TL;DR
This paper introduces a robust framework for estimating causal effects and designing experiments in networks where the full "who-knows-whom" map is missing. By combining structural causal models with random graph theory, the authors demonstrate that cheap, partial network data (like survey summaries) can be nearly as effective as expensive, full-network data for understanding spillover effects and optimizing treatment seeding.
What Problem Does This Paper Address?
In classical statistics, we often rely on the Stable Unit Treatment Value Assumption (SUTVA)—the idea that my treatment (e.g., taking a vaccine) doesn't affect your outcome. In the real world, this is rarely true. We live in networks where "interference" or "spillover effects" are the norm: information diffuses, diseases spread, and social behaviors are contagious.
To account for these spillovers, researchers typically need a complete map of the network. However, collecting a full network census is logistically nightmarish, prohibitively expensive, and often raises privacy concerns. Most researchers instead have access to partial network data, such as:
- Aggregated Relational Data (ARD): Asking "How many people do you know who are farmers?"
- Egocentric Sampling: Asking a few individuals about their immediate friends.
- Respondent-Driven Sampling: Tracking referrals.
The gap this paper fills is providing a mathematically rigorous way to perform inference and experimental design using these "coarsened" snapshots of a network rather than the whole graph.
Methods
The authors propose a dual-layered approach:
- Structural Causal Models (SCMs): They model the potential outcome of an individual as a function of their own treatment, their "exposure" (how many of their neighbors are treated), and unobserved noise. This allows them to define the Global Average Treatment Effect (GATE)—the difference between treating everyone versus treating no one.
- Graphon-Based Estimation: Since the full network $G$ is unknown, they treat the network itself as a random variable drawn from a model (like a Stochastic Blockmodel). They use an iterated expectation method to "average out" the missing links. Essentially, if we know the rate at which different types of people connect, we can estimate the expected exposure of any individual without knowing their specific friends.
For experimental design, they use Bayesian Optimization to find the optimal "saturation levels"—deciding what percentage of different social groups should be treated to either minimize the variance of their estimates or maximize the spread of a behavior (optimal seeding).
Datasets
The authors validate their framework using semi-synthetic experiments built on real-world social networks:
- India (Banerjee et al., 2013, 2019): Data from 70 rural villages focused on microfinance and information diffusion ("gossip" models).
- Malawi (Beaman et al., 2021): A network of Malawian farmers used to study the adoption of new agricultural technologies (complex contagion).
They use these real network structures but simulate the outcomes to test how well their partial-data methods recover the "true" causal parameters.
Key Insights and Ideas
- The "Coarsened" Advantage: You don't need to know exactly who is connected to whom to get valid causal estimates. Knowing the connection probabilities between latent groups (e.g., "farmers connect to shop owners 10% of the time") is often sufficient.
- Shorter Surveys, Better Science: The authors extended previous methods to allow for non-mutually exclusive traits. Previously, you had to ask about "Farmers who own houses" as a single category. Now, you can ask about "Farmers" and "Homeowners" separately, drastically reducing the number of questions needed in a survey.
- Transportability: Because the method relies on a model of the process of interference rather than just the specific links of one graph, the insights are more "transportable" to new populations where the network hasn't been mapped yet.
Key Results
- Accuracy: In estimating the GATE, the partial-data approach (using ARD) performed comparably to regression models that had access to the full network.
- Variance Reduction: Using their Bayesian Optimization for experimental design, the authors reduced the Root Mean Square Error (RMSE) of their estimates by 38% to 45% compared to standard uniform random designs.
- Superior Seeding: In the Malawi technology adoption case, their "optimal seeding" strategy (using only partial data) increased technology adoption by 1.5 times compared to the common practice of simply targeting high-degree "influencers."
Limitations and Future Work
- Model Dependence: The framework's success hinges on the "correct" specification of the causal model and the graph model. If the underlying diffusion process is fundamentally different from the one assumed (e.g., a simple contagion vs. a complex one), the results may be biased.
- Static vs. Dynamic: The current work focuses on a single point in time. The authors note that future research should extend this to panel data and staggered rollouts, where the network and the treatments evolve over time.
- Sensitivity Analysis: More work is needed to understand how sensitive the results are to "hidden" traits that aren't captured in the partial network data.
Summary updated