Abstract <p> The consistency of new universal kernel estimators for conditional variance function in aheteroscedastic nonparametric regression model has been proven. The new estimators areinsensitive to the nature of the design dependence. For design that can be either fixed or random,only the following condition is used: the design points densely fill the domain of regressionfunction. As a consequence, we consider the problem of constructing a confidence region for aregression function under the above-mentioned very general conditions on the design in terms ofdense data.</p>