tests/manual/test-largecounts.R

library(COMPoissonReg)
set.seed(1234)

n = 400
X = matrix(1, n, 1)
beta.true = log(2)
Xbeta.true = X %*% beta.true
lambda.true = exp(Xbeta.true)
gamma.true = log(0.2)
nu.true = exp(gamma.true)
y = rcmp(n, lambda.true, nu.true)
dat = data.frame(x = X[,1], y = y)

options(COMPoissonReg.optim.method = "L-BFGS-B")
glm.out = glm.cmp(formula.lambda = y ~ 1,
	formula.nu = ~1,
	data = dat,
	beta.init = beta.true, gamma.init = log(nu.true))
print(glm.out)

res = residuals(glm.out, type = "quantile")
qqnorm(res); abline(c(0,1), lty = 2, col = "red")


# Some combinations of parameters move the mass of CMP to very large values.
# This can lead to functions like rcmp taking a huge amount of computation
# and memory if we're not careful. We set a global ymax parameter to protect
# against this.

n = 5
offx = rep(5, n)
offs = rep(-5, n)
lambda.true = rep(3, n)
nu.true = rep(0.1, n)

options(COMPoissonReg.ymax = 1e3)
y = rcmp(n, lambda.true, nu.true)

options(COMPoissonReg.ymax = 1e4)
y = rcmp(n, lambda.true, nu.true)

options(COMPoissonReg.ymax = 1e5)
y = rcmp(n, lambda.true, nu.true)

options(COMPoissonReg.ymax = 1e6)
y = rcmp(n, lambda.true, nu.true)
lotze/COMPoissonReg documentation built on Feb. 11, 2024, 12:03 p.m.