Reducing Neonatal Mortality at Scale: Lessons for Targeting

Neonatal sepsis kills over 400,000 children annually. Experimental estimates of the preventive use of chlorhexidine vary widely, leading to external validity concerns. We provide the first quasi-experimental estimates of the effect of chlorhexidine in “real life” conditions

and apply machine-learning (ML) to analyze treatment effect heterogeneity in a nationallyrepresentative, Nepalese observational dataset. We find that chlorhexidine decreases neonatal mortality by 43% and that a simple targeting policy leveraging heterogeneous treatment effects improves neonatal survival relative to WHO recommendations. Heterogeneous treatment effects extrapolated from our ML analysis are broadly in line with experimental findings across five countries despite significant implementation differences.