Sensitivity check: regenerate target statistics at multiple N values

[smjenness]

Context

Even though N is foundational rather than calibratable, it is worth documenting how ARTnet-generated network targets change at different plausible values. Some target statistics scale linearly with N (edges, nodefactor counts); others are invariant (mean degree, mixing proportions); some may behave non-trivially (concurrency, durations under TERGM dissolution coefs).

This issue produces the empirical sensitivity evidence needed to make a defensible call once the literature and surveillance audits close. It can proceed in parallel with the other popsize2026 issues since it does not depend on their conclusions.

Investigation tasks

• Run build_netstats() for Atlanta at N values: 53K (Islek), 72K (Bennett-derived), 102K (Grey), and any other defensible candidates that emerge from the literature/PHSKC issues.
• Repeat for Seattle at the corresponding values: 38K (Islek), Bennett-derived if available, 82K (Grey), and PHSKC-implied.
• Tabulate edges, mean degree per layer, concurrency, momentary degree distributions, and key mixing matrices side-by-side.
• Identify which target statistics scale linearly with N, which are invariant, and which behave non-trivially.
• Document the scaling behavior for the future-self / collaborators audience: what changes by simply rescaling existing targets vs what would require re-fitting upstream models.
• Optionally: run a short ERGM estimation pass at each N to confirm convergence behavior is stable across the range.

Priority

Medium. Larger time investment but produces the empirical sensitivity numbers needed for a defensible decision.

Source

Plan: inst/popsize/msm_population_size_discussion.md, §Q5.