Much of the statistical networks literature is dedicated to methodological development for a restricted class of models, namely the stochastic blockmodel and graphon models. But broad consensus agrees that these models fail to reflect even the most basic properties of real world networks, raising questions about the practical use of methods developed for these models.
I discuss some basic aspects of network modeling, beginning with the observation that many network datasets are sampled from a larger population network. In such cases, the sampling scheme provides the critical link between the data and the population, which is needed to carry out sound inferences.
I show some examples of how ignoring or improperly describing the sampling scheme can adversely affect inferences. A particular instance of this is on display when comparing the characteristics of network data obtained by vertex sampling and edge sampling. I then argue that an adequate network model ought to account for both the structure of the population network as well as the sampling scheme that relates observed data to the population network.
|