NCJ Number
248233
Journal
Journal of Quantitative Criminology Volume: 30 Issue: 3 Dated: September 2014 Pages: 373-387
Date Published
September 2014
Length
15 pages
Annotation
Egocentric measures of peer delinquency, obtained through a census of a social network, have become the preferred operationalization for examining the relationships between social influence and delinquency.
Abstract
Egocentric measures of peer delinquency, obtained through a census of a social network, have become the preferred operationalization for examining the relationships between social influence and delinquency. Studies regressing ego's delinquency on the delinquency of nominated friend/s (i.e. alter/s) conclude that a statistically significant coefficient provides evidence of social influence. However, the inferences drawn from these studies may be biased by the introduction of artificial statistical dependence as a consequence of using social network data in a regression framework. Recent work (Shalizi and Thomas Sociol Methods Res 40:211-239, 2011) shows that latent homophily, or unmeasured confounding of observables, may lead to nonzero estimates of social influence, even if there is no causal significance. To examine this possibility, sensitivity analyses have been created (e.g. VanderWeele and Arah Epidemiology 22:42-52, 2011; VanderWeele Sociol Methods Res 40:240-255, 2011) to determine the robustness of an estimated coefficient to latent homophily. In this research note, I examine the robustness of estimates for social influence from two articles (Haynie Am J Sociol 106:1013-1057, 2001; Meldrum et al. J Res Crime Delinq 46:353-376, 2009) using egocentric measures of peer delinquency. Findings indicate that for large, precise point estimates, highly improbable conditions are needed to explain away the effects of social influence. However, less precise point estimates (i.e. large standard errors) are more sensitive to latent homophily. The analyses indicate that studies using egocentric measures should conduct sensitivity tests, particularly when the estimated effect is weak and/or has a relatively large standard error. Scripts written in the free programming language R (R Core Team R: a language and environment for statistical computing. R Foundation for Statistical Computing, Vienna, 2012) are provided for researchers to conduct such analyses. Abstract published by arrangement with Springer.