I am reading the book Bayesian data analysis written by Gelman. And I do not confirm the answer of this question:
suppose it is known a priori that the $2J$ parameters $\theta_1,\ldots,\theta_{2J}$ are clustered into two groups, with exactly half being drawn from a $N(1,1)$ distribution, and the other half being drawn from a $N(-1,1)$ distribution, but we have not observed which parameters come from which distribution.
$(a)$ Are $\theta_1,\ldots,\theta_{2J}$ exchangeable under this prior distribution?
$(b)$ Show that this distribution cannot be written as a mixture of $iid$ components.
$(c)$ Why can we not simply take the limit as $J\rightarrow \infty$ and get a couterexample to $de \space Finetti's$ theorem?

IMO, we have no any information about ordering or grouping of the parameters can be made, so the parameters $\theta_1,\ldots,\theta_{2J}$ are exchangeable in this prior distribution?
Besides, not all parameters are identity distributed, can we simply say this distribution cannot be written as a mixture of $iid$ components?
We cannot take the limit as $J\rightarrow \infty$ and get a couterexample to $de \space Finetti's$ theorem since this distribution is not exchangeable. Right?

If you have any idea please write your comment. Let me know I do a right answer or not. Thanks :)


(a) The question doesn't seem to provide you with the distribution of $\theta_1, \ldots, \theta_{2J}$, so you can't say for sure whether or not the distribution is exchangeable. E.g. the possibility that $E(\theta_1) = \ldots = E(\theta_{J}) = -1$ seems to be consistent with the wording of the question, but if this possibility obtains then $\theta_j$'s are not exchangeable. However, it seems likely that Gelman intends the following: 1) let $(R_1, \ldots, R_{2J})$ have uniform distribution on the set of $\{1, -1\}$-valued sequences having sum 0. 2) let $(W_1, \ldots, W_{2J})$ be IID N(0,1) random variables independent of $(R_1, \ldots, R_{2J})$. 3) let $X_j := W_j + R_j$ for $1 \leq j \leq 2J$.

In this setup, $(X_1, \ldots, X_{2J})$ is an exchangeable sequence because the pairs $(R_1, W_1), \ldots, (R_{2J}, W_{2J}) $ are exchangeable (because .... [you fill in the blanks] ).

(b) Try computing $E(X_1X_2)$ for $(X_j)$ defined as above by conditioning on $(R_1, R_2)$. You'll find that the expectation is negative. Now assume that these $X_j$'s are derived by taking IID draws from a random distribution $\mu$ as in the classical de Finetti theorem, and try to say something about the sign of $E[X_1X_2]$ by conditioning on $\mu$. You should get an easy contradiction showing that $(X_1, \ldots, X_{2J})$ are not derived by conditional IID sampling. (Now how do you argue that \emph{Gelman's} $X_j's$ are not derived by conditional IID sampling?)

(c) Drawing n things u.a.r. without replacement from N things is essentially the same as drawing n things u.a.r. with replacement from N things provided n is small compared to N, and this becomes more and more true if n is fixed and N gets large. If you properly formalize this you may be able to show that when you take an appropriate limit as $J \to \infty$ in the setup above, the limit is exchangeable and satisfies de Finetti's theorem.


I don't know my explanation is correct or not but I try to explain it in detail. Here is the story, there is $\frac{1}{\binom{2J}{J}}$ probability to find the $J$ parameters that follow the distribution $N(1,1)$. Hence the remaining $J$ parameters must be sampled from $N(-1,1)$. Use Bayes theorem, $p(\theta_1,\ldots,\theta_{2J})=p(\theta_{p(1)},\ldots,\theta_{p(J}))p(\theta_{p(J+1)},\ldots,\theta_{p(2J})|\theta_{p(1)},\ldots,\theta_{p(J}))$ $=\sum_{p}\left(\left(\dfrac{1}{\binom{2J}{J}}\prod_{j=1}^{J}N(\theta_{p(j)}|1,1)\right)\prod_{j=J+1}^{2J}N(\theta_{p(j)}|-1,1)\right)$

feel free to write your comment if you have any idea. :D


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.