I'm working on a question with simultaneous learning about an underlying population and individual members of the population. The basic setup is:

Let $N_g$ be the size of a population.

At any point in time $t$, each object $i \in \{1, ... , N_g\}$ has been sampled $\{n_1, ..., n_{N_g}\}$ times.

For each object $i$, we observe $r_i$ successful realizations out of the $n_i$ observations. With this in mind, I'm using the Beta-binomial distribution so that: \begin{align*} r_i &\sim Bin(n_i, \theta_i) \\ \theta_i &\sim Beta(\alpha,\beta) \\ (\alpha,\beta) &\sim p_0(\alpha,\beta) \end{align*}

Given a sequence of observations $\{r_i,n_i\}_{i=1}^{N_g}$, I'm hoping to calculate the marginal posterior expectations: \begin{align*} \mathbb{E}(\theta_i \vert \{r_i,n_i\}_{i=1}^{N_g}) \end{align*}

That is, I'm hoping to calculate the expected probability of success for object $i$ given the entire sample from objects $1$ to $N_g$.

My logic in proceeding has been to calculate: \begin{align*} Pr(\{\theta_{j},r_j,n_j\}_{j^{\prime}}, \alpha,\beta) &= \left[\prod_{j=1}^{N_{g}} {n_{j}\choose r_{j}} \theta_{j}^{r_{j}}(1-\theta_{j})^{n_{j}-r_{j}}\right] \cdot \left[\prod_{j=1}^{N_{g}} \frac{\Gamma(\alpha+\beta)}{\Gamma(\alpha)\Gamma(\beta)}\theta_{j}^{\alpha-1}(1-\theta_{j})^{\beta-1} \right] \cdot p^{g}_{i}(\alpha,\beta) \\ \text{given a uniform prior } p_0(\alpha,\beta): \\ Pr(\alpha,\beta \vert \{r_j,n_j\}_{j^{\prime}}) & \propto \left[\prod_{j=1}^{N_{g}} \frac{\Gamma(\alpha+\beta)}{\Gamma(\alpha)\Gamma(\beta)} \frac{\Gamma(\alpha + r_{j})\Gamma(\beta+n_{j}-r_{j})}{\Gamma(\alpha+\beta + n_{j})} \right] \\ \text{so:} \\ Pr(\theta_{j} \vert \{r_j,n_j\}_{j^{\prime}}) &= \int_{(\alpha,\beta)} Pr(\theta_{j} \vert \alpha,\beta\{r_j,n_j\}_{j^{\prime}}) \cdot Pr(\alpha,\beta \vert \{r_j,n_j\}_{j^{\prime}}) d(\alpha,\beta) \\ \Rightarrow \mathbb{E}(\theta_{j} \vert \{r_j,n_j\}_{j^{\prime}}) &= \int_{\alpha,\beta} \left[\frac{\alpha + r_j}{\alpha+\beta+n_j}\right] \cdot \prod_{j=1}^{N_{g}} \frac{\Gamma(\alpha+\beta)}{\Gamma(\alpha)\Gamma(\beta)} \frac{\Gamma(\alpha + r_{j})\Gamma(\beta+n_{j}-r_{j})}{\Gamma(\alpha+\beta+n_{j})} \cdot Cd(\alpha,\beta) \end{align*}

Unfortunately, I'm getting stuck in this calculation. I've been trying to find literature where someone has already worked this out (I do Economics, and we're often 40+ years behind math, so I'm sure it's been done.), but I'm coming up empty.

I've been trying to read through discussions of the Dirichlet distribution (insofar as the beta-binomial is a subcase of Dirichlet), but I can't seem to be able to find anything about this type of posterior.

Does anyone have any idea / recommendation where I could look?

Thank you!


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.