Study: Parental Background Beats GNNs in Predicting University Completion

A new study, 'Beyond Parents? Prediction Gaps in University Completion Using Population-Scale Networks and Flexible Machine Learning,' published on arXiv, examines how much children's educational attainment can be predicted from their wider social networks once parental background is known. The researchers—Javier Garcia-Bernardo, Eva Jaspers, Weverthon Machado, Samuel Plach, and Erik Jan van Leeuwen—used population-scale administrative data from Statistics Netherlands. They constructed a network linking a full cohort of children aged 11–12 to their parents, extended kin, classmates, household members, and neighbors. They then predicted university completion at ages 24–25 using logistic regression, gradient boosting (which uses individual-level aggregates of these contexts), and graph neural networks (GNNs) operating directly on the network structure. The study interprets differences in out-of-sample performance as 'prediction gaps.'

The results show that parental socioeconomic background captures most of the predictable variation—even GNNs add little predictive power once parents are known. The largest prediction gaps were found among children without a registered father, particularly girls and those with less-educated mothers. The authors argue that prediction gaps can support sociological theory-building: small gaps indicate where current theory-based models already explain what can be measured, while large gaps identify where targeted mechanism-focused research is needed. This challenges the assumption that broader social contexts (schools, neighborhoods, extended kin) are independent sources of inequality, suggesting they primarily reproduce parental advantage. The study offers a novel methodological approach to evaluating sociological theories using machine learning interpretability.