Balancing Interference and Correlation in Spatial Experimental Designs: A Causal Graph Cut Approach

Jin Zhu, Jingyi Li, Hongyi Zhou, Yinan Lin, Zhenhua Lin, Chengchun Shi · 2025

Read Balancing Interference and Correlation in Spatial Experimental Designs: A Causal Graph Cut Approach by Jin Zhu, Jingyi Li, Hongyi Zhou, Yinan Lin, Zhenhua Lin, Chengchun Shi (2025) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2505-20130.txt.

Introduction

Jin Zhu and five co-authors tackle how to lay out spatial experiments when treatment spills over between nearby units, in a 2025 paper. Spillover, called interference, muddies causal estimates, yet units close together also tend to behave alike, which helps precision. Good design must balance the two forces rather than simply separating treated and control areas.

The team built a stand-in formula for the estimator's mean squared error that can be optimized with classic graph cut methods. The approach handles moderate to large spillover, adapts to different patterns of spatial similarity, and stays fast to compute. Tests on synthetic settings and a city-scale ridesharing simulator supported the design, and the authors released code.

Catalog record

CREATORJin Zhu, Jingyi Li, Hongyi Zhou, Yinan Lin, Zhenhua Lin, Chengchun Shi
YEAR2025
COLLECTIONPAPERS
FORMATTXT
FILE SIZE67 KB
VERSION1.0
LICENSEhttp://creativecommons.org/licenses/by/4.0/
RIGHTSCC BY
SOURCEarXiv

Source: https://arxiv.org/abs/2505-20130

Browse the full library · About the Tower of Babel

Full text

The opening of the work, from the library's hosted text:

Balancing Interference and Correlation in Spatial Experimental Designs: A Causal Graph Cut Approach Jin Zhu, Jingyi Li, Hongyi Zhou, Yinan Lin, Zhenhua Lin, Chengchun Shi Abstract This paper focuses on the design of spatial experiments to optimize the amount of information derived from the experimental data and enhance the accuracy of the resulting causal effect estimator. We propose a surrogate function for the mean squared error (MSE) of the estimator, which facilitates the use of classical graph cut algorithms to learn the optimal design. Our proposal offers three key advances: (1) it accommodates moderate to large spatial interference effects; (2) it adapts to different spatial covariance functions; (3) it is computationally efficient. Theoretical results and numerical experiments based on synthetic environments and a dispatch simulator that models a city-scale ridesharing market, further validate the effectiveness of our design. A python implementation of our method is available at https://github.com/Mamba413/CausalGraphCut . † † affiliationnotice: Equal contribution 1 Introduction Background . Before deploying any policy in practice, it is essential to evaluate its impact. This renders randomized control trials or online experiments extensively used for accurate policy evaluation. In numerous applications, the experiments involve multiple units that are either distributed across different spatial regions or connected through a network. The spatial or network dependencies inherent in these experiments pose two unique challenges for policy evaluation: (i) the existence of interference effects where interventions applied to one experimental unit may influence others, leading to the violation of the fundamental stable unit treatment value assumption (SUTVA) for causal inference ( Imbens & Rubin, 2015 ) ; (ii) the strong outcome correlations among units, which increase the variance of the causal effect estimator. We present the following examples to illustrate. Example 1: A/B testing in marketplaces. A/B testing is frequently used in two-sided markets such as Uber, Lyft, Airbnb and eBay to assess the impact of a newly developed product relative to a standard control. It has become the gold standard for companies in conducting data-driven decision making ( Johari et al., 2022 ) . Spatial interference is ubiquitous in marketplaces. For instance, ridesharing companies constantly offer subsidizing policies to drivers or passengers across different regions in a city. Applying a subsidizing policy in one location can attract drivers from neighboring regions to this area, potentially decreasing the driver supply in other regions. Thus, a subsidizing policy in one location can affect outcomes at other locations, inducing interference over space ( Shi et al., 2023a ) . Meanwhile, market features like online driver numbers and call order numbers form interconnected spatial networks, leading to considerable spatial correlation ( Luo et al., 2024 ) . Example 2: Environmental and epidemiological applications. Environmental and epidemiological studies frequently utilize spatial data to examine causal effects of certain interventions on health outcomes in different geographical regions ( Reich et al., 2021 ) . Spatial correlation and interference are likely to arise in these applications. For instance, in evaluating the impact of vaccination campaigns on disease incidence, infection rates exhibit strong correlations within regions. Furthermore, the vaccination coverage in a given region not only influences its own infection rate but also those in the neighboring regions ( VanderWeele et al., 2012 ; Perez-Heydrich et al., 2014 ) . Example 3: Experimentation in social networks. Social network sites such as Facebook and LinkedIn extensively conduct online experiments to evaluate user engagement or satisfaction from a new service or feature. In these experiments, the treatment applied to a given user may spill over to other users via underlying social connections (like friends), leading to network interference. Additionally, since connected users often exhibit similar behaviors, their outcomes are likely correlated ( Gui et al., 2015 ) . Contributions . This paper primarily focuses on the first application, aiming to identify the optimal design that minimizes the mean squared error (MSE) of the resulting average treatment effect (ATE) estimator. Nonetheless, the theories and methods developed herein can be adapted to the other two examples. Methodologically , our contribution lies in the development of a causal graph cut algorithm for designing experiments with spatial interference and correlation. The advances of our algorithm are summarized as follows: • Flexibility : Our proposal is more flexible than existing works that assume weak interference ( Viviano et al., 2023 ) . It accommodates moderate to large interference effects . • Adaptivity : Instead of employing existing minimax approaches ( Leung, 2022 ; Viviano et al., 2023 ; Zhao, 2024b ) that minimize the same worst-case MSE across applications with different covariance structures and can be overly conservative, our algorithm adapts to different correlation structures , providing a tailored solution to each individual application. • Computational efficiency : The design problem requires allocating treatment for each region and is inherently NP-hard, as the number of treatment allocation rules grows exponentially fast with respect to the number of regions. Unlike existing solutions that rely on computationally intensive integer programming ( Zhao, 2024b ) , we propose a new surrogate function for the MSE, which enables the use of spectral clustering-based graph cut algorithms for minimization , thus substantially enhancing computational efficiency. Theoretically , our analysis reveals that the two pivotal factors — interference and correlation — drive the optimal design in opposite directions . Specifically, experiments with large interference effects are best designed by assigning same policies to neighboring regions whereas those with strong spatial correlations benefit from allocating different policies. By incorporating the effects of the two factors into our surrogate function, the proposed algorithm effectively address both challenges, achieving desirable properties. Empirically , based on synthetic environments and a city-level simulator — constructed using physical models and a real dataset from a ridesharing company to realistically simulate driver and passenger behaviors, we demonstrate the superior performance of our design compared to existing state-of-the-art methods. Notably, in the simulator, the proposed ATE estimator achieves an MSE that is 3.5 times smaller than those of the benchmark approaches . 1.1 Related works This section reviews related works on spatial causal inference, experimental design and off-policy evaluation. Spatial causal inference . Our proposal is closely related to a growing line of research on causal inference in the presence of spatial interference. Depending on the underlying assumptions, the methodologies developed within this area can be broadly grouped into three categories: 1. The first category assumes partial interference where units are partitioned into clusters and interference is restricted to within these clusters, not between them ( Halloran & Struchiner, 1995 ; Sobel, 2006 ; Hudgens & Halloran, 2008 ; Tchetgen & VanderWeele, 2012 ; Crépon et al., 2013 ; Liu et al., 2016 ; Zigler & Papadogeorgou, 2021 ) . 2. The second category assumes neighborhood interference where interference is limited to a unit’s neighbors or a predefined local region ( Verbitsky-Savitz & Raudenbush, 2012 ; Bhattacharya et al., 2020 ; Fatemi & Zheleva, 2020 ; Ma & Tresp, 2021 ; Gao & Ding, 2023 ; Dai et al., 2024 ; Jiang et al., 2024 ; Yang et al., 2024 ) . 3. The last category considers more general interference structures ( Aronow & Samii, 2017 ; Forastiere et al., 2021 ; Puelz et al., 2022 ; Song & Papadogeorgou, 2024 ; Zhang et

Read the complete text