Limitations

Data

• n = 52 cases is small. Posteriors are tight for the well-anchored delays (onset → admit n = 40; onset → death via convolution n ≥ 22) and wide for sparse or heavy-tailed ones (admit → discharge n = 15; onset → notification has max observed 86 d).

• No transmission pairs in the deposit. The serial and generation intervals cannot be estimated from this dataset — they need transmission-pair information we do not have.

• No exposure dates. The incubation period cannot be fitted from this deposit.

• 5 admission-date encoding outliers (−89, −5, −4, −1, 328720 days from onset) and 1 notification-date outlier (−62 days) are set to missing during loading. These look like onset/secondary date swaps but we cannot verify the direction with the custodian, so they are dropped rather than swapped.

• 3 of 52 cases have missing age, imputed at the sample mean before standardisation.

• Notification has a long right tail (47, 50, 53, 60, 65, 75, 86 days after onset) — likely retrospective notifications of community deaths investigated post hoc. Gamma handles this better than LogNormal; the mean (≈ 20 d) is still well above the median (≈ 11 d).

Model

• The HCW-stratified delay model (bdbv_model_stratified) still uses the marginalised double_interval_censored formulation and so treats per-case censoring as independent across delays. The main bdbv_model was converted to per-case latents (see "Model" in MODEL.md); the stratified variant is a smaller analysis used only for the HCW-vs-non-HCW odds-ratio table and has not yet been ported.

• No latent severity links delays and mortality. The delay fits and the CFR fits do not share information about case severity.

• Age effect is linear in standardised age. Nonlinearities (e.g. U-shape) are not captured.

• Sex is not in the CFR model. Rosello records sex (40 F / 12 M in Isiro 2012) but Kratz 2015's community-vs-ETC stratification is not reproducible from the deposit.

• 5 community deaths (died without recorded admission) contribute to the CFR block but not to the delay block; the model has no pathway for non-admitted deaths.

• The stratified delay model is restricted to the Gamma family. The LogNormal parametrisation in the stratified model would need a different SD prior to be consistent with the unstratified fit, and the Weibull shape prior needed extra truncation that complicated the stratified parametrisation. Gamma also fits the pooled data best (lowest WAIC), so reusing it here is a reasonable convenience; we have not tested whether each stratum is best fit by Gamma on its own.

Inference

• All reported results come from a single random seed (20260519). The convergence checks look strong, but we have not done a multi-seed run to confirm that.

Generalisability

BDBV is the third Ebolavirus species, with ~250 cases globally across three outbreaks (Uganda 2007–08, DRC 2012, DRC/Uganda 2026). Isiro is the only delay parametrisation we have. The Isiro outbreak was rural, mining-influenced, with MSF ETC support. The 2026 Ituri outbreak has urban (Bunia) and conflict-zone components that may push delays longer.

Out of scope

• Whether the HCW protective effect is causal (faster access to care) or selection (HCWs were preferentially confirmed). The probable-vs-confirmed adjustment partially controls for the latter but does not eliminate it.

• Reservoir or spillover dynamics.

Downstream priors

For downstream re-applications anchored on the 2012 Isiro outbreak (e.g. epiforecasts/BVDOutbreakSize), the exploratory exponential fit in Analysis walkthrough § Early-phase growth rate is the recommended source of a growth-rate prior. As a one-line summary, use Normal(0.008, 0.005) on r per day (posterior mean and SD from a Poisson regression on weekly onset counts from week 1 through the peak week). The 95% CrI on r covers zero — Isiro was a slow, noisy rise — so treat this as a weakly-informative prior rather than a tight constraint, and re-run fit_growth_rate(ll) if you want the live posterior summary.