inst/dev/design_obs.md

In epiforecasts/EpiNow2: Estimate Real-Time Case Counts and Time-Varying Epidemiological Parameters

Combining multiple observation models

The overall premise is that there is a single latent process (modelled as GP or with the renewal equation) that informs multiple observation processes. Each of these can have:

• a delay or set of delays
• truncation
• scaling
• accumulation
• week(end) effects
• different weights
• overdispersion

The delays need to be identified in case they are common across different observations, e.g.

• incubation period + reporting delay for symptomatic cases
• incubation period + hospital admission + another reporting delay for hospitalisations

We're assuming that delays can be equal but the observations are independent (not e.g. hospitalisation conditional on case report)

This means that the current options specified in delay_opts(), trunc_opts(), obs_opts(), the accumulation argument in forecast_opts() are all specific to the particular observation type, except delays that can be shared.

Implementation

Stan

Observations are converted from an array to a ragged array. The main required changes here will be lookups so that observations are correctly associated with different observation models. Some variables e.g. for day of the week effects will have to become 2D arrays.

R

The observation options mentioned above will have to be combined in a single interface. The proposed solution is that the current data, delays, truncation and future_accumulation become folded into a new obs_model() (based on obs_opts() and these can be passed as obs.

Independent observations

estimate_infections(
  obs = c(
    "Confirmed cases" = obs_model(
      data = reported_cases,
      delays = incubation_period + reporting_delay,
      truncation = trunc_opts(Lognormal(1, 2)),
      family = "poisson",
      accumulation = 7
    ),
    "Hospital admissions" = obs_model(
      data = admissions,
      delays = incubation_period + admission_delay,
      truncation = trunc_opts(Gamma(1, 1)),
      family = "negbin",
      weight = 0.5
    )
  ),
  generation_time = generation_time_opts(generation_time)
)

Dependent observations

This covers the case where only confirmed cases are ever admitted to hospital.

estimate_infections(
  ## "Confirmed cases" is the default name
  obs = obs_model(
      data = reported_cases,
      delays = incubation_period + reporting_delay,
      truncation = trunc_opts(Lognormal(1, 2)),
      family = "poisson",
      accumulation = 7,
      obs = c(
        "Hospital admissions" = obs_model(
          data = admissions,
          delays = admission_delay,
          truncation = trunc_opts(Gamma(1, 1)),
          family = "negbin",
          weight = 0.5
        )
      )
    ),
  ),
  generation_time = generation_time_opts(generation_time)
)

epiforecasts/EpiNow2 documentation built on June 9, 2025, 3:51 p.m.