For samples of $M$ Poisson-distributed datapoints $X_{1, r},...,X_{M,r}$ $\sim$ $Pois(\lambda_{r})$, and $N$ such distributions ($1 < r < N$), I have defined a likelihood function to describe my data as:

\begin{equation}\mathcal{L}\left(data \mid H_{}\right)=\prod_{r=1}^{N} \prod_{i=1}^{M} \frac{e^{-\lambda_{r}} \lambda_{r}^{X_{i, r}}}{X_{i, r} !}\end{equation}

I am attempting to find the uniformly most powerful (UMP) statistic using the Neyman-Pearson lemma, to test for the hypothesis $H_{a}: \lambda_{r} = \lambda_{a, r}, \lambda_{a, r} > \lambda_{0, r} \ \forall \ r \in (1, 2, .. N)$ against the null $H_{0}: \lambda_{r} = \lambda_{0, r}$. The trouble is, when I find the likelihood ratio,

$$\Lambda=\frac{\mathcal{L}\left(data \mid H_{0}\right)}{\mathcal{L}\left(data \mid H_{a}\right)}=\prod_{r=1}^{N} \frac{e^{M} \lambda_{a, r}}{e^{M} \lambda_{0, r}}\left(\frac{\lambda_{0, r}}{\lambda_{a, r}}\right)^{\sum_{i=1}^{M}X_{i, r}} \leq c$$

After taking the log of both sides, I get what appears to be a weighted combination of a Poisson-distributed variable whose distribution I am having difficulty describing:

$$\sum_{r=1}^{N} \sum_{i=1}^{M}{X_{i, r}} \log \left(\frac{\lambda_{0}, r}{\lambda_{1, r}}\right) \geq \log (c)-\sum_{r=1}^{N}\left(\lambda_{a, r}-\lambda_{0, r}\right) M$$

I recognize that the sum of Independent Poisson random variables should also be Poisson distributed: $$\sum_{r=1}^{N} \sum_{i=1}^{M}{X_{i, r}}$$

but I can't seem to deconvolve this expression to find a neat statistic. I came across a great post by whuber that describes one approach to approximate the behaviour of this statistic using Cornish-Fisher expansion, but I wonder if there is a more practical way to find the likelihood ratio test for this case?


Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Browse other questions tagged or ask your own question.