I'm struggling with question 6 in the Exercises to Chapter 3 (page 80) of Bayesian Data Analysis by Andrew Gelman.


We have data Y modeled as independent binomial data, with both $N$ and $θ$ unknown, as per Raftery's 1988 paper "Inference for the binomial N parameter: A hierarchical Bayes approach".

$Y∼Bin(N,θ)$ and

$N∼Poisson(μ)$, where $λ=μθ$

The (noninformative) prior distribution of $λ,θ$ is $p(λ,θ) \propto λ^{-1}$

The question 6(a) asks you to transform to determine $p(N,θ)$.

It's similar to the following question, but I haven't been able to use that to get to the answer.

Bayesian Aproach: Infering the N and $\theta$ values from a binomial distribution

  • $\begingroup$ I might start by looking for an expression for the distribution of $\mu, \theta$ i.e. the prior $p(\mu, \theta)$ as a change of variable, remembering $\theta \in [0,1]$ $\endgroup$ – Henry Dec 14 '20 at 10:05

Here is what I got (I'm not very sure about it). I think in that exercise, $N$ is supposed to follow a Poisson distribution with random expectation $\mu$. The (improper) joint distribution of $\mu, \theta$ is defined on the transformation $(\lambda = \mu \theta, \theta)$ by $$p(\mu, \lambda) \propto 1/\lambda .$$ In order to get the joint distribution of $(\mu, \theta)$ you would need to use the fact that $$p(\mu, \theta) = p(\lambda, \theta) \mid\det\frac{\partial(\lambda, \theta)}{\partial(\mu, \theta)}\mid$$

Here, $\mid\det\frac{\partial(\lambda, \theta)}{\partial(\mu, \theta)}\mid = \theta$ such that the improper distribution of $(\mu, \theta)$ is $p(\mu, \theta) \propto 1 / \mu$ so the prior is : $$\begin{array}{lcl} p(\mu) &\propto & 1 / \mu\\ N & \sim & \mathcal{P}(\mu) \\ \theta & \sim & \mathcal{U}([0, 1]) \end{array}$$


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.