EpiNow2 1.9.0

New features

• Added support for all dist_spec delay families and obs_opts() observation model selection (Poisson or negative binomial) to estimate_truncation(). The existing observation model functions (report_lp, report_rng) are now reused internally.
• Wired the likelihood and return_likelihood settings of obs_opts() through estimate_truncation() so that prior-only fits and loo-compatible log-likelihood output are available, matching estimate_infections() and estimate_secondary().
• Added a noise argument to estimate_truncation() that takes a dist_spec controlling the additive noise term, replacing the previously hardcoded sigma ~ normal(0, 1) T[0,].
• estimate_truncation() now warns via check_truncation_obs_opts() when obs_opts() settings unused by the truncation model (weight, week_effect, week_length, scale) are set to non-default values.
• Added estimate_dist() for fitting delay distributions from interval-censored linelist data using Bayesian inference, with support for lognormal, gamma, normal, exponential, and Weibull distributions.
• Added Exp(), Weibull(), and Normal() distribution constructors.
• Added Dirichlet() and updated NonParametric() to support estimating nonparametric delay distributions using a Dirichlet prior, as an alternative to fixed NonParametric() or parametric distributions. This uses a gamma normalisation trick for efficient sampling with ragged simplex support in Stan.
• Added a shared internal helper check_simulation_input() that bundles the repeated input data frame validation in simulate_infections() and simulate_secondary().
• Added as_forecast_sample() S3 methods for epinow, estimate_infections, forecast_secondary, and estimate_truncation objects, allowing direct conversion to forecast_sample objects via scoringutils::as_forecast_sample() for evaluation.

Breaking changes

• Removed deprecated arguments that have been erroring since v1.7.0/v1.8.0:
  • gp_opts(ls_mean, ls_sd, ls_min, ls_max): use ls instead
  • gp_opts(alpha_mean, alpha_sd): use alpha instead
  • obs_opts(phi): use dispersion instead
  • obs_opts(na): use fill_missing() instead
  • estimate_infections(filter_leading_zeros, zero_threshold, horizon)
  • estimate_secondary(filter_leading_zeros, zero_threshold)
  • epinow(filter_leading_zeros, zero_threshold, horizon)
  • regional_epinow(horizon)
  • format_fit(burn_in, start_date)
• Removed the internal function default_fill_missing_obs().
• The pop argument in rt_opts() and simulate_infections() now errors when passed a numeric value. Use Fixed(pop) instead.
• Deprecated accessors on model objects now error instead of warning:
  • estimate_infections(): $samples, $summarised
  • estimate_secondary(): $predictions, $posterior, $data
  • estimate_truncation(): $dist, $obs, $data, $last_obs, $cmf
  • epinow(): $estimates, $estimated_reported_cases, $summary, $plots, $estimate_infections
• summary.epinow(output) and summary.estimate_infections(type = 'samples') now error.
• Removed internal function extract_parameter_samples(). Use format_simulation_output() instead.

Deprecations

• estimate_delay() is soft-deprecated in favour of estimate_dist().

Model changes

• Changed the non-stationary Gaussian process used to model Rt to sample in a mean-centred form internally, eliminating the (R0, drift) ridge in the joint posterior that caused stuck chains and catastrophic R-hat values on some seeds. The user-facing interpretation is unchanged: the prior argument in rt_opts() is still the prior on the initial Rt. Implemented by sampling the trajectory mean internally and applying the user prior to the derived initial Rt with the Jacobian-correct change of variables, so the joint prior matches the pre-change model.
• Updated the breakpoint random walk to be centred over the observation window, removing the (R0, drift) ridge in the joint posterior alongside the equivalent fix for the non-stationary Gaussian process.
• Delay distribution discretisation now properly accounts for primary event censoring during model fitting, matching the correction already applied on the R side since v1.8.0. This improves accuracy for short delays where the observation window is large relative to the delay.
• Left truncation of delay distributions (e.g. excluding generation times of zero) is now handled analytically rather than by zeroing and renormalising, giving more accurate PMFs near the truncation point.
• Refactored the Gaussian process, convolution, and observation model Stan functions for efficiency and readability, including a shared matern_indices() helper for the Matern spectral densities, outer-product basis function construction in PHI() and PHI_periodic(), and a shared reporting_phi() helper for the negative binomial overdispersion.

Bug fixes

• Fixed a bug where the target_date argument to epinow() was silently ignored: the recomputed forecast horizon from update_horizon() was never propagated into the call to estimate_infections().
• Fixed a bug in forecast_infections() where the summary call to extract dates was using modified args instead of the original fit dimensions, causing a date-dimension mismatch when extending the R trajectory beyond the original observation period.
• Centralised the gating of dispersion on family in obs_opts(): dispersion is now NULL whenever family != "negbin", with a warning if the caller supplied one explicitly. As a result, reporting_overdispersion is no longer sampled from its prior in estimate_infections(), estimate_secondary(), simulate_infections() and estimate_truncation() when a Poisson observation model is used.
• Fixed a bug where estimate_dist() would fail with a "model fitting timed out or failed" error when observed delays had near-zero variance. Scale parameters are now initialised from the prior when variance cannot be estimated from the data, and a warning is issued.
• Fixed a bug where example_truncated was generated with the old discrete-CDF PMF while estimate_truncation() now uses the primarycensored-based PMF, causing biased parameter recovery. The dataset has been regenerated for consistency.
• Fixed a bug where get_parameters() failed with $ operator not defined for this S4 class on a fit with an estimated nonparametric (Dirichlet) delay when using the rstan backend. Posterior draws are now extracted in a backend-agnostic way.
• Fixed a bug where forecast_infections() errored when called on an estimate_infections object fitted with the cmdstanr backend (it was calling rstan::extract() directly on the CmdStanMCMC fit; now uses the backend-agnostic internal helper).
• A bug was fixed where the trailing entries of imputed_reports in estimate_infections() output were undefined due to a size mismatch in the Stan generated quantities block.

Documentation

• Added a vignette demonstrating delay distribution fitting workflows and posterior validation.
• Added a model definition vignette for estimate_dist().
• Added a model overview vignette with an architecture diagram showing how the package's models connect.
• Added a model features vignette providing a quick reference to all modelling options with links to detailed documentation.
• The delay distribution fitting vignette has been renamed from vignette("estimate-dist") to vignette("estimate_dist_workflow") for consistency with the other workflow vignettes; the model definition is at vignette("estimate_dist").
• Added a cross-reference from the estimate_truncation() vignette to estimate_dist(), summarising when to use each.
• Expanded the case studies vignette with additional literature references and public health surveillance examples.
• The lower-accuracy GP example in the estimate_infections_options vignette now uses basis_prop = 0.15 instead of 0.1, which fits cleanly under the new centred parameterisation.

Package changes

• The touchstone continuous benchmarks now also cover estimate_truncation(), estimate_secondary() and estimate_dist() alongside the existing epinow() configurations, giving a single standardised view of wall-clock performance across the main user-facing fitting functions.
• Updated the vendored primarycensored Stan code to version 1.5.0 (CRAN release 2026-06-04).