**Freakonometrics - Tag - R-english**, and kindly contributed to R-bloggers)

Here

and there, I mentioned two codes to generated quasiPoisson random

variables. And in both of them, the negative binomial approximation

seems to be wrong. Recall that the negative binomial distribution is

where

and in R, a negative binomial distribution can be parametrized using

two parameters, out of the following ones

- the size,
- the probability,
- the mean,

Recall also that the variance for a negative binomial distribution

should be

Here, we consider a distribution such that

and . In the previous posts, I used

i.e.

rqpois = function(n, lambda, phi) {

mu = lambda

k = mu/(phi * mu - 1)

r1 = rnbinom(n, mu = mu, size = k)

r2 = rnbinom(n, size=phi*mu/(phi-1),prob=1/phi)

k = mu/phi/(1-1/phi)

r3 = rnbinom(n, mu = mu, size = k)

r4 = rnbinom(n, size=mu/phi/(1-1/phi),prob=1/phi)

r = cbind(r1,r2,r3,r4)

return(r)

}

but as we can see below, none of those two functions work,

> N=rqpois(1000000,2,4)

> mean(N[,1])

[1] 2.001992

> mean(N[,2])

[1] 8.000033

> var(N[,1])/ mean(N[,1])

[1] 7.97444

> var(N[,2])/ mean(N[,2])

[1] 4.002022

with the first one, the expected value is correct, while it is the

overdispersion parameter for the second. Now, if we do the maths

correctly, it comes

which should now work,

> mean(N[,3])

[1] 2.001667

> mean(N[,4])

[1] 2.002776

> var(N[,3])/ mean(N[,3])

[1] 3.999318

> var(N[,4])/ mean(N[,4])

[1] 4.009647

So, finally it is better when we do the maths well.

**leave a comment**for the author, please follow the link and comment on their blog:

**Freakonometrics - Tag - R-english**.

R-bloggers.com offers

**daily e-mail updates**about R news and tutorials on topics such as: Data science, Big Data, R jobs, visualization (ggplot2, Boxplots, maps, animation), programming (RStudio, Sweave, LaTeX, SQL, Eclipse, git, hadoop, Web Scraping) statistics (regression, PCA, time series, trading) and more...