Ground suspected-case case-finding in contact tracing (headline; merges #389)

[seabbs-bot]

Ground suspected-case case-finding in contact tracing (headline)

This began as an investigation into why the joint fit's reproduction number is
pulled down relative to the single-stream fits (leave-one-out traced it to the
DRC suspected surveillance reconciling a suspected-vs-confirmed growth conflict
through the reproduction-number walk, because the suspected stream had only a
constant ascertainment to absorb a change in case-finding effort). It now
merges and unifies with the contact-tracing work in #389 into a single headline
change: the observed contact-tracing follow-up rate drives both channels
case-finding acts through.

Model

• Shared latent follow-up-rate process (contact_tracing_model). The daily
  "taux de suivi des contacts" is modelled as a logistic random walk q_t ∈ (0,1) over the whole grid, fitted to the reported rate where it exists.
  Because it is latent everywhere it is inferred before the reported window and
  projects into forecasts under its own dynamics rather than a fixed
  covariate stopping at the data boundary (so the case-finding anchor does not
  drop out when the data stop).
• Non-BVD background scales with it, exp(β_contact·(q−q̄))
  (contact_background_model, from feat(model): add contact-tracing background covariate (#374) #389).
• BVD suspected ascertainment carries a reporting-effort multiplier anchored
  to the same rate plus a random-walk deviation, exp(β_asc·(q−q̄) + w)
  (reporting_effort_walk_model). Both touch the suspected likelihood only, not
  the shared onsets or p_drc, so neither moves the outbreak size (Testing & case-finding operations vs model: retesting/rule-out and background scaling #374).
• Headline, not opt-in: the live joint and every frozen re-fit pass the
  follow-up history with suspected_reporting_effort = true. Empty history /
  flag off recover the prior model. Also carries main's confirmed-case
  BetaBinomial overdispersion.

What the fits show (300 draws × 2, confirmatory)

Fit	R_T	R_T/R0	C_T	β_contact	β_asc	divergences
headline (both channels)	1.32	0.87	4981	3.46	0.53	4
− ascertainment effort	1.35	0.90	5142	2.64	—	5
− contact grounding (free-walk effort)	1.41	0.93	5316	—	—	22

• Both elasticities are positive and data-supported: contact tracing measurably
  scales detection of non-BVD (β_contact ≈ 3.5) and BVD (β_asc ≈ 0.5)
  suspects.
• The outbreak size C_T is unchanged across the three (~5000), so the
  case-finding channels re-shape the suspected reporting without moving the
  size — no ascertainment / outbreak-size confounding.
• Grounding the ascertainment in the observed rate collapses the divergences
  from 22 (free walk) to 4: the observed anchor identifies the ascertainment
  change far better than a free walk, so the grounded model is both more
  defensible and better-behaved.

Analysis and sensitivity

• The main analysis reports the contact-tracing grounding of the reproduction
  number and the two elasticities (background_contact_coef, β_asc).
• Two gated sensitivity re-fits on the sensitivity page: sens_no_contact
  (drop the grounding) and sens_no_effort (drop the ascertainment effort),
  each compared to the headline for R_T and C_T.

Files

• src/models/priors.jl — contact_tracing_model (latent rate),
  contact_background_model (feat(model): add contact-tracing background covariate (#374) #389), reporting_effort_walk_model (anchored).
• src/models/observations.jl — unified latent-q scaling of background and
  ascertainment in reported_cases_model.
• src/models/joint.jl — contact_followup_history +
  suspected_reporting_effort wiring; background_contact_coef deterministic.
• src/data.jl, data/observations.toml — the follow-up-rate stream (feat(model): add contact-tracing background covariate (#374) #389).
• docs/fits/registry.jl, docs/examples/{_setup,analysis,sensitivity}.jl,
  docs/src/news.md, test/test_{contact_background,reporting_effort}.jl.

Supersedes the fixed-covariate form in #389 with the forecast-safe latent
process; #389 can close in favour of this.

---

This was opened by a bot. Please ping @seabbs for any questions.

[codecov]

Codecov Report

✅ All modified and coverable lines are covered by tests.
✅ Project coverage is 88.33%. Comparing base (39e8a6c) to head (0fd40d8).
⚠️ Report is 2 commits behind head on main.

Additional details and impacted files

@@            Coverage Diff             @@
##             main     #400      +/-   ##
==========================================
+ Coverage   87.50%   88.33%   +0.83%     
==========================================
  Files          13       13              
  Lines        2433     2530      +97     
==========================================
+ Hits         2129     2235     +106     
+ Misses        304      295       -9     

☔ View full report in Codecov by Harness.
📢 Have feedback on the report? Share it here.

🚀 New features to boost your workflow:

• ❄️ Test Analytics: Detect flaky tests, report on failures, and find test suite problems.

[github-actions]

📖 Documentation preview is ready!

View the docs for this PR at: http://epiforecasts.io/BVDOutbreakSize/previews/PR400/

This preview will be updated automatically when you push new commits.

[seabbs]

This doesn't converge so we can't merge.