1
$\begingroup$

I am trying to understand the following:

Suppose that the number of people who visit the grocery store on any given day is Poisson($\lambda$) and the parameter of the Poisson distributed has a $\Gamma(3,15)$ distribution. On Monday 62 people visited the grocery store. Derive the posterior distribution $P(\Lambda\mid\vec{X}=\vec{x})$, where $\vec{X}=(X_1,X_2,\ldots,X_n)$ and $\vec{x}$ is an observed sample from the distribution of $X$. Then compute the posterior mean using the information provided. Note the sum of iid Poisson r.v. is also Poisson.

I computed the posterior distribution for a general poisson parameter with $\Gamma(a,b)$. I found this to be $\Gamma\left(a+n,1+\frac{1}{b}\right).$ Note, the shape and scale parameterization of $\Gamma$ was used. I don't understand Why I need the fact that the sum of iid poisson r.v.'s is also poisson. Can't I just substitute $a=3$, $b=15$, $n=62$ into my posterior mean formula?

$\endgroup$
1
$\begingroup$

For a Poisson likelihood, the gamma distribution is a conjugate prior, so the posterior is also gamma, as you have found. This lets you calculate the posterior in two different ways: one is to compute $f_{\Lambda}(\lambda \mid x)$ for a single observation, recognize this as gamma with updated hyperparameters, and then use an inductive argument to get $f_{\Lambda}(\lambda \mid \boldsymbol x)$ for $\boldsymbol x = (x_1, x_2, \ldots, x_n)$ for any number of observations. Alternatively, you could note that $n \bar x = \sum x = x_1 + x_2 + \cdots + x_n$ is itself Poisson with parameter $n \lambda$, hence the posterior for $\lambda$ satisfies $$f_{\Lambda}(\lambda \mid n\bar x) = f_{\Lambda}(\lambda \mid \boldsymbol x).$$ Explicitly, $$f_{\Lambda}(\lambda \mid x) \propto \Pr[X = x \mid \lambda] f_{\Lambda}(\lambda) = e^{-\lambda} \frac{\lambda^x}{x!} \frac{b^a \lambda^{a-1} e^{-b \lambda}}{\Gamma(a)} \propto \lambda^{a+x-1} e^{-(b+1)\lambda},$$ which is clearly proportional to a gamma density with shape $a^* = a+x$ and rate $b^* = b+1$; thus for $n$ iid observations, we can extend this to get $$f_{\Lambda}(\lambda \mid \boldsymbol x) \propto \lambda^{a + \sum x - 1} e^{-(b+n) \lambda},$$ or $$\Lambda \mid \boldsymbol x \sim \operatorname{Gamma}(a + n\bar x, b + n).$$ But we could have also gotten this by writing $$f_{\Lambda}(\lambda \mid n\bar x) \propto \Pr[{\textstyle\sum} X = n\bar x \mid \lambda]f_\Lambda(\lambda) = e^{-n \lambda} \frac{(n \lambda)^{n \bar x}}{(n \bar x)!} \frac{b^a \lambda^{a-1} e^{-b \lambda}}{\Gamma(a)} \propto \lambda^{a + n\bar x - 1} e^{-(b+n)\lambda}.$$ The result is the same.

$\endgroup$
3
  • $\begingroup$ So $n$ in my case is the number of observations, which is 62. Just seems kind of weird that the mean of the posterior is then $(a+n\vec{x})*(b+n)$, because that number is a lot larger than the observed sample... $\endgroup$
    – user75514
    Jun 15 '15 at 12:51
  • $\begingroup$ No. $n$ is the number of days you recorded; because the Poisson rate parameter represents the mean number of customers arriving per day. So if you observe $x = 62$ customers arriving in $n = 1$ day, that is a single observation with $\bar x = 62$ (the sample mean of one observation is simply that observation), and $n = 1$. $\endgroup$
    – heropup
    Jun 15 '15 at 13:52
  • $\begingroup$ Be careful to distinguish the sample mean $\bar x$ and the vector of observations $\boldsymbol x = (x_1, \ldots, x_n)$. Where I use a bar atop $x$, I refer to the sample mean. Where I use boldface, I refer to the vector of observations. Also, you are asked to compute the general case of the posterior distribution if you were to observe the arrival of $x_1$ customers on day 1, $x_2$ on day 2, etc., until $x_n$ customers on day $n$, but you were only asked to use this formula for $n = 1$. $\endgroup$
    – heropup
    Jun 15 '15 at 13:56

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.