team_name: Epiforecasts / London School of Hygiene and Tropical Medicine
model_name: EpiNow2 (epiforecasts)
model_abbr: epiforecasts-EpiNow2
model_contributors:
  - name: Nikos Bosse
    email: nikos.bosse@lshtm.ac.uk
    affiliation: London School of Hygiene and Tropical Medicine
    twitter: ftargument
  - name: Sam Abbott
    email: sam.abbott@lshtm.ac.uk
    affiliation: London School of Hygiene and Tropical Medicine
    twitter: seabbs
  - name: Sebastian Funk
    affiliation: London School of Hygiene and Tropical Medicine
    twitter: sbfnk
website_url: https://epiforecasts.io/EpiNow2
repo_url: https://github.com/epiforecasts/europe-covid-forecast
license: mit
team_funding: Wellcome Trust via a Senior Research Fellowship to Sebastian Funk (210758/Z/18/Z)
  and the Health Protection Research Unit (grant code NIHR200908)
team_model_designation: secondary
data_inputs: ECDC deaths and cases
methods: Semi-mechanistic estimation of the time-varying reproduction number for latent
  infections mapped to reported cases/deaths.
methods_long: EpiNow2 implements a Bayesian latent variable approach using the probabilistic
  programming language Stan, which works as follows. For an initial, unobserved, seeding
  time infections were estimated using an exponential model with priors based on observed
  growth. For each subsequent time step, previous imputed infections were summed,
  weighted by an uncertain generation time probability mass function, and combined
  with an estimate of Rt to give the incidence at that time. The infection trajectories
  were then mapped to mean reported case counts by convolving over an uncertain incubation
  period and report delay distribution. Observed reported case counts were then assumed
  to be generated from a negative binomial observation model with overdispersion,
  multiplied by a day of the week effect with an independent parameter for each day
  of the week. Temporal variation was controlled using an approximate Gaussian process
  with a Matern 3/2 kernel. Rt was assumed to be constant over the forecast horizon
  although a correction was applied to adjust for the time-varying proportion of the
  population that was susceptible. Deaths were then modelled as a convolution of forecast
  cases combined with some scaling factor, a day of the week effect and a negative
  binomial observation model.
citation: https://doi.org/10.5281/zenodo.3957489