Take the 2-minute tour ×
Cross Validated is a question and answer site for people interested in statistics, machine learning, data analysis, data mining, and data visualization. It's 100% free, no registration required.

On the Wikipedia page for the Binomial distribution, the following property is mentioned (under the related distribution section): (paraphrased)

If $X\sim \text{Bin}(n,p)$ and $Y|X \sim \text{Bin}(X,q)$, then $Y \sim \text{Bin}(n,pq)$

I interpret this the following way. The probability mass function for $X$ is: $$P(X=x) = \binom{n}{x}p^x(1-p)^{n-x}$$ The conditional mass function for $Y$ given $X=x$ is: $$P(Y=y|X=x) = \binom{x}{y}q^{y} (1-q)^{x-y}$$ The mass function of $Y$ is: $$P(Y=y) = \binom{n}{y} (pq)^y (1-pq)^{n-y}$$

There is no citation for this particular property. I have tried to prove it, but to no avail. I wrote the following R code to get a sense of the veracity of the claim.

# Observations of X & Y to be generated
obs <- 10000

n <- 10
p <- 0.6
q <- 0.4

X <- rbinom(obs, n, p)
Y <- X

for( i in 1:obs)
{
  Y[i] <- rbinom(1, X[i], q)
}

# Simulated pmf of Y 
hist(Y, breaks=obs)

# Theoretical/claimed pmf
Y_theoretical <- rbinom(obs, n, p*q)
hist(Y_theoretical, breaks=obs)

The two histograms generated are shown below: (The simulated pmf) Simulated distribution (Claimed pmf) Claimed distribution

Both seem identical for the choice of $p$ and $q$.

Can a proof of this claim be provided?

share|improve this question

3 Answers 3

up vote 4 down vote accepted

Let $X = \sum_{i=1}^{n} X_i$, with $X_i \overset{iid}{\sim} Bin(1, p)$, and $Z = \sum_{i=1}^{n} Z_i$, with $Z_i \overset{iid}{\sim} Bin(1, q)$. If all the $X_i$ and $Z_i$ are mutually independent, then $Z_i | X_i \overset{iid}{\sim} Bin(1, q)$.

Now to construct $Y$ we want to throw out all the $(X_i, Z_i)$ pairs where $X_i=0$ and then count the number of times $Z_i=1$ in the remaining pairs. That makes $Y | X \sim Bin(x, q)$. We can also write $Y = \sum_{i=1}^{n} Y_i$ with $Y_i = X_i Z_i$. We know $X_i Z_i=1$ if $X_i=1$ and $Z_i=1$, otherwise it is 0. Thus $Y_i \overset{iid}{\sim} Bin(1, pq)$, and $Y \sim Bin(n, pq)$.

share|improve this answer
    
Why is $Y_i=X_i Z_i$? –  Comp_Warrior Aug 19 at 20:45
    
Does the edit make it clearer? To get $Y$ we want to take all the times $X_i=1$ and then among those count how many times $Z_i=1$ as well. –  P Schnell Aug 19 at 20:47
    
Yes, it does now. A very elegant solution! –  Comp_Warrior Aug 19 at 20:49

Looking at the same thing in two different but equivalent ways offers insight.

A Binomial$(n,p)$ variable is the sum of $n$ independent Bernoulli$(p)$ variables. A Bernoulli variable works exactly like drawing one ticket from a box in which all tickets have either a $0$ or $1$ written on them; the proportion of the latter is $p$.

To say that $X=x$ means that $n$ such tickets were drawn from such an "$X$ box" (with replacement each time) and $x$ of them had a $1$ on it. To say that $Y$ has a Binomial$(X,q)$ distribution amounts to performing a second follow-on experiment in which $x$ draws (with replacement) are made from a separate box, the "$Y$ box," in which the proportion of tickets with $1$s is $q$. The value of $Y$ is the count of the $1$s that are drawn.

An alternative way to carry out the same procedure is not to wait until all $n$ tickets are drawn from the $X$ box. Instead, after drawing each ticket, immediately read its value. If it says $X=0$, do nothing more. If it says $X=1$, though, immediately draw a ticket from the $Y$ box and read its value.

This alternative procedure can be described by drawing a single ticket from a new box. Up to two numbers are written on each ticket, called "$X$" and "$Y$", to record a single sequence of up to two draws. According to the foregoing description, which has three outcomes, there must be three kinds of corresponding tickets:

  1. $X=0$. These tickets model drawing a value of $0$ from the $X$ box. Their proportion within the new box, in order to emulate the properties of the first step, must equal $1-p$. Don't bother to write any value for $Y$, because $Y$ will not be observed when such a ticket is drawn.

  2. $X=1, Y=0$. These tickets model drawing a $1$ from the $X$ box and then a $0$ from the $Y$ box.

  3. $X=1, Y=1$. These tickets model drawing a $1$ from the $X$ box and then a $1$ from the $Y$ box.

The total proportion of tickets of types (2) and (3) must equal the proportion of $1$s in an $X$ box, namely $p$. Since $Y$ is drawn independently of $X$, the fraction of the tickets with $X=1$ for which $Y=1$ must be $q$. The fraction of the tickets with $X=1$ for which $Y=0$ similarly must be $1-q$.

To summarize, the three tickets and their proportions in the new box must be

  • $X=0$, proportion $p$.

  • $X=1, Y=0$, proportion $p(1-q)$.

  • $X=1, Y=1$, proportion $pq$.

What kind of variable is $Y$? According to our new (but equivalent) description, it is obtained by drawing $n$ tickets from the new box (with replacement) and counting the number of times a value of $1$ for $Y$ is observed. The only way this can happen is when the third type of ticket is drawn. These occupy a fraction $pq$ of all the tickets. This exhibits $Y$ as the sum of $n$ independent Bernoulli$(pq)$ variables, whence $Y$ has a Binomial$(n, pq)$ distribution.

share|improve this answer

Have you tried calculating the marginal distribution? In general, for a discrete random variable the following is true:

\begin{align*} p(y) &= \sum_x p(y,x) dx\\ &=\sum_x p(y|x)p(x)dx \end{align*}

so all you need to do is show the following:

\begin{align*} p(y) &= \sum_{x=0}^{\infty} \binom{x}{y}q^y(1-q)^{x-y}\binom{n}{x}p^x(1-p)^{n-x}\\ &=\,\,\,\vdots\\ &=\binom{n}{y} (pq)^y (1-pq)^{n-y} \end{align*}

Is this a trivial problem, I don't think so.

share|improve this answer
    
I am not sure, but the sum should be $\sum_{x=y}^n$ –  niandra82 Aug 19 at 20:40
    
Shouldn't we be summing over the plausible values of x? Think of the continuous case when you want to integrate out x from f(x,y). We integrate with respect to the support of x. And in this case the support of x does not start at y (although I suppose it should be bounded at n). –  Dan Aug 19 at 20:43
    
if $x<y$ then what does mean $P(Y|X)$? maybe you have to assume that it is zero –  niandra82 Aug 19 at 20:45
    
$x<y$ can't happen since $y$ is dependent upon $x$. Right? We are choosing $y$ "successes" out of $x$ trials. So the relationship should really be $y\leq x$ –  Dan Aug 19 at 20:46
    
I was taking essentially the same approach but the summation seems hard to get through. –  Comp_Warrior Aug 19 at 20:50

Your Answer

 
discard

By posting your answer, you agree to the privacy policy and terms of service.

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