On the validity of perceived social structure

摘要

The validity of survey-based reports of social relationships is a critical assumption for much social network research. Research on informant accuracy has shown that observational data and recalled behavior by informants are imperfectly correlated, which calls into question whether complex relations like friendship and advice-seeking can be accurately measured from individual reports. A class of network inference models, the Bayesian Network Accuracy Models, growing out of the pioneering work of Batchelder and Romney on inference from informant reports, provides a principled basis for inferring network structure given such error-prone data. Using these models, we can gain insight into the accuracy of informants' self and proxy reports of social ties, and more broadly, the reliability and validity of respondents' reports of informal social relations. While existing data does not provide a criterion validity check for inferring most relationships, other notions of validity and/or reliability can still be applied. For instance, if friendship reports are generated from a common underlying network that is perceivable (albeit imperfectly) by all actors, then random subsets of actors should produce estimates that should agree (i.e., split-half reliability). Using informant reports on friendship and advice-seeking networks from four different organizations, we show substantially higher levels of split-half reliability than can be explained by chance, suggesting that models are indeed estimating a common underlying relation. We also show that informants' errors appear to be structured in ways that are consistent with cognitive models of social perception, with greater accuracy on average for large-scale network features rather than fine details, for own versus others' ties, and for core-periphery structures versus bipartitions. Evidence from construct validity checks further suggests the that common networks underlying informants' reports have properties that would be expect