Primary Submission Category: Causal Inference and SUTVA/Consistencies Violations
Outcome Modeling in Design-Based Inference for Spatial Experiments
Authors: Arisa Sadeghpour, Erin Hartman,
Presenting Author: Arisa Sadeghpour*
In many spatial settings, randomized treatments have effects that bleed out, violating the standard non-interference assumption, and researchers often want to estimate how these spillover effects decay in space. Wang et al. (2023) present a design-based estimand for the average marginalized effect at a specific distance, which we call the “circle average marginalized effect” (CAME). The CAME is the average effect of switching intervention nodes from treatment to control on points along a circle a specific distance away from those nodes, marginalizing over possible realizations to other nodes. Since it is impossible to observe outcomes at every point in space, we argue that estimating CAME necessitates using outcome models. Through simulations, we explore how the choice of outcome model impacts the bias and variance of the Horvitz-Thompson and Hajek estimators for the CAME in different scenarios. We find that when using modeled outcomes as in practice, even the Horvitz-Thompson estimator accrues some bias, under design-based inference, as a result of the modeling. We illustrate the importance of outcome modeling in estimating the CAME using randomized field experiments on policing and crime.