Causal Inference in Connected Populations with Contagion
Subhankar Bhadra, Department of Statistics, The Pennsylvania State University · 2025
Read Causal Inference in Connected Populations with Contagion by Subhankar Bhadra, Department of Statistics, The Pennsylvania State University (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-2504-06108.txt.
Introduction
Subhankar Bhadra and Michael Schweinberger study cause and effect when outcomes spread like a contagion in a 2025 paper. In linked groups, one person's result can sway a neighbor's, so a treatment's effect leaks and grows through the web. Past work on linked cause mostly set that spread aside, leaving its mark on estimates unknown.
The pair derive closed forms for effects under spillover and contagion in simple setups. The forms show main, spillover, and contagion effects braid together, and contagion can shrink or swell the total effect. Model based reads that ignore contagion drift off target in the long run. Design based reads break too when contagion runs free. The authors sketch fixes for both camps.
Catalog record
| CREATOR | Subhankar Bhadra, Department of Statistics, The Pennsylvania State University |
|---|---|
| YEAR | 2025 |
| COLLECTION | PAPERS |
| FORMAT | TXT |
| FILE SIZE | 74 KB |
| VERSION | 1.0 |
| LICENSE | http://creativecommons.org/licenses/by/4.0/ |
| RIGHTS | CC BY |
| SOURCE | arXiv |
Source: https://arxiv.org/abs/2504-06108
Browse the full library · About the Tower of Babel
Full text
The opening of the work, from the library's hosted text:
Causal Inference in Connected Populations with Contagion Subhankar Bhadra, Department of Statistics, The Pennsylvania State University Abstract We address a gap in the literature on causal inference in connected populations: while there is a growing body of work on estimating causal effects in connected populations, little is known about how contagion and other network processes impact causal effects and inference. Contagion and other network processes imply that the outcomes of units affect the outcomes of other units, which enables the effects of interventions to propagate throughout connected populations and complicates insight into causal effects and inference. We offer novel insight into how contagion and other network processes impact causal effects and inference based on closed-form expressions for causal effects under spillover and contagion. These closed-form expressions reveal that the main effects of interventions, spillover, and contagion are intertwined even in the simplest possible settings, and that contagion can decrease or increase causal effects. We discuss statistical implications, including asymptotic bias of model-based estimators ignoring contagion, the violation of neighborhood exposure assumptions by unrestricted contagion and its effect on design-based estimators, and possible remedies. Keywords: Interference, Markov Random Field, Peer Influence, Random Graph, Spillover 1 Introduction Insight into the effects of public health, economic, and financial interventions in connected populations is complicated by spillover, contagion, and other network processes that allow the effects of interventions to propagate throughout connected populations.For example, advertisers may be interested in the effect of exposing teenagers on social media to advertisements of designer clothes ( X\in\{0,1\} ) on purchases of designer clothes ( Y\in\mathbb{R} ). The effect of advertisements on purchases of designer clothes can spill over from exposed teenagers to other teenagers, as demonstrated in Figure 1 . Y_{1} Y_{2} X_{1} X_{2} Y_{1} Y_{2} X_{1} X_{2} Y_{1} Y_{2} X_{1} X_{2} Main Effect of Intervention + Effect of Spillover + Effect of Contagion Figure 1 : Interventions X_{1},X_{2} and outcomes Y_{1},Y_{2} of connected units 1,2 , which can be viewed as a single observation from the stationary distribution of a time-indexed stochastic process ( Lauritzen & Richardson 2002 ) . The main effect of interventions and spillover are represented by directed lines. Contagion is represented by an undirected line: if Y_{1},Y_{2} are observed at a single time point, we cannot distinguish whether Y_{1} affects Y_{2} or vice versa. First, if exposed teenagers share the advertisements with friends, their friends may purchase the advertised designer clothes (spillover). Second, if exposed teenagers purchase the advertised designer clothes, their friends may observe them and may purchase them as well (contagion). Both spillover and contagion can increase purchases of designer clothes and are of interest to advertisers and other decision makers. Contagion implies that the outcomes of units can affect the outcomes of other units, enabling the effects of interventions to propagate throughout connected populations. While the effects of interventions can be estimated using existing methods ( Sobel 2006 , Rosenbaum 2007 , Hudgens & Halloran 2008 , Toulis & Kao 2013 , van der Laan 2014 , Ugander et al. 2013 , Aronow & Samii 2017 , Choi 2017 , Tchetgen Tchetgen et al. 2021 , Sävje et al. 2021 , Forastiere et al. 2021 , Li & Wager 2022 , Ogburn et al. 2024 , e.g.,) , little is known about how contagion and other network processes impact causal effects and inference. 1.1 Contributions An empirical study by Lee & Ogburn (2021) highlights the consequences of ignoring contagion and other network processes. We provide theoretical insight into how network processes—including spillover and contagion—can impact causal effects and inference by deriving closed-form expressions for the causal effect \tau_{T}\coloneqq\tau_{E}+\tau_{A} composed of \begin{array}[]{llllllllll}\tau_{E}&\coloneqq&\lim\limits_{N\rightarrow\infty}\dfrac{1}{N}\displaystyle\sum\limits_{i=1}^{N}\mathbb{E}[Y_{i}(x_{i}=1;\bm{X}_{-i})-Y_{i}(x_{i}=0;\bm{X}_{-i})]\\ \tau_{A}&\coloneqq&\lim\limits_{N\rightarrow\infty}\dfrac{1}{N}\displaystyle\sum\limits_{i=1}^{N}\displaystyle\sum\limits_{j=1:j\neq i}^{N}\mathbb{E}[Y_{j}(x_{i}=1;\bm{X}_{-i})-Y_{j}(x_{i}=0;\bm{X}_{-i})],\end{array} where \bm{Y}(\bm{x})\in\mathbb{R}^{N} are potential outcomes given interventions \bm{x}\in\{0,1\}^{N} , and \bm{x}_{-i}\coloneqq(x_{1},\ldots,x_{i-1},x_{i+1},\ldots,x_{N})\in\{0,1\}^{N-1} . The terms \tau_{E} and \tau_{A} quantify the effects of interventions x_{i} on the potential outcomes Y_{i}(x_{i};\bm{x}_{-i}) of units i (ego effect) and the potential outcomes Y_{j}(x_{i};\bm{x}_{-i}) of other units j (alter effect), averaged over all \bm{x}_{-i}\in\{0,1\}^{N-1} and i . Closed-form expressions for \tau_{E} , \tau_{A} , and \tau_{T} exist in the absence of contagion ( Hu et al. 2022 ) , but are challenging and unavailable when there is contagion along with spillover. We make the following contributions: 1. We obtain closed-form expressions for \tau_{E} , \tau_{A} , and \tau_{T} in the presence of spillover and contagion, which are the first such results to the best of our knowledge (Section 2 ). 2. The closed-form expressions for \tau_{E} , \tau_{A} , and \tau_{T} reveal that the main effects of interventions, spillover, and contagion are intertwined even in the simplest possible scenarios, and that contagion can decrease or increase \tau_{T} , but it cannot cancel it or change its sign. If the effects of interventions vary across communities, then larger communities can contribute more than smaller ones (Sections 3 ). 3. We discuss statistical implications, including asymptotic bias of model-based estimators of \tau_{E} , \tau_{A} , and \tau_{T} ignoring contagion, the violation of neighborhood exposure assumptions by unrestricted contagion and its effect on design-based estimators of the total causal effect \tau_{T} , and possible remedies (Sections 4 and 5 ). These results suggest that contagion poses challenges, but offers opportunities: e.g., if there is contagion, policy makers can recruit influencers and offer incentives for building communities to amplify the effects of interventions. 1.2 Notation The set \bm{Y}(.)\coloneqq\{\bm{Y}(\bm{x}):\,\bm{x}\in\{0,1\}^{N}\} denotes the set of all potential outcomes. By convention, the observed outcomes \bm{Y} are related to the potential outcomes \bm{Y}(\bm{x}) as follows: \begin{array}[]{llllllllll}\bm{Y}=\displaystyle\sum\limits_{\bm{x}\in\{0,1\}^{N}}\bm{Y}(\bm{x})\,\mathbb{I}(\bm{X}=\bm{x}),\end{array} where \mathbb{I}(.) is an indicator function. The N\times N matrix \bm{Z}\in\{0,1\}^{\binom{N}{2}} consists of connection indicators Z_{i,j} , where Z_{i,j}=1 indicates that i and j are connected and Z_{i,j}\coloneqq 0 otherwise. The connection indicators satisfy Z_{i,i}\coloneqq 0 and Z_{i,j}=Z_{j,i} , and the degree of unit i is N_{i}\coloneqq\sum_{j=1}^{N}Z_{i,j} . By convention, C/N_{i}\coloneqq 0 when N_{i}=0 for all C\in\mathbb{R} . Probability measures and expectations are denoted by \mathbb{P} and \mathbb{E} . We denote the \ell_{2} -norm of vectors \bm{a}\in\mathbb{R}^{d} ( d\geq 1 ) by |\!|\bm{a}|\!|_{2}\coloneqq\sqrt{\sum_{i=1}^{d}a_{i}^{2}} and the spectral norm of matrices \bm{A}\in\mathbb{R}^{d\times d} by |\!|\!|\bm{A}|\!|\!|_{2}\coloneqq\sup_{\bm{u}\in\mathbb{R}^{d}:\;|\!|\bm{u}|\!|_{2}=1}\,|\!|\bm{A}\,\bm{u}|\!|_{2} . The vector \bm{1}_{d}\in\mathbb{R}^{d} is the d -vector of 1 ’s and \bm{I}_{d}\in\mathbb{R}^{d\times d} is the identity matrix of order d . The matrix \mbox{diag}\left(r_{1},\ldots,r_{d}\right)\in\mathbb{R}^{d\times d} is the d\times d diagonal matrix with diagonal elements r_{1},\ldots,r_{d}\in\mathbb{R} . A sequence of random variables R_{N} is said to be of order o_{P}(1) if R_{N} converges to 0 in probability as N\to\infty . 2 Closed-Form Expressions for Causal Effects To provide insight into interventions in connected populations in the presence of spillover and contagion, we provide closed-form expressions for \tau_{E} , \tau_{A} , and \tau_{T} under two conditions. Condition 1 . \bm{X}\mathchoice{\mathrel{\hbox to0.0pt{$\displaystyle\perp$\hss}\mkern 2.0mu{\displaystyle\perp}}}{\mathrel{\hbox to0.0pt{$\textstyle\perp$\hss}\mkern 2.0mu{\textstyle\perp}}}{\mathrel{\hbox to0.0pt{$\scriptstyle\perp$\hss}\mkern 2.0mu{\scriptstyle\perp}}}{\mathrel{\hbox to0.0pt{$\scriptscriptstyle\perp$\hss}\mkern 2.0mu{\scriptscriptstyle\perp}}}\bm{Y}(.)\mid\bm{Z} and \mathbb{P}(X_{i}=x_{i},\bm{X}_{-i}=\bm{x}_{-i}\mid\bm{Z}=\bm{z})>0 whenever \mathbb{P}(\bm{X}_{-i}=\bm{x}_{-i}\mid\bm{Z}=\bm{z})>0 for all (\bm{x},\bm{z})\in\{0,1\}^{N+\binom{N}{2}} . Condition 2 makes the standard assumption of ignorability and can be extended to conditional ignorability given covariates. To capture contagion along with spillover and provide insight into interventions in connected populations, we need an additional assumption. Condition 2 . The conditional expectations of outcomes Y_{i}\mid(\bm{X},\bm{Y}_{-i},\bm{Z})=(\bm{x},\bm{y}_{-i},\bm{z}) are \begin{array}[]{llllllllll}\mathbb{E}[Y_{i}\mid(\bm{X},\bm{Y}_{-i},\bm{Z})=(\bm{x},\bm{y}_{-i},\bm{z})]&=&\alpha_{i}+\beta_{i}\,x_{i}+\gamma_{i}\,\dfrac{1}{N_{i}}\displaystyle\sum\limits_{j=1}^{N}x_{j}\,z_{i,j}+\delta_{i}\,c_{N,i}(\bm{z})\displaystyle\sum\limits_{j=1}^{N}y_{j}\,z_{i,j},\end{array} where \alpha_{i}\in\mathbb{R} is an intercept, \beta_{i}\in\mathbb{R} is the main effect of interventions, \gamma_{i}\in\mathbb{R} is the effect of spillover, \delta_{i}\in\mathbb{R} is the effect of contagion, and c_{N,i}(\bm{z})\in(0,\infty) are scaling constants ( i=1,\ldots,N ). Without loss \alpha_{i}\coloneqq 0 , because \alpha_{i} does