Network Super Learner

• In Event: New Advances in Social Network Analysis

Abstract

Much of sociology's causal vocabulary—contagion, diffusion, peer influence—describes outcomes that travel through social ties. Yet estimating these effects confronts a double bind: the ties that carry influence also encode the homophilous sorting that confounds it. Modern methods use machine learning to adjust for this sorting but require analysts to choose a specific algorithm. I show this choice is consequential: because the best-performing algorithm shifts with network structure, a fixed choice introduces hidden bias that standard errors will not reveal. To resolve this, I develop the Network Super Learner (NSL), which sidesteps the choice by adapting to the network at hand. Across diverse network structures and confounding regimes, the NSL performs at or near the best available method in estimating both direct and peer effects, where individual algorithms are erratic. I map the limits that unobserved homophily imposes on any such method and illustrate the approach with political expression during the 2020 Black Lives Matter protests.

Author

• Yuanhao (Colin) Liu, Johns Hopkins University