Pilot Analysis for: Learning to Encode Multi-level Dynamics in Effect Heterogeneity Estimation
Keywords: Causal inference, Randomized controlled trials, Earth observation, Causal representation learning
TL;DR: A novel multi-scale approach using satellite imagery at different sizes improves causal effect estimation in poverty alleviation studies across Peru, Honduras, and Ghana by capturing both household and neighborhood-level information.
Abstract: Earth Observation (EO) data are increasingly used in policy analysis by enabling granular estimation of treatment effects. However, a challenge in EO-based causal inference lies in balancing the trade-off between capturing fine-grained individual heterogeneity and broader contextual information. This paper introduces algorithms that address this challenge by combining satellite images of different sizes to estimate Conditional Average Treatment Effects (CATEs). This multi-scale approach employs Vision Transformer (ViT) models fine-tuned on satellite images, then applied to images of varying patch sizes to capture both household- and neighborhood-level information. We first perform simulation studies, showing how a multi-scale approach captures multi-level dynamics that single-scale ViT models fail to capture. We then apply the multi-scale method to two randomized controlled trials (RCTs) conducted in Peru and Uganda using Landsat satellite imagery. The Rank Average Treatment Effect (RATE) measure is employed in this pilot analysis to assess performance without ground truth individual treatment effects, with this analysis indicating that our dual-size inference technique improves the performance of deep learning models in EO-based CATE estimation.
Submission Number: 45
Loading