📊 statistics

Optimizing Treatment Allocation in Experiments with Network Interference

This paper proposes a network-aware treatment allocation framework that optimizes experimental design under network interference by balancing allocation and topology via a Fisher information-based criterion, solved through a scalable local search algorithm and validated through simulations and real-world applications.

Original authors: Zuhra F. S. Lebbe, Asim K. Dey

Published 2026-08-25

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Zuhra F. S. Lebbe, Asim K. Dey

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

In the world of scientific experiments, researchers often rely on a simple rule: what happens to one person should not affect the outcome for another. This idea, known as the stable unit treatment value assumption, works well when testing a new drug on isolated patients or a new fertilizer on separate plots of land. However, the real world is rarely so isolated. People live in communities, animals in herds, and devices in networks where actions ripple outward. When a person gets vaccinated, their neighbors are safer. When a social media user sees an ad, their friends might see it too. This phenomenon, called interference, breaks the old rules of experimental design because the result for any single individual depends not just on their own treatment, but on what their neighbors receive. Designing experiments in these connected environments is incredibly difficult, as researchers must figure out how to assign treatments to maximize the clarity of their results without letting the network's structure muddy the data.

A team of researchers at Texas Tech University has tackled this challenge by creating a new way to plan experiments on complex networks. Instead of treating connections as a nuisance to be ignored, they built a system that uses the network's shape to guide decisions. Imagine trying to place two different types of signs on a map of a city to see which one gets more attention. If you place them randomly, you might accidentally put all the "Type A" signs in one neighborhood and all the "Type B" signs in another, making it impossible to tell if the difference in attention is due to the signs or the neighborhood itself. The researchers developed a mathematical method to find the perfect arrangement of signs that balances the two types while respecting the city's layout. They used a computer algorithm that acts like a careful editor, constantly swapping the positions of treatments on the network to see if the new arrangement provides clearer information. This process is guided by a specific measure of how much information the experiment will yield, ensuring that the final design is robust against the confusing effects of neighbors influencing one another.

The team tested their method on a variety of simulated networks that mimic different real-world structures. Some were like random webs of connections, others were based on physical distance, and some had distinct communities or "hubs" where many connections met. In every case, their optimized design outperformed standard methods like random assignment or grouping people into clusters. The new approach consistently produced a more balanced distribution of treatments, ensuring that neither option was overrepresented in any specific part of the network. When they looked at the results, they found that while the overall effect of the treatment could be measured reliably, the specific effects of a treatment on an individual versus the effect of a neighbor's treatment were harder to pin down. The accuracy of these specific measurements depended heavily on how the network was shaped; in some structures, the interference was so complex that it introduced significant uncertainty into the estimates.

To prove their method worked in the real world, the researchers applied it to two actual datasets. The first was a network of college students living in shared housing, where 278 individuals were connected by 1,193 shared living arrangements. The second was a social network of 220 Facebook users and their 576 friendship ties. In both cases, the algorithm generated a specific map of who should receive which treatment to get the best possible data. For the college network, the method distributed the treatments across the dense dormitory clusters and the sparse individual rooms in a way that no random method could achieve. Similarly, for the Facebook network, the design navigated the mix of tight-knit friend groups and long chains of acquaintances to ensure a fair and informative spread. The results showed that by accounting for the network's geometry, researchers could reduce the confusion caused by interference and get a clearer picture of how treatments work in connected populations.

The study concludes that while the total impact of a treatment can be measured with stability, understanding the precise mechanics of how a treatment affects an individual versus their neighbors remains a complex puzzle. The researchers found that the structure of the network itself dictates how much information can be extracted. In networks with highly uneven connections, such as those with a few very popular hubs and many isolated individuals, the estimates for specific effects can become quite unstable. The authors acknowledge that their method is a powerful tool for finding near-perfect arrangements, but it is a heuristic approach, meaning it finds the best solution it can through a smart search rather than guaranteeing the absolute mathematical optimum. They also note that their current work focuses on static, two-way connections, leaving open the question of how to handle networks that change over time or have one-way influences. Despite these limitations, the work provides a practical blueprint for designing better experiments in our interconnected world, moving beyond the old assumption that people act in isolation to embrace the reality that we are all part of a larger, influencing web.

Technical Summary: Optimizing Treatment Allocation in Experiments with Network Interference

Problem Statement

Experimental design in modern epidemiology, social media, and online platforms is increasingly complicated by network interference, where a unit's outcome depends not only on its own treatment but also on the treatments assigned to its neighbors. This phenomenon violates the classical Stable Unit Treatment Value Assumption (SUTVA), which posits that potential outcomes are independent of other units' treatments. While existing literature has developed frameworks for network experiments (e.g., graph cluster randomization, rerandomization), these approaches are predominantly evaluated on small or simplified networks. Consequently, their applicability to complex, large-scale real-world networks with varying topologies (e.g., sparsity, community structure) remains insufficiently established. The core challenge is to design treatment allocation strategies that jointly optimize for allocation balance and network topology to improve estimation efficiency under interference.

Methodology

1. Network Model and Interference Framework

The authors model the experimental setting as an undirected network $G = (V, E)$ with $n$ nodes. They introduce a treatment-treatment edge incidence matrix $L$ , where entry $L_{j\ell}$ counts the number of edges connecting nodes assigned treatment $t(j)$ and $t(\ell)$ . This matrix serves as a metric for network interference.

The network structure is modeled using a random graph logistic model. The probability of an edge existing between nodes $i$ and $k$ is defined as:
$P(A_{ik} = 1 | X_i, X_k) = \frac{\exp(s_{ik} + \theta L_{t(i)t(k)})}{1 + \exp(s_{ik} + \theta L_{t(i)t(k)})}$
where $s_{ik}$ represents covariate similarity and $\theta$ controls the influence of treatment similarity on edge formation.

2. Optimality Criterion

To determine the optimal treatment assignment, the authors derive the Fisher information matrix $J(\alpha)$ for the unknown parameters $\alpha = (\beta^T, \sigma, \theta)^T$ . The optimality criterion is defined as the determinant of this matrix, $\tau = \det(J(\alpha))$ .

Following Chen and Chang [2023], the determinant is approximated (considering the limiting case where $\theta \to 0$ to ensure analytical tractability) as:
$\tau = \left( \prod_{j=1}^m n_j \right) \left[ \sum_{h=1}^m \binom{n_h}{2} \frac{L_{hh}^2 \exp(1 + \theta L_{hh})}{[1 + \exp(1 + \theta L_{hh})]^2} + \sum_{j=1}^{m-1} \sum_{\ell > j}^m n_j n_\ell \frac{L_{j\ell}^2 \exp(\theta L_{j\ell})}{[1 + \exp(\theta L_{j\ell})]^2} \right]$
This criterion balances two components:

Treatment Balance: Promoted by the product of treatment counts $\prod n_j$ .
Structural Interaction: Captured by the weighted sum of squared edge incidences between treatment groups, reflecting how treatments are distributed relative to the network structure.

3. Optimization Algorithm

Since maximizing $\tau$ is a combinatorial optimization problem, the authors propose a local search algorithm (adapted from Cook and Nachtsheim [1980] and Koutra et al. [2021]). The procedure involves:

Generating $R$ random initial treatment assignments.
Iteratively swapping the treatment of a single node with an alternative treatment.
Accepting the swap if it increases the criterion value $\tau$ .
Repeating until convergence for each initialization and selecting the global maximum $\tau^*$ across all runs.

4. Causal Estimation

The paper evaluates the causal properties of the resulting designs by estimating:

Total Treatment Effect (TTE): The average impact of universal treatment.
Average Treatment Effect (ATE): The direct effect of treatment on a unit, assuming neighbors are untreated.
Average Interference Effect (AIE): The effect of neighbors' treatments on a unit.
Estimators are derived based on a heterogeneous linear outcome model, accounting for neighborhood exposure probabilities ( $\rho_i$ ).

Key Contributions and Results

Simulation Studies

The framework was tested across four random graph models: Erdős–Rényi (ER), Geometric Random (GR), Preferential Attachment (PA), and Stochastic Block Models (SBM) (both balanced and unbalanced).

Optimality: The proposed local search algorithm consistently achieved the highest criterion values ( $\tau$ ) compared to Randomized Design, Graph Cluster Randomization, and Rerandomization across all network types.
Balance: The optimized designs maintained treatment proportions close to 0.5, demonstrating robustness against the instability often seen in cluster randomization, particularly in unbalanced SBMs.
Topology Influence: The optimal assignments adapted to network geometry:
- In ER graphs, treatments were well-mixed.
- In GR graphs, spatial clustering led to localized treatment pockets.
- In PA models, high-degree hubs often shared the same assignment.
- In SBMs, community structures drove distinct treatment groupings.

Causal Estimation Performance

TTE Estimators: Remained stable across all models, as they do not depend on neighborhood exposure patterns.
ATE and AIE Estimators: Exhibited bias driven by the heterogeneity of neighborhood exposure ( $\rho_i$ $ρ_{i}$ ).
- In ER and GR graphs, large biases were observed, attributed to control nodes with near-zero exposure (affecting AIE) or high variability in treated node exposure (affecting ATE).
- In PA and SBM models, biases were generally smaller, suggesting that the optimized allocations in these structures generated more uniform exposure levels.

Real-World Applications

The framework was applied to two real-world datasets:

College Housing Network (278 nodes): The optimal design distributed treatments across dense residence hall clusters and sparse isolated pairs, avoiding dominance by a single treatment in any subgraph.
Ego-Facebook Network (220 nodes): The design allocated treatments to mix within dense cliques while maintaining homogeneity along long, sparse peripheral branches.

In both cases, the optimized designs successfully balanced treatment allocation while respecting local connectivity, though estimator biases for ATE and AIE persisted due to structural heterogeneity in neighborhood exposure.

Significance and Claims

The paper claims to provide a network-aware design framework that effectively bridges the gap between theoretical optimality criteria and practical, large-scale network experiments. By integrating the Fisher information matrix with a scalable local search algorithm, the approach offers a computationally feasible method to identify treatment assignments that account for complex interference patterns.

The authors emphasize that their work demonstrates how network topology fundamentally shapes optimal treatment allocations. They argue that ignoring these structural dependencies leads to suboptimal designs and biased causal estimates. However, the paper maintains a modest tone regarding its limitations:

The optimization is heuristic, yielding near-optimal rather than globally optimal solutions.
The current framework is restricted to undirected, static networks, whereas real-world networks may be directed, weighted, or dynamic.
Extending the framework to multi-treatment settings complicates the definition of standard causal estimands (TTE, ATE, AIE), which are inherently tied to binary treatment-control paradigms.

The study concludes that while the proposed method significantly improves design efficiency and balance, the inherent complexity of network interference necessitates careful consideration of exposure heterogeneity when interpreting causal effects.