[Phase 1.5] Validation suite: marginal vs joint_gcomp on ARTnet 2017-18

[smjenness]

Context

Before releasing the joint g-computation method in production, we need a rigorous comparison to the current marginal approach to:

1. Quantify the bias in current marginal-method netstats.
2. Validate that joint g-comp reproduces marginal-method behavior when the target distribution matches ARTnet (internal consistency check).
3. Identify which ERGM terms are most affected by the correction.
4. Generate the comparison figures/tables that will anchor the methods paper.

Scope

Run both methods on ARTnet 2017-18 data with:

• Scenario 1 — ARTnet-self target (sanity): target population = ARTnet 2017-18 respondents themselves. Joint and marginal should give similar results (any difference is pure marginal-bias artifact).
• Scenario 2 — Current ARTnet default target (NCHS age + national race): what build_netstats() currently does. Joint method may differ if interactions matter.
• Scenario 3 — Shifted target (e.g., older population or different race mix): explicitly stress-test the correction. This is where joint-vs-marginal should diverge most.

For each scenario, report:

• All netstats values side-by-side (marginal vs joint).
• % difference per target stat.
• Which interactions drive the differences (via ablation).

Tasks

• Write validation script in tests/testthat/ or inst/validation/ directory.
• Produce a side-by-side comparison table for main / casl / inst × all target stats × marginal / joint.
• Identify stats where joint-marginal difference is >5% — flag as "materially affected by method choice."
• Write up validation results (will feed methods paper + vignette).
• Sanity check: on Scenario 1 (ARTnet-self target), all differences should be negligible (<2%).

Acceptance criteria

• Reproducible validation script.
• Comparison table for all ERGM target stats.
• Clear interpretation of which stats shift how much under which scenarios.

Related

• Blocked by: [Phase 1.1] Refactor build_netparams() to fit joint Poisson GLM across attributes #61, [Phase 1.2] Refactor build_netstats() to use g-computation predictions from joint model #62, [Phase 1.3] Joint modeling for nodematch + absdiff + dissolution (dyad-level target stats) #63, [Phase 1.4] Support post-stratification via user-supplied target population distribution #64
• Feeds into: methods paper drafting ([ARTnetPredict/issue on paper])
• Reference doc: ARTnetPredict/Plan/Step09-Plan.md §2.4

Stretch

Extend validation to a second dataset if available (e.g., a city-specific subset). Internal consistency check: the bias magnitude should correlate with how much the target distribution differs from ARTnet.