This proposal examines which school-level conditions are most strongly linked to mathematics success in Türkiye in PISA 2022. It uses the PISA 2022 School Questionnaire and supervised machine learning to model school success from leadership practices, school climate, resources, and family–school relations.

Research question

Which school-level factors best predict whether a school in Türkiye is “successful” in mathematics in PISA 2022, and how well can a supervised model classify school success using school-level indicators?

Objective

The objective is to build an interpretable predictive model and to identify actionable predictors that school leaders and policymakers can strengthen, while keeping sight of schools’ historical and social mission, as emphasized by Guevara-Reyes and his colleagues (2025).

Conceptual and theoretical framework
The framework combines leadership, school climate, and structural equity perspectives:

·       Leadership and administration. School leaders shape learning conditions by setting goals, managing resources, and building a climate conducive to learning. Leadership influences student outcomes directly and indirectly through teacher quality, instructional coherence, and equitable learning climates (Leithwood & Jantzi, 2006; Robinson et al., 2008). In mathematics, instructional leadership, distributed leadership, and data-driven decision making are positively related to student performance (Hallinger, 2011). In developing countries, leaders often work under limited resources, unclear policy mandates, and fragmented teacher training systems, so leadership capacity matters for linking policy and classroom practices to improve attainment (Bush, 2009). Administrators can also support teacher collaboration and community engagement, which is an indirect predictor of mathematics performance (OECD, 2023).

·       School climate. School climate (relationships, safety, and institutional values) shapes engagement and academic outcomes (Thapa et al., 2013). Self-Determination Theory and Stage-Environment Fit Theory emphasize how supportive contexts can meet students’ needs for competence, autonomy, and relatedness; positive teacher–student relationships can also buffer bullying and academic risks (Ren, Chen, & Zhao, 2025). Bullying undermines achievement by weakening students’ social bonds and sense of safety (Ren et al., 2025; Graham, 2016). International PISA evidence also shows that chronic absenteeism rose to 7.6% in PISA 2022, up from 4.7% in 2018 (OECD, 2023).

·       Resources and equity. Cross-country PISA research suggests that school-to-school achievement gaps are smaller where resources are distributed more equally (Beese & Liang, 2010). School material resources and the quality of the learning environment show meaningful links to achievement in PISA-based studies (Trinidad, 2020). School-level safety, instructional leadership, and parental support also predict mathematics performance (Wardat et al., 2022; Tomul, Önder, & Taşlıdere, 2021).

European/international dimension

PISA is designed for international comparison, so identifying school-level predictors in Türkiye contributes to European and international discussions on equity, governance, and school improvement. The study defines school “Success” using an OECD-linked mathematics proficiency threshold used in OECD contexts (European Schools, 2022, p. 35), supporting interpretation across systems. The predictors prioritized by the model—parental involvement, disciplinary climate, technology use, leadership practices, student–teacher relationships, student absenteeism, and bullying—are widely relevant for equity-oriented school improvement efforts.

Method overview

The study analyzes 81 school-level features from 196 Turkish schools in the PISA 2022 School Questionnaire. To address class imbalance and small sample size, it applies Synthetic Minority Over-sampling Technique (SMOTE) and creates 3,564 balanced observations (Chawla et al., 2002). It tests supervised classifiers and reports logistic regression results; the final sample meets recommended minimums for stable modeling (Silvey & Liu, 2024). Generalizability is assessed using the training–test accuracy gap criterion (Géron, 2022).

Expected contribution

The study shows how interpretable machine learning can produce policy-relevant evidence from large international assessment data and highlight institutional levers linked to mathematics success (Acısıl-Çelik & YeşilKana, 2022; Erdoğan &Taştan, 2024).

This study employed a predictive modeling research design to identify school-level factors associated with mathematics achievement in Türkiye, analyzing publicly available OECD PISA 2022 School Questionnaire data. The Türkiye sample included 196 schools, each described by 81 school-level features. Due to the small sample size and class imbalance, the SMOTE was applied to augment and balance the data, resulting in 3,564 school-level observations (Chawla et al., 2002). Data preprocessing was used to improve data quality and suitability for modeling. Missing data were handled through imputation: discrete variables were imputed using the mean when values were close to the median, and the median when distributions were skewed. For ordinal Likert-type variables, the mode was used to preserve the original category order. Outliers in numeric variables were detected with the Interquartile Range (IQR) method. Except for the outcome variable (“Success”), outliers were capped through Winsorization to reduce skewness. After cleaning, box plots were generated for each numeric feature to review distributions and check whether outlier treatment worked as intended. The outcome variable “Success” was coded as 1 for scores above 65 and as 0 for scores below 65, in line with the 65% math proficiency threshold in OECD countries (European Schools, 2022, p. 35). After balancing with SMOTE, the dataset was randomly divided into training (80%) and test (20%) sets. For modeling, two supervised classification algorithms were tested: Random Forest (RF) and Logistic Regression (LR). Model performance was evaluated using accuracy and weighted F1-score. LR outperformed RF in this context, so only the LR results were reported. The final sample size (3,564 cases) exceeds recommended minimums for stable modeling (696 for LR and 3,404 for RF) (Silvey & Liu, 2024). Potential overfitting was evaluated by examining the accuracy gap between the training and test sets; a difference below 5% was treated as evidence of satisfactory generalizability (Géron, 2022).

This study shows that school-level indicators from PISA 2022 can predict mathematics “Success” in Türkiye with high accuracy. After cleaning the data and balancing the classes with SMOTE (Chawla et al., 2002), the Logistic Regression (LR) model achieved training accuracy of 0.9161 and test accuracy of 0.9038, with weighted F1-scores of 0.9160 (training) and 0.939 (test). The gap between training and test accuracy was well below 5%, indicating strong generalizability and no evidence of overfitting (Géron, 2022). In terms of findings, parental involvement emerged as the strongest predictor of mathematics achievement, followed by disciplinary climate. Other important predictors included technology use, school background, student–teacher relationships, student absenteeism, administrative behavior, student support, bullying, leadership activities, and leadership practices. Overall, the results underline that mathematics success is linked to a combination of relational, organizational, and resource-related conditions inside schools. Expected outcomes include a clearer, evidence-based picture of which institutional levers matter most for school improvement efforts in Türkiye. By highlighting actionable predictors, the study offers a practical basis for school reform, leadership capacity building, and targeted interventions. More broadly, it demonstrates how machine learning can turn large international datasets into scalable, policy-relevant insights for educational decision-making (Acısıl-Çelik & YeşilKana, 2022; Erdoğan &Taştan, 2024).

Acıslı-Celik, S., & Yesilkanat, C.M. (2023). Predicting science achievement scores with machine learning algorithms: A case study of OECD PISA 2015–2018 data. Neural Computing and Applications,35(28),21201-21228. Beese, J.A., & Liang, X. (2010). Do resources matter? PISA science achievement comparisons between students in the United States, Canada and Finland. Improving Schools,13(3),266–279.https://doi.org/10.1177/1365480210378941 Bush, T. (2009). Leadership development and school improvement: Contemporary issues in leadership development. Educational Review,61(4),375–389.https://doi.org/10.1080/00131910903403956 Chawla, N.V., Bowyer, K.W., Hall, L.O., & Kegelmeyer, W.P. (2002). SMOTE: Synthetic Minority Over-sampling Technique. Journal of Artificial Intelligence Research,16,321-357.https://doi.org/10.1613/jair.953 Erdoğan, S., & Taştan, H. (2024). Predicting Student Achievement via Machine Learning: Evidence from Turkish Subset of PISA. Yildiz Social Science Review,10(1),7-27. European Schools. (2022). PISA-based Test for European Schools: Group Report 2022. https://www.eursc.eu/Documents/Group%20Report_PISA_based_Test_for_European_schools_2022_en.pdf Géron, A. (2022). Hands-on machine learning with scikit-learn, keras and tensorflow: Concepts, tools, and techniques to build Intelligent Systems. O’Reilly Media, Inc. Guevara-Reyes, R., Ortiz-Garcés, I., Andrade, R., Cox-Riquetti, F., & Villegas-Ch, W. (2025). Machine learning models for academic performance prediction: Interpretability and application in educational decision-making. Frontiers in Education,10.https://doi.org/10.3389/feduc.2025.1632315 Hallinger, P. (2011). Leadership for learning: Lessons from 40 years of empirical research. Journal of Educational Administration,49(2),125–142.https://doi.org/10.1108/09578231111116699 Leithwood, K., & Jantzi, D. (2006). Transformational School Leadership for large-scale reform: Effects on students, teachers, and their classroom practices. School Effectiveness and School Improvement,17(2),201–227.https://doi.org/10.1080/09243450600565829 OECD. (2023). PISA 2022 results: Learning during and after COVID-19. OECD Publishing. Ren, Y., Chen, X., & Zhao, J. (2025). The protective role of teacher-student relationships and school belonging against bullying in Chinese middle schools. Journal of School Psychology,93,112–124. Robinson, V.M., Lloyd, C.A., & Rowe, K.J. (2008). The impact of leadership on student outcomes: An analysis of the differential effects of leadership types. Educational Administration Quarterly,44(5),635–674.https://doi.org/10.1177/0013161x08321509 Silvey, S., & Liu, J. (2024). Sample size requirements for popular classification algorithms in tabular clinical data: Empirical study. Journal of Medical Internet Research,26.https://doi.org/10.2196/60231 Thapa, A., Cohen, J., Guffey, S., & Higgins-D’Alessandro, A. (2013). A review of school climate research. Review of Educational Research,83(3),357–385. Tomul, E., Önder, E., & Taşlıdere, E. (2021). The impact of school-related variables on mathematics achievement in Turkey: A regional comparison. Educational Studies,47(6),637–655. Trinidad, J. E. (2020). Material Resources, school climate, and achievement variations in the Philippines: Insights from Pisa 2018. International Journal of Educational Development,75,102174.https://doi.org/10.1016/j.ijedudev.2020.102174 Wardat, Y., Almekhlafi, A., & BinSubaih, A. (2022). School factors influencing students' mathematics performance in the UAE. International Journal of Educational Research,112,101912.