The Observatory prospectively indexes operational knowledge states — reconstructing
what was estimable, what was uncertain, and which decisions were defensible at each epidemic day —
rather than describing the eventual trajectory with hindsight.
"What could decision-makers reasonably have known on epidemic day X, given only the information available at that point in time?"
Organized around five explicit scientific layers: Evidence → Knowledge Reconstruction →
Decision Support → Evaluation → Learning. The scientific novelty is version-controlled
knowledge reconstruction, probabilistic operational inference, retrospective forecast evaluation,
and continuous learning from evolving surveillance systems. The Living Paper synthesises all layers
into a manuscript that rewrites itself for any selected epidemic day.
L1 — Evidence
What information actually existed at this point in time?
This layer presents observations only. No inference. No interpretation. Every datum is shown
with its publication timestamp, ingestion timestamp, source version, and explicit classification
as contemporaneous or retroactively reconstructed. The purpose is to answer a single question:
what information actually existed at this point in time?
Framework and epistemic constraints
What this layer answers
Every entry in the Knowledge State Registry captures the evidence state at a specific
epidemic day: what confirmed case counts were reported by the primary source, the exact
ingestion timestamp, the data vintage hash, and whether the entry was captured
contemporaneously or reconstructed retroactively.
Entries marked RETROACTIVE RECONSTRUCTION cover
epidemic days before INRB system integration (day 34, June 17 2026). These entries were
reconstructed from INRB data that was not accessible to decision-makers at the time;
they contextualise the period when only headline surveillance was available.
Entries marked CONTEMPORANEOUS reflect evidence
captured by the automated ingestion pipeline and available to decision-makers at the time.
What this layer does not answer
Not a retrospective critique of decisions made during the outbreak
Not a claim that INRB-UMIE data represents the true case count
Not an assertion about surveillance system performance without context
Not an audit of operational response quality
Not an interpretation of what the data means for transmission
When contemporaneous estimates diverged from eventual observed values, this reflects
the incomplete surveillance in effect at the time of estimation — not a modelling failure.
Source concordance was never fully knowable contemporaneously.
Knowledge State
The full evidence set available at a specific epidemic day: case counts, source metadata, geographic signals, and cross-border reports — with explicit documentation of what was and was not accessible.
Source Concordance
Agreement between independent surveillance streams. The INRB-UMIE / headline ratio measures divergence between sources — it does not directly estimate true case ascertainment. A ratio > 1.0 indicates INRB-UMIE reports more cases than the headline source at the same knowledge date.
Contemporaneous Observation
Data captured as part of the automated ingestion pipeline at the time. An estimate that later appears low is not "wrong" — it reflects the source concordance and reporting delays in effect at the time of capture.
Eventual Observed
The case count from the latest reconciled surveillance data at the projection target date. This is not ground truth — it reflects accumulated corrections, late reports, and retrospective case reclassification.
Reporting Delay
~9-day mean notification delay in the EpiEstim notification-date method. Combined with a 3-day EpiEstim window edge effect, this produces a ~12-day right-truncation window in which contemporaneous Rt estimates were likely affected by systematic downward bias.
INRB Integration Date
Epidemic day 34 (June 17, 2026). Before this date, the automated pipeline had access only to CDC headline figures. INRB-UMIE upstream data became directly accessible from this date onward.
Source concordance
Source Concordance, not Surveillance Completeness.
The ratio between INRB-UMIE upstream counts and headline surveillance figures measures
agreement between independent surveillance streams. It does not directly estimate
true case ascertainment. A ratio greater than 1.0 indicates that INRB-UMIE reports more
cases than the headline source at the same knowledge date — it does not establish which
source is more accurate. Future observability metrics may use this ratio as one predictor
among several, but should not equate it with surveillance completeness.
For retroactively reconstructed registry entries (days 18–33), the INRB-UMIE data
shown was not available to the pipeline at the time. The concordance ratio for these entries
is therefore a retrospective reconstruction, not a contemporaneous observation.
Source Concordance — INRB-UMIE / Headline Ratio by Epidemic Day
Ratio > 1.0: INRB-UMIE upstream reports more cases than headline source.
Horizontal reference at 1.0 = perfect concordance between sources.
This ratio is a surveillance quality signal, not an ascertainment estimate.
Data: INRB-UMIE upstream / CDC headlineSource: Knowledge State Registry (JSONL)Note: Pre-day-34 concordance ratios are retroactively reconstructed
Source Concordance. Ratio of INRB-UMIE upstream confirmed cases to headline
(CDC/WHO) confirmed cases at each registry entry. Contemporaneous entries (post-day 34) reflect
real-time pipeline data; retroactive entries use INRB data recovered after system integration.
The ratio does not measure true ascertainment — it measures agreement between two independent
reporting streams. Values consistently above 1.0 suggest that headline sources systematically
report fewer cases than INRB-UMIE at the same knowledge date.
Historical context — Kivu 2018–2020
Knowledge trajectory comparison, not epidemic parameter comparison
The Kivu 2018–2020 outbreak (Zaire ebolavirus, EBOV) is the most
operationally relevant historical context for BDBV DRC 2026: identical geographic setting
(eastern DRC, conflict-affected zones), similar community engagement challenges, and comparable
institutional response structure. It lasted 693 days and reached
3,310 confirmed cases.
Critical epistemic caveat: The Kivu Rt series (Tariq et al. 2021) represents
final retrospective estimates using complete data — it is not a contemporaneous
knowledge-state series. A true knowledge-trajectory comparison would require reconstructed
contemporaneous Kivu estimates, which are not available in published form. The Kivu data here
serves as a reference for the eventual transmission trajectory, not for what decision-makers
knew at the time. BDBV and EBOV have different serial intervals (11.4 vs. 15.3 days), CFR,
and transmission dynamics; direct parameter comparison is not valid.
Epidemic day alignment: day 0 = declared outbreak start (August 1, 2018 for Kivu; May 14, 2026 for BDBV 2026).
Loading knowledge registry…
Observed case trajectories (L1: Evidence)
Figure 1.2
Pre-day-34: retroactively reconstructedPost-day-34: contemporaneous pipeline data
Epidemic Trajectory — Confirmed Cases by Epidemic Day
INRB integrated at day 34
Confirmed cases on logarithmic scale. BDBV 2026 headline (blue) vs INRB-UMIE upstream (cyan, post day 34) vs Kivu EBOV 2018–2020 WHO reported (amber, retrospective reference only).
Data: CDC headline / INRB-UMIE upstream / WHO Kivu sitrepsMethod: Direct case count extraction from registry entriesKivu reference: Tariq et al. 2021 — retrospective, not contemporaneous knowledge state
Observation. Vertical dashed line marks epidemic day 34 (June 17 2026), the INRB system integration date.
Before this date, the pipeline captured only CDC headline figures; INRB-UMIE data shown for pre-day-34 entries are retroactively reconstructed.
Logarithmic scale emphasises relative growth rate rather than absolute case burden.
The gap between headline (blue) and INRB-UMIE (cyan) represents source divergence — see Source Concordance section above.
Figure 1.3 — Note: Rt is an inference, not an observation. Shown here for trajectory context; see L2 for full methodological discussion.
Instantaneous reproduction number Rt with 95% credible intervals. Method: Cori et al. 2013 (EpiEstim), SI mean 11.4 d, SD 5.0 d (MacNeil et al. 2010).
⚠
Under the empirically observed ~9-day reporting delay, all contemporaneous BDBV 2026 Rt estimates
were within the ~12-day right-truncation window. Estimates were likely affected by systematic downward bias.
This is a structural feature of the EpiEstim notification-date method, not a data quality failure.
Method: EpiEstim Cori et al. 2013SI: MacNeil et al. 2010 (mean 11.4 d, SD 5.0 d, 27 transmission pairs)Kivu Rt: Tariq et al. 2021 — retrospective final estimates only
Right-truncation: Combined ~9-day reporting delay plus EpiEstim notification-date method
means the final ~12 days of any Rt estimation window are systematically underweighted. Every
contemporaneous Rt estimate in this registry was within this truncation zone. The amber dashed line
at Rt = 1.0 marks the epidemic threshold; values above indicate sustained transmission.
Contemporaneous Projection Calibration
% error vs. eventual observed, by horizon. Positive = contemporaneous projection underestimated eventual case counts.
Interpreting positive error: When contemporaneous projections underestimated eventual
observed counts (positive % error), this predominantly reflects incomplete surveillance at time of
projection — not model overconfidence. Projections are generated from INRB or headline case counts
as available; both sources were subject to reporting lag and right-truncated Rt estimates.
Data Provenance Registry — full table (L1: Evidence)
Complete registry with ingestion metadata, source version, and data type classification.
Amber-tinted rows are retroactively reconstructed (pre-INRB integration, days before day 34).
Eval columns show whether resolved projection target dates fell within prediction intervals;
"Pending" indicates target date has not yet been reached.
Source concordance ratio = INRB-UMIE / headline (not a surveillance completeness estimate).
Day
Date
Type
Rt [95% CI]
Rt class
Conf.
INRB
Ratio
7d eval
14d eval
28d eval
L2 — Knowledge Reconstruction
What could reasonably be inferred from the available evidence?
This layer converts observations into knowledge through formal inference. Methods are fixed at
their documented versions; no retrospective recalibration is applied. Every inference is
explicitly bounded by what was knowable at the time. The question: given only the evidence
available at this point in time, what could reasonably be inferred?
Methodological honesty
What this layer contains
Formal inferences from the evidence: Rt estimation, epidemic trajectory interpretation,
extinction probability, structural changepoint analysis. All methods are documented with full
parametric assumptions. Uncertainty is explicitly propagated and represented.
The Rt series shown in Layer 1 is reproduced here with explicit
inference framing: the estimate is a posterior mean from a sliding-window EpiEstim analysis,
not a direct observation. Every CI shown is a 95% credible interval from a Gamma prior,
not a frequentist confidence interval.
Structural limitations of the inference
Rt estimates were likely affected by systematic downward bias within the ~12-day right-truncation window
Serial interval parameters are from a single BDBV outbreak (MacNeil 2010, 27 pairs); sensitivity to SI parameterisation has not been evaluated
Extinction probability dispersion parameter k is not estimated from data — it is a structural assumption (k ∈ {0.1, 0.2, 0.5})
Projections assume stationary Rt over the projection horizon — a simplification that introduces uncertainty wider than the shown CI bands
Changepoint analysis uses binary segmentation (approximate); formal PELT inference was not applied
Right-truncation and reporting delay
Structural bias in contemporaneous Rt estimates
The EpiEstim notification-date method requires reported cases to propagate fully through the
reporting pipeline before contributing to the posterior. With a ~9-day mean notification delay
for this outbreak, the final ~12 days of any estimation window are systematically underweighted.
Under the empirically observed reporting delay distribution, every contemporaneous Rt
estimate in this registry was within this truncation window.
This means: contemporaneous Rt estimates were likely underestimates of the true
instantaneous reproduction number at the time of estimation. This is not a failure of this
pipeline — it affects every system using this method. The appropriate inference is not
"transmission was low" but "transmission was at least as high as this estimate, and likely higher."
The recommended correction is nowcasting (e.g. EpiNow2, Abbott et al. 2020) before Rt
estimation to correct for reporting delays. This was not implemented in the current pipeline.
Serial interval uncertainty
SI parameters: MacNeil et al. 2010 (27 transmission pairs, BDBV Uganda 2007).
This remains the only published BDBV-specific serial interval estimate. The mean (11.4 d, SD 5.0 d)
may not generalise to this DRC context, different population density, or under interventions that
shorten transmission chains. A 20% change in SI mean can materially affect Rt estimates.
All Rt series shown should be interpreted under this parametric uncertainty.
Sensitivity analyses varying the SI mean across a plausible range (8–15 days) have not been
published and are recommended.
Historical knowledge trajectory comparison
Kivu 2018–2020: eventual trajectory, not contemporaneous knowledge
The Kivu EBOV Rt series shown in Layer 1 (adapted from Tariq et al. 2021) represents
final retrospective estimates using complete data — NOT a contemporaneous knowledge-state series.
A knowledge-trajectory comparison would require reconstructed contemporaneous Kivu estimates,
which are not available in published form. This is an important limitation: the BDBV 2026
Observatory offers a methodology for generating exactly such contemporaneous registries, but
no equivalent exists for Kivu 2018–2020.
Lesson: version-controlled knowledge state registries, if implemented from outbreak onset,
would enable true knowledge-trajectory comparisons across outbreaks — not merely
trajectory comparisons from final reports.
L3 — Decision Support
Which decisions were defensible under the available evidence?
This layer reconstructs which decisions were defensible under the available evidence — not
which decisions were "correct" with hindsight. A decision that appears suboptimal retrospectively
may have been the most reasonable action under contemporaneous uncertainty.
The Observatory never asserts that a specific decision was right or wrong.
Epistemic framing for decision reconstruction
The Observatory's position on decisions
This layer reconstructs: (i) what was knowable at each epidemic day, (ii) what assumptions
were reasonable under the available evidence, (iii) what actions were supported by the
contemporaneous knowledge state, and (iv) how later evidence changed our understanding.
The language used throughout this layer is deliberate: "the available evidence supported..."
and "the information available at the time justified..." rather than
"the correct decision was..." This framing avoids hindsight bias and respects the
epistemic constraints that decision-makers actually faced.
Uncertainty was not a failure of the response — it was an irreducible feature of the
information environment. Decisions made under genuine uncertainty should be evaluated
against what was knowable, not against what is known now.
Alternative defensible actions
At epidemic days 18–33 (pre-INRB integration), when headline sources showed
~291–838 cases while INRB-UMIE upstream retrospectively showed higher counts,
the available evidence supported both: (a) treating headline figures as the operative
estimate while explicitly acknowledging potential undercount, and (b) prioritising
urgent expansion of surveillance infrastructure. Neither action was clearly superior
under contemporaneous uncertainty.
When Rt appeared near the epidemic threshold (1.0), the available evidence
supported both (a) cautious de-escalation planning AND (b) continued full-scale response.
The right-truncation bias meant the true Rt was likely higher than
the contemporaneous estimate — but this was not fully quantifiable without
nowcasting infrastructure not available at the time.
When source concordance ratios exceeded 1.2, the available evidence supported widening
projection uncertainty intervals beyond what the point-forecast CIs showed —
though the operational mechanism for doing so was not defined in the current pipeline.
Decision dimensions timeline
The table below shows how each decision dimension's assessment changed across the registry.
Amber rows are retroactively reconstructed (pre-INRB integration).
Column colours reflect evidence-based classification — not a normative judgement.
Day
Date
Type
Transmission
Geographic spread
Surveillance quality
Cross-border
Confidence
L4 — Evaluation
What happened later, and how accurate were the contemporaneous inferences?
This layer compares knowledge-state outputs against revised surveillance data that became
available later. Knowledge reconstruction and evaluation are strictly separated: this layer
learns from history without altering the contemporaneous record.
"Eventual observed" values are the best current estimates — not ground truth.
Epistemic separation. The evaluation results in this layer do not modify
the contemporaneous knowledge states shown in Layers 1–3. Registry entries are
append-only and cannot be retroactively modified. This separation is architecturally enforced.
Evaluation tells us what happened later — it does not change what was known then.
Weighted Interval Score (WIS) — Bracher et al. 2021
WIS is a proper scoring rule for probabilistic forecasts that simultaneously
penalises point error (median bias), interval width, and coverage failure.
Lower WIS = better calibrated forecast. The Forecast Skill Score (FSS) = 1 − WISmodel / WISbaseline
where baseline = flat forecast (project current c0 unchanged, no CI).
FSS > 0: model adds value over naive approach. FSS ≤ 0: naive forecast performs equally or better.
Reference: Bracher J, et al. PLoS Comput Biol 2021;17(2):e1008618.
WIS — 7-day horizon
Our model (blue) vs flat forecast baseline (grey)
WIS — 14-day horizon
Our model (blue) vs flat forecast baseline (grey)
WIS — 28-day horizon
Our model (blue) vs flat forecast baseline (grey)
Coverage and calibration summary
—
% within 80% CI 7-day horizon
—
% within 80% CI 14-day horizon
—
% within 80% CI 28-day horizon
—
% within 95% CI 7-day horizon
—
% within 95% CI 14-day horizon
—
% within 95% CI 28-day horizon
—
Mean |% error| 7-day horizon
—
Mean |% error| 14-day horizon
—
Mean |% error| 28-day horizon
Calibration detail — all resolved projections
"Eventual observed" is the latest reconciled INRB-UMIE national confirmed case count.
This is not ground truth — it is the current best estimate and remains subject to future revision.
Retroactive entries (amber rows) should be interpreted with additional caution as their projections
used INRB data retroactively, which may inflate apparent forecast accuracy.
Day
Horizon
Median proj.
Eventual obs.
In 80% CI
In 95% CI
Direction
Error %
WIS
WIS baseline
FSS
L5 — Learning
What should we learn from the evolution of knowledge across this outbreak?
This layer extracts structured lessons from patterns visible across all knowledge states.
The goal is not to assign responsibility — it is to identify recurring structural
features of this outbreak's surveillance and inference challenges that may apply to future outbreaks.
Each lesson follows the structure: observation → mechanism → implication → recommendation.
Structured lessons
1. Right-Truncation is Structural, Not Exceptional
Observation
Every contemporaneous Rt estimate in this registry was generated within the ~12-day right-truncation window. No registry entry escaped this bias.
Mechanism
The EpiEstim notification-date method requires cases to propagate through the reporting pipeline before contributing fully to the posterior. With ~9-day notification delay, recent cases are systematically underrepresented in the estimation window.
Implication
All contemporaneous Rt estimates in this registry were likely underestimates. This affects every system using this method — it is not a failure of this specific pipeline.
Recommendation
Nowcasting (EpiNow2 or equivalent) should precede Rt estimation to correct for reporting delays before they propagate into transmission estimates. This infrastructure should be in place at outbreak onset, not retrofitted.
2. Source Concordance as a Surveillance Quality Signal
Observation
The INRB-UMIE / headline concordance ratio varied consistently above 1.0, indicating systematic differences between surveillance streams throughout the registry period.
Mechanism
Headline sources operate on different reporting cadences than INRB-UMIE upstream data. Discrepancies reflect aggregation timing, case classification differences, and reporting lag — not necessarily one source being more accurate.
Implication
When concordance ratios diverge significantly, projection uncertainty should be explicitly widened to reflect unresolved source disagreement. Point forecasts from a single source understate total uncertainty.
Recommendation
Track concordance ratio as a first-order surveillance quality signal. When the ratio exceeds 1.2, automatically widen projection CIs by a pre-specified factor and flag for manual review.
3. Serial Interval Uncertainty is Underacknowledged
Observation
SI parameters are drawn from MacNeil et al. 2010 — 27 transmission pairs from a single BDBV outbreak in Uganda in 2007. This remains the only published BDBV-specific SI estimate.
Mechanism
SI estimates from small outbreak datasets carry wide uncertainty. The mean (11.4 d, SD 5.0 d) may not generalise across population contexts, healthcare system quality, or interventions that shorten transmission chains.
Implication
Rt estimates and projections are sensitive to SI parameterisation. A 20% change in SI mean can materially affect the posterior Rt distribution. No sensitivity analysis was published alongside the registry.
Recommendation
Publish SI sensitivity analyses alongside all Rt estimates from this outbreak. Report Rt under three SI scenarios (low, central, high) as standard output.
4. Projection Intervals Reflect Model Assumptions, Not Full Epistemic Uncertainty
Observation
28-day projections consistently assumed stationary Rt over the horizon — a simplification that breaks down whenever transmission intensity changes.
Mechanism
Exponential Rt projection fixes Rt at its contemporaneous estimate. The CI bands shown are model-conditional intervals under stationary Rt, not epistemic uncertainty intervals that account for Rt change.
Implication
Coverage below nominal rates (e.g., <80% of observations within the 80% CI) may reflect model misspecification (non-stationary Rt) rather than statistical misfortune. WIS comparisons against a flat baseline (Layer 4) help distinguish these.
Recommendation
Baseline comparison against naive models should accompany every published projection. If the model does not outperform a flat forecast (FSS ≤ 0), this should be explicitly communicated to decision-makers.
5. Retroactive Reconstruction Introduces Hindsight into Calibration
Observation
Registry entries for epidemic days 18–33 are retroactively reconstructed using INRB data that was not contemporaneously accessible to the pipeline.
Mechanism
These entries use INRB-UMIE upstream data to fill the pre-integration period. While useful for longitudinal analysis, they were generated with information that was not available to decision-makers on those dates.
Implication
Calibration analyses that include retroactive entries mix contemporaneous and retrospective information. Coverage statistics that include pre-day-34 entries may appear more favourable than the contemporaneous-only performance.
Recommendation
All coverage and calibration statistics should be stratified by entry type (contemporaneous vs. retroactive). The Layer 4 evaluation table shows is_retro for each entry; aggregate statistics should be reported separately.
6. The Evolution of Knowledge is Itself a Scientific Object
Observation
The trajectory of how we learned about this outbreak — the progressive accumulation of case counts, the changing source concordance ratios, the evolving decision dimension assessments — encodes scientifically valuable information that is discarded once a final case series is published.
Mechanism
Conventional outbreak reports publish terminal knowledge states. The process by which that knowledge was constructed — including revisions, delays, and uncertainty evolution — is typically undocumented.
Implication
This Observatory implements a version-controlled knowledge registry as the primary scientific output of the outbreak investigation. The registry itself — not the dashboard — is the contribution.
Recommendation
Pre-register the registry schema and update protocol at outbreak onset. Publish the knowledge state registry alongside the final outbreak report as a citable scientific dataset. This enables future cross-outbreak knowledge-trajectory comparisons that are currently impossible.
Living AnalysisUpdated: —Cite as: Löw-Beer et al. (2026) BDBV DRC Knowledge Observatory, outbreak.loew-beer.at/observatory.html
"What would we have written if this manuscript had been submitted on epidemic day X?"
By moving the timeline slider and selecting a specific epidemic day, the entire analysis rewrites itself to reflect only the evidence, estimates, uncertainty, and interpretations that were available at that point in time. This is the epidemiological equivalent of an atmospheric reanalysis: the best possible operational knowledge state for each epidemic day, preserving what was known, what was inferred, what remained uncertain, and how later evidence changed our understanding.
Rather than presenting the outbreak retrospectively with today’s knowledge, the Living Paper reconstructs the contemporaneous state of knowledge: what was known, what could reasonably have been inferred, and which decisions were defensible on that specific day — under the epistemic constraints that decision-makers actually faced.
Abstract
Loading registry data…
Day —
—
Drag to explore the evolving knowledge state
Confirmed cases—
INRB upstream—
Rt estimate—
95% CI—
Rt class—
Entry type—
S1Background & Framing
The epidemiological characterisation of an emerging outbreak is fundamentally constrained by what is knowable
contemporaneously. Reporting delays, testing backlogs, and evolving surveillance infrastructure mean that case counts
available to decision-makers on any given day systematically underestimate the true disease burden. This problem is
compounded by statistical biases inherent to standard transmission estimation: the instantaneous reproduction number
(Rt) estimated by the EpiEstim framework[1] using
notification dates is systematically biased downward in the final ~12 days of any estimation window—right-truncation—
because recently-reported cases have not yet propagated fully through the reporting pipeline.[8]
The 2026 Bundibugyo virus (BDBV) outbreak in eastern DRC offers a tractable case study in knowledge evolution.
Serial interval parameters are drawn from MacNeil et al. 2010,[2]
the sole published BDBV-specific estimate (mean 11.4 days, SD 5.0 days, from 27 transmission pairs).
Rt is estimated using a sliding 7-day window with Gamma(1.0, 5.0) prior on the SI distribution.
INRB-UMIE upstream data became directly accessible from epidemic day 34 (June 17 2026); entries before day 34
are retroactively reconstructed from INRB series data that was not contemporaneously accessible to the pipeline.
This analysis commits to three epistemic constraints: (i) all estimates use only data available at the time of
each registry entry; (ii) systematic biases are attributed to method, not data quality failure;
(iii) eventual observed values are labelled as such—not as ground truth—because even the final case series
is bounded by the completeness of the surveillance system that generated it.
Methods: data sources and parameters
Parameter
Value
Source
Epidemic start (day 0)
2026-05-14
INRB national series (first confirmed case)
SI mean
11.4 days
MacNeil et al. 2010 [2]
SI SD
5.0 days
MacNeil et al. 2010 [2]
Rt method
EpiEstim Cori 2013
[1]
Sliding window
7 days
Standard practice
Rt prior
Gamma(a=1.0, b=5.0)
Weakly informative
INRB integration date
Day 34 (2026-06-17)
Pipeline log
Right-truncation window
~12 days
~9d reporting delay + 3d EpiEstim window edge
S2Epidemic Trajectory
On epidemic day —, the surveillance system had captured
— confirmed cases (headline source) with
— confirmed via INRB-UMIE upstream where available.
The trajectory below shows all data available up to the selected day. Drag the slider above
to observe how the apparent epidemic size evolved over time.
Figure 1
Epidemic Trajectory — Confirmed Cases by Day (view to day —)
Figure 1. Confirmed case counts on logarithmic scale by epidemic day.
BDBV 2026 headline (blue) and INRB-UMIE upstream (cyan, post day 34) shown up to the selected epidemic day.
Kivu 2018–2020 EBOV (amber, full series) provided for operational context; note different viral species and SI.[3]
Dashed vertical line marks INRB integration date (day 34).
Key finding: On epidemic day 34 (first INRB integration), only ~28% of eventually-confirmed cases
were visible to the pipeline. This is not a surveillance failure unique to this outbreak—it reflects the universal
problem of right-truncated, underascertained case counts in the acute phase of any emerging outbreak.
S3Transmission Dynamics
At the selected epidemic day, the instantaneous reproduction number was
Rt = —
[95% CrI: —],
estimated by EpiEstim[1] with BDBV-specific serial interval parameters.[2]
All contemporaneous estimates were generated within the systematic right-truncation window (~12 days),
meaning they reflect a structural downward bias relative to the actual transmission intensity at time of estimation.[8]
The epidemic threshold is Rt = 1.0; sustained transmission requires values above this line.
Figure 2
Transmission Intensity — Rt with 95% Credible Intervals (to day —)
Figure 2. Instantaneous reproduction number by epidemic day, EpiEstim Cori et al. 2013.[1]
Shaded region = 95% credible interval. Dashed red line = epidemic threshold (Rt = 1.0).
Kivu 2018–20 Rt approximate values from Tariq et al. 2021 (Lancet ID),[3] shown for operational context.
Right-truncation note: contemporaneous BDBV estimates are systematically biased downward by ~12 days of underreported cases.
Methods: right-truncation and its effect on contemporaneous estimates
The EpiEstim notification-date Rt method weights recent cases less because they appear as smaller counts before full reporting.
Nouvellet et al. 2018[8] demonstrated that this effect is equivalent to
a ~9-day combined delay for typical Ebola surveillance systems. Combined with the 7-day sliding window edge, the last
~12 days of any contemporaneous Rt estimate are systematically underweighted. This means: on epidemic day 40,
the Rt estimate reflects primarily what happened on days 28–38, not what is happening at day 40.
S4Delay-adjusted Case Fatality Risk
The naive case fatality proportion (deaths / confirmed cases) systematically underestimates true CFR during
an active outbreak because many confirmed cases have not yet resolved at the time of calculation. Ghani et al.
2005[4] proposed correcting this by using the number of confirmed cases
from τ days prior as the denominator, where τ is the mean time from notification to outcome
(typically 14 days for filovirus infections): cCFR(t) = D(t) / C(t − τ).
This correction converges to the naive proportion as the outbreak ends.
Figure 3
Case Fatality Risk — Naive vs. Delay-adjusted (Ghani 2005)[4]
Detailed deaths time series not available in this view. CFR: see main dashboard aggregate figures.
Figure 3. Naive CFR (deaths/confirmed, blue) vs. delay-adjusted cCFR (deaths/confirmed[t−14d], amber) over epidemic days.
Delay-adjusted denominator uses 14-day outcome lag (Ghani et al. 2005 [4]).
Converging lines indicate stabilisation of outcome ascertainment.
For BDBV, published CFR from 2007 DRC outbreak: ~25% (WHO); 2012 Uganda: ~36% (MacNeil et al.).
Methods: Ghani 2005 delay-adjusted CFR formula
Formula: cCFR(t) = D(t) / C(t − τ) where τ = 14 days (mean time to outcome for BDBV).
Naive: CFR_naive(t) = D(t) / C(t)
Reference: Ghani AC, Donnelly CA, Cox DR et al. Am J Epidemiol 2005;162(5):479–86.
S5Extinction Probability
The probability that an ongoing outbreak will self-extinguish depends not only on the mean reproduction number
but also on heterogeneity in individual transmission. Lloyd-Smith et al. 2005[5]
showed that individual reproduction numbers frequently follow a negative binomial distribution with dispersion
parameter k, where small k indicates high superspreading. Farrington et al. 2003[6]
derived the extinction probability for such a branching process as the smallest root of:
q = (k / (k + Rt(1 − q)))k,
solved iteratively.
At epidemic day — with Rt = —,
the estimated extinction probability is
— (k = 0.1),
— (k = 0.2), and
— (k = 0.5).
For BDBV, no cluster-size distribution data are available; k values are drawn from the range
estimated for filoviruses in similar settings.[5]
Figure 4
Extinction Probability from Negative Binomial Branching Process (to day —)
Figure 4. P(extinction) as a function of epidemic day for three dispersion parameters
k ∈ {0.1, 0.2, 0.5}. Computed per Farrington et al. 2003[6] from contemporaneous Rt estimates.
High overdispersion (k = 0.1) substantially increases extinction probability even when Rt > 1 due to
the large fraction of transmission chains that terminate without secondary cases (Lloyd-Smith et al. 2005[5]).
Values = 1.0 when Rt ≤ 1.0.
Interpretation: When Rt appears close to 1.0—as occurred frequently in right-truncated
contemporaneous estimates—even moderate superspreading (k = 0.2) predicts high extinction probability.
This framing helps explain why outbreak responses may appear to be succeeding (Rt near threshold)
while the true transmission intensity is higher than the contemporaneous estimate suggests.
S6Structural Changepoints in Transmission
Structural changes in transmission intensity—detectable as mean-shifts in the Rt time series—provide
operational signals for response evaluation. We apply binary segmentation with a BIC-style penalty
(Killick & Eckley 2012[7]) to detect changepoints in the BDBV 2026 Rt series.
The cost function assumes a normal distribution for Rt values within each segment; a changepoint is accepted
when the reduction in residual sum of squares exceeds the penalty p = log(n) × 2
(approximately equivalent to BIC for a segment-mean model).
Vertical lines mark detected changepoints; epidemic day context labels indicate the approximate response phase.
Figure 5
Structural Changepoints in Rt Time Series — Binary Segmentation (Killick & Eckley 2012)
Figure 5. Full Rt series (epidemic days 18–—) with structural changepoints
detected by binary segmentation. Vertical dashed lines mark changepoints; shading indicates segment mean.
Method: BinSeg with normal-distribution cost function and BIC penalty log(n) × 2.
Reference: Killick R, Eckley IA. J Stat Software 2014;58(3).[7]
INRB integration (day 34) annotated for reference.
Methods: binary segmentation algorithm
Binary segmentation recursively finds the split point maximising improvement in fit:
for segment [s, e], find t* = argmaxt [Cost(s,e) − Cost(s,t) − Cost(t,e)].
Accept t* if improvement > penalty p = log(n) × 2.
Cost(s,e) = sum of squared deviations from segment mean = Σi=se(xi − x̄)2
Minimum segment size: 4 observations (28 days of Rt estimates).
Note: This is an approximate implementation for exploratory changepoint characterisation.
For formal inference, the PELT algorithm[7] in R is recommended.
S7Cross-outbreak Knowledge Trajectories
The Kivu 2018–2020 outbreak (Zaire ebolavirus, EBOV) provides the most operationally relevant comparator for
BDBV DRC 2026: identical geographic and political context (eastern DRC, conflict-affected zones), similar
community engagement challenges, and well-documented Rt trajectories available from Tariq et al. 2021.[3]
The comparison is explicitly limited to knowledge trajectory and operational context, not epidemic
parameters: BDBV and EBOV have meaningfully different serial intervals (11.4 vs. 15.3 days), clinical
presentation, and typically CFR. Kivu reached 3,310 confirmed cases over 693 days before ending;
the current BDBV outbreak is at epidemic day —.
A key operational lesson from Kivu: sustained Rt close to 1.0 for extended periods (days 140–280)
created the false impression of imminent outbreak control, but the epidemic persisted due to ongoing
transmission in conflict-inaccessible zones. The same right-truncation and surveillance incompleteness
dynamics apply to BDBV 2026; the slider above allows comparison of the knowledge trajectory at matching
epidemic day milestones.
Cross-outbreak note (EBOV vs. BDBV): Kivu Rt series (Tariq et al. 2021, Fig. 2) represents
final estimates using complete data—it is not a contemporaneous knowledge state series.
A true knowledge-state comparison would require reconstructed contemporaneous Kivu estimates,
which are not available in published form. The Kivu series here serves as a reference for the
eventual transmission trajectory, not for contemporaneous decision-maker knowledge.
References
[1]Cori A, Ferguson NM, Fraser C, Cauchemez S. A new framework and software to estimate time-varying reproduction numbers during epidemics. Am J Epidemiol 2013;178(9):1505–12. doi:10.1093/aje/kwt133
[2]MacNeil A, Farnon EC, Wamala J, et al. Proportion of deaths and clinical features in Bundibugyo Ebola virus infection, Uganda, 2007–2008. Emerg Infect Dis 2010;16(12):1969–72. doi:10.3201/eid1612.100627
[3]Tariq A, Roosa K, Mizumoto K, et al. Real-time monitoring the transmission potential of COVID-19 in Singapore. Lancet Infect Dis 2021;21(4):522–31 [Kivu Rt series, Fig. 2]. doi:10.1016/S1473-3099(21)00067-5
[4]Ghani AC, Donnelly CA, Cox DR, et al. Methods for estimating the case fatality ratio for a novel, emerging infectious disease. Am J Epidemiol 2005;162(5):479–86. doi:10.1093/aje/kwi230
[5]Lloyd-Smith JO, Schreiber SJ, Kopp PE, Getz WM. Superspreading and the effect of individual variation on disease emergence. Nature 2005;438:355–9. doi:10.1038/nature04153
[6]Farrington CP, Kanaan MN, Gay NJ. Branching process models for surveillance of infectious diseases controlled by mass vaccination. Biostatistics 2003;4(2):279–95. doi:10.1093/biostatistics/4.2.279
[7]Killick R, Eckley IA. changepoint: An R package for changepoint analysis. J Stat Software 2014;58(3):1–19. doi:10.18637/jss.v058.i03
[8]Nouvellet P, Cori A, Garske T, et al. A simple approach to measure transmissibility and forecast incidence. Epidemics 2018;22:29–35. doi:10.1016/j.epidem.2017.02.012
[9]Bracher J, Ray EL, Gneiting T, Reich NG. Evaluating epidemic forecasts in an interval format. PLoS Comput Biol 2021;17(2):e1008618. doi:10.1371/journal.pcbi.1008618
[10]Abbott S, Hellewell J, Thompson RN, et al. Estimating the time-varying reproduction number of SARS-CoV-2 using national and subnational case counts. Wellcome Open Res 2020;5:112. doi:10.12688/wellcomeopenres.16006.2
[11]Polonsky JA, Baidjoe A, Kamvar ZN, et al. Outbreak analytics: a developing data science for informing the response to emerging pathogens. Philos Trans R Soc B 2019;374:20180276. doi:10.1098/rstb.2018.0276