[Phase 1.3] Joint modeling for nodematch + absdiff + dissolution (dyad-level target stats)

[smjenness]

Context

After #61/#62, all ego-level target statistics (edges, nodefactor_*, concurrent) are produced under method = "joint" via g-computation on joint Poisson / binomial GLMs. What remains are the dyad-level target statistics — every ERGM term whose value depends on both partners' attributes:

1. Nodematch stats (nm.age.grp, nm.race, nm.race_diffF, nm.role.class) are currently estimated via univariate logistic GLMs on partnership-level data, e.g. same.age.grp ~ index.age.grp.
2. Absdiff stats (absdiff.age, absdiff.sqrt.age) are currently a single lm(ad ~ 1) scalar scaled by edges at netstats time.
3. Durations / dissolution (durs.<layer>.byage) are currently empirical means/medians stratified only by (age-match × index.age.grp) — no regression structure at all.

All three suffer the same marginal-vs-joint bias as the ego-level target stats did before #61: when the target population joint distribution differs from ARTnet's, these statistics carry ARTnet's conditional dyad attribute distribution baked in.

This issue is the dyad-level counterpart to #61/#62. Originally scoped to nodematch only; expanded on 2026-04-19 to cover durations after PI review of #68 noted the gap.

Proposed approach

All three sub-areas share the same long-form data unit (ARTnet partnership records with (ego_attrs, alter_attrs) pairs) and the same g-computation template:

1. Fit a joint regression on partnership-level data with both ego and partner attributes on the RHS.
2. Predict per-dyad for the synthetic population's implied mixing structure.
3. Aggregate to target statistics.

1. Nodematch

Fit joint logistic models:

m_nm_age <- glm(same.age.grp ~ index.age.grp + index.race.cat.num +
                  part.age.grp + part.race.cat.num + ...,
                data = lmain, family = binomial())

Then predict on synthetic partnership pairs. Generating synthetic pairs is more complex than synthetic nodes — need the joint distribution of (ego, alter) attributes. Two options:

Option A (simpler): keep partnership-pair modeling marginal (current approach), but ensure ego-side attributes come from the corrected target joint distribution.

Option B (fully joint): fit joint ego-alter pair model; generate synthetic pairs from p(ego, alter) in the target population; predict match probabilities.

Recommend Option A for first pass; escalate to Option B if Option A still shows substantial bias after joint-nodefactor correction.

2. Absdiff

Joint regression on the partnership age gap:

m_ad_age <- lm(ad ~ index.age.grp + index.race.cat.num +
                 part.age.grp + part.race.cat.num + ...,
               data = lmain)

Then under joint netstats: predict per-dyad |age_i - age_j| for the synthetic mixing structure and aggregate. Same infrastructure choice (Option A vs Option B) as nodematch — probably share whichever approach we pick.

3. Durations / dissolution (new sub-scope)

Currently netparams$<layer>$durs.<layer>.byage is a summary stat:

durs.main.byage <- lmain |>
  filter(ongoing2 == 1) |>
  group_by(index.age.grp) |>
  summarise(mean.dur = mean(duration.time), median.dur = median(duration.time))

No adjustment for race, HIV concordance, or other dyad attributes. Marginal-vs-joint problem at a dyad level.

Joint analog:

m_dur_main <- lm(log(duration.time) ~ index.age.grp + index.race.cat.num +
                   part.age.grp + part.race.cat.num +
                   hiv.concord + same.race + same.age.grp,
                 data = lmain[lmain\$ongoing2 == 1, ])

Under method = "joint" in build_netstats: predict expected log-duration per edge given the synthetic network's mixing structure, exponentiate to get mean duration, then run through the existing geometric-distribution rate math (rates.adj = 1 - 2^(-1/median), mean.dur.adj = 1/(1 - 2^(-1/median))) to produce dissolution coefs.

Open design question: log-linear on duration.time assumes multiplicative attribute effects on mean duration. Alternative: Weibull or other survival models on the censored data (many partnerships are ongoing, so ongoing2 == 1 filtering to get durations is already dropping right-censored observations — that's a pre-existing issue with the current summary-stat approach too).

Tasks

• Joint logistic for same.age.grp, same.race in build_netparams under method = "joint". Store at netparams$<layer>$joint_nm_age_model, joint_nm_race_model.
• Joint lm for ad (age absdiff) in build_netparams. Store at netparams$<layer>$joint_absdiff_age_model.
• Joint lm for log(duration.time) among ongoing partnerships. Store at netparams$<layer>$joint_duration_model.
• [~] Under method = "joint" in build_netstats: predict dyad-level stats on synthetic pairs, aggregate to nodematch_*, absdiff_*, diss.<layer>.byage. Retire the univariate-ratio × new-edges shortcut currently in PR Use joint GLM g-computation for build_netstats target stats (#62) #68. (partial: nodematch and absdiff done in PR Joint dyad-level modeling: nodematch + absdiff (#63 phases 1 & 2) #69 with ego-attr aggregation over the synthetic population. Duration model fit in PR Duration methods: empirical + joint_lm (#63 phase 3) #71 but its consumption on the synthetic population is deferred to issue Duration g-computation: predict joint_lm per-dyad on synthetic population in build_netstats #73.)
• Decide Option A vs Option B for generating synthetic partnership pairs (probably same answer for all three sub-areas).
• Unit tests: on ARTnet-own distribution, recover marginal observed same-race / same-age rates, mean absdiff, and mean duration per stratum.
• End-to-end EpiModelHIV-Template estimation with method = "joint". Verified on PR Duration methods: empirical + joint_lm (#63 phase 3) #71: all 6 ERGMs (3 layers x default + joint methods) converge cleanly under Stochastic-Approximation; netdx static diagnostics on the default main model show all target stats matched within |Z| <= 2.05 and |% diff| <= 4.2% across 1000 sims.

Validation suite work — comparing univariate vs joint across multiple target-population scenarios, ablating interaction terms, generating the methods-paper figures — is handled separately on #65 (Phase 1.5), which is explicitly blocked by this issue and lives downstream.

Acceptance criteria

• Under method = "joint", all of nodematch_*, absdiff_*, and diss.<layer>.byage come from joint models rather than univariate marginals.
• Internal consistency: Σ_k nodematch_age.grp[k] ≤ Σ_k nodefactor_age.grp[k] / 2 etc.
• End-to-end netest() run on EpiModelHIV-Template converges and diagnostics look reasonable (dx_main, dx_casl, dx_inst).

Related

• Blocked by: [Phase 1.2] Refactor build_netstats() to use g-computation predictions from joint model #62 (build_netstats joint infra) — now in PR Use joint GLM g-computation for build_netstats target stats (#62) #68.
• May spawn: follow-up issue for full ego-alter pair joint generation (Option B) if Option A bias residual is large.