In this release model run times have been reduced using a combination of code optimisation and testing to reduce the likelihood of long running edge cases. Model flexibility has also been increased, particularly for the back calculation approach which now supports an increased range of prior choices. A significant development in this release is the edition of the experimental estimate_secondary model (and supporting forecast and plot functions). This allows a downstream target to be forecast from an observation. Example use cases include forecasting deaths from test positive cases and hospital bed usage from hospital admissions. This approach is intended to provide an alternative to models in which multiple targets are estimated jointly.
prior, to backcalc_opts. This allows the use of different priors for the underlying latent infections when estimating using deconvolution/back-calculation rather than the package default of using a generated Rt model (enable this option by setting rt = NULL). The default prior remains smoothed mean delay shifted reported cases but optionally no prior can now also be used (for scenarios when the data is very untrustworthy but likely to perform extremely poorly in real time).In addition, the previously estimated infections can be used (i.e infections[t] = infections[t-1] * exp(GP)) with this being an approximate version of the generative Rt model that does not weight previous infections using the generation time.prior = "reports" to be a partial centred moving average rather than a right aligned moving average. This choice means that increasing the prior window does not alter the location of epidemic peaks as when using a right alighted moving average.init_fit to stan_opts() that enables the user to pass in a stanfit to use to initialise a model fit in estimate_infections(). Optionally init_fit = "cumulative" can also be passed which first fits to cumulative data and then uses the result to initialise the full fit on incidence data. This approach is based on the approach used in epidemia authored by James Scott. Currently stan warnings from this initial fit are broadcast to the user and may cause concern as the short run time and approximate settings often lead to poor convergence.estimate_secondary and forecast_secondary along with a plot method and a new option function (secondary_opts). These functions implement a generic model for forecasting a secondary observation (such as hospital bed usage, deaths in hospital) that entirely depends on a primary observation (such as hospital admissions) via a combination of convolving over a delay and adding/subtracting current observations. They share the same observation model and optional features used by estimate_infections and so support data truncation, scaling (between primary and secondary observations), multiple log normal delays, a day of the week effect, and various error models. stationary_opts allows for easy specification of the most common use cases (incidence and prevalence variables). See the documentation and examples for model details.discretised_gamma_pmf (discretised truncated Gamma PMF) and discretised_lognormal_pmf (discretised truncated lognormal PMF) to limit/clip the values of the parameters by prespecified lower and upper bounds.estimate_infections by reducing all standard deviations used by a scaling factor of 0.1 in create_initial_conditions.gt_mean (the mean of the generation time) to reject samples with a mean greater than gt_max (the maximum allowed generation time). Adds boundary checking to reject standard deviations that are negative. Adds a boundary check on R values to reject them if 10 times greater than the mean of the initial prior. In some scenarios this will require users to supply a prior not is not completely misspecified (i.e if the prior has a mean of 1 and the posterior has a mean of 50).update_rt (an internal stan function found in inst/stan/functions/rt.stan) to be vectorised. This change reduces run times by approximately 1- ~ 20% (though only tested on a small subset of examples) and opens the way for future model extensions (such as additive rather than multiplicative random walks, and introducing covariates).This release focusses on model stability, with a functional rewrite of the model implementation, finalising the interface across the package, and introducing additional tooling. The additional tooling includes: support for adjusting for and estimating data truncation, multiple approaches for estimating Rt (including the default generative Rt approach, de-convolution coupled with Rt calculation, and EpiEstim like estimation on observed cases but with a robust observation model), optional scaling of observed data, and optional adjustment of future forecasts based on population susceptibility. The examples have also been expanded with links out to Covid-19 specific work flows that may be of interest to users. The implementation and model options are now considered to be maturing with the next release planned to contain documentation on the underlying approach, case studies, validation, evaluation the various supported options, and tools for dealing with secondary reports that are dependent on a primary report (i.e hospital admissions and hospital bed usage). If interested in contributing to any of these features please contact the package authors or submit a PR. User contributions are warmly welcomed.
estimate_infections to be divided into calls to _opts functions. Options are now divided by type for delays (delay_opts), Rt (rt_opts), backcalculation (backcalc_opts), the Gaussian process (gp_opts), and stan arguments (stan_opts). This has resulted in a larger number of the arguments from estimate_infections being folded into the related _opts function. Please see the function documentation and examples for details._opts function in regional_epinow using the helper functions opts_list and update_list or alternatively by constructing a named list with an entry for each region to be estimated.{EpiEstim} (here with a default window of 1 but this can be updated using backcalc_opts(rt_window)) but this approaches incorporates uncertainty from all inputs in a single estimate.epinow and estimate_infections.simulate_infections function from @sbfnk which allows the simulation of different Rt traces when combined with estimates as produced by estimate_infections. This function is likely to form the basis for moving all forecasting out of estimate_infections which may improve model stability.gp_opts. The length scale is now defined using a log normal truncated prior with a mean of 21 days and a standard deviation of 7 days truncated at 3 days and the length of the data by default. This prior is an area of active research and so may change in future releases.1 / sqrt(half_normal(0, rho_prior)) based on this advice and as the over dispersion being measured is in reports and not infections and hence a priori there is not strong evidence for over dispersion (which may be the case for infections) so the previous prior was overly weighted towards this.obs_opts. This removes week_effect and family from the main argument list which will allow for future extensions. Also adds a new argument scale which controls the uncertain fraction of cases that are eventually observed (defined as normally distributed). Setting this parameter will not impact Rt estimates.rt, using rt_opts, this includes the initial prior,use_breakpoints, and future. Adds a new helper argument rw which enables easy parameterisation of a fixed length random walk. These changes also help make it clear that these arguments only impact the Rt generative model and not the back calculation model.{epidemia} when Rt is fixed into the future (set by passing a population to rt_opts(pop = initial susceptible population). Note this only impacts case forecasts and not output Rt estimates and only impacts estimates at all beyond the forecast horizon as those based on data already account for population susceptibility by definition. The impact of this assumption can be explored using simulate_infections (by updating est$arg$pop in the example).truncation as a new argument to estimate_infections and higher level functions. This takes output from trunc_opts and allows for internally adjusting observed cases for truncation. A new method estimate_truncation has also been added to support estimating a log normal truncation distribution from archived versions of the same data set though this method is currently experimental.estimate_delay as a user friendly wrapper around bootstrapped_dist_fit.epinow and in regional_epinow for large batch runs.regional_epinow) to be based on whether or not usage is interactive.burn_in argument of estimate_infections as updates to model initialisation mean that this feature is likely no longer needed. Please contact the developers if you feel you have a use case for this argument.convert_to_logmean and convert_to_logsd).expose_stan_fns that exposes the internal stan functions into R. The enables unit testing, exploration of the stan functionality and potentially within R use cases for these functions.warmup to be 250 samples and the default adapt_delta to be 0.98.{lifecycle} badges to all functions to indicate development stage.This release introduces multiple breaking interface changes. Please see the README for examples of the new interface. It adds a range of quality of life improvements including updating the stan interface to support fitting each chain independently and offering variational inference as an alternative, experimental, fitting option. Notably it also adds support for nesting logging and a parallel enabled progress bar via the progressr package. Minor bugs have been fixed in the core model implementation focussing on stability and several already implemented features have been extended. Major model developments are planned for the next release of EpiNow2.
method = "exact") or Variational inference (method = "approximate").get_dist, get_generation_time, get_incubation_period based on ideas from @pearsonca. (This leads to breaking changes with the removal of covid_generation_times and covid_incubation_periods).setup_logging to enable users to specify the level and location of logging (wrapping functionality from futile.logger). Also added setup_default_logging to give users sensible defaults and embedded this function in regional_epinow and epinow.setup_future to making using nested futures easier (required when using future = TRUE).progressr.regional_epinowregional_epinowregional_runtimes that summarises the run time across regions.estimate_infections interface and expanded the range of options for the future_rt argument. Users can now choose to set Rt from any time point referenced to the forecast date.plot_summary.discretised_gamma_pmf (discretised truncated Gamma PMF) to constrain gamma shape and (inverse) scale parameters to be positive and finite (alpha > 0 and beta > 0).readLines incomplete final line warnings.fit_chain function where an interaction between rstan and timing out may have introduced an exception that caused whole regions to fail. This did not show on current unit tests or exploration using examples indicating a gap in testing.create_stan_args with the option to override using stan_args. This leads to breaking changes - see the examples for details of the new interface.samples argument to get_regional_results to make loading in samples optional. This also allows samples to be dropped when using regional_epinow which reduces RAM usage.global_map and country_map expanded by @ellispglobal_map from @ellispregional_summary now saves input reported cases data reported_cases.csv.