Sliced inverse regression (SIR): An appraisal of small sample alternatives to slicing (Q5906409)

The authors consider the semiparametric regression model \(y=g(x'\beta_1,\dots, x'\beta_k,\epsilon)\), where the dependent variable \(y\) depends on the explanatory variables through so-called indexes, i.e. linear combinations \(x'\beta_i\), \(i=1,\dots,k\). Here the error term \(\epsilon\) is independent of \(x\), and the link function \(g\) and the parameters \(\beta_i\) are unknown. The main attention is concentrated on the estimation of the directions of the vectors \(\beta_i\) in single or multi-indexed models. The known methods based on sliced inverse regression are reviewed. Some new methods preferable for small samples are proposed, and their asymptotic behaviour is also studied. Small sample properties of existing and the new methods are compared by simulation.