6
$\begingroup$

I am doing a homework about infering the N value of a binomial distribution for my Bayesian Statistics Course and I have seen a paper in Biometrika magazine published in 1988 for doing so. The question is how to simulate the N and $\theta$ values based on their respective marginal posteriors. The properties about the model are explained below:

$$X\sim Binom(N,\theta)$$ The priori distribution of N:$$ N\sim Poisson(\mu)$$ $$\lambda = \mu\theta$$(The logic here is that it may be easier to formulate a prior considering the unconditional expectation of the observations, rather than the mean of the unobserved N).
The $\lambda$ and $\theta$ a prior are distributed as:$$\lambda \sim Gamma(\kappa_{1},\kappa_{2})$$ $$\theta\sim Beta(\alpha,\beta)$$

After doing the procedings found on the paper which this link is below, the N marginal posterior is the following one: $p(N|x_{1},x_{2},..,x_{n})\propto(N!)^{-1}\Gamma(N+\kappa_1)\{\prod_{i=1}^5\dbinom{N}{x_i}\} \int_0^1 \! \frac{\Gamma(\alpha+\beta)}{\Gamma(\alpha)\Gamma(\beta)}\theta^{-N+S+\alpha-1}(1-\theta)^{nN-S+\beta-1}(\theta^{-1}+\kappa_{2})^{(N+\kappa_{1})} \, \mathrm{d}\theta $.

How can create a random variable generator from that marginal probability function?

Then, I suppose that this integral:$ \int_0^1 \! \frac{\Gamma(\alpha+\beta)}{\Gamma(\alpha)\Gamma(\beta)}\theta^{-N+S+\alpha-1}(1-\theta)^{nN-S+\beta-1}(\theta^{-1}+\kappa_{2})^{(N+\kappa_{1})} \, \mathrm{d}\theta $ represents part of the marginal posterior of $\theta$ but I don´t know if this is true, My second question is that: Which could be the marginal distribution function for $\theta$? and How can I create a random variable function for getting $\theta$ values?.

I attached the paper I 've working with, the only thing I added is a Beta prior distributión for $\theta$.

http://ftp.stat.washington.edu/raftery/Research/PDF/bka1988.pdf

$\endgroup$
5
  • 2
    $\begingroup$ Your post doesn't contain a clearly defined question. $\endgroup$
    – Glen_b
    Nov 25, 2013 at 4:36
  • $\begingroup$ I agree with Glen. Please, give a better description of the problem. Like $X\mid N=n,\Theta=\theta\sim\mathrm{Bin}(n,\theta)$. Which priors have you used for $N$ and $\Theta$? $\endgroup$
    – Zen
    Nov 25, 2013 at 4:42
  • $\begingroup$ Thank you for taking care of my issue. $N \sim Poisson(\mu)$ where $\mu$ is unknown, so $\lambda = \mu\theta$ (The logic here is that it may be easier to formulate a prior considering the unconditional expectation of the observations, rather than the mean of the unobserved N). $\theta ~ Beta\sim(\alpha,\beta)$. The model is hierarchical because $\lambda ~Gamma(\kappa_{1},\kappa_{2})$. The thing is that I don´t know how to obtaina random variable generator function using the inverse of this marginal posterior distribution. I don´t know either if the integral contained there would be $\endgroup$
    – Georgy boy
    Nov 25, 2013 at 4:58
  • $\begingroup$ a marginal posterior for $\theta$ $\endgroup$
    – Georgy boy
    Nov 25, 2013 at 5:02
  • $\begingroup$ Please update the text of your question to contain the query that you want answered. It's confusing to read through a comment thread to understand how to answer the question. $\endgroup$
    – Sycorax
    Nov 25, 2013 at 15:09

1 Answer 1

5
$\begingroup$

The hierarchical model seems to be $$ \begin{eqnarray} X_i\mid N=n,\Theta=\theta,\Lambda=\lambda &\sim& \mathrm{Bin}(n,\theta) \qquad\qquad\qquad i=1,\dots,m \\ N\mid\Theta=\theta,\Lambda=\lambda &\sim& \mathrm{Poisson}(\lambda/\theta) \\ \Theta &\sim& \mathrm{Beta}(\alpha,\beta) \\ \Lambda &\sim& \mathrm{Gamma}(\kappa_1,\kappa_2) \, , \\ \end{eqnarray} $$ in which $\alpha,\beta,\kappa_1$, and $\kappa_2$ are known. Defining $X=(X_1,\dots,X_m)$ and $x=(x_1,\dots,x_m)$, and using Bayes' Theorem, we have $$ \begin{eqnarray} f_{N,\Theta,\Lambda\mid X}(n,\theta,\lambda\mid x) &\propto& f_{X\mid N,\Theta,\Lambda}(x\mid n,\theta,\lambda)\; f_{N,\Theta,\Lambda}(n,\theta,\lambda) \\ &=& \left(\prod_{i=1}^m f_{X_i\mid N,\Theta,\Lambda}(x_i\mid n,\theta,\lambda)\right) \; f_{N\mid\Theta,\Lambda}(n\mid\theta,\lambda) \; f_\Theta(\theta) \; f_{\Lambda}(\lambda) \, .\\ \end{eqnarray} $$ Therefore, $$ \begin{eqnarray} f_{N\mid X}(n\mid x) &=& \int_0^1 \int_0^\infty f_{N,\Theta,\Lambda\mid X}(n,\theta,\lambda\mid x)\,d\lambda\,d\theta \\ &\propto& \frac{1}{n!} \left( \prod_{i=1}^m {n\choose x_i}\right) \int_0^1 \int_0^\infty \theta^{t-n+\alpha -1} (1-\theta)^{mn-t+\beta-1} \\ && \qquad\qquad\qquad\qquad\qquad\times\lambda^{n+\kappa_1-1} \exp\left(-\left(\frac{1+\kappa_2\theta}{\theta}\right)\lambda\right)\,d\lambda\,d\theta \, , \end{eqnarray} $$ in which $t=\sum_{i=1}^m x_i$. The integration in $\lambda$ is easy. Just identify the kernel of a $$ \mathrm{Gamma}\left( n+\kappa_1, \frac{1+\kappa_2\theta}{\theta} \right) $$ distribution, and adjust the normalization constant. Doing this, we have $$ f_{N\mid X}(n\mid x) \propto \frac{\Gamma(n+\kappa_1)}{n!} \left( \prod_{i=1}^m {n\choose x_i}\right) \int_0^1 \frac{\theta^{t+\alpha-mn-1}(1-\theta)^{mn-t+\beta-1}}{(1+\kappa_2\theta)^{n+\kappa_1}} \,d\theta \, . $$ Of course, the last expression holds for $n\geq \max\{x_1,\dots,x_m\}$. If this is not clear, write the indicators $I_{\{0,\dots,n\}}(x_i)$ in the binomials explicitly.

If we use an improper prior $f_\Lambda(\lambda)\propto 1/\lambda$, the integrations are easy. Using an uniform prior $(\alpha=\beta=1)$ for $\Theta$, we find the closed form $$ f_{N\mid X}(n\mid x) \propto \frac{n\,\Gamma(t+1)\Gamma(mn-t+1)}{\Gamma(mn+2)} . $$

How do we normalize this? What is the Bayes estimate with quadratic loss $\mathrm{E}[N\mid X=x]$? Can we find analytically at least the posterior mode? Can we sample from this distribution to find an estimate and a credible set?

I didn't check the algebra. Please, do it.

$\endgroup$
7
  • $\begingroup$ THANK YOU FOR THAT Zen, It can can not be described better. The thing is how can create values of N based on the marginal posterior distribution stated above? and how can I create values of $\theta$ based on its marginal posterior distribution? $\endgroup$
    – Georgy boy
    Nov 25, 2013 at 16:23
  • $\begingroup$ One approach, which eschews integration entirely, would just compute the bivariate posterior density across a grid of points $(\theta, N)$ and then numerically sum over $N$ or $\theta$ to get the desired marginal distribution. $\endgroup$
    – Sycorax
    Nov 25, 2013 at 16:29
  • $\begingroup$ Use the results in the answer to compute the full conditionals $N\mid\Theta,\Lambda,X$, and $\Theta\mid N,\Lambda,X$, and $\Lambda\mid N,\Theta,X$. If you can sample from these full conditionals, code a Gibbs sampler. Discarding the simulated $\lambda$'s you will have a sample from $N,\Theta\mid X$, and so on. $\endgroup$
    – Zen
    Nov 25, 2013 at 16:55
  • $\begingroup$ I think the original right question would have been, how can I obtain full conditionals $N\mid\Theta,\Lambda$ , $\Theta\mid N,\Lambda,X$, $\Lambda\mid N,\Theta,X$? , I feel that the integral $ \int_0^1 \frac{\theta^{t+\alpha+\kappa_1-1}(1-\theta)^{mn-t+\beta-1}}{(1+\kappa_2\theta)^{n+\kappa_1}} \,d\theta \,$ is very difficult to compute. Moreover, Is this integral useful to OBTAIN the full conditional $\Theta\mid N,\Lambda,X$? $\endgroup$
    – Georgy boy
    Nov 25, 2013 at 18:09
  • $\begingroup$ Check with your teacher if you can use an improper prior for $\Lambda$. I've just added that to the answer. $\endgroup$
    – Zen
    Nov 25, 2013 at 18:15

Your Answer

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

Not the answer you're looking for? Browse other questions tagged or ask your own question.