Efficient estimation in the two-sample semiparametric location-scale models (Q1822862)

Let \(X=(Y,Z)\). Our basic model is \(Y=\mu +ae_ 1\), \(Z=\Delta +\rho (\mu +ae_ 2)\) where \(a,\rho >0\), \(e_ 1\) and \(e_ 2\) are independent and have unknown common density g with respect to Lebesgue measure. Let \(\theta =(\Delta,\rho)\) and \(\eta =(\mu,a,\Delta,\rho).\) We construct an efficient estimator \({\hat \theta}_ n=({\hat \Delta}_ n,{\hat \rho}_ n)\) based on a sample of size n such that for any sequence \((\eta_ n,g_ n)\) with \(n^{1/2}| \eta_ n-\eta | \leq M\) for some positive M and \(\| n^{1/2}(g_ n^{1/2}-g^{1/2})-h\| \to 0\) for some \(h\in L^ 2(m)\) where m is Lebesgue measure, \(n^{1/2}({\hat \theta}_ n-\theta_ n)\to N(0,\Gamma^{-1})\) under \((\eta_ n,g_ n)\) where \(\Gamma^{-1}\) is the information bound which depends on (\(\eta\),g) but not on the particular choice of \((\eta_ n,g_ n)\), \(| \cdot |\) is Euclidean norm and \(\| \cdot \|\) is the usual \(L^ 2\) norm with respect to Lebesgue measure. It is noticeable that we show the asymptotic efficiency of \({\hat \theta}{}_ n\) using only the minimal conditions on the density g, which has never been done before. This estimator is an adaptive estimator in the sense that its asymptotic variance is the same as that of an efficient estimator in the case we know g completely. Small sample properties of \({\hat \theta}{}_ n\) are investigated through Monte Carlo simulations. We examine the performance of \({\hat \theta}{}_ n\) when the smoothing parameters (used in estimating the unknown common density) are pre-selected or, alternatively, are determined by a data-based bootstrap method.