I have been studying a Bayesian hierarchical model. In that model all I am dealing is with the estimation of parameters. In Bayesian analysis, loosely speaking, we update our prior knowledge (in light of new evidences/data) to posterior knowledge. But in hierarchical model I don't see any prior knowledge or any prior distribution.

My question is what is the relation between Bayesian analysis and Bayesian hierarchical analysis?

I suppose the latter is a subset of former but I am still confused how are these two related? Is it enough for any statistical model which uses Bayes theorem to be categorized under Bayesian analysis/statistics?

http://www.stat.missouri.edu/~wikle/WikleBerlinerCressie1998.pdf - I have been studying a hierarchical model from this paper.


1 Answer 1


In my view, hierarchical modeling in a Bayesian setting mainly refers to the building of a complex prior structure. Consider a parameter of interest $\theta_{0}$ and your observation $(x_i)$.

Now, consider for example that you are adding a supplemental layer to your model $p(\theta_0|\theta_1)$ through hyperprior $p(\theta_1)$ on $\theta_1$, then $p(\theta_0)$ writes: $$ p(\theta_0)=\int_R p(\theta_0|\theta_1)p(\theta_1)d\theta_1, $$ and so on for $\theta_2, \ldots$.

The same for the observation model : consider that your parameter of interest $\theta_0$ is not directly related to the observation but to an other parameter $\theta_1$ that is itself related to the observations: $$ p((x_i)|\theta_0)=\int_R p((x_i)|\theta_1)p(\theta_1|\theta_0) d\theta_1. $$

To sum up, in principle, you can always (to the best of my knowledge) marginalize the hierarchical structures to get something as $p(\theta_0|x) \propto p(x|\theta_0) \cdot p(\theta_0)$ so to the simplest Bayes formulation. However, most of the time, integrals are intractable and we need to work with all the latent variables of the prior structure. So, IMHO, a hierarchical Bayesian model is only a decomposed Bayesian model (assuming that we call a Bayesian model something of the simplest form $p(\theta_0|x) \propto p(x|\theta_0) \cdot p(\theta_0)$).

Finally to answer your last question: "Is it enough for any statistical model which uses Bayes theorem to be categorized under Bayesian analysis/statistics" I would say no. A model can be qualified as a Bayesian model if it relies on Bayesian interpreation of probability https://en.wikipedia.org/wiki/Bayesian_probability and in particular if the posterior $p(\theta|x_i)$ makes any sense. Bayes theorem can be used in other contexts. See the related answer to the question Can frequentists use Bayes theorem?.


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.