Average versus expected causal effects under interference: Modeling assumptions and estimation bottlenecks
Authors: Vishesh Karwa, Edoardo M. Airoldi, Donald B. Rubin
Organizations: Department of Statistics, Operations, and Data Science, Temple University · Data Science Institute, Temple University, Philadelphia, PA 19122, USA · Department of Statistics, Harvard University, Cambridge, MA 02138, USA
Causal inference under treatment interference requires assumptions to reduce the exponentially large number of potential outcomes. We develop hierarchical finite-population models for potential outcomes by separating these assumptions into three layers: interference neighborhoods, exposure mappings, and outcome models. Given an interference neighborhood and exposure mapping, we show that potential outcomes admit a nonparametric decomposition into direct, spillover, and treatment--spillover interaction components. This decomposition provides a common coordinate system for nonparametric, semiparametric, and parametric models of finite-population potential outcomes. For a given potential-outcome model, we define two classes of estimands. Average causal effects (ACEs), such as direct and total treatment effects, contrast potential outcomes under fixed treatment--exposure conditions. Expected causal effects (ECEs), such as expected direct and indirect effects, contrast expectations of potential outcomes under a specified allocation policy. For both classes, we bound the variance of the canonical Horvitz--Thompson estimator by three terms: a diagonal weight term, a term from overlapping interference neighborhoods, and a residual design-dependence term. Our upper and lower bounds show that under graph-unaware designs, nonparametric design-unbiased estimation of ACEs can require sample sizes growing exponentially in the interference degree, whereas polynomial growth can suffice for design-aligned ECEs. Under neighborhood interference, we construct graph-aware independent-set designs with variance bounds polynomial in the maximum degree for direct and total treatment effects on every graph, whereas graph cluster randomization can incur worst-case variance exponential in the degree.
Figures & tables
Representation
Target conditions
HT H
LB H
HT variance scale
Arbitrary interference
(1,0−i) vs. (0,0−i)
eΘ(n)
eΘ(n)
Vn(eΘ(n),n)
Neighborhood interference
(1,0Ni) vs. (0,0Ni)
eΘ(dmax)
eΘ(dmax)
Vn(eΘ(dmax),b2,avg)
Count/fraction exposure
(Zi,Si)=(1,0) vs. (0,0)
eΘ(dmax)
eΘ(dmax)
Vn(eΘ(dmax),b2,avg)
Binary any-treated exposure
(Zi,Ui)=(1,0) vs. (0,0)
eΘ(dmax)
eΘ(dmax)
Vn(eΘ(dmax),b2,avg)
Table 1: Direct-effect ACEs under Bernoulli assignment. The rows show the DTE target under arbitrary interference, neighborhood interference, and count/fraction or binary exposure representations.
Representation
Target conditions
HT H
LB H
HT variance scale
Arbitrary interference
(1,1−i) vs. (0,0−i)
eΘ(n)
eΘ(n)
Vn(eΘ(n),n)
Neighborhood interference
(1,1Ni) vs. (0,0Ni)
eΘ(dmax)
eΘ(dmax)
Vn(eΘ(dmax),b2,avg)
Count/fraction exposure
(Zi,Si)=(1,di) vs. (0,0)
eΘ(dmax)
eΘ(dmax)
Vn(eΘ(dmax),b2,avg)
Binary any-treated exposure
(Zi,Ui)=(1,1) vs. (0,0)
eΘ(dmax)
eΘ(dmax)
Vn(eΘ(dmax),b2,avg)
Table 2: Total-effect ACEs under Bernoulli assignment. Exposure-mapping rows show representative treated-exposure versus no-exposure targets.
Figure 1: Design root-mean-square error of the Horvitz–Thompson estimator under a Bernoulli design ( p=1/2 ) against the mean interference degree, on Erdős–Rényi graphs ( n=2000 ); the vertical axis is logarithmic. The direct treatment effect (fixed-exposure ACE, red) follows the exponential inverse-propensity scale (red dashed; Proposition 20 ); the expected direct effect (aligned ECE, blue) is horizontal; the expected indirect effect (graph-aware ECE, green) grows polynomially.
Figure 2: The difference-in-means estimator against two targets as the sample grows (Erdős–Rényi, mean degree 8 , p=1/2 ). Its error for the expected direct effect decays at the parametric rate; its error for the direct treatment effect settles at the policy-averaged interaction floor Cˉπ set by the treatment–exposure interaction that own-treatment contrasts cannot remove.
Figure 3: Estimand-aligned design for the direct treatment effect (Erdős–Rényi, n=2000 , p=1/2 ). Under a Bernoulli design the Horvitz–Thompson error grows exponentially in the mean interference degree; under an independent-set design on the same graphs, which reveals Yi(1,0) and Yi(0,0) for ego units, it stays nearly flat.
Figure 4: Heavy-tailed interference (Barabási–Albert, n=3000 ). Left: the inverse propensity of the direct-effect exposure condition grows exponentially in a unit’s degree, exceeding 1070 for the highest-degree unit. Right: the probability that a unit ever reveals the relevant direct-effect potential outcome, in 104 assignment draws, collapses to zero for high-degree units, whereas the expected direct effect uses every unit.
Appendix figures & tables1 asset
Supplementary material from the paper’s appendix.
Appendix
Graph sequence
ACE scale An
EDE dependence cavg
Bounded degree
O(n−1)
O(1)
Second-neighborhood regular, degree d
eγd/n
d2
Sparse Erdős–Renyi, mean degree λ
approximately eγλ/n
approximately λ+λ2
Dense block model, block size s
approximately eγs/n
approximately s
Star graph
dominated by hub term eγn/n2
n
Complete graph
eγn/n
n
Appendix
Table 3: Generic comparison of the ACE inverse-propensity scale and the EDE dependence scale under Bernoulli assignment. The ACE column corresponds to a direct-effect target combination such as (1,0Ni) . The EDE column gives the order of cavg .
Estimating causal effects under network interference typically assumes that the network used for training and the network used for deployment coincide. In practice, an intervention is run on one population while the question of interest concerns a different population, and the two generally differ in topology, node-covariate composition, and spillover pathways. Transporting a causal effect across networks is therefore a data-fusion problem that no existing algorithm solves. We employ a selection diagram, extended to the network setting so that covariate shift and structural network shift enter as separate selectors, and derive from it a transport formula for the direct, spillover, and total effects in the deployment population. Each formula makes explicit which interventional mechanism is assumed invariant and which observational distribution must be reweighted. We then turn the formulas into TranCE (Transported Causal Effects), a doubly-robust algorithm combining an interventional outcome model, a domain density-ratio correction, and cross-fitted inference. Extensive experiments on two semi-synthetic benchmarks derived from real-world social networks and on a fully real weather-insurance field experiment, where the transported effects are checked against held-out randomized estimates, confirm the effectiveness of our approach. Our findings have the potential to improve intervention strategies in networked systems, particularly in social networks and public health.
Xiaojing Du, Jiuyong Li, Lin Liu +3
Adelaide University, Adelaide, Australia · Hainan University, Haikou, China
Estimating causal effects from observational network data requires accounting for both individual treatment and neighborhood exposure. Existing outcome models commonly incorporate these intervention variables into a joint predictor and estimate causal effects through prediction contrasts, leaving the corresponding response components implicit. We propose the Network Effect Decomposition Network (NEDNet), which parameterizes the conditional outcome response as a query-invariant pre-treatment baseline plus separate treatment- and exposure-response pathways. Path-specific neighborhood contextualization supplies local pre-treatment information to the two pathways. We characterize the fitted response through a unique reference-normalized decomposition, which is equivalent to a zero mixed treatment--exposure contrast. Experiments on semi-synthetic benchmarks derived from BlogCatalog and Flickr show that NEDNet achieves the lowest PEHE among the compared methods for individual, peer, and total effects across homogeneous and covariate-heterogeneous response settings.
Zhirong Huang, Wenzhen Zhang, Guangquan Lu +3
School of Computer Science and Engineering, Guangxi Normal University; the Guangxi Key Lab of MIMS; and the MOE Key Lab of EBIT, Guilin, Guangxi 541004, China · School of Computer Science and Technology, Hainan University, Haikou, Hainan 570228, China
We study causal inference under outcome interference for sequential, observational settings. Specifically, we consider settings where the binary outcomes over N units are Markovian across T time steps. At each time step, the outcomes of N units have dependencies captured through an Ising model; each outcome is also impacted through an external field capturing the effects of its treatment as well as latent confounders. Similar to panel data literature, these latent confounders are modeled to have a low-rank factor structure. Our data is a single sample from this high-dimensional distribution. To estimate causal quantities of interest, we provide a computationally efficient method based on Maximum Pseudo-Likelihood Estimation (MPLE) for learning the model parameters. Under mild assumptions, we establish non-asymptotic consistency for parameter estimation and show this translates to faithful estimation of causal quantities of interest after sampling from the learned model. We demonstrate the efficacy of the method through synthetic experiments as well as a real-world case-study investigating causal effects of vaccine rates on COVID-19 death rates within US counties nationwide.
Phevos Paschalidis, Constantinos Daskalakis, Devavrat Shah
EECS, LIDS, CSAIL, MIT · EECS, CSAIL, MIT · EECS, IDSS, LIDS, SDSC, MIT