Prediction Gaps as Pathways to Explanation: Rethinking Educational Outcomes through Differences in Model Performance

Abstract

Social contexts -- such as families, schools, and neighborhoods -- shape life outcomes. The key question is not simply whether they matter, but rather for whom and under what conditions. Here, we argue that prediction gaps -- differences in predictive performance between statistical models of varying complexity -- offer a pathway for identifying surprising empirical patterns (i.e., not captured by simpler models) which highlight where theories succeed or fall short. Using population-scale administrative data from the Netherlands, we compare logistic regression, gradient boosting, and graph neural networks to predict university completion using early-life social contexts. Overall, prediction gaps are small, suggesting that previously identified indicators, particularly parental status, capture most measurable variation in educational attainment. However, gaps are larger for girls growing up without fathers -- suggesting that the effects of social context for these groups go beyond simple models in line with sociological theory. Our paper shows the potential of prediction methods to support sociological explanation.