Rerandomization and high-dimensional adjustment with assignment-dependent auxiliary variables under network interference
Abstract
Network experiments are widely used to evaluate interventions in settings such as vaccine trials and social networks, where interference is often present. In these settings, predictive auxiliary variables, such as treatment-weighted neighbor summaries, depend on treatment assignments and generally exhibit nonzero exposure effects. In this paper, we propose rerandomization and covariate adjustment methods based on estimation errors in these auxiliary exposure effects to improve precision. For Mahalanobis distance-based rerandomization, we derive the asymptotic distributions of the Horvitz-Thompson and Hájek estimators under approximate neighborhood interference and quantify the resulting precision gains. At the analysis stage, we propose a covariate adjustment method that acts directly on these estimation errors. This method requires no normalization of the auxiliary variables and is optimal within the corresponding class of linear adjustments. We show that rerandomization and covariate adjustment become asymptotically equivalent as the rerandomization threshold approaches zero, extending this equivalence to settings with interference. For high-dimensional auxiliary variables, we propose Lasso-adjusted estimators and establish their joint asymptotic normality under suitable sparsity conditions. These estimators retain the same asymptotic distribution when combined with fixed-dimensional rerandomization. Simulations and a social-network application illustrate the precision gains from the proposed methods.
Related
- Rerandomization under Interference
- Data-Adaptive Rerandomization for 2K Factorial Designs
- High-Dimensional Change Point Analysis for Temporally Dependent Data
- Penalizations of Two-Dimensional Simple Random Walk by Direction-Dependent Excursion Counts and Edge-Crossing Counts
- Frequency-dependent capacitance matrix formulation for Fabry--Pérot resonances. Part II: One-dimensional periodic media
BibTeX
@misc{zhang2026rerandomization,
title = {{Rerandomization and high-dimensional adjustment with assignment-dependent auxiliary variables under network interference}},
author = {Honghao Zhang and Yanghao Wei and Hanzhong Liu},
howpublished = {arXiv},
year = {2026},
month = oct,
doi = {10.48550/arxiv.2610.27741},
eprint = {2610.27741},
archivePrefix = {arXiv},
url = {https://arxiv.org/abs/2610.27741},
}
RIS
TY - UNPB
TI - Rerandomization and high-dimensional adjustment with assignment-dependent auxiliary variables under network interference
AU - Honghao Zhang
AU - Yanghao Wei
AU - Hanzhong Liu
PY - 2026
DA - 2026/10/20
PB - arXiv
DO - 10.48550/arxiv.2610.27741
UR - https://arxiv.org/abs/2610.27741
AB - Network experiments are widely used to evaluate interventions in settings such as vaccine trials and social networks, where interference is often present. In these settings, predictive auxiliary variables, such as treatment-weighted neighbor summaries, depend on treatment assignments and generally exhibit nonzero exposure effects. In this paper, we propose rerandomization and covariate adjustment methods based on estimation errors in these auxiliary exposure effects to improve precision. For Mahalanobis distance-based rerandomization, we derive the asymptotic distributions of the Horvitz-Thompson and Hájek estimators under approximate neighborhood interference and quantify the resulting precision gains. At the analysis stage, we propose a covariate adjustment method that acts directly on these estimation errors. This method requires no normalization of the auxiliary variables and is optimal within the corresponding class of linear adjustments. We show that rerandomization and covariate adjustment become asymptotically equivalent as the rerandomization threshold approaches zero, extending this equivalence to settings with interference. For high-dimensional auxiliary variables, we propose Lasso-adjusted estimators and establish their joint asymptotic normality under suitable sparsity conditions. These estimators retain the same asymptotic distribution when combined with fixed-dimensional rerandomization. Simulations and a social-network application illustrate the precision gains from the proposed methods.
ER -