ajrominger
commented
6 years ago yo yo i like it! the traditional choice would be to sample 1/sigma^2 from a gamma. that's the conjugate prior. but in a world where we're already simulating (i.e. not trying to get the analytical solution) there are those who advocate sampling sigma from a uniform. the reason being: our intuitive prior on sigma is often that it should be flat-ish (all values equally likely) and trying to get the square root of an inverse gamma distribution to be flat-ish can be hard. but what do you all think?
or something fancy, as a nuisance parameter.