There are many different formulations of the negative binomial distribution (NBD).
Default Formulation
In the default formulation, it is the number of failures to occur in a series of Bernoulli trials before a target number of successes is reached.
For example consider the number of failures that occur before 10 successes (\(N\)) are achieved where the probability of success (\(\rho\)) is 0.5.
With R this can be achieved as follows.
N <- 10
rho <- 0.5
hist(rnbinom(1e+05, size = N, prob = rho))
Ecological Formulation
A common use of the NBD in ecology is to model over-dispersion in count data. In this situation a more useful formulation is in terms of the mean number of counts (\(\mu\)) and the dispersion (\(\phi\)) where the standard deviation of the distribution is given by
\[\sigma = \sqrt{\mu + \mu^2 \cdot \phi}\]
It is worth noting that if \(\phi = 0\) then the NBD is equivalent to the Poisson distribution.
R
In R, the alternative ecological formulation is parameterised in terms of size
(\(1/\phi\))
mu <- 100
phi <- 10
x <- rnbinom(1e+05, mu = mu, size = 1/phi)
mean(x)
## [1] 99.69504
sd(x)
## [1] 315.0446
TMB
In TMB, the default formulation dnbinom()
follows the default convention of using \(N\) (size
) and \(\rho\) (prob
).
However, an alternative formulation, dnbinom2()
, uses the mean (\(\mu\)) and variance (\(\sigma^2\)).
To reparameterise in terms of \(\mu\) and \(\phi\), the variance should be set to be \(\mu + \mu^2 \cdot \phi\).
STAN
In STAN, there is a formulation neg_binomial_2()
which uses the alternative R formulation except the second parameter (which is equivalent to \(\phi^{-1}\)) is called phi
.
For examples of the NBD in action see https://github.com/joethorley/bioRxiv-028274/blob/master/model-lek.R.
JAGS
In JAGS, the only formulation is the default formulation but with \(N\) (refered to as \(r\)) and \(\rho\) (refered to as \(p\)) switched to give dnegbin(p, r)
.
In order to reparameterise the default formulation in terms of \(\mu\) and \(\phi\), \(r\) should be \(1/\phi\) and \(p\) should be \(r/(r + \mu)\).