3.3.2 Temporal Networks

Network data are increasingly gathered across multiple points in time.67

Panel data is among the most common forms of longitudinal network data. This entails gathering network data from the same population at multiple time windows. For a simple example of this type of network data, see Figure 3.6. This is sometimes referred to as multi-slice data (e.g., Add Health gathered friendship nomination data in the in-school Wave, and each of the in-home Waves).68 A common limitation of panel network data is the churn of network members across waves (with some members leaving or being added to the population across the observation windows); I discuss the implications of this in Chapter 5. Other data sources—especially the types of digital trace data described above—may represent relatively continuous streams of data, without clear distinctions between panel waves. Each of these forms of data lead to distinctly different types of questions and corresponding analytic methods; for a review see Marcum and Schaefer (2019).

Figure 3.6: Heuristic Longitudinal Network. Three waves of network data represented in graph and adjacency matrix form. Note that in this example, the nodeset remains the same across all three time periods, while some ties change across the slices.
Figure 3.6: Heuristic Longitudinal Network. Three waves of network data represented in graph and adjacency matrix form. Note that in this example, the nodeset remains the same across all three time periods, while some ties change across the slices.

While gathering longitudinal network data occurs with virtually any active or passive design strategy, the varying forms found in the longitudinal nature of network data are sometimes more common in one or the other. In particular, active strategies are more likely to be split into discrete rounds, while passive data (especially digital trace data) may provide opportunities for more or less continuous data streams. Longitudinal network data can leverage advantages over approaches designed specifically with cross-sectional data in mind. Researchers could, for example, seed a data collection instrument in later waves with information about ties from previous ones (e.g., asking whether ties still exist, rather than generating relationships from scratch at each wave). Obviously, this approach also has its strengths and potential drawbacks. On the positive side, this approach is more logistically expedient to collect and would facilitate linking data across waves, it could have the drawbacks of potentially overestimating network continuity, or reinforcing edge-level sampling biases introduced in an early wave (Brewer 2000).