Interactive demo · Statistical Methods for Modern Biology · Chapter 4

Gamma-Poisson counts and variance-stabilising transformations

Draw a rate λ from a gamma distribution with mean μ and shape a, then a count X ~ Poisson(λ): the count has the gamma-Poisson (negative binomial) distribution, with mean μ and variance v(μ) = μ + αμ2, α = 1/a (left). Its spread grows with its mean. A transformation g with g′(μ) ∝ 1/√v(μ) makes the spread nearly constant (the delta method): the square root for Poisson counts, and gα(x) = arcsinh(√(αx))/√α for gamma-Poisson counts. The right panel shows the exact standard deviation of each transformed count against the mean.

Count distribution

Gamma shape a (α = 1/a)

Mean μ shown on the left

Highlight

Scoreboard (at the μ shown left)

Try these

Every standard deviation on the right is exact: it is computed by summing the probability mass function over the counts up to six standard deviations above the mean and beyond, until the probabilities fall below 10−18; there is no simulation. The dotted line is the large-mean limit √(aψ′(a))/2 of the arcsinh transformation, ψ′ being the trigamma function, which differs from the delta-method value 1/2 when a is small (see the notes).

In the lecture notes