Causal inference for dyadic data in randomized experiments with interference

Statistical Seminar

Organizer：

Yunan Wu 吴宇楠（YMSC）

Speaker：

Wang Miao 苗旺

(PKU 北京大学概率统计系)

Time：

Fri., 16:00-17:00, Mar. 28, 2025

Venue：

C654, Shuangqing Complex Building A

Title：

Causal inference for dyadic data in randomized experiments with interference

Abstract:

Estimating the treatment effect in a network is of particular interest in online experimentation conducted everyday in social media companies. We investigate a novel setting where the outcome of interest comprises a series of dyadic outcomes, such as forwarding a message or sharing a link between friends and international trade relation between countries.Dyadic outcomes are pervasive in many social network sources and of particular interest in online experimentation (A/B testing).We propose a causal inference framework for dyadic outcomes in randomized experiments in the presence of network interference, and develop consistent estimators of the global average causal effect.We derive the convergence rate and variance bound of the proposed estimators,and provide a variance estimator that is conservative for quantifying the estimation uncertainty.We illustrate with a variety of numerical experiments and apply our approach to an online experiment in Wechat Channels.