Modelling of Dynamic Networks based on Egocentrically-Sampled Data
Pavel Krivitsky (University of Wollongong)
Dynamic network models—models for network evolution—have manifold applications, but inference often suffers from
challenges of data availability. In epidemiology, in particular, of interest is not just the presence of relationships of
interest but their timing, yet the data available are often limited to an observation at one point in time, and even there,
rather than observing the whole network, only an egocentric sample of it is observed: a random sample of individuals
is suveyed for the demographic information about themselves and their partners, often with partners not identifiable.
For such surveys, temporal information is limited to duration of extant and recent ties, often censored and truncated.
We develop one practical approach to fitting dynamic models to such data: to find what the dynamic network model
parameters had to have been in order to induce, in the long run (i.e., its equilibrium), the network with properties
implied by the egocentric data: a form of generalized method of moments estimation (GMME). We apply this process
to fitting the Separable Temporal Exponential-Family Random Graph Models (STERGMs) to such data by deriving
their long-run properties, and quantify the uncertainty (e.g., standard errors) associated with the resulting parameter
estimates by leveraging design-based inference results for cross-sectional exponential-family random graph models
(ERGMs). We demonstrate its use via application to egocentrically sampled sexual partnership network data.
Last reviewed: 10 June, 2016