Are efficient estimators in single-indexed models really efficient? A computational discussion (Q1424611)

The author investigates the performance of a semiparametric M-estimator of a finite-dimensional parameter \(\theta\) in the single-index regression model defined by \(E(Y\mid X)=E(Y\mid X\theta_0)=r_0(X\theta_0)\), where \(\theta_0\) belongs to a compact subset of \(R^d\), \(X\theta_0\) denotes the usual inner product in \(R^d\) and \(r_0: R \to R\) is an unknown function. A weighted version of the semiparametric M-estimator defined via maximizing a pseudo-likelihood based on linear exponential families is of the main interest. Such type of estimators were shown to be asymptotically efficient [see \textit{M. Delecroix} and \textit{M. Hristache}, Bull. Belg. Math. Soc. - Simon Stevin 6, No. 2, 161--185 (1999; Zbl 0951.62032)]. The optimal choice of the density to construct the pseudo-likelihood and practical efficiency of the proposed estimators for a finite number of observations are studied through computational experiments comparing the performance of several estimators on the same simulation samples. It is shown that for a large but reasonable sample size the asymptotically efficient M-estimator works better than the usual one.