When I study the Bayesian econometrics, the book firstly introduces Gamma-Normal distribution as (conjugate) prior, then the posterior will have the same distribution as the prior. But my question is, why can't we use other simple distribution as prior for the introduction example, such as normal distribution, why do we have to use a such as strange 'compound' distribution?


There is no reason to use any particular prior in a Bayesian analysis. Conjugate priors are chosen out of convenience so that it is easy to express the posterior distribution in a simple form as a member of the same family as the prior. With the introduction of Markov Chain Monte Carlo methods investigators use whatever priors they find to be realistic for the model parameters. Now for an itnroductory course it may be simpler to get the concepts across when it si easy to compute the posterior distribution. Simple priors will not always lead to nice posterior distributions but with conjugate priors you are safe. That is probably why it was done that way.

  • $\begingroup$ Thanks for the answer. If we use gamma-normal prior, the posterior will have closed form density function, right? Then we don't need to use MCMC to simulate the posterior? $\endgroup$ – Flying pig Jul 19 '12 at 19:47
  • $\begingroup$ Yes but the prior is restrictive and you may not want to do this in practice. $\endgroup$ – Michael R. Chernick Jul 19 '12 at 20:02
  • $\begingroup$ Good answer. It might be good to point out that the likelihood has to be gamma-normal if the estimated distribution is normal. $\endgroup$ – Neil G Jul 20 '12 at 8:00

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Not the answer you're looking for? Browse other questions tagged or ask your own question.