Sampling large-scale social networks: Insights from simulated networks

We conduct a detailed simulation study to assess how well various sampling techniques recover network characteristics such as degree, clustering coefficient, and path length distributions of several simulated population networks that have the high clustering tendency characteristic of social networks but vary in terms of degree distribution and density. We consider several alternative sampling procedures tailored to the context of social network sampling, including random-node and random-edge sampling, egocentric sampling, and several variations of graph-exploration-based sampling methods (random walk, forest fire, and snowball methods). Our main findings are that for networks with Poisson degree distribution the snowball method is overall the best while for networks of power-law degree distribution random walk is the best when the network is sparse and the forest fire method is the best when the network is dense.