Faranak Alikhah | Tor Vergata
A Design-Based Approach to Network Interference under Egocentric Sampling
Wednesday, October 28, 2026 h. 13:00-14:00
Room F (Building B, 2nd Floor)
Abstract
(joint work with Tiziano Arduini, Federico Belotti, and Laura Forastiere)
Many approaches to causal inference with network interference treat the observed network as a random draw from a larger population of networks and assume that the network is fully observed. This perspective is less natural in empirical settings where the object of interest is a particular fixed and finite network, and the researcher observes only the part of the network revealed by the sampling design. In this paper, we develop a finite-population least-squares framework for observational settings with network interference in which treatment is realized at the population level, while researchers observe the network partially through egocentric sampling. Under an individualistic treatment-selection mechanism, regression coefficients are shown to admit a weighted-average interpretation in terms of heterogeneous direct and spillover effects. We identify a distinct source of contamination bias that arises when the selection mechanism for individuals depends on the treatment status of their peers, resulting in direct and spillover coefficients to load on each other's causal effects. The paper establishes a finite population central-limit-theorem and law-of-large-numbers for Bernoulli-sampled, locally dependent arrays, unifying superpopulation and finite population inference within our single framework. These results yield asymptotic normality and a variance decomposition that clarifies when conventional EHW standard errors are sufficient and when network-robust standard errors are required. The paper develops finite population network robust covariance estimators. To ensure valid inference, we introduce adjustments that guarantee the conservativeness of the proposed estimators, alongside covariate-based refinements that improve on the conservativeness while preserving valid inference.

