Headteacher experience, management quality and pupils’ learning in Sub-Saharan African schools: evidence from PASEC 2019

Abstract School governance is widely promoted as a low-cost lever for learning in Sub-Saharan Africa, yet most evidence rests on in-sample prediction or on estimates that ignore the clustering of pupils in schools. Using the 2019 PASEC survey (25,850 pupils in 14 countries), we relate sixteen governance indicators to a continuous learning score in all schools, public schools and private schools. We combine OLS with HC3 errors, double machine learning (DML) with school-clustered inference and Holm’s correction, decile quantile regression and a penalised-spline additive model (GAM), and we benchmark their predictive accuracy against XGBoost, Random Forest and Extra Trees on one common set of held-out schools. Three findings emerge. First, flexible models do not beat linear ones on unseen schools: in all schools OLS and median quantile regression reach a test R² of 30.6%, the GAM 29.2% and the tree-based models 24.4% to 26.2%; in private schools no model exceeds 7%. Second, once the sixteen tests of each sample are corrected, only the headteacher’s years of teaching is robustly associated with learning in all schools (DML + 34.3 points across its observed range) and, in public schools, together with management quality (+ 62.9 and + 10.0 points, the latter at the 10% level only); nothing survives in private schools. Third, the association of teaching experience is positive at every decile of the score distribution and peaks near the median, and equality of slopes across deciles is rejected for it in all and public schools. The evidence concerns conditional associations in cross-sectional data, not causal effects. JEL codes: I21, I25, O15, C21