Accounting for correlation and censoring in Bayesian Network Scale-up Method Models

Abstract

The Network Scale-up Method (NSUM) estimates the size of hard-to-reach populations using survey data on individuals’ social networks. Existing NSUM models incorporate correlation across groups in the responses. We propose a generalized model that improves NSUM accuracy by addressing data censoring and accounting for the relationship between social network size and the likelihood of knowing individuals in different groups. Correlations are directly estimable from NSUM survey data, and simulations show that subpopulation estimates are biased when censoring and correlations are ignored. We analyze two data sets, yielding both population size estimates and novel insights into social network structures in these communities.