epidemia (version 1.0.0)

epirt: Model Reproduction Rates

Description

epirt defines a model for reproduction rates. For more details on the model assumptions, please read the model description vignette.

Usage

epirt(
  formula,
  link = "log",
  center = FALSE,
  prior = rstanarm::normal(scale = 0.5),
  prior_intercept = rstanarm::normal(scale = 0.5),
  prior_covariance = rstanarm::decov(scale = 0.5),
  ...
)

Arguments

formula

An object of class formula which determines the linear predictor for the reproduction rates. The left hand side must take the form R(group, date), where group and date variables. group must be a factor vector indicating group membership (i.e. country, state, age cohort), and date must be a vector of class Date. This is syntactic sugar for the reproduction number in the given group at the given date.

link

The link function. This must be either "identity", "log", or a call to scaled_logit.

center

If TRUE then the covariates for the regression are centered to have mean zero. All of the priors are then interpreted as prior on the centered covariates. Defaults to FALSE.

prior

Same as in stan_glm. In addition to the rstanarm provided priors, a shifted_gamma can be used. Note: If autoscale=TRUE in the call to the prior distribution then automatic rescaling of the prior may take place.

prior_intercept

Same as in stan_glm. Prior for the regression intercept (if it exists).

prior_covariance

Same as in stan_glmer. Only used if the formula argument specifies random effects.

...

Additional arguments for model.frame

Value

An object of class epirt.

Details

epirt has a formula argument which defines the linear predictor, an argument link defining the link function, and additional arguments to specify priors on parameters making up the linear predictor.

A general R formula gives a symbolic description of a model. It takes the form y ~ model, where y is the response and model is a collection of terms separated by the + operator. model fully defines a linear predictor used to predict y. In this case, the <U+201C>response<U+201D> being modeled are reproduction numbers which are unobserved. epirt therefore requires that the left hand side of the formula takes the form R(group, date), where group and date refer to variables representing the region and date respectively. The right hand side can consist of fixed effects, random effects, and autocorrelation terms.

Examples

# NOT RUN {
library(epidemia)
library(ggplot2)
data("EuropeCovid")
options(mc.cores = 2)

data <- EuropeCovid$data
data$week <- lubridate::week(data$date)

# collect arguments for epim
args <- list(
  inf = epiinf(gen = EuropeCovid$si),
  obs = epiobs(deaths ~ 1, i2o = EuropeCovid$inf2death, link = scaled_logit(0.02)),
  data = data, 
  algorithm = "fullrank", # For speed - should generally use "sampling"
  iter = 2e4,
  group_subset = "France",
  seed = 12345,
  refresh = 0
)

# a simple random walk model for R
args$rt <- epirt(
  formula = R(country, date) ~ rw(time = week),
  link = scaled_logit(7)
)

fm1 <- do.call(epim, args)
plot_rt(fm1) + theme_bw()

# Modeling effects of NPIs
args$rt <- epirt(
  formula = R(country, date) ~ 1 + lockdown + public_events,
  link = scaled_logit(7)
)

fm2 <- do.call(epim, args)
plot_rt(fm2) + theme_bw()


# shifted gamma prior for NPI effects
args$rt <- epirt(
  formula = R(country, date) ~ 1 + lockdown + public_events,
  link = scaled_logit(7),
  prior = shifted_gamma(shape = 1/2, scale = 1, shift = log(1.05)/2)
)

# How does the implied prior look?
args$prior_PD <- TRUE
fm3 <- do.call(epim, args)
plot_rt(fm3) + theme_bw()
# }