Variable Selection in Nonparametric Regression with Categorical Covariates (Q4031129)

Suppose we observe an i.i.d. sequence \((X_ i,Y_ i)\) of random vectors, where each \(X_ i\) is categorical and \(Y_ i\) is real-valued. A histogram-type estimator \(\hat m_ d\) of the regression function \(m(x)=E[Y| X=x]\) is proposed which, for each finite \(d\), involves only the first \(d\) components of each \(X_ i\). A cross-validated minimizer \(\hat d\) is proposed which serves as an estimator for the true model dimension \(d_ 0\). It is shown that for increasing sample size, \(\hat d\) is consistent if and only if \(d_ 0=\infty\). Furthermore, a so- called accumulated prediction error criterion is shown to yield estimators which are consistent also when \(d_ 0<\infty\).