Which disadvantaged students beat the odds in STEM? A design-consistent, fairness-audited machine learning study of 143,573 adolescents in 79 PISA 2022 education systems
This study employs a design-consistent, fairness-audited machine learning approach on PISA 2022 data from 143,573 disadvantaged adolescents across 79 education systems to identify that while mathematics self-efficacy is the strongest universal predictor of STEM resilience, predictive models exhibit significant transportability limitations and systematic biases against girls and rural students, underscoring the need for context-specific and equitable screening strategies.
Original paper licensed under CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/). This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer
Every year, millions of fifteen-year-olds around the world sit for a standardized test that measures how well they understand mathematics and science. For decades, researchers have known a simple, stubborn truth about these results: a student's family wealth is one of the strongest predictors of their score. Children from poorer backgrounds tend to score lower, while those from wealthier homes tend to score higher. This gap matters deeply because proficiency in these subjects is the gateway to careers in science, technology, engineering, and mathematics. Yet, the story is not entirely deterministic. In every country that participates in these tests, there are students from the poorest quarter of households who defy the odds. They achieve high marks in both math and science, proving that disadvantage does not make success impossible. These students are often called "academically resilient." The question that has long puzzled educators and policymakers is not just who these students are, but what factors help them succeed. Is it their own confidence? The support of their teachers? The resources available in their schools?
For a long time, answering this question relied on traditional statistics, which often struggle to untangle the complex web of factors that influence a student's life. Recently, researchers have begun using machine learning, a type of computer program that can find patterns in vast amounts of data, to look for answers. However, many of these early attempts made shortcuts that ignored the specific way the test data was collected. They treated the data as a simple list of names rather than a carefully designed survey, and they often tested their models only on the same data they used to build them, without checking if the results would hold up in a different country. A new study published in 2026, led by Omar Allohibi of Taibah University, sets out to fix these problems. By analyzing data from 143,573 disadvantaged adolescents across 79 different education systems, they built a machine learning model designed to predict which students would beat the odds. Crucially, they built the model with extreme care to respect the survey's design, tested it on countries it had never seen before, and checked whether it treated different groups of students fairly.
The study began by defining exactly what "beating the odds" meant. They focused only on students from the bottom quarter of the socio-economic scale in each country. They then looked for those who reached a specific level of proficiency in both mathematics and science, a standard high enough to suggest they are ready for advanced study. They found that, globally, only about 12.5 percent of these disadvantaged students met this high bar. The rate varied wildly from country to country, ranging from less than one-tenth of a percent in some systems to 62 percent in Singapore. This variation suggested that the environment of the school system itself plays a massive role in whether a student can thrive.
When they trained their machine learning model to predict which students would be resilient, it performed with remarkable accuracy. The model could distinguish between struggling and thriving students far better than a simple guess based on the country's average performance. The most powerful predictor the model found was not a school resource or a family income level, but the student's own belief in their ability to do math. This feeling, known as self-efficacy, was the single strongest factor across all 79 countries. Students who believed they could solve math problems were far more likely to succeed, even when they faced significant disadvantages. Other factors, such as having access to computers at home and the ability to resist stress, also played important roles. Interestingly, the model also found that students who worked paid jobs outside of school were less likely to be resilient, likely because they had less time to study.
However, the study revealed a critical limitation in how these models are often used. When they took a model trained on data from many countries and tried to use it to predict outcomes in a single, unseen country, its accuracy dropped significantly. While the model was still useful, it was not as reliable as the global numbers suggested. More importantly, the specific factors that mattered most in one country were often different from the factors that mattered most in another. The only exception was the belief in one's own math ability, which remained the top predictor everywhere. This finding challenges the idea that there is a single, universal list of "key ingredients" for success that applies to every culture and education system. Instead, it suggests that while confidence is a universal engine for success, the other supporting factors depend heavily on the local context.
They also conducted a fairness audit to see if the model treated all students equally. They discovered that while the model was accurate on average, it made different kinds of mistakes for different groups. It was significantly less likely to correctly identify resilient girls than resilient boys. It was also much less likely to spot resilient students living in rural areas compared to those in cities. When they removed gender from the model's inputs, the gap for girls disappeared, suggesting the model had learned to use gender as a shortcut that worked against girls. However, removing the rural location from the model did not fix the gap for rural students. This indicated that the issue for rural students was not the model's bias, but rather the fact that the overall conditions in rural areas made success harder to predict with a single standard.
The study concludes that while machine learning can help them understand the factors behind academic resilience, they must be careful about how they use these tools. The most important takeaway is that confidence in one's own ability to learn math is a powerful, universal force that can help disadvantaged students succeed. But relying on a single global model to make decisions for every student in every country is risky. The factors that drive success are not the same everywhere, and a model that works well in one place might miss the mark in another. Furthermore, if these tools are used to identify students for support programs, they must be checked carefully to ensure they are not systematically overlooking girls or students in rural communities. By respecting the complexity of the data and testing models in the real world where they will be used, researchers can build tools that truly help equity in education rather than accidentally widening the gaps they aim to close.
Drowning in papers in your field?
Get daily digests of the most novel papers matching your research keywords — with technical summaries, in your language.