Software Repository for Calibration-to-Deployment Mismatch in HIV Prevention Trials: How Structural Censoring Biases Counterfactual Incidence Estimates

```markdown **Background.** Cross-sectional HIV incidence estimation using recent-infection testing algorithms (RITAs) underpins counterfactual-controlled pre-exposure prophylaxis (PrEP) efficacy trials. The Kassanjee estimator assumes closed-system observability: that every recently-infected individual within the recency window is equally present at screening. Populations experiencing structural censoring—competing-risk hazards from overdose, incarceration, displacement, and related mechanisms—violate this assumption in a systematic and directional manner. **Methods.** We derive the effective mean duration of recent infection under structural censoring, Ω*(γ) = ∫₀ᵀ P_R(t) S_c(t) dt, and the joint bias factor on the reported incidence rate ratio (IRR) incorporating both screening-cohort and intervention-arm observation probabilities. We identify the standard 90-day no-prior-testing eligibility criterion, common to Phase 3 PrEP trials including PURPOSE 1 (NCT04994509) and PURPOSE 2 (NCT04925752), as a selection mechanism operating on the same axis that drives the competing-risk hazard. We apply the framework to 34 high-burden US metropolitan areas using AIDSVu 2023 surveillance data with late-diagnosis percentage as the empirical proxy for structural hazard, and empirically anchor the structural-functions reframing using the 2014–2023 AIDSVu county-level panel. **Results.** The structural-functions reframing is empirically anchored against a 10-year AIDSVu county-level panel (35 EHE-priority MSAs × 2014–2023, N = 350). Four independent longitudinal analyses converge on the reframing: temporal stability of the surveillance variables (detrended MSA-level AR(1) φ = +0.04; test-retest r = 0.999 for total diagnoses, r = 0.89 for IDU share); a pre-EHE vs post-EHE within-MSA Wilcoxon break-point on total HIV diagnoses (median Δ = −14.8 cases/year, p < 0.0001); stratum-level geographic redistribution of IDU diagnoses 2014–2023 (Stratum A EHE-priority MSAs −3.0%, Stratum C all other counties +18.5%) against null aggregate state-level and MSA-level IDU-share slopes (+0.12 and −0.05 pp/yr, p = 0.45 and 0.47); and a Kassanjee correction invariance test across 34 high-burden MSAs showing identical optimal cascade policies with and without correction (Spearman ρ = 0.9979, 34/34 cities). A complementary COVID-era counterfactual analysis shows stratum-divergent disruption: Stratum A absorbed and overshot its declining pre-COVID trajectory while Strata B and C sustained cumulative IDU-diagnosis deficits significantly different from zero against pre-COVID counterfactual extrapolations (+185 [+74, +294] and +361 [+18, +528] cases, respectively, over the 2020–2022 caveat window). **Conclusions.** The Kassanjee/Gao cross-sectional incidence estimator produces systematically biased point estimates when applied to populations with elevated structural censoring, and the bias is structurally guaranteed rather than merely possible when trial-design eligibility criteria select the Incidence Phase cohort on the testing-engagement axis. Correction requires explicit modeling of population-specific hazard using surveillance data; the framework presented here is one such correction, fully reproducible from public data, empirically anchored against a decade of US HIV surveillance, and generalizes to any RITA-based trial with analogous eligibility structure. --- **Changes from v4 (10.5281/zenodo.19796212):** - **Canonical numeric refresh** following correction of a county-name normalization bug in `build_xlxs.py` that had silently zeroed Connecticut Planning Region (Hartford, Bridgeport, New Haven) contributions to the 2014–2023 longitudinal MSA panel. Cross-sectional values (Figure 1, Table 1) were never affected. - Stratum A IDU trajectory: 636→631 (−0.8%) corrected to **656→636 (−3.0%)** - Stratum C IDU trajectory: 561→646 (+15.2%) corrected to **541→641 (+18.5%)** - Wilcoxon Δ: −9.6 → **−14.8 cases/yr** (p < 0.0001) - AR(1) φ: +0.05 → **+0.04** - Test-retest r (IDU share): 0.94 → **0.89** - MSA IDU-share post-EHE slope: −0.08 → **−0.05 pp/yr** - **New Figure 3** replacing v4's invariance scatter: COVID-era stratum-divergent counterfactual analysis. v4 Figure 3 (Kassanjee invariance scatter, numerically unchanged) relocated to supplement as **Figure S3**. - **New supplement §S8.5** (COVID-era stratum-divergent disruption): - §S8.5.1: Counterfactual methodology + per-stratum deficit derivation with bootstrap 95% CIs - §S8.5.2: Comparison with Viguerie et al. 2024 HOPE counterfactual model — both share the calibration-to-deployment epistemic mismatch when applied to populations whose latent barrier substrate has shifted between calibration and deployment - **§1 Introduction closing paragraph** elevates the recency-alone → stratified-longitudinal thesis (the cross-sectional Kassanjee correction is one instance of a more general calibration-to-deployment mismatch). - **§5.7 Conclusion closing paragraph** generalizes the framing to its full structural-functions / stratified-longitudinal surveillance scope. - **§5.1 closing paragraph** extended with COVID partition findings as a third axis of operational concern. - **Bibliography expanded** from 37 to 47 entries (Viguerie 2024, Randall 2022, DiNenno 2022, Hassan 2022, Carrico 2020, Cranston 2019, Hershow 2022, Cohen 2022). - **One-command reproducibility suite** added: `reproduce_v9.py` orchestrates raw AIDSVu → workbook → figures → COVID counterfactual → regression test; `verify_v9.py` validates 22 load-bearing statistics against canonical reference values within per-statistic tolerances. See `REPRODUCE.md` for reviewer-facing reproduction instructions. - **Workspace consolidation**: deprecated parallel pipelines (`build_longitudinal_panel.py`, `build_figure.py`, `Fig_4_stratum_trajectories.py`) removed; canonical pipeline (`build_xlxs.py`) is the sole aggregator.