Estimation and hypothesis test for partial linear single-index multiplicative models (Q2183764)

The partial linear single-index multivariate model is studied, \[Y=\exp(\alpha^\top X+g(\beta^\top Z))\epsilon,\] where \(Y\) is the response variable, \(X\) and \(Z\) are observable random vectors, \(g\) is an unknown smooth link function, \(\alpha\) and \(\beta\) are unknown index parameters. Random error \(\epsilon\) is positive with (a) \(E(\ln(\epsilon)|X,~Z)=0\), and (b) \(E(\epsilon-\epsilon^{-1}|X,~Z)=0\). Condition (a) makes the link function unique, and condition (b) is used for the least relative error estimation. The index parameters and the link function are estimated using the profile least product relative error criterion [\textit{K. Chen} et al., J. Multivariate Anal. 144, 91--98 (2016; Zbl 1328.62146)], which is equivalent to minimize \(\sum _{i=1}^n (\epsilon_i^{-1}+\epsilon_i)\). The criterion is combined with a leave-one-component-out method. This way in some cases more efficient estimators are obtained than by the logarithmic transformation method [\textit{H. Liang} et al., Ann. Stat. 38, No. 6, 3811--3836 (2010; Zbl 1204.62068)]. A Wald-type test is used to test a linear hypothesis on the parametric components. To select relevant variables, It is adopted the smoothly clipped absolute deviation approach due to \textit{J. Fan} and \textit{H. Peng} [Ann. Stat. 32, No. 3, 928--961 (2004; Zbl 1092.62031)]. A version of the integrated conditional moment test statistic is proposed to study model checking problem, and a bootstrap procedure is suggested to calculate critical values. Simulation study demonstrates the performance of the proposed procedure and a real example is analyzed.