Beyond Parents? Prediction Gaps in University Completion Using Population-Scale Networks and Flexible Machine Learning

Abstract

How much of children's educational attainment remains predictable from the wider social contexts in which they grow up, once parental background is known? Sociological research places households, schools, neighborhoods, and extended kin at the center of intergenerational reproduction, yet whether these contexts add predictive information beyond parental background is rarely tested directly. This matters because the added value of social contexts helps distinguish whether they operate as independent sources of inequality or as channels through which parental advantage is reproduced. Using population-scale administrative data from Statistics Netherlands, we construct a network linking a full cohort of children aged 11-12 to parents, extended kin, classmates, household members, and neighbors, and predict university completion at ages 24-25. We compare logistic regression and gradient boosting, which use individual-level aggregates of these contexts, with graph neural networks (GNNs) operating directly on the network, interpreting differences in out-of-sample performance as prediction gaps. Parental socioeconomic background captures most predictable variation; even GNNs add little once parents are known. Prediction gaps are largest among children without a registered father, especially girls and those with less-educated mothers. Methodologically, we argue that prediction gaps can support sociological theory-building: small gaps show where current theory-based models already explain what can be measured; large gaps identify where targeted mechanism-focused research is warranted.