Can Satellite Images Predict Treatment Effect Heterogeneity? A Robust Inference Approach Restricted; Files Only

Abstract

Randomized controlled trials (RCTs) are widely used in the social sciences to study the effect of policy interventions. When participants reside across separate geographic areas, it might be reasonable to expect the effects of interventions to differ because of factors such as local infrastructure, topography, and neighborhood amenities. Satellite images present an exciting low cost source of geographic information to study heterogeneity. We use experimental data from the Youth Opportunities Program in Uganda, and apply convolutional neural networks to map participant geo-referenced satellite images to predictions of hours worked for treated and control individuals, respectively. We use the model to estimate Conditional Average Treatment Effects (CATEs). Our key contribution is to propose a robust inference approach that can be used to (i) test whether CATEs differ by geography, (ii) compare the variance of image-based CATEs against those obtained from pre-randomization survey data. Our testing approach has conservative coverage guarantees, even under misspecification, which means that researchers can obtain valid tests of homogeneity (if slightly underpowered) when the neural network model is tuned incorrectly.

Table of Contents

1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1

2 Background . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 5

2.1 Framework . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 5

2.2 Neural Network . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 6

2.2.1 Convolutional Neural Network . . . . . . . . . . . . . . . . . . . 7

2.2.2 Model Architecture . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 8

2.3 Estimating Debiased Average Effects . . . . . . . . . . . . . . . 8

2.4 Methodology . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 9

2.5 Data . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 10

3 Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 13

3.1 Regression on YOP Study . . . . . . . . . . . . . . . . . . . . . . . . . 13

3.2 VCATE Approach . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .15

3.2.1 Further Analysis . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 15

4 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 18

Bibliography . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 19

About this Honors Thesis

Rights statement

• Permission granted by the author to include this thesis or dissertation in this repository. All rights reserved by the author. Please contact the author for information regarding the reproduction and use of this thesis or dissertation.

School	Emory College
Department	Quantitative Science
Degree	B.S.
Submission	Honors Thesis
Language	English
Research Field	Computer Science Information Science Statistics
Palavra-chave	Convolutional Neural Network Treatment Effect Heterogeneity Conditional Average Treatment Effect Satellite Images
Committee Chair / Thesis Advisor	Alejandro Sanchez-Becerra, Emory University
Committee Members	Zhiyun Gong, Emory University Kevin McAlister, Emory University

Última modificação

Primary PDF

Thumbnail	Title	Date Uploaded	Actions
	File download under embargo until 22 May 2027	2025-04-08 22:59:56 -0400	File download under embargo until 22 May 2027

Supplemental Files

Thumbnail	Title	Date Uploaded	Actions