dnbmix: Simulation-Based Forecast Distributions

In HIDDA/forecasting: Forecasting Based on Surveillance Data

Simulation-Based Forecast Distributions

Description

The function dnbmix() constructs a (vectorized) probability mass function from a matrix of (simulated) means and corresponding size parameters, as a function of the time point (row of means) within the simulation period. The distribution at each time point is obtained as a mixture of negative binomial (or Poisson) distributions.

Usage

dnbmix(means, size = NULL)

Arguments

means	a n.ahead x n.sim matrix of means.
size	the dispersion parameter of the dnbinom() distribution or NULL (Poisson forecasts). Can also be time-varying (of length n.ahead).

Value

a ⁠function(x, tp = 1, log = FALSE)⁠, which takes a vector of counts x and calculates the (log-)probabilities of observing each of these numbers at the tp'th time point of the simulation period (indexing the rows of means).

Author(s)

Sebastian Meyer

See Also

logs_nbmix() where this function is used.

Examples

## a GLARMA example
library("glarma")
y <- as.vector(CHILI)

## fit a simple NegBin-GLARMA model
X <- t(sapply(2*pi*seq_along(y)/52.1775,
              function (x) c(sin = sin(x), cos = cos(x))))
X <- cbind(intercept = 1, X)
fit <- glarma(y = y[1:883], X = X[1:883,], type = "NegBin", phiLags = 1)

## simulate the last four weeks (only 500 runs, for speed)
set.seed(1)
means <- replicate(500, {
    forecast(fit, n.ahead = 4, newdata = X[884:887,], newoffset = rep(0,4))$mu
})

## derive the weekly forecast distributions
dfun <- dnbmix(means, coef(fit, type = "NB"))
dfun(4000, tp = 1)
dfun(4000, tp = 4)
curve(dfun(x, tp = 4), 0, 30000, type = "h",
      main = "4-weeks-ahead forecast",
      xlab = "No. infected", ylab = "Probability")

HIDDA/forecasting documentation built on Jan. 11, 2024, 1:11 a.m.