# Jeffreys prior for inverse gamma distribution

Does anybody have the experience of dealing with Jeffreys prior?

I am working with hierarchical model at the moment where the parameter σ^2 from normal distribution is said to be chosen according to inverse gamma distribution from Jeffreys prior.

How should it work? And why are inverse gamma distributions used so often as the prior for variances?

I found in some presentations that I can choose Jeffreys prior as, for example, IG(0.001,0.001) But it looks strange cause when both alpha and beta are <1 the gamma function behaves in a strange manner. If I try to generate values from IG(0.001,0.001) in R with code like

a=rgamma(1000,0.001,0.001)
aa=1/a
mean(aa)


I recieve Inf value since all values of a are close to 0 and some of them are exactly 0.

So how should one choose Jeffreys prior?

• There is a good answer here about the use of an IG(0.001,0.001) on variance parameters. – Robert Long Apr 10 '13 at 11:46

Because they're conjugate - at least for $\sigma^2$ in normal models.
You seem to have some confusion between Jeffreys' prior and (more-or-less) "uninformative" conjugate priors more generally. If you are taking a Jeffreys' prior, there's no choice involved: $p(\theta) \propto \sqrt{I(\theta)}\,$
where $I(.)$ is the Fisher information.