Added post2prior(), which allows passing the posterior from one model fit into priors of a second model fit. This allows the model to be fit in blocks. There is a new vignette showing a full example.
Removed the f_ramp_rate argument and option from fit_seir().
Allow not fitting the initial social distancing ramp and allow specifying the NB2 phi prior as a log normal distribution to allow passing the posterior as a prior.
Decrease default time_increment to 0.25 and make ODE control more precise again after issues in leeway paper.
Add initialization option (MAP or random).
Stop making projections parallel for now (possible issue in leeway paper).
Change default time_increment to 0.5 now that the pandemic has been going on for many months. You could even increase this to 1 for more speed.
Remove estimation of fraction distancing (e and e_prior). There isn’t generally information to inform this parameter and it slows everything down.
Some of the default arguments have changed to accommodate the new estimated parameters and to make the model easier to work with. In particular, population size is now specified via the N_pop argument and i0 is now i0_prior and requires a vector of length 2 to specify the prior.
Add estimation of the fraction distancing (parameter e). In most cases there will not be information in the data to inform this parameter; however, the prior allows for the inclusion of uncertainty on this fraction.
Add estimation of i0, the total number of infected people at the initial point in time (default -30 days), rather than requiring it to be a fixed value.
Add estimation of the starting and ending dates of the initial ramp in of physical_distancing.
forecast_seir() has been renamed to project_seir().
plot_projection() and tidy_sier() have been added. See examples in project_seir().
fit_seir() now allows for multiple estimated blocks (segments) of fractions of normal contacts (f) through time. The f values are indexed by s for segment. I.e., f_s. They currently all share the same prior.
fit_seir() now allows for estimated blocks (segments) of sampling fractions for the first data type. The intent is that this is used for the reported cases in which the sampling fraction is unknown and can be estimated if another type of case data (e.g., hospitalizations) is included with a much better-known fraction of positive cases represented.
Forecasting is now done with the same Stan model as the fitting via forecast_seir(). This eliminates approximately half of the code base (the R reimplementation), means that all of the model fitting options also work with the forecasting, makes it much easier to add functionality and harder to make mistakes, and generates much faster forecasts. The f values are represented as a vector and so can represent any desired pattern.
fit_seir() now allows for NA values in the case data. These are omitted from the likelihood. If no NA values are present then a vectorized likelihood is used, which is slightly faster.
The output from forecast_seir() includes helpful metadata such as which time steps are forecasted and which time steps use a fixed (vs. estimated) f value.
fit_seir() now uses a time step of 0.2 by default, which seems to give similar results with faster fits.
Fixed f values can be set to start on an arbitrary date in the future.
project_fit() has been renamed to forecast_seir().
Arguments to the functions and within the Stan model have been standardized and cleaned up.
Added continuous integration testing via Travis https://travis-ci.org/github/seananderson/covidseir/builds.
Added many examples to fit_seir() and forecast_seir(). Removed the vignette for now as it was out of date and is currently replaced by those examples.
The previous plotting function has been removed for now. It was bloated and included many dependencies and wasn’t working with all options. A simpler version will likely be added back later.
Documentation has been cleaned up throughout.