Non- and semiparametric alternatives to generalized linear models (Q1966020)

The author discusses additive models \(y_i=g(x_i)+\varepsilon, \;i=1,\dots,n\), where \(\varepsilon\) are i.i.d. errors and \(g\) is an unknown function from \(R^d\) to \(R^1\) that is approximated by the form \(g(X)\sim g_0 +\sum_{j=1}^d g_j(X_j)\) with the smooth univariate functions \(g_j(x)\) obtained from a linear scatterplot smoother. Generalized version of these models are discussed, too. The single index model \(y=f(\beta^T x)+\varepsilon\) as well as multiple index models of the form \(y=f(\theta_1^T,\dots,\theta_m^T)+\varepsilon\) are discussed. Semiparametric models with a single and multiple nonparametric component are considered. The penalized least squares technique is compared to the Speckman approach to partial linear models with one unparameterized explanatory variable; see, e. g., \textit{P. Speckman}, J. R. Stat. Soc., Ser. B 50, No. 3, 413-436 (1988; Zbl 0671.62045). Further, generalized partial linear models are briefly mentioned.