Effects of design and error on normal convergence rates in regression problems (Q2640286)

This paper describes the part played by experimental error along with the design, which are the two sources of variation in a linear regression model, in determining the convergence rates to normality for the regression estimators. It is shown that if the convergence rate for errors in the central limit theorem is of the order of \(n^{-\alpha}\), \(0<\alpha <1/2\), n being the sample size, then this rate of \(n^{- \alpha}\) is maintained for the regression estimators if the distribution generating design has moments of the \((2\alpha +2)th\) order. The author also considers the following situations: (i) when the design variables are bounded and (ii) when the design variables are bounded and the error distributions symmetric. ``Characterizations'' of convergence rates are also developed.