Hyperpriors

Bayesian methods

Since sigma (the standard deviation) cannot be negative, a prior on sigma should follow the positive half of any distribution, or a distribution that is always positive, like the exponential. McElreath often uses HalfCauchy.

Using a Exp(1) distribution is nice and intuitive for a first-timer IMO. "Values of sigma further from zero are exponentially less likely". A power law is very similar, but with a fatter tail.

What links here

Created (3 years ago)