I'm reading Casella-Berger's Statistical Inference and trying to follow along in example 9.2.15, which constructs an exact confidence interval for a Poisson rate. In this example, the authors solve the sum

$$ \frac{\alpha}{2}=\sum_{k=0}^{y_0} \frac{e^{-n\lambda}(n\lambda)^k}{k!} $$ by using a link between the gamma and Poisson distributions. Namely, they use the fact $$ P(Y\geq \alpha)=P(X\leq x) $$ for $Y\sim Poisson(\frac{x}{\beta})$ and $X\sim Gamma(\alpha, \beta)$. Additionally, it helps to note $Gamma(p/2,\beta=2)$ is a $\chi^2$ random variable with $p$ degrees of freedom.

They then conclude $$ \sum_{k=0}^{y_0} \frac{e^{-n\lambda}(n\lambda)^k}{k!}=P(Y\leq y_0)=P(\chi^2_{2(y_0+1)}>2n\lambda). $$ My question is why this is true. Using the link between the Poisson and Gamma mentioned above, I keep finding the degrees of freedom of the $\chi^2$ to be $2y_0$ and not $2(y_0+1)$. I feel like this might just be a simple calculation mistake, but I'm not seeing it.


Given \begin{equation} X \sim {{\chi}_{2n}}^2 \end{equation} and \begin{equation} Y \sim P(\lambda) \end{equation} Let's prove (the complement of your expression) \begin{equation} P(X < 2\lambda) = P(Y \ge n) \end{equation} Notice that $X \sim \chi_{2n}^{2}=\Gamma(n,2)$. In addition, \begin{equation} P(X<2\lambda)=\frac{1}{\Gamma(n)2^n} \displaystyle\int_{0}^{2\lambda}x^{n-1}e^{-\frac{x}{2}}dx \end{equation} and \begin{equation} P(Y\geq n)=1-P(Y \leq n-1)=1-\displaystyle\sum_{i=0}^{n-1}\frac{e^{-\lambda}\lambda^i}{i!} \end{equation} Do the derivative on both sides with respect to $\lambda$ \begin{equation} \frac{\partial}{\partial \lambda}P(X<2\lambda) = \displaystyle\frac{2}{\Gamma(n)2^n}(2\lambda)^{n-1}e^{-\lambda}=\frac{\lambda^{n-1}e^{-\lambda}}{(n-1)!} \end{equation} and \begin{equation} \frac{\partial}{\partial \lambda} P(Y \ge n) = \displaystyle\sum_{i=0}^{n-1}\frac{e^{-\lambda}\lambda^i}{i!}-\displaystyle\sum_{i=1}^{n-1}\frac{e^{-\lambda}\lambda^{i-1}}{(i-1)!}=\frac{\lambda^{n-1}e^{-\lambda}}{(n-1)!} \end{equation} Notice that the derivatives are the same, hence \begin{equation} P(X<2\lambda) = P(Y \ge n) + C \end{equation} where $C$ is a constant. Plug $\lambda = 0$, you'll get $C =0$. This means that ( by inverting both inequalities) \begin{equation} P(X>2\lambda) = P(Y < n) \end{equation} or \begin{equation} P(X>2\lambda) = P(Y \leq n-1) \end{equation} You've got an $n \lambda$ instead of my $\lambda$ and a $y_0+1$ instead of my $n$

| cite | improve this answer | |
  • $\begingroup$ Perhaps this should be edited to have $G(n,2)$ instead of $\Gamma(n,2)$ that way the usual notation for the function $\Gamma(\cdot)$ isn't confused with the Gamma distribution. Just a suggestion. $\endgroup$ – SecretAgentMan Sep 13 '18 at 19:39

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.