Attrition in Randomized Control Trials: Evidence From Early Education Interventions in Sub-Saharan Africa

This study addresses the problem of attrition bias in longitudinal education RCTs by comparing multiple attrition correction methods. Our analysis of three early reading experiments in Zambia, Ethiopia, and South Africa shows that treatment status is generally not associated with attrition, while baseline reading skills are negatively correlated with it. We estimate bounds under alternative assumptions about the attrition process. Our main contribution lies in proposing guidelines for selecting appropriate correction methods based on data characteristics, including the extent of treatment-effect heterogeneity and the correlation of attrition with baseline characteristics. These guidelines incorporate a novel diagnostic for assessing the plausibility of the stochastic-dominance assumption. The findings offer practical direction for researchers on how to design, implement, and interpret evaluations where attrition poses risks to causal validity.