The researchers, led by Souhail Hermassi of Qatar University together with colleagues from the University of Bern in Switzerland and Nord University in Norway, recruited 126 children with a mean age of 12.2 years from a single urban school in Doha. Participants were divided into two equal groups of 63 based on self-reported moderate-to-vigorous physical activity (MVPA), measured with the International Physical Activity Questionnaire Short Form (IPAQ-SF). Children reporting at least 300 minutes of MVPA per week were classified as physically active, while those below that threshold were classified as non-active. The cutoff was chosen as a pragmatic threshold used in previous questionnaire-based pediatric research, and the authors caution that it should not be read as equivalent to meeting international guidelines, which recommend about 60 minutes of MVPA per day.

The most striking result concerned mathematics. Active boys scored an average of 90.2 percent on their official mathematics assessments, compared with 76.3 percent for non-active boys. Active girls scored 89.5 percent on average, against 77.1 percent for non-active girls. The statistical effect size for this difference was large, with Cohen’s d reaching 1.7, and the group effect was highly significant (F = 96.8, p < 0.001). Notably, no significant sex effects or group-by-sex interactions emerged for mathematics scores, meaning that the academic advantage associated with being active appeared similar for boys and girls. Mathematics was chosen as the academic outcome because it relies heavily on executive cognitive functions such as attention, working memory, and problem-solving, and because it is widely used as an achievement indicator in this research area.

Physical fitness told a consistent story. The children completed three tests: a Stork balance test measuring static postural control, a medicine ball throw assessing upper-body muscular power with a 3-kilogram ball, and a handgrip strength test using a digital dynamometer. Active children outperformed non-active children across all three measures, with significant group effects (F ≥ 3.98, p ≤ 0.048, d ≥ 0.36). Sex effects were significant for the medicine ball throw and handgrip strength, with boys generally outperforming girls, but there were no significant group-by-sex interactions. In post hoc comparisons, active boys threw the medicine ball farther than all other groups and held the Stork balance longer than non-active boys and girls, although non-active boys showed greater handgrip strength than active girls.

Anthropometric and maturation measures added nuance. Active children had significantly lower body mass than their non-active peers (boys: 47.2 versus 53.8 kilograms; girls: 50.6 versus 53.8 kilograms), with a small-to-moderate effect size. Arm span was nearly significantly shorter in active boys. However, height, body mass index, sitting height, and estimated biological maturity, calculated using Moore’s equations for predicting the age at peak height velocity, did not differ meaningfully between groups. The authors suggest that the narrow age range and the cross-sectional design may explain why growth parameters showed few differences.

The activity data revealed clear sex-specific patterns. Active boys reported the highest vigorous activity of any subgroup, averaging 1,209 MET-minutes per week, and the lowest sedentary time, at 2.6 hours per day, significantly less than all other groups. Active girls also spent less time sedentary than non-active girls (3.7 versus 5.3 hours per day). Moderate-intensity activity and walking, by contrast, did not differ significantly between groups, suggesting that the differences in overall energy expenditure and fitness were driven primarily by vigorous-intensity movement rather than routine lower-intensity activity. The authors argue that this pattern underscores the potential value of structured, high-intensity activities within school and extracurricular programs.

Correlational analysis sharpened the picture further. Sedentary time showed a moderate negative correlation with mathematics scores (r = −0.48, p < 0.001), meaning that children who sat longer tended to score lower. Total weekly physical activity volume, expressed in MET-minutes per week, showed only a trivial positive correlation with mathematics achievement (r = 0.18, p = 0.049). The authors interpret this asymmetry as indicating that overall activity volume alone is unlikely to explain academic performance, whereas reducing prolonged sitting may be more meaningfully associated with better mathematics outcomes.

The study’s authors are careful to frame these findings as associations rather than causal relationships. Because the design is cross-sectional, it cannot establish whether physical activity drives better mathematics achievement, whether academically successful children are simply more likely to be active, or whether unmeasured factors influence both. Biological maturation was estimated but not included as a statistical covariate, and confounders such as nutrition, sleep, socioeconomic status, parental education, prior academic achievement, and classroom factors were not measured. Self-reported activity data are also subject to recall and social desirability bias, and the Arabic version of the IPAQ has limited validation evidence for children in this age range. The single-school sample further limits generalizability, and the study was not powered to detect small interaction effects.

The researchers also emphasize conceptual distinctions that are often blurred in public discussion. Physical activity is a modifiable behavior, physical fitness is a physiological attribute that may result from habitual activity, and academic achievement is a multidetermined educational outcome. Physical fitness has been proposed in the literature as a potential correlate linking activity to academic outcomes, but the study did not perform a mediation analysis, so any mediating role remains hypothetical. Similarly, references to neurocognitive mechanisms such as improved executive function or cerebral blood flow draw on prior literature rather than on measurements made in this study, which did not directly assess cognitive function.

Despite these caveats, the findings carry practical weight. They align with a broad body of evidence, including meta-analyses showing small positive effects of physical activity interventions on academic performance and of resistance training on cognition in youth. The authors suggest that schools may consider providing regular opportunities for moderate-to-vigorous activity, incorporating active learning strategies, and reducing prolonged sedentary time, while ensuring that physical education programs are gender-sensitive so that girls have equitable opportunities to develop strength and power. At the policy level, collaboration between education and health authorities could support evidence-based school activity initiatives. The authors stress, however, that the present findings should not be interpreted as proof that increasing physical activity will directly improve academic achievement; establishing causality will require well-designed longitudinal and intervention studies, ideally with objective activity measurement, larger and more diverse samples, and statistical control for maturation and other confounders.

Subject of Research: Associations between physical activity, physical fitness, sedentary behavior, and mathematics achievement in school-aged children

Article Title: Associations between physical activity, physical fitness, sedentary behavior, and mathematics achievement in school-aged children: a cross-sectional study

Article References: Hermassi, S., Alshaer, A., Alneama, A., Ketelhut, S., & van den Tillaar, R. (2026). Associations between physical activity, physical fitness, sedentary behavior, and mathematics achievement in school-aged children: a cross-sectional study. Frontiers in Cognition, 5, Article 1903208. https://doi.org/10.3389/fcogn.2026.1903208

Image Credits: AI Generated

DOI: 10.3389/fcogn.2026.1903208

Keywords: physical activity, physical fitness, sedentary behavior, mathematics achievement, schoolchildren, MVPA, Qatar, cross-sectional study, executive function, pediatric health, IPAQ-SF, sex differences

The research team, led by Jiawei Shao of the Centre for Instructional Technology and Multimedia at Universiti Sains Malaysia, together with Ju Jinming and Shupeng Tang, recruited 174 Chinese fifth-grade students. Using individual-level random assignment, the researchers divided the children into three groups. One group learned with the full AI-supported Gamified Learning Program, which the authors abbreviate as AI-GILP. A second group used a gamified learning program without AI support, known as GILP, while a third received traditional lecture-based instruction. The design is notable for its rigor: rather than comparing a new technology against nothing, the study isolated the specific contribution of the AI layer by holding the gamification constant across two of the three conditions.

Mathematics anxiety, learning motivation, and achievement were measured at three points: before the intervention, immediately after the two-week program, and again four weeks later at a delayed posttest. The researchers analyzed the data using analysis of covariance and mixed-design analyses of variance, standard statistical techniques for detecting group differences while accounting for baseline performance. The results were striking on the emotional and motivational front. The AI-supported condition produced significantly lower mathematics anxiety and significantly higher learning motivation than both the gamified-only condition and traditional instruction, with overall group effects corresponding to partial eta-squared values of .208 for anxiety and .184 for motivation, both of which represent substantial effects in educational research.

The achievement results told a more nuanced story. Both gamified conditions outperformed traditional lecture-based instruction on mathematics achievement, confirming that gamification itself carries real academic benefits. However, the AI-supported group did not significantly outperform the gamified-only group on achievement, with a Cohen’s d of 0.19, a small difference that did not reach statistical significance. In other words, the AI companion appeared to work primarily on how children felt about mathematics rather than on how much mathematics they learned, at least over the timescale of this study. At the delayed posttest, the descriptive means for anxiety and motivation retained the same favorable ordering for the AI-supported group, although the authors note that pairwise significance at that assessment was not formally tested.

Perhaps the most revealing detail lies in where the motivational advantage came from. The researchers found that the AI-supported program’s motivational benefit was concentrated in intrinsic motivation and perceived competence, two dimensions closely tied to whether learners engage with a subject because they genuinely enjoy it and believe they can succeed. Notably, there was no significant difference in the effort/importance dimension. This pattern suggests that the AI companion did not simply push children to work harder; instead, it appears to have nurtured a sense of capability and genuine interest, which many learning scientists consider more durable drivers of long-term engagement.

The technology behind these effects is documented in unusual detail in the study’s appendix, and it is the engineering that makes the findings credible. The AI Learning Companion, as the system is called, is built on a modular workflow architecture with four functional modules: a dialogue module, a hint-hierarchy module, a behavior-triggered support module, and a personality-configuration module. When students watch recorded instructional videos or complete gamified mathematics tasks, they can activate the companion through a robot icon and ask questions by speech or text. An Intent Classifier categorizes each input as conceptual, procedural, help-seeking, or another predefined interaction type, and a Knowledge Router retrieves relevant concepts, common misconceptions, and illustrative examples from an embedded knowledge base.

The hint system follows a deliberately escalating structure based on a clue-reasoning-example sequence. At the first level, clue-based prompts direct students’ attention to key information and activate prior knowledge without giving anything away. If a learner makes three consecutive errors on the same task or remains inactive for a predefined period, the system advances to the second level, where reasoning-oriented prompts identify the relevant mathematical relationships and guide the student toward the required problem-solving procedure. Only at the final level does the system present a worked example or visual demonstration. This design gives children the chance to attempt tasks independently before receiving the most explicit assistance, a structure informed by Self-Determination Theory’s principles of competence support and learner choice.

Crucially, the system does not claim to read minds. The Behavior-Triggered Support Module does not measure students’ emotional states or diagnose anxiety and frustration. Instead, it monitors observable interaction indicators such as error frequency, time-on-task, repeated inactivity, and the frequency of hint requests. When predefined combinations of these indicators are reached, the companion offers supportive messages through voice or text, with examples including encouragement to try again together or reminders of what the learner has already mastered. The system also delivers gamified rewards, adding points and displaying badges after correct responses, and it allows students to choose among three interaction personalities: scholarly, humorous, or adventurous, each with a different communication tone that children can switch at will.

The authors are refreshingly careful about what their study does and does not demonstrate. The proposed mechanisms involving psychological-need support, emotional regulation, adaptive feedback, and motivational internalization were not directly tested, and the study did not isolate the effects of individual companion modules or measure their mediating processes. The supportive messages should not be interpreted as evidence that the system detected a learner’s emotional state or that children perceived the messages as empathic. What the experiment does establish is that the complete AI-supported package, compared against both a well-matched gamified control and traditional teaching, produced meaningful reductions in anxiety and gains in specific motivational dimensions, while the achievement gains were attributable primarily to gamification itself.

For educators and developers watching the rapid arrival of AI tutors in classrooms, the study offers a valuable calibration. The biggest wins from adding an AI layer to gamified learning may not show up immediately in test scores but in the emotional climate of the mathematics classroom, where anxiety quietly erodes participation and confidence. If future process-oriented research confirms that adaptive feedback, tiered hints, and configurable companionship are the active ingredients, the blueprint documented here, from intent classification to escalating hints to behavior-triggered encouragement, could become a reference architecture for learning technologies that treat children’s feelings as seriously as their answers.

Subject of Research: Effects of an AI-supported gamified learning program on elementary students' mathematics anxiety, motivation, and achievement

Article Title: The effects of an AI-supported gamified learning program on elementary students' mathematics anxiety, learning motivation, and achievement

Article References: Shao, J., Jinming, J., & Tang, S. (2026). The effects of an AI-supported gamified learning program on elementary students' mathematics anxiety, learning motivation, and achievement. BMC Psychology. https://doi.org/10.1186/s40359-026-05367-8

Image Credits: AI Generated

DOI: 10.1186/s40359-026-05367-8

Keywords: artificial intelligence, gamified learning, mathematics anxiety, learning motivation, elementary education, randomized controlled trial, Self-Determination Theory, educational technology, mathematics achievement, AI learning companion, intrinsic motivation, BMC Psychology

The research, conducted by Artemis Eleftheriadi and Konstantinos Lavidas of the University of Patras together with Panagiotis Gridos of the University of the Aegean, set out to rethink how spatial skills in young children are assessed. Rather than treating spatial ability as a single, monolithic capacity measured by paper-and-pencil tests, the team built a typology grounded in spatial relations and informed by the widely used learning trajectories framework of Clements and Sarama. The central idea is simple but powerful: what children can do with space depends critically on the type of spatial relation involved—positional, directional, referential, or representational—and assessment should capture those distinctions rather than collapse them into one score.

To test whether such a typology could actually work in a real classroom, the researchers carried out a qualitative case study with twenty-four children aged four to six years in a kindergarten setting. The design was deliberately close to everyday early childhood practice. Instead of standardized testing conditions, children engaged in a sequence of structured, play-based activities built around a farm scene, featuring animals, a barn, a windmill, fences, a maze, and a friendly farmer character named Dimitris. Through a Simon-says-style game, children were asked to place animals in front of themselves or behind themselves, on their own right or left, or relative to Dimitris and other objects. They then navigated a three-dimensional maze, giving verbal directions such as take two steps forward and turn right. Finally, they worked with maps and a gridded coordinate surface, reading locations from a floor plan, transferring objects onto a model, and even drawing their own maps of the farm.

This approach allowed the researchers to combine task performance data with a fine-grained qualitative examination of children’s responses. Crucially, the activities crossed two dimensions that spatial cognition researchers have long considered essential: the reference system involved and the kind of representation demanded. Egocentric tasks required children to encode space from their own point of view, while allocentric tasks required them to adopt the perspective of another person or object. Some tasks relied on spatial language alone, while others demanded the interpretation and construction of spatial representations such as maps and models. The typology organized these demands into a coherent framework that could be operationalized by teachers, not just by laboratory researchers.

The findings paint a strikingly consistent picture of the preschool spatial landscape. Basic positional relations were largely developed: most children could reliably identify what was inside or outside, above or below, in front of or behind, and near or far. But directional relations were a different story. Right and left, in particular, remained problematic across the sample, echoing a long-standing finding in developmental psychology that the distinction between these two directions is one of the last spatial concepts to consolidate and one of the most persistent sources of error in early childhood.

The contrast between egocentric and allocentric performance was equally clear. Children were markedly more successful when tasks let them anchor spatial judgments in their own bodies. When the reference system shifted—to the farmer Dimitris, to an animal on the model, or to an abstract coordinate grid defined by numbers and colors—performance dropped. This result aligns with decades of theory distinguishing perspective-taking from self-based coding, but the study’s classroom setting shows that the gap is visible in ordinary educational activities, not just in controlled laboratory experiments. For teachers, it means that a child who seems spatially competent during movement games may genuinely struggle when the same relations must be computed from another viewpoint.

Perhaps the most revealing difficulties emerged when children had to coordinate multiple spatial relations at once. Navigation through the maze required chaining directional decisions—move forward, then turn, then move again—into a coherent sequence. Map-based activities required mapping relations on a two-dimensional plan onto a three-dimensional model and vice versa, holding the relational structure stable across surfaces. Here, many children who could recognize a map as a map nonetheless failed to maintain the relational structure it encoded: they could identify the depicted objects but misplace them, swap directions, or lose track of the correspondence between the plan and the model. Recognition, in other words, far outpaced relational construction.

That distinction may be the study’s most important conceptual contribution. The researchers observed that although many children recognized spatial representations, fewer were able to maintain their relational structure. This suggests that early spatial assessment, if it stops at recognition, will systematically overestimate what children actually understand. A relational approach—asking whether a child can use a representation to carry out a transformation, preserve correspondences, or coordinate two reference systems—gives a far more accurate and educationally useful picture. It also connects directly to the broader evidence base: spatial skills are strongly linked to later mathematics achievement, and meta-analyses of training studies show that spatial ability is highly malleable in early childhood. Identifying precisely which relational skills are lagging gives educators a concrete target for intervention rather than a vague deficit label.

The study also carries practical implications for curriculum design. Its activities were aligned with widely adopted standards, such as describing, naming, and interpreting relative positions in space and finding and naming locations in coordinate systems such as maps. The typology offers teachers a diagnostic lens: a class may be fluent with positional language but shaky on direction, or confident with egocentric commands but lost with maps. Because the framework is grounded in a respected learning trajectories approach, it can guide not only assessment but also the sequencing of instruction, moving children from body-anchored relations toward increasingly abstract and coordinated systems.

Limitations remain, as the authors would acknowledge. The sample was small, the setting a single kindergarten, and the design qualitative—so the findings are a proof of concept for the typology rather than a population-level description. Data from the study are available from the corresponding author upon reasonable request, and the research was supported by the Andreas Mentzelopoulos Foundation. Still, the message is clear and, for a field increasingly focused on early STEM foundations, timely: spatial relations are the load-bearing structure of preschool spatial thinking, and assessments that ignore that structure risk misreading what young children know. The left–right confusions, the egocentric crutch, and the fragile relational grip on maps are not failures to be stamped out but signposts pointing to the next step in each child’s spatial learning trajectory.

Subject of Research: A relational typology for assessing spatial skills in preschool children

Article Title: Rethinking the Assessment of Preschoolers’ Spatial Skills: A Typology Based on Spatial Relations

Article References: Eleftheriadi, A., Gridos, P., & Lavidas, K. (2026). Rethinking the Assessment of Preschoolers’ Spatial Skills: A Typology Based on Spatial Relations. International Journal of Early Childhood. https://doi.org/10.1007/s13158-026-00554-5

Image Credits: AI Generated

DOI: 10.1007/s13158-026-00554-5

Keywords: spatial skills, preschool education, spatial relations, spatial assessment, egocentric reference frame, allocentric reference frame, spatial language, navigation, map use, early childhood, learning trajectories, mathematics achievement

The ten systems examined were Argentina, Chile, Chinese Taipei, Finland, Hungary, Italy, Japan, Korea, the Philippines, and the United States. The authors stress that this was a purposive comparative sample, chosen to span geographic regions, OECD membership status, and a wide range of average performance, from a weighted mean mathematics score of 352.57 in the Philippines to 531.14 in Chinese Taipei. The design supports comparisons across these ten systems but does not license claims about all countries or continents. Within the sample, 2,617 schools contributed students, with country sample sizes ranging from 4,838 in the United States to 11,975 in Argentina.

What distinguishes this study from much of the educational data-mining literature is its fidelity to PISA’s assessment architecture. Mathematics achievement in PISA is not a single number but a set of ten plausible values, multiple imputed estimates that represent uncertainty in each student’s latent proficiency. The researchers fitted and evaluated every model across all ten plausible values rather than relying on just the first. They also incorporated the final student weight, W_FSTUWT, in descriptive statistics, model fitting, and test-set evaluation, and used the 80 student replicate weights to estimate sampling variability. Total variance for each performance metric combined the mean replicate-weight sampling variance with the between-plausible-value variance, following the formula U-bar plus (1 + 1/M) times B, so that confidence intervals reflected both complex sampling uncertainty and achievement-scaling uncertainty. Few machine-learning studies of PISA data go to these lengths.

The predictor side of the analysis was equally disciplined. The authors constructed a transparent primary candidate pool of 33 predictors from PISA 2018 student questionnaire variables and OECD-derived indices, organized into six substantive domains: student demographics, family socioeconomic status and home resources, engagement and learning time, peer climate, belonging and parent support, and classroom climate. Rather than assuming all 33 variables mattered everywhere, they performed country-specific stability feature selection using only training schools. Three complementary methods were applied: mutual information, which captures general dependence; weighted elastic net regression, a regularized linear method that handles correlated predictors through combined L1 and L2 penalties; and weighted random forest ranking, which captures nonlinear and interaction-based structure. Feature rankings were repeated across ten plausible values and five internal school-level folds, yielding 50 selection runs per country. A predictor was deemed stable if selected in at least half of those runs. The resulting stable sets contained between 11 and 15 predictors, averaging 14.0 per country.

Four models were then compared under country-specific school-level holdout evaluation, with roughly 70 percent of students in training and 30 percent in test sets. Splitting by school rather than by student reduced leakage from students in the same school appearing on both sides. The contenders were weighted linear regression, weighted linear regression augmented with pairwise interactions, random forest, and CatBoost, a gradient-boosted tree ensemble designed for structured tabular data. Averaged across countries, CatBoost was the clear winner, achieving a mean weighted R-squared of 0.358 and a mean mean absolute error of 57.29 score points. Random forest followed with a mean weighted R-squared of 0.313 and an MAE of 59.30. The interaction-augmented linear model barely improved on the additive linear benchmark, with mean R-squared values of 0.290 versus 0.286. CatBoost posted the lowest weighted MAE in all ten systems, and its advantage over random forest was consistent but moderate: roughly 2.01 MAE points and 0.045 in R-squared on average. Its country-level R-squared ranged from 0.210 in Italy to 0.443 in Hungary.

The interpretive core of the study used SHAP, or Shapley additive explanations, a technique that decomposes each model prediction into additive contributions from individual features. The authors were careful to frame SHAP summaries as model-based explanations of predictive associations, not causal effects. Because CatBoost performed best, interpretation focused on its SHAP values, computed on held-out test schools and averaged across all ten plausible values to produce PV-robust stability summaries. The headline finding is that books at home was the most stable predictor of all, ranking among the top ten SHAP predictors in all ten countries on average and among the top five in 8.7 countries on average. Highest parental occupational status was nearly as stable, appearing in the top ten in 9.2 countries and the top five in 7.1. Socioeconomic status itself appeared in all ten country-specific models and ranked in the top ten in 8.3 countries on average, though its average top-five count was lower at 4.3.

Beyond the socioeconomic core, the picture became more heterogeneous. Grade placement entered the stable feature sets of seven countries and ranked in the top ten in all seven, but it was not selected in Chinese Taipei, Japan, or Korea. Mathematics learning time appeared in six countries and ranked in the top ten in all six, with an average top-five count of 5.5, and it ranked especially highly in Chinese Taipei, Japan, and the United States. Test effort was retained in only four systems but was consistently influential there, ranking in the top five in all four and even first in Finland, Italy, and Korea in the diagnostic visualization. Directed instruction and disciplinary climate also showed cross-national reach, ranking in the top ten in 7.0 and 5.8 countries on average respectively. The authors argue that this mix of stability and country-specificity is precisely why a single pooled importance ranking would be misleading.

Sensitivity analyses reinforced the robustness of these conclusions. Comparing SHAP summaries based on the first plausible value with summaries averaged across all ten showed only small differences, with maximum absolute gaps of 0.8 countries for top-ten counts, 1.2 countries for top-five counts, and 1.25 rank positions for mean rank, confirming that the main pattern was not an artifact of PV1MATH. A second sensitivity analysis removed grade placement from the stable feature sets where present and refitted the CatBoost models. Performance dropped modestly, with a mean R-squared change of -0.028 and a mean MAE increase of 1.17 score points, the largest decline occurring in Chile, where R-squared fell by 0.091 and MAE rose by 3.92 points. Crucially, the SHAP stability pattern remained broadly similar without grade: books at home, parental occupational status, learning time, socioeconomic status, test effort, directed instruction, and disciplinary climate all stayed among the most stable predictors.

The authors are explicit about the limits of their contribution. They do not claim to have discovered new determinants of mathematics achievement; domains such as socioeconomic resources, home literacy environments, learning time, and classroom climate are already well established. The value lies in showing how these familiar predictors behave under a rigorous, country-specific, survey-weighted, plausible-value-aware explainable machine-learning framework, and in separating three questions that are often conflated: which models predict best, which predictors are stable across systems, and which are context-specific. They also caution that the analysis is predictive rather than causal, that the ten-system sample is not statistically representative of global education, that the 33-predictor pool cannot exhaust all relevant influences such as school policies or teacher characteristics, and that cross-national questionnaire comparisons may be affected by response styles, translation, and measurement comparability.

The implications reach well beyond this dataset. The framework, with its train-only feature selection, school-level holdout evaluation, replicate-weight variance estimation, and all-plausible-value pooling, offers a reproducible template that could be extended to more PISA systems, to other assessments such as TIMSS and PIRLS, to school-level and system-level predictors, and to fairness analyses examining whether prediction errors or SHAP patterns differ across gender, socioeconomic, immigrant, or language groups. Repeated-cycle analyses could test whether cross-national stability patterns persist as PISA evolves. For a field increasingly drawn to black-box prediction, the study makes a compelling case that explainability and methodological rigor are not optional extras but the foundation of responsible machine learning in comparative education.

Subject of Research: Explainable machine learning applied to predicting and interpreting mathematics achievement in PISA 2018 across ten education systems.

Article Title: Explainable AI for predicting and interpreting mathematics achievement: a cross-national analysis of PISA 2018

Article References: Liu, L., & Dai, R. (2026). Explainable AI for predicting and interpreting mathematics achievement: a cross-national analysis of PISA 2018. Large-scale Assessments in Education, 14(1), Article 46. https://doi.org/10.1186/s40536-026-00320-y

Image Credits: AI Generated

DOI: 10.1186/s40536-026-00320-y

Keywords: PISA 2018, mathematics achievement, explainable artificial intelligence, SHAP, CatBoost, machine learning, survey weights, plausible values, large-scale assessment, socioeconomic status, comparative education, feature selection

The study was motivated by a practical question with high stakes. Policymakers across the world rely on measures of socio-economic status, or SES, for two main purposes. First, they use them to decide which schools should receive additional resources aimed at levelling the playing field for students from poor backgrounds. Second, researchers use them to understand how family and school circumstances shape achievement, and to monitor whether equity reforms are working. If the underlying measures are flawed, both the flow of funding and the scientific conclusions built on them are compromised.

The research team, led by Lorraine Gilleece of the Educational Research Centre in Dublin, together with Anastasios Karakolidis and Aidan Clerkin, compared four sets of indicators. Two came from international large-scale assessments: the 2019 Trends in International Mathematics and Science Study, or TIMSS, which tested Grade 8 students, and the 2022 Programme for International Student Assessment, or PISA, which tested 15-year-olds. From each assessment, the researchers extracted a school-average SES index, aggregated from individual student responses, together with principals’ own estimates of the percentage of their students coming from disadvantaged homes. These were set against two national indicators: the Pobal HP Deprivation Index, a continuous measure derived from census data on small residential areas, and a binary flag for whether a school participates in Ireland’s Delivering Equality of Opportunity in Schools programme, known as DEIS, which directs extra resources to schools with high concentrations of disadvantage.

Technically, the analyses were extensive. The team computed correlations, ran linear and logistic regressions, and fitted two-level multilevel models using survey weights and plausible values for achievement, in line with the statistical protocols of the International Association for the Evaluation of Educational Achievement and the OECD. The TIMSS sample covered 4,118 students in 149 schools, while the PISA 2022 sample included 5,569 students in 170 Irish schools. The HP index, built from census indicators such as educational attainment, social class composition, unemployment rates and demographic vitality, was matched to schools using student home addresses, meaning school-average values were based on almost the entire enrolled cohort rather than only the assessed students.

The convergent validity results were emphatic. Correlations between the international and national indicators were large by conventional statistical standards, with coefficients of roughly 0.7 among indicators derived from aggregated student-level data, such as school-mean SES and school-mean HP scores, and around 0.6 among inherently school-level indicators such as principals’ reports and DEIS status. When both TIMSS-derived indicators were entered together in a regression predicting the school-mean HP index, they explained 55 percent of its variance; the equivalent PISA model explained 50 percent. DEIS schools were also shown to differ markedly from non-DEIS schools on every international indicator, with large effect sizes throughout.

One of the most reassuring findings concerned school principals. Because principals’ estimates of the share of disadvantaged students rest on impressions that cannot be directly verified, they have long been treated with suspicion. Yet in this study, principals’ judgements aligned well with both the census-based HP index and the aggregated student-level SES indices. In the logistic regression predicting DEIS status, schools whose principals reported that more than half of their students came from disadvantaged homes had 65 times higher odds of DEIS categorisation than schools reporting up to a quarter, a pattern the authors read as reflecting the policy’s deliberate targeting of schools with the highest concentrations of disadvantage. Principals, the study concludes, appear to know their schools remarkably well.

The picture changed when the researchers turned to criterion-related validity, asking how well each family of indicators predicted mathematics achievement. The international indicators were consistently stronger. Together, school-mean SES and the principal-reported share of disadvantaged students explained 16 percent of the variance in TIMSS mathematics scores and 9 percent in PISA. The national indicators on their own explained less, at 11 percent for TIMSS and 6 percent for PISA. Crucially, in a combined model, only the ILSA-derived school-mean SES remained a statistically significant predictor; adding the national indicators to the international ones barely improved predictive power at all. Multilevel models told a consistent story, with the final predictor sets explaining 87 percent of between-school variance in TIMSS and 76 percent in PISA, despite relatively modest intraclass correlations of 21 percent and 11 percent respectively.

The authors are careful to note that the national indicators should not be dismissed as redundant simply because they add little explanatory power once international measures are controlled for. Loss of statistical significance in a combined model does not mean two variables measure the same construct, and the national indicators each capture unique variance consistent with their different conceptual origins. The PISA ESCS index reflects a gradient approach to SES, combining parental occupation, parental education and home possessions; the TIMSS Home Educational Resources scale draws on books in the home, study supports and parental education; the HP index is an area-level census composite; and DEIS status is an administrative designation shaped by policy history. Each is a different lens on the same underlying phenomenon.

Nevertheless, the findings carry a clear practical message for governments: wherever feasible, SES data should be collected at the student level rather than relying solely on school-level designations. Indicators aggregated from individual data tracked each other more tightly and predicted achievement more powerfully than inherently school-level labels. This matters because higher-level designations can blur over time, as Ireland’s own experience with successive waves of DEIS designation illustrates, complicating efforts to monitor outcomes. The lesson also echoes international research: in Sweden, register-based SES explained less variance in mathematics achievement than TIMSS measures and added almost nothing beyond them, while French work on TIMSS 2023 found the national Social Position Index correlated less strongly with performance than the international home resources indices.

For a field increasingly dependent on large-scale assessment data to guide equity policy, this study provides a methodological template. By linking TIMSS and PISA data to national administrative records, countries can audit their own indicators of disadvantage and decide, on evidence rather than habit, which measures deserve a place in funding formulae and research. Given that the OECD has urged Ireland to refine and validate the indicators underpinning its resource allocation model, and that similar debates over free school meals, Universal Credit eligibility and register-based measures are playing out in the United Kingdom, the Netherlands and Belgium, the Irish findings arrive at an opportune moment. Principals’ perceptions, census indices and international survey composites all tell substantially the same story about which schools serve disadvantaged communities. But if the goal is to understand and predict how socio-economic context shapes learning, the richest measures remain those built from the experiences of students themselves.

Beyond its immediate policy relevance, the study speaks to a long-standing methodological puzzle in comparative education research: the fact that no single operational definition of socio-economic status exists. Because SES is a latent construct, every measure — whether built from survey questionnaires, census small-area statistics, or administrative designations — is an imperfect proxy. This is precisely why validity evidence matters more than any one indicator’s apparent sophistication. A census-based index can be statistically elegant yet blunt at the level of individual schools, while a survey composite can capture family circumstances that area-level averages smooth over.

The choice of mathematics achievement as the criterion variable is also worth noting. Mathematics performance is often regarded as particularly sensitive to family resources, since out-of-school support — private tutoring, study materials, parental familiarity with the curriculum — tends to track household advantage more closely than in some other subjects. This may partly explain why the home-resources-based international measures showed such consistent predictive strength across both TIMSS and PISA cohorts in the study.

The multilevel modelling approach adopted by the researchers also deserves attention from an international audience. Because achievement data in assessments such as TIMSS and PISA are nested within schools, analysing them with single-level regression risks overstating the precision of school-level effects. By explicitly modelling between-school variance and reporting intraclass correlations alongside explained variance at each level, the study offers a transparent account of how much of the socio-economic gradient operates between rather than within schools. The relatively high proportion of between-school variance explained suggests that school-level SES indicators, whatever their construction, capture a substantial share of the structural differences among institutions.

Finally, the study’s timing is significant. As education systems grapple with rising data expectations, integrating international assessment data with national administrative records remains rare in practice, often blocked by technical or legal barriers rather than analytical ones. The Irish example demonstrates that such linkage is feasible and can be undertaken with existing public-use and administrative datasets, lowering the threshold for other countries to subject their own equity instruments to comparable empirical scrutiny.

Subject of Research: Comparing the validity of school socio-economic status measures from TIMSS and PISA against Irish national administrative indicators and their power to predict mathematics achievement

Article Title: The validity of school socio-economic status measures: how do indicators from TIMSS and PISA compare to national administrative data sources?

Article References: Gilleece, L., Karakolidis, A., & Clerkin, A. (2026). The validity of school socio-economic status measures: how do indicators from TIMSS and PISA compare to national administrative data sources?. Large-scale Assessments in Education, 14(1), Article 45. https://doi.org/10.1186/s40536-026-00305-x

Image Credits: AI Generated

DOI: 10.1186/s40536-026-00305-x

Keywords: socio-economic status, TIMSS, PISA, educational disadvantage, Ireland, DEIS, Pobal HP Deprivation Index, mathematics achievement, validity, large-scale assessments, education policy, administrative data