Spillover Gridlock:

The presented study addresses the challenge of identifying spillovers in spatial interventions through the generation of counterfactual rasters—an innovative method that leverages convolutional neural networks to evaluate the impact of spatial policies. The thesis proposes a synthetic control methodology that generates counterfactual images from a data grid, enabling the observation of treatment effects (e.g., fertilizer application) and their spatial diffusion across pixels—without the need to predefine transmission patterns of the spillovers. This work falls within the literature of spatial econometrics, where treatment effects are often biased due to strong assumptions about spatial propagation, which this approach seeks to avoid through the flexibility of convolutional neural networks. To assess the effects of interest, a potential outcomes framework is introduced that distinguishes between total effects (a combination of direct and spillover effects) and emphasizes the importance of appropriately selecting treatment matrices and rules. Finally, a comparative simulation with traditional methods such as difference-in-differences is conducted, highlighting the advantages of the proposed method in avoiding assumptions about spillover structure, thereby enabling more accurate estimation of spatial effects in controlled settings.