A Causal Machine Learning approach for estimating the heterogeneous effects of fertilizer dose and timing on smallholder maize yields in Tanzania

Smallholder farmers are responsible for 69% of the food produced in Tanzania, yet their productivity remains constrained by low soil fertility and limited economic access to inputs. While fertilizers are essential for achieving higher yields, suboptimal management can lead to environmental degradation and economic losses for the farmer. Therefore, optimizing the agronomic efficiency of fertilizers, specifically the question of the ideal dose and timing, is critical for the sustainable intensification of smallholder agriculture.  While on-farm field experiments are the gold standard to address this question, they are often prohibitively expensive, labor-intensive, geographically limited, and unable to account for farmer management differences. Causal Machine Learning offers a robust alternative that uses observational data by integrating the rigor of causal inference with the flexibility of Machine Learning. This approach is designed to overcome the selection bias present in observational data and some of the restrictive assumptions of standard statistical approaches. 
In this study, we analyze observational survey data from smallholder maize farmers in Tanzania (2023-24 season) using a Double Machine Learning approach to estimate conditional average treatment effects, identifying how Nitrogen and Phosphorus fertilizer response varies across different dose and timing regimes. Our findings show an average agronomic efficiency of 18 kg grain/kg of applied Nitrogen and 60 kg grain/kg of applied Phosphorus; results that closely align with established benchmarks from regional field trials. More importantly, our model captures management-driven heterogeneity. The results demonstrate that split applications of  Nitrogen –at planting/emergence and two times before silking– are more likely to provide higher efficiencies, while Phosphorus reaches peak efficiency when applied during the earliest development stage. Furthermore, the estimated dose-response curves exhibit characteristic diminishing returns; this showcases the framework’s ability to recover complex non-linear biophysical patterns. The successful recovery of these well-known agronomic insights from noisy observational data serves as a validation of the Causal Machine Learning framework for this specific context. This success demonstrates the potential to address increasingly complex agronomic challenges, utilizing existing datasets to identify site-specific patterns that provide a robust foundation for personalized optimal management.