Estimation of heteroscedasticity in regression analysis (Q1095531)

The authors consider the regression model \(Y_ i=g(t_ i)+\epsilon_ i\), \(1\leq i\leq n\), with design variables \(t_ i\) (nonrandom) and measurements \(Y_ i\), g being the unknown regression function. The errors are assumed to be heteroscedastic. The problem is to estimate the variance \(\sigma^ 2(t)\) of the observations at any given design point t. The authors consider the case where \(\sigma^ 2(t)\) is assumed to be a smooth function of t. A class of initial estimators \({\tilde \sigma}^ 2(t_ i)\) for the local variances at \(t_ i\) is given and it is shown that they are asymptotically unbiased but not consistent. Hence it is proposed to smooth neighboring values of \({\tilde \sigma}^ 2(t_ i)\) using kernel methods. The resulting smoothed estimators are shown to be uniformly consistent. Several applications of local variance estimation are discussed. The needed results on uniform convergence rates for weighted averages of m-dependent random variables are discussed in a separate section.