I have two Bayesian regression models: both use the same data $D$ and all same specification $M$ but the different priors $p_1(\theta)$ and $p_2(\theta)$.

Can I interpret the posteriors from the two models, $p(\theta|D,M,p_1(\theta))$ and $p(\theta|D,M,p_2(\theta))$ independently and compare them? One example of comparing the results is something like "one model results in an expected loss of 1 while the other results in an expected loss of 2, so that given the different priors, the range of the expected loss is between 1 and 2." Or do I actually have to condition each model not only on its own prior but also on the prior of the other model, in order to estimate the single posterior that captures the uncertainty of the prior choice, using a hyperprior on the prior choice, such that the posterior to be estimated is $p(\theta|D,M,p(p_i(\theta)))$? My intuition is that interpreting each posterior independently means I implicitly assume that the hyperprior is a uniform distribution.

I have come across this question, which talks about estimating a posterior using all different datasets together while using the same prior. My setting is opposite, using the same dataset while using different priors.

I would appreciate your suggestions.

  • $\begingroup$ From a purely Bayesian point of view, there cannot be several priors co-existing at the same time! $\endgroup$
    – Xi'an
    Jul 17, 2019 at 17:08
  • $\begingroup$ Xi'an, thank you for your reply. This is an interesting point; would you say so even if we assumed there is a hyperprior on the choice of priors? $\endgroup$
    – Patrick
    Jul 22, 2019 at 8:47


Your Answer

By clicking “Post Your Answer”, you agree to our terms of service and acknowledge you have read our privacy policy.

Browse other questions tagged or ask your own question.