Normal-approximation tests for a Poisson exposure model (source code)

= Normal-approximation tests for a Poisson exposure model

In the <Poisson exposure model>, write $S_r=\sum_i x_i^r$ with positive total exposure. Under $H_0:\theta=1$, the total count has the exact <Poisson distribution> $N=\sum_iY_i\sim\operatorname{Poisson}(S_1)$. Thus an exact upper-tail test chooses the smallest <integer> $k$ satisfying $\Pr(N\geq k)\leq\alpha$; discreteness can make its <type I error> smaller than $\alpha$.

When $S_1$ is large, the <normal approximation to the Poisson distribution> yields the approximate level-$0.05$ rejection rule
$$
\widehat\theta_{\rm MLE}>1+1.645/\sqrt{S_1}.
$$
The <least-squares estimator> instead has null <variance> $S_3/S_2^2$, giving
$$
\widehat\theta_{\rm LS}>1+1.645\sqrt{S_3}/S_2.
$$
This second approximation needs a <central limit theorem> for the weighted sum $\sum_i x_i(Y_i-x_i)$, such as a suitable <Lindeberg condition>. A large unweighted total exposure alone does not establish that weighted approximation. Both displayed thresholds use the null <variance>, rather than a response-dependent plug-in <standard error>.