Estimation of multiple Poisson means: A compromise between maximum likelihood and empirical Bayes estimators (Q578776)

Let \(X_ 1,...,X_ p\) \((p\geq 2)\) be independent Poissons with respective means \(\theta_ 1,...,\theta_ p\) which are to be estimated with loss \(L({\underset\sim \theta},\underset \sim a)=\sum^{p}_{i=1}(\theta_ i-a_ i)^ 2/\theta_ i\). \(\underset \sim X=(X_ 1,...,X_ p)\) is inadmissible for the stated problem [\textit{M. L. Clevenson} and \textit{J. V. Zidek}, J. Am. Stat. Assoc. 70, 698-705 (1975; Zbl 0308.62018)]. Set \(\hat u(t)=(p-1)/(t+p-1)\) and \(t=\sum^{p}_{i=1}x_ i\). The estimator \({\underset \sim \delta}^*(\underset \tilde{} x)=(1-\hat u(t))\underset \sim x\) is an improvement over \(\underset \tilde{} x\), but, as measured by maximum risk, is poor for the component problems. The authors suggest \[ \delta_{M,i}(\underset \sim x) = \begin{cases} {\underset \sim \delta}^*(\underset \sim x) &\text{ if \(0\leq x_ i\leq M/\hat u(t),\)} \\ x_ i-M &\text{ otherwise} ,\end{cases} \] where \(M\geq 0\) is chosen so as to strike the desired balance between the overall and component problems. Large M emphasizes the former.