Suppose I have six data points (n,x):







which are realizations of binomial experiments on n trials with x successes respectively..

And I assume I have a prior on the proportion of successes, p.

Such that $ 0 \le p_{1} \lt p_{2} \lt p_{3} \lt p_{4} \lt p_{5} \lt p_{6} \le 1$

For which we have an observed initial value from the above data.

So for each $p_{i}$, we have that it is conditional uniform on the two other close.

For example, $Pr(p_{3}|p_{1},p_{2},p_{4},p_{5},p_{6}) \sim U(p_{2},p_{4})$

I want to gibbs sample, so I need the conditional distributions,

my question is do I need to include the binomial portion in each for example would I need to do

$Pr(p_{3}|p_{1},p_{2},p_{4},p_{5},p_{6},X)$ ,ie with X included. If so, how would I incorporate it , would I only use that binomial corresponding to the third collection of data points? Or would I multiply by binomial for the first , second fourth and fifth collection

Someone suggested I find the joint posterior before full. I do not understand what is meant by that.

For the binomial for example, how do I incorporate it? Like for the first and second

${14 \choose x_{1}}p_{1}^{x_{1}}(1-p_{1})^{14-x_{1}}$

${13 \choose x_{2}}p_{2}^{x_{2}}(1-p_{2})^{13-x_{2}}$

and so forth. Would I just multiply these?

$Pr(p_{1},p_{2},p_{3},p_{4},p_{5},p_{6})= c I(0 \le p_{1} \lt p_{2}...\lt p_{6} \le 1)$ but I don't know how to find c, nor what order to approach the problem

So would the joint posterior then be $p(p_{1:6}|x_{1:6}) $

proportional to

$f(x_{1:6}|p_{1:6})I(0 \le p_{1} \lt p_{2})I(p_{2} \lt p_{3})...I(p_{6} \le 1)$

Hence it seems it would be product of the binomials times the indicator type function.What distribution would this be though? Where do I incorporate the observed data?

It seems like I need to have $Pr(p_{1}|p_{2},p_{3},p_{4},p_{5},p_{6},X)$ , where X is the data, but i have literally no idea at all how to do this. I am completely lost. Would I multiply by all the other binomial probabilities? But which do I plug in? I don't know how I would implement this in R or a program to do gibbs sampling

Anyone can please help at all?

I really appreciate any help at all, I have been trying to get some help here for many days, any hints or ideas that I can understand

  • $\begingroup$ I suggest you start by writing the joint posterior of $p_{1:5}$ before asking for the full conditional. $\endgroup$ – Xi'an Dec 2 '18 at 13:45
  • $\begingroup$ Is what I added what you meant by joint posterior? I am a bit confused what you mean by finding that before the full conditional. What is the difference? $\endgroup$ – Learning Dec 2 '18 at 20:39
  • $\begingroup$ I am really trying to understand, but I just cant seem to. Can you please try to help explain a bit more? I want to work through it but I am stuck. I dont know how to condition. Do we take the values from the data given and condition on that or what? I am totally lost $\endgroup$ – Learning Dec 2 '18 at 20:52
  • $\begingroup$ @Xi'an . I think I see what you meant, I updated again. Hopefully someone can help me to understand what I am missing or need to do. Thank you very much for your time $\endgroup$ – Learning Dec 2 '18 at 21:58
  • $\begingroup$ Been working all day on it and I am simply stuck. I have tried all I know. $\endgroup$ – Learning Dec 3 '18 at 4:02

Likelihood $$\prod_{i=1}^6 p_i^{x_i}(1-p_i)^{n_i-x_i}$$ Prior $$\mathbb{I}_{0 \le p_{1} \lt p_{2}...\lt p_{6} \le 1}$$ Posterior $$\prod_{i=1}^6 p_i^{x_i}(1-p_i)^{n_i-x_i} \times \mathbb{I}_{0 \le p_{1} \lt p_{2}...\lt p_{6} \le 1}$$ Conditional $$p_i|p_{-i},\mathbf{x} \sim p_i^{x_i}(1-p_i)^{n_i-x_i}\times \mathbb{I}_{p_{i-1}\le p_i\le p_{i+1}}$$ Hence this is a truncated Beta distribution that can be easily simulated in R with the inverse cdf method and the qbeta R function.

  • $\begingroup$ Thank you. I think in my post I had tried to get close to this. But my other issue is implementing it to gibbs sample for the parameters p.. Is there a way to combine in R code? Or do I need to do each individually? $\endgroup$ – Learning Dec 3 '18 at 7:43

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.