Suppose I have $X \sim \text{Poisson}(\lambda_x)$ and $Y \sim \text{Poisson}(\lambda_y)$ and they are independent. Suppose $H_0: \lambda_x =\lambda_y$ and $H_A: \lambda_x\ne\lambda_y$. My likelihood ratio is


The MLE for the null is $\hat{\lambda}=\frac{x+y}{2}$ and for the alternative, $\hat{\lambda}_x=x,\hat{\lambda}_x=y$. After canceling, I have

$$\Lambda = \frac{(\frac{x+y}{2})^{x+y}}{x^x y^y}$$

I want to now test whether this is statistically significant. But to do that, I usually use the result of Wilk's theorem which is

$$-2\log(\Lambda)\sim \chi^2_{d-d_0}$$

where $d = \dim(\Omega)$ and $d_0 = \dim(\Omega_0)$. Note, $\Omega_0$ is the set of possible values for the numerator maximization problem and $\Omega$ for the denominator.

Can I still apply this rule given that $n=2$?

I know that the sum of independent Poisson distributions is also Poisson distributed. Could I consider $X$ and $Y$ to be the sums of independent Poisson, which would be like having a large sample size?


You should not in general trust that asymptotic result when $n=2$, at least not without further information. If both $\lambda_x$ and $\lambda_y$ are large, so that the Poisson distributions can be approximated by normals, maybe. You could investigate it yourself by simulation.

But, in this case, why bother? There are non-asymptotic results you can use, see Checking if two Poisson samples have the same mean.

| cite | improve this answer | |

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.