📊 statistics

Heterogeneous Treatment Effects under Network Interference: A Nonparametric Approach Based on Node Connectivity

This paper proposes KECENI, a doubly robust nonparametric framework that estimates node-wise counterfactual means to uncover heterogeneous treatment effects driven by network structure under interference, demonstrating its efficacy through microfinance data analysis.

Original authors: Heejong Bong, Colin B. Fogarty, Elizaveta Levina, Ji Zhu

Published 2026-08-07

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Heejong Bong, Colin B. Fogarty, Elizaveta Levina, Ji Zhu

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you are trying to figure out why some people in a town get sick while others stay healthy. In the old days, scientists might have looked at each person in isolation, asking, "Did this person eat the bad apple?" But in the real world, people are connected like a giant, invisible web of friendships, family ties, and shared spaces. If your best friend eats the bad apple, you might get sick too, even if you never touched it. This is called network interference: your outcome depends not just on your own choices, but on what your neighbors are doing.

To make sense of this, researchers usually look for the "average" effect. They ask, "On average, does the medicine work?" But averages can hide the truth. Just like a weather report saying "it's 70 degrees" doesn't tell you if you need a coat or sunglasses, an average effect might miss the fact that the medicine works wonders for some people but does nothing for others. The big question is: how does the shape of your social web change the result? Does being the most popular person in the group make you more likely to catch a trend? Does being isolated protect you? Answering this is crucial for policymakers who want to target help to the people who need it most, rather than just throwing a blanket over the whole crowd.

Enter a new team of statisticians who decided to stop looking at the crowd as a blurry blob and start looking at the individual threads of the web. They developed a clever new tool called KECENI (Kernel Estimator of Causal Effect under Network Interference). Think of KECENI as a super-powered microscope that doesn't just look at one person, but at their entire neighborhood of friends, neighbors, and acquaintances all at once.

Usually, figuring out what would have happened to a specific person if their friends had acted differently is a nightmare. It's like trying to predict the weather in your town by only looking at the sky in a town 100 miles away; the conditions are too different. The authors realized that because there are so many possible combinations of friends and treatments, you will rarely find two people with the exact same situation. So, instead of demanding an exact match, KECENI uses a "smoothing" technique. It's like blending colors on a palette: if you can't find a perfect shade of blue, you mix the closest blues you have to get the right color. KECENI blends data from people who have similar social circles and treatment histories to estimate what would happen to a specific individual under a specific scenario.

The paper shows that this method is not just a guess; it is mathematically robust. The authors proved that as you get more data, the estimates get closer and closer to the truth, and they can even calculate how sure they are about those numbers. They tested this in two ways. First, they ran computer simulations where they knew the answer beforehand. In these tests, KECENI successfully found the right answers, while older methods that ignored the complexity of the network got confused and gave wrong results. Second, they applied KECENI to real-world data from rural villages in India, studying how joining a savings group (a self-help group) affected people's financial risk.

Here is what they found: On average, the savings group didn't seem to change much for the whole village. But when KECENI looked at the details, a hidden story emerged. For people with very few connections (isolated nodes), the group had almost no effect. But for people with a small number of connections (1 to 4 friends), joining the group had a significant positive effect. It seems that having just a few trusted neighbors made the difference between success and failure, while having a huge, crowded network diluted that benefit.

The paper also highlights a major advantage of their approach: it doesn't force you to guess the "right" way to measure similarity between people beforehand. Instead, it lets the data decide which similarities matter most, using a smart cross-validation process (like trying on different pairs of glasses to see which one makes the world clearest). This means researchers don't have to rely on shaky assumptions about how the network works; they can let the data speak.

In short, this paper suggests that when we study how things spread through a network—whether it's a virus, a new idea, or a financial habit—we need to stop treating everyone as an average statistic. By using KECENI, we can see the unique impact of the network on every single person, revealing that the "best" intervention might look very different depending on who you are and who you know. It's a shift from asking "What works for everyone?" to "What works for you, given your place in the web?"

Technical Summary: Heterogeneous Treatment Effects under Network Interference

Problem Statement
Causal inference in network settings is complicated by network interference, where an individual's outcome depends not only on their own treatment but also on the treatments received by their neighbors. While existing literature extensively addresses the estimation of population-averaged treatment effects (such as Average Treatment Effects, Average Direct Effects, and Average Spillover Effects), these approaches often obscure heterogeneity induced by local network structures. Furthermore, identifying node-specific counterfactual means is theoretically challenging because, without additional assumptions, a single observation of network-linked units does not provide sufficient data to nonparametrically identify the counterfactual outcome for a specific node under a specific neighborhood intervention. Previous attempts to address this, such as autoregressive chain graph models, rely on restrictive parametric assumptions.

Methodology: KECENI
The authors propose KECENI (Kernel Estimator of Causal Effect under Network Interference), a nonparametric framework designed to estimate node-wise counterfactual means, denoted as $\theta_{i^*}(t^*_{N_{i^*}})$ . This estimand represents the expected outcome for a target node $i^*$ under a hypothetical intervention assignment $t^*_{N_{i^*}}$ to its neighborhood, conditional on the observed network $G$ .

The methodology proceeds in three main steps:

Pseudo-Outcome Construction: The authors adapt the pseudo-outcome approach from Kennedy et al. (2017) to the network setting. They define a pseudo-outcome $\xi_i$ that combines an outcome regression function $\mu_i$ and a propensity score $\pi_i$ . This construction ensures double robustness: the pseudo-outcome is an unbiased estimator of the node-wise counterfactual mean if either the outcome model or the propensity score model is correctly specified.
Kernel Smoothing: To address the "limited overlap" problem—where exact matches for a target node's neighborhood configuration are rare in the observed data—KECENI pools information from nodes with similar neighborhood configurations. Similarity is quantified by a user-specified dissimilarity metric $\Delta$ (e.g., based on exposure mappings or Gromov-Wasserstein distances). The estimator applies kernel smoothing to the pseudo-outcomes of these similar units.
Inference: The paper develops new empirical process theory to handle the dependence structure inherent in network data. This theory establishes the consistency and asymptotic normality of the KECENI estimator, accounting for the fact that the same data is used to estimate nuisance parameters and the causal estimand. A dependence-adjusted empirical sandwich variance estimator is proposed to construct valid confidence intervals.

Key Contributions

Shift in Estimands: The paper shifts the focus from population-averaged estimands to node-wise counterfactual means as the primary targets of inference. This allows for the preservation and analysis of heterogeneity driven by local network features (e.g., node degree, clustering).
Nonparametric Identification: Unlike previous methods that rely on parametric models (e.g., autoregressive structures) to identify node-wise effects, KECENI offers a nonparametric approach. It leverages the connection between node-wise estimation and dose-response function estimation, using smoothing to borrow strength across similar neighborhood configurations.
Double Robustness under Interference: The method extends the double robustness property to network interference settings, protecting against model misspecification in either the outcome or treatment assignment mechanism.
New Theoretical Toolkit: The authors develop novel empirical process theory for weighted averages under network dependence. This theory controls stochastic terms in the presence of network dependence and increasing network complexity, a necessary step for establishing valid asymptotic inference in this setting.
Data-Driven Tuning: The framework facilitates data-driven selection of the dissimilarity metric, kernel, and bandwidth via cross-validation, treating the choice of similarity metric as a model selection problem.

Results

Simulations:
- Consistency and Normality: Simulations demonstrate that KECENI estimates are consistent and asymptotically normal, with confidence intervals achieving nominal coverage rates.
- Double Robustness: In misspecification scenarios, KECENI maintains low root-mean-square error (RMSE) as long as one of the nuisance models (outcome or propensity) is correctly specified, whereas G-computation estimators fail if the outcome model is misspecified.
- Comparison with Existing Methods: When the true data-generating process involves complex interactions that are unknown or misspecified in parametric models (e.g., AUTOGNET and TMLENET), KECENI successfully identifies treatment effects using nonparametric smoothing and fused Gromov-Wasserstein distances. Conversely, when parametric models are correctly specified, KECENI performs comparably to these methods.
Empirical Application (Indian Microfinance):
- Applied to microfinance data from rural India (Banerjee et al., 2013), the method estimated the impact of joining savings self-help groups (SHGs) on financial risk tolerance.
- While the average spillover effect (ASpE) across the population was found to be statistically insignificant (aligning with previous studies), the node-wise analysis revealed significant heterogeneity.
- Specifically, a significant positive spillover effect was detected for treated nodes with few connections (low degree), whereas highly connected nodes showed no such effect. This suggests that less connected individuals are more susceptible to spillover effects from their neighbors' participation, a nuance masked by population-averaged estimates.

Significance and Claims
The paper claims that KECENI provides a flexible framework for understanding how treatment effects vary with network connectivity, offering granular insights for intervention targeting and policy evaluation. By moving beyond average effects, the method reveals that spillover effects are not uniform but depend critically on local network characteristics.

The authors emphasize that their approach addresses a practical limitation of existing methods: the ad hoc selection of exposure mappings. By allowing for data-driven selection of similarity metrics through cross-validation, KECENI offers a principled path toward causal model selection under network interference.

The paper is modest in its scope, noting that while it identifies heterogeneity associated with network structure, it does not claim to disentangle causal effects of the network structure itself from homophily (selection effects), which would require stronger assumptions or longitudinal network data. The theoretical guarantees rely on specific regularity conditions regarding network dependence (e.g., bounded local dependence radius) and the smoothness of the underlying functions.

More like this