Modelling options for reporting delays

InSight Net nowcasting strike team

Sam Abbott

2026-04-26

From three examples to a methods view

You have just seen three concrete cases:

Different surveillance streams, different operational responses, but a similar underlying data picture.

All three examples share right truncation: most events have happened, but some are not yet visible to us.

They also have other things going on — data revisions, sites coming and going, missing or messy fields — that need different kinds of methods.

Approaches

Simple: pruning recent dates, caveats, switching to report date
Individual-level delay estimation: fit a delay distribution to paired event/report records (line list); feed into other methods
Multiplicative (chain ladder): scale partial counts by historical completion rates
Regression: joint model of epidemic curve and delay surface, with covariates
Generative: joint model of counts and delay distribution, with explicit structure (parametric delays, pooling, mechanism)
Other statistical: machine learning, time-series methods

What you get: estimates of current levels and trends, usually with uncertainty.

Preliminary — we need your input.

Approaches

Nowcasting methods that do not assume monotonic (increasing) reporting, so they can absorb downward revisions (e.g. deduplication, reclassification)
Direct modelling of the revision process across snapshots

What you get: estimates that do not flip when a later release revises earlier values down.

Preliminary — we need your input.

Approaches

Pooling across sites, with explicit handling of which sites contributed at each time
Reporting-population models that separate “what happened” from “who was reporting”

What you get: jurisdiction-wide estimates that stay coherent as sites join, leave, or report intermittently.

Preliminary — we need your input.

Approaches

Pre-processing: deduplication, format harmonisation, strata cleaning
Generative / Bayesian methods can be adapted to many of these — e.g. epinowcast does missing-date imputation alongside the nowcast
More custom or ad-hoc work where no standard tool fits

What you get: principled handling of messy data, in the same model as the rest of the analysis where possible.

Preliminary — we need your input — please share your own approaches.

More detail in the guide’s modelling options section.

Most of the methods we work on sit in the epinowcast community:

epinowcast — modular Bayesian framework for nowcasting, Rt estimation, delay estimation, and forecasting from delayed data
epidist, primarycensored, baselinenowcast — R packages for delay distribution estimation, primary event censored distributions, and simple baseline nowcasts
EpiNow2 — R package for Bayesian Rt estimation, nowcasting, delay estimation, and forecasting with reporting delays
The EpiAware org for the same ideas in Julia, including CensoredDistributions.jl

If any of these are relevant to your setting, please grab me after the session.

Course — free, open-source, on nowcasting and forecasting:

NFIDD — Nowcasting and forecasting of infectious disease dynamics, available online any time
Taught in person at SISMID 2026 (Emory, July 2026): course site

Seminar series — talks from the wider community:

epinowcast seminars — recurring online seminar series on nowcasting, forecasting, and real-time methods

Materials are open and reusable; happy to point you to specific sessions if a topic from today is useful for your team.

As you listen, things to note: