Sub-national heterogeneity in the time-varying reproduction number during the 2026 Bundibugyo virus disease outbreak in the Democratic Republic of the Congo: a hierarchical Bayesian analysis
Listed in
This article is not in any list yet, why not save it to one of your lists.Abstract
National-level estimates of the time-varying reproduction number (Rt) for the 2026 Bundibugyo virus disease (BDBV) outbreak in the Democratic Republic of the Congo (DRC) declined from a value close to the epidemic threshold in early August 2026 to a modestly sub-threshold value by late August, consistent with an independent case-count series compiled separately from the same underlying situation reports. A single national Rt, however, can obscure divergent sub-national epidemic trajectories, particularly across an outbreak that had by late August reached six provinces, ranging from a declining original epicentre to recently-seeded fronts. We estimated Rt at national, provincial, and — where case volume allowed — health-zone level, using both a standard sliding-window (Cori) estimator and a hierarchical Bayesian renewal model with partial pooling across spatial units, fitted by Hamiltonian Monte Carlo (No-U-Turn Sampler), with a province- or zone-specific (rather than shared) dispersion parameter retained on the basis of leave-one-out cross-validation. As of the week of 20–26 August 2026, Ituri — the outbreak’s original epicentre — had a hierarchical median Rt of 0.79 (95% credible interval [CrI] 0.61–1.04), Nord-Kivu remained above threshold (1.10, [0.89–1.39]), and Haut-Uele, now individually resolved at health-zone level for the first time, sat near threshold (1.01, [0.69–1.38]). Health-zone disaggregation, feasible in Ituri, Nord-Kivu, and now Haut-Uele given case volume, showed the provincial picture itself masked further heterogeneity, and that this heterogeneity is not static: Mongbwalu, the zone in which the outbreak began, remained in clear decline (Rt 0.37, [0.19–0.74]), while Rwampara — elevated three weeks earlier — had reversed to a clearly sub-threshold trajectory (0.53, [0.35–0.80]); conversely Butembo, previously in decline, had reversed to an elevated trajectory (1.36, [0.85–2.07]) comparable to the persistently-elevated Katwa (1.09, [0.82–1.41]). An initial disagreement between the sliding-window and hierarchical provincial estimates was traced to a data-reconstruction artefact (forward-filling, rather than interpolating, multi-day gaps in health-zone reporting) rather than a genuine methods disagreement; a targeted validation experiment, injecting an equivalent reporting gap into an independent historical outbreak’s case series (Beni health zone, 2018–2020 North Kivu/Ituri epidemic), confirmed that linear interpolation reduces reconstruction error roughly five-fold relative to forward-fill under the same failure mode, indicating the correction generalises beyond this specific outbreak. The hierarchical model achieved 96.1% pooled 95% posterior-predictive interval coverage, and the provincial ranking was unchanged across a generation-interval sensitivity grid (Spearman ρ = 1.0). Aggregation masks meaningful heterogeneity in transmission intensity at every spatial resolution examined, and that heterogeneity itself shifts over periods as short as two to three weeks; response prioritisation based on a single national, provincial, or even a single dated health-zone snapshot risks directing attention away from where transmission is currently supercritical.
Author Summary
During an Ebola outbreak, public health teams track a number called the reproduction number, or Rt — roughly, how many new people each infected person goes on to infect. When Rt is above 1, an outbreak is growing; below 1, it is shrinking. For the 2026 Bundibugyo virus disease outbreak in the Democratic Republic of the Congo, the national Rt has moved from close to 1 in early August 2026 to modestly below 1 by late August. We show that this single national number hides very different, and rapidly changing, realities in different places. The province where the outbreak started, Ituri, has an estimated Rt below 1 overall — but even within Ituri, specific towns move independently: one area that was still growing three weeks ago has since turned around, while the original outbreak site continues to decline. In Nord-Kivu province, one town that had been declining has since reversed and is now growing again, while a neighbouring town remains elevated throughout. We used a statistical technique called hierarchical Bayesian modelling, which allows data-poor areas to “borrow strength” from data-rich ones rather than producing wild, uninterpretable estimates, and which we validate three separate ways, including by deliberately reproducing a real data problem in a completely different, older outbreak’s data to check that our fix for it generalises. We also found and corrected a data-processing error that had made a simpler method appear to disagree with our more careful one — a reminder that methodological rigour includes checking the data pipeline, not just the statistical model. Our results suggest that outbreak response resources should be targeted at the specific health zones still driving transmission, and reassessed frequently, since a national, provincial, or even a single dated health-zone reading can be overtaken by events within a few weeks.
