Uncovering Bullying Influence Networks and Peer Effects in Chinese Junior High Schools: A High-Dimensional Sparse Machine Learning Approach

• In Event: Advances in Quantitative Methods Panel 1: Machine Learning and AI for Crime Analysis

Abstract

Objectives: Understanding the mechanisms behind peer effects is key to revealing the spread patterns and root causes of school bullying. While applying standard network models is useful, it suffers from limitations such as network mismeasurements or the unavailability of network data. The current research aims to: (1) develop a novel estimation approach to recover the peer influence network using observed behavioral data without requiring prior network data, based on the Social Interaction Model; (2) define the causal interpretations of different types of peer influences and disentangle the peer influence effect from the selection effect; and (3) uncover the bullying influence networks and estimate bullying peer effect in Chinese junior high schools using the newly developed method.

Methods: A high-dimensional sparse machine learning algorithm is proposed to estimate the peer influence network. The algorithm is executed on supercomputers or high-performance computational clusters. We then adopt a cross-sectional global difference technique to disentangle the peer influence effect from the selection effect, and the potential endogeneity problem is addressed using instrumental variable estimation.

Results: Peer influence effect can be categorized into two types: the complier effect, which is based on social learning; and the defier effect, which is based on strategic substitutability or adversarial social interaction. In Chinese junior high schools, the overall peer influence effect is positive (δ ̂=0.055, SE=0.014, p<0.001), implying that the complier effect predominates bullying influence networks.

Conclusions: The proposed data-driven machine learning approach can uncover the underlying peer influence network and estimate endogenous peer influence effect, even when prior network data are unavailable.

Authors

• Zhong Wang, Hunan University, China

• Shiming Hao, Yunnan University

• Yan Zhang, Sam Houston State University