Rank-based estimation of the ratio of scale parameters and applications (Q1193797): Difference between revisions

The authors study estimation in the two-sample scale problem. Let \(X_ 1'=(X_{11}',\dots,X_{1n_ 1}')\) and \(X_ 2'=(X_{21}',\dots,X_{2n_ 2}')\) be two independent samples from respective continuous distribution functions \(F_ 1\) and \(F_ 2\) satisfying one of the following conditions: \(F_ 1(u)=F(u/\sigma_ 1)\) and \(F_ 2(u)=F(u/\sigma_ 2)\), for some distribution function \(F\), with support in a subset of \((0,\infty)\). \(F_ 1(u)=F((u-\nu_ 1)/\sigma_ 1)\) and \(F_ 2(u)=F((u-\nu_ 2)/\sigma_ 2)\) for some distribution function \(F\) symmetric about zero or skewed with median zero; \(\nu_ 1\) and \(\nu_ 2\) are unknown, possibly equal. Let \(\theta=\sigma_ 1/\sigma_ 2\) and \(h(X_ 1',X_ 2')\) some rank- based test statistic that will be defined later. If \(\mu=E[h(X_ 1',X_ 2')]\) under the null hypothesis \(H_ 0:F_ 1=F_ 2\) define: \[ \hat\theta_ 1'=\inf\{\rho:h(x_ 1',x_ 2'/\rho)<\mu\},\quad\hat\theta_ 2'=\sup \{\rho:h(x_ 1',x_ 2'/\rho)>\mu\}, \] and consider \(\hat\theta'=0.5(\hat\theta_ 1'+\hat\theta_ 2')\) and \(\hat\theta\) the estimator based on \(h\), as \(\theta'\), with observations rounded-off to the nearest integers. The average scores method for ties is used. The principal results in this article show that \(\hat\theta'\), based on adaptive procedures, is a good estimator and that \(\hat\theta\) is close to \(\hat\theta'\), even when ties occur profusely. The proofs are nontrivial and break down if the midrank method for ties is used. They also need a precise construction of \(h\)'s modification, which is more tractable in the presence of ties. In general \(\hat\theta\) is not asymptotically normal, but an approximate large sample distribution is obtained. By Monte Carlo study an asymptotically distribution free property is established and confidence intervals are constructed. For the statistic \(h\), the authors use a modification of Hogg and Gastwirth procedures. Let \(\delta_ i=1\) if the \(i^{th}\) element of the order statistic of \((X_ 1',X_ 2')\) is an \(X_ 2'\) observation and \(\delta_ i=0\) otherwise, \(i=1,\dots,N=n_ 1+n_ 2\); \(\delta=1\) if \(N\) is odd and \(\delta=0\) if \(N\) is even, and \(P\) the nearest integer \(\leq Np+1\), for \(p\in(0,1)\). Define \[ T_{p,\gamma}=\sum^ N_{i=N+1- P}(i-N+P-0.5(1-\delta))^ \gamma\delta_ i,\quad B_{p,\gamma}=\sum^ P_{i=1}(P+0.5(1+\delta)-i)^ \gamma\delta_ i, \] with \(\gamma=1,2\). The authors propose the following scheme: if \(F\) has light tail, \(h=T_{0.1,2}+B_{0.1,2}\), (or \(T_{0.05,2}+B_{0.05,2}\) if \(N\geq 40)\); if \(F\) has normal tail, \(h=T_{0.125,2}+B_{0.125,2}\); if \(F\) has heavy tail, \(h=T_{0.2,1}+B_{0.2,1}\); if \(F\) has very heavy tail, \(h=\) the quadratic statistic. As measure of tail-weight of \(F\), a classical expression of averages of some order statistics of the two samples is used. The authors apply their procedure to various practical problems: earth velocity, animal carcinogenesis, meteorology and measurement of sodium in erythrocytes and whole blood.