On hierarchical Bayesian estimation and selection for multivariate hypergeometric distributions

From MaRDI portal





For each \(i=1,\dots, n\) the random variables \(S_i=(s_{i1},\dots,s_{ik})\) are i.i.d. with multinomial \((M_i,p)\) distribution, where \(p=(p_1,\dots,p_k)\), \(0\leq p_i\leq 1\), \(\sum p_i=1\). The observed vectors \(X_i=(X_{i1},\dots,X_{ik})\) have a multivariate hypergeometric distribution \[ P\{X_{ij}=x_{ij},\;i=1,\dots,n,\;j=1,\dots,k\} =\left.\left\{\prod_{j=1}^k\left({s_{ij}\atop x_{ij}}\right)\right\}\right/ \left({M_i\atop m_i}\right), \] \(m_i=\sum_{j=1}^k x_{ij}\), conditionally on \(S_i\). The problem is to estimate the number of the most probable event, i.e. such \(j_i\) for which \(s_{ij_i}=\max_{1\leq l\leq k} s_{il}\). If \(j^*_i\) is some estimator (the selection rule), then the loss is \(L(j^*_i)=s_{ij_i}-s_{ij_i^*}\) and the risk is \(R(j^*)=E\sum_{i=1}^n L(j_i^*)\). The Bayesian selection rule \(j^B\) for known \(p\) and the hierarchical Bayesian selection rule \(j^H\) for unknown \(p\) with Dirichlet prior are considered (\(j^H\) is obtained from \(j^B\) by replacing of true \(p\) by its Bayes estimator). It is proved that the relative risk of the hierarchical selection rule is \[ \rho(j^H)={R(j^H)-R(j^B)\over R(j^B)} =O(e^{-2(m_1+\dots+m_n)\beta}) \] for some \(\beta\) independent of \(n\). For the entire collection see [Zbl 0990.62506].











This page was built for publication: On hierarchical Bayesian estimation and selection for multivariate hypergeometric distributions

Report a bug (only for logged in users!)Click here to report a bug for this page (MaRDI item Q3843161)