2
$\begingroup$

Consider two independent discrete random variables $y_1$ and $y_2$, both distributed with a Beta-Binomial distribution, with different number of successes $n_1$ and $n_2$ but the same parameters $a$ and $b$

$ p(y_1|n_1,a,b) = {n_1 \choose y_1} \dfrac{B(y_1 + a,n_1 -y_1 +b)}{B(a,b)} $

$ p(y_2|n_2,a,b) = {n_2 \choose y_2} \dfrac{B(y_2 + a,n_2 -y_2 +b)}{B(a,b)} $

Consider a discrete variable $Z = y_1 + y_2$. Is $Z$ distributed as well as a Beta-Binomial (with parameters $n_1+n_2$, $a'$ and $b'$?

I could not prove it in an analytic form so far, but I have been trying it out with some simulations, at least to check whether the assumption is wrong in some cases. Reparametrising $a$ and $b$ as $\mu = \dfrac{a}{a+b}$ and $\rho = \dfrac{1}{a+b+1}$, $y_1$ and $y_2$ have mean $\mu n_1$ and $\mu n_2$ respectively and variance $\mu(1-\mu)n_1(1 + (n_1-1)\rho)$ and $\mu(1-\mu)n_2(1 + (n_2-1)\rho)$ respectively.

Based on independence, the mean of $Z$ is $\mu(n_1+n_2)$ and the variance of $Z$ is $\mu(1-\mu)(n_1(1 + (n_1-1)\rho) + n_2(1 + (n_2-1)\rho))$. If $Z$ was distributed according to a beta-binomial distribution, then it would have paramters $\mu'$ and $\rho'$, with $\mu'= \mu$ and

$\rho' = \dfrac{\dfrac{n_1(1 + (n_1-1)\rho) + n_2(1 + (n_2-1)\rho)}{n_1+n_2} - 1}{n_1+n_2-1} = \rho \dfrac{n_1(n_1-1)+n_2(n_2-1)}{(n_1+n_2)(n_1+n_2 -1)} $

Here is some code to generate $Z$ as a sum of two independent Beta-Binomials (sorry about the code, R is not my main language)

n1 = 20
n2 = 50
mu = .6
k  = 20

p1  = rbeta(1e6,mu*k,(1-mu)*k)
y1  = rbinom(1e6,n1,p1)

p2  = rbeta(1e6,mu*k,(1-mu)*k)
y2  = rbinom(1e6,n2,p2)

z   = y1+y2

rho  = 1/(k+1)
rho1 = rho*(n1*(n1-1)+n2*(n2-1))/((n1+n2)*(n1+n2-1))
k1   = 1/ rho1 - 1
p3  = rbeta(1e6,mu*k1,(1-mu)*k1)
z1  = rbinom(1e6,n1+n2,p3)

print(c(var(z),var(z1)))
plot(density(z,width= 3))
lines(density(z1,width = 3))

I have been trying this code for different values of $n_1$, $n_2$, $\mu$ and $k$, but in all the cases the variances using the sum of two beta-binomials or an appropriately tuned beta-binomial are very similar (the densities look indistinguishable)

$\endgroup$
  • 1
    $\begingroup$ So what conclusion did you reach based on your simulations? Why not show a diagram of the resulting plot rather than this plethora of code? In fact, I would suggest that you delete everything in your question after the line: Is Z distributed as well as a Beta-Binomial (with parameters n1+n2, a and b? $\endgroup$ – wolfies Oct 18 '13 at 14:34
  • $\begingroup$ I edited your question, please check if it is still correct. If not, you can rollback. $\endgroup$ – Momo Oct 18 '13 at 15:03
  • $\begingroup$ Sorry for not describing the results of the simulations. I could not find a combination of the free parameters for which the two variables $z$ (sum of beta-binomials) and $z1$ (beta-binomial with adjusted $a$ and $b$) have different densities (given the same mean and variance). I edited the text to include this. @wolfies: you are indeed right, the question could stop after the line you mention, I just wanted to point out that I have strong suspicions that the sum is distributed as a Beta-Binomial, but I am looking for a proof. $\endgroup$ – Giancarlo Oct 18 '13 at 16:22
1
$\begingroup$

I know this question was posted a while ago, but I was looking for an answer to that question myself and stumbled upon this post. So here are my thoughts on it: in general, the answer is “no”. Consider an example of two iid random variable with beta-binomial distributions BB(1,1,n). BB(1,1,n) is the same as U(0,n) – the discrete uniform distribution on the interval [0,n]. The sum of two discrete uniforms is a triangular distribution (of discrete variety), which is not the same as the BB with parameters that you suggest.

$\endgroup$

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.