraayenoord/publication/46401318_The_role_of_metacognition_in_reading_comprehension/links/0deec51e455db7bef7000000/The-role-of-metacognition-in-reading-comprehension.pdf

80.

Van Zyl

Turpin

Matthee

(2020). How can critical thinking be used to assess the credibility of online information? Responsible Desgn, Implementation and Use of Information and Communication Technology, 12067, 199–210. https://doi.org/10.1007%2F978-3-030-45002-1_17

81.

Wang

C. L.

Liou

P. Y.

(2018). Patterns of motivational beliefs in the science learning of total, high-, and low-achieving students: Evidence of Taiwanese TIMSS 2011 data. International Journal of Science and Mathematics Education, 16(4), 603–618. https://doi.org/10.1007/s10763-017-9797-3

82.

Chang

H. H.

Kong

Zhang

(2020). International comparative study on PISA mathematics achievement test based on cognitive diagnostic models. Frontiers in Psychology, 11, 1–13. https://doi.org/10.3389/fpsyg.2020.02230

83.

Zhang

Chang

H. H.

(2021). A comparative study on cognitive diagnostic assessment of mathematical key competencies and learning trajectories. Current Psychology, 41, 7854–7866. https://doi.org/10.1007/s12144-020-01230-0

84.

(2005). Generalized linear models in family studies. Journal of Marriage and Family, 67(4), 1029–1047. https://doi.org/10.1111/j.1741-3737.2005.00192.x

85.

Zhou

Lafontaine

(2020). Cross-cultural comparability and validity of metacognitive knowledge in reading in PISA 2009: A comparison of two scoring methods. Assessment in Education: Principles, Policy & Practice, 27(6), 635–654. https://doi.org/10.1080/0969594X.2020.1828820

Which noncognitive features provide more information about reading performance? A data-mining approach to big educational data

Abstract

Keywords

Introduction

Background of the study

Why is reading performance taken as the dependent variable?

Why was PISA data chosen instead of TIMSS?

Research questions

Method

Participants

Variables/features

Data analysis procedure

Information gain

Decision tree

Random forest

Model evaluation measures

Structural equation modeling

Results

Information gain analysis

Decision tree analysis

Random forest analysis

Path analysis

Discussion

Limitations of the study

Suggestions for future studies

Footnotes

Acknowledgements

Authors' Note

Declaration of conflicting interests

Ethics statement

Funding

ORCID iD

Appendix

Appendix B. Link of whole decision tree structure

References