Average derivatives are the mean slopes of regression functions. In pr
actice they are estimated via a nonparametric smoothing technique. Eve
ry smoothing method needs a calibration parameter that determines the
finite sample performance. In this paper we use the kernel estimation
method and develop a formula for the bandwidth that describes the sens
itivity of the average derivative estimator. One can determine an opti
mal smoothing parameter from this formula which tries out to undersmoo
th the density of the regression variable.