I have a Bayesian hierarchical model with datapoints $y_{ij}$ which are samples from distributions with parameters $\theta_j$. For each distribution parameter $\theta_j$, there are $n_j$ datapoints generated. The distribution parameters themselves are independent samples from another distribution with parameter $\psi$. Both $\theta=\{\theta_1,...,\theta_j\}$ and $\psi$ are unobservable. I am trying to get the posterior distribution of $\psi$ given the datapoints $y = \{y_{ij}\}\forall i,j$.

I started with the basic Bayesian hierarchical model formula: $$ P(\psi|y)\propto P(y|\psi)P(\psi)=P(\psi|y)\propto \sum_{\theta}P(y|\theta)P(\theta|\psi)P(\psi) $$ I understand that both $y$ and $\theta$ are vectors, therefore I have a few questions:

  1. How to compute $P(y|\theta)$ since I only have $P(y_{ij}|\theta_j)$? I can compute $P(y_j|\theta_j)=\prod_i P(y_{ij}|\theta_j)$ but I don't think $P(y|\theta)=\prod_j P(y_j|\theta_j)$, right?
  2. If $\theta$ is a vector of length $n$ (there are $n$ distributions generating datapoints) an state sapce size of each $\theta_j$ is 3, does the total number of available $\theta$ to be marginalized over be $3^n$? I am trying to figure out how many permutations of $\theta$ I need to sum over here, but I realize there is the exchangability condition which I am not sure if it has an effect on this or not.

Any help is deeply appreciated! Thanks!



Your Answer

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

Browse other questions tagged or ask your own question.