For order q kernel density estimators we show that the constant bq in bias=bqhq+o(hq) can be made arbitrarily small, while keeping the variance bounded. A data-based selection of bq is presented and Monte Carlo simulations illustrate the advantages of the method.