Changelog

Source: NEWS.md

composableIDModelR (development version)

Breaking changes

• The package has been renamed from EpiAwareR to composableIDModelR, and epiaware_setup_julia() and epiaware_available() are now cidm_setup_julia() and cidm_available().
• The package now wraps ComposableTuringIDModels.jl, the successor to EpiAware.jl, and the interface has been rewritten around it. Code written for earlier versions needs updating.
• Components mirror their Julia constructors: argument names follow Julia (e.g. AR(damp, init, epsilon_t), Renewal(generation_time, rt, initialisation), NegativeBinomialError(cluster_factor)), and distributions come from distspec (Normal(), Gamma(), LogNormal(), Exponential(), Weibull(), Beta(), Fixed(), NonParametric(), re-exported), replacing norm(), gamma_dist(), lognorm() and exponential(). HalfNormal() (parameterised by its mean) and truncated() replace halfnorm() and truncnorm().
• The latent process now belongs to the infection model (e.g. Renewal(rt = AR())), and IDModel(infection, observation) replaces EpiProblem(). There is no tspan: pass the observations to fit directly.
• fit(model, y, method = nuts()) takes a numeric vector of observations and replaces nuts_sampler() with nuts(). Pathfinder initialisation has been removed.
• epiaware_call() has been replaced by julia_spec() and julia().
• plot(type = "posterior") has been removed; use bayesplot on posterior::as_draws_array(fit) instead.
• The Julia setup dependency is now juliabridge, which juliaready was renamed to. An install from GitHub picks it up through Remotes; an existing checkout needs the dependency reinstalled.
• julia_spec(), julia() and as_julia() come from juliabridge, which composableIDModelR re-exports. The code that renders a component to Julia source moved there so that other packages can build Julia calls the same way. The role of a component is given as .role, leaving every Julia keyword name free; components carry classes such as julia_latent and julia_model; and one prints as <julia model spec>. A component built without a role is accepted wherever one is expected, since the role is what the caller says the component is.

New features

• Components are lazy R objects rendered to Julia code, so models can be built and printed without starting Julia. as_julia() shows the Julia code.
• New components: MA(), RandomWalk(), IID(), Intercept(), FixedIntercept(), HierarchicalNormal(), DiffLatentModel(), DirectInfections(), ExpGrowthRate(), PoissonError(), NormalError(), and Ascertainment().
• Delays and generation times with uncertain distspec parameters (e.g. LogNormal(meanlog = Normal(1.6, 0.2), sdlog = Normal(0.4, 0.05), max = 15)) are inferred.
• simulate() draws from a model’s prior.
• predict() returns posterior predictive draws and forecasts.
• plot() gains type = "infections", type = "latent" and a horizon argument for forecasts.

Bug fixes

• Forecasting from a fit whose Julia session has gone says so, rather than setting Julia up again and then failing inside Julia.