# Why do we use separate priors or joint priors?

I'm studying prior choice, and as far as I understand, when more than one parameter in a distribution is unknown, it is both possible to place one prior for each of the parameters in the likelihood, as well as to place a joint prior over all of said parameters.

For example, for a normal distribution $N(\mu,\sigma^2)$, it is possible to place, say, a normal prior on $\mu$ and another normal prior on $\sigma^2$, but it is also possible to place a bivariate normal distribution as a prior on both $\mu$ and $\sigma^2$.

One of the reasons for using joint priors, I'm guessing, is that we can construct conjugate joint priors for some distributions. However, when not using a conjugate prior, like above, what is the difference between using a joint prior vs separate priors? Is it to encode our belief in the correlation between the parameters? Beyond that, is there any reason to prefer joint priors over separate priors?

• Quick note - A normal distribution on $\sigma^2$ isn't conjugate and would be an odd choice. – Dason Jul 10 '17 at 13:12
• Thanks for the note! I deliberately avoided inverse-gamma (or normal-inverse-gamma for the joint version) to bring up a situation where conjugate priors aren't used. Of course, we can use conjugate priors for the normal distribution, but I'm also very interested in the case for a distribution where the conjugate prior doesn't exist or is computationally intractable. – peco Jul 10 '17 at 18:30
• I believe that copulas can be hard to specify reasonably – user795305 Jul 12 '17 at 18:11

I think the correct way to phrase this is whether the priors are independent or not. The priors can always be described as (for example in your Normal example) $p(\mu, \sigma^2)$, but the question is does that joint prior factorize as $p(\mu, \sigma^2) = p(\mu)p(\sigma^2)$ or not.
All the priors you mention are "joint" priors in that they define a joint distribution on the parameter vector $\mathbf{\theta}=(\theta_1,\ldots,\theta_p)$. When the prior writes down as $$\prod_{i=1}^p \pi_i(\theta_i)$$ each component $\pi_i(\theta_i)$ can also be interpreted as a (marginal) prior on the component $\theta_i$ [provided all components are proper] and the components are independent a priori. Since all priors are acceptable within the Bayesian paradigm, there is no foundational reason to favour independent priors over dependent priors.