In general, there will be visible differences between a parametric and
a nonparametric curve estimate. It is therefore quite natural to comp
are these in order to decide whether the parametric model could be jus
tified. An asymptotic quantification is the distribution of the integr
ated squared difference between these curves. We show that the standar
d way of boot-strapping this statistic fails. We use and analyse a dif
ferent form of bootstrapping for this task. We call this method the wi
ld bootstrap and apply it to fitting Engel curves in expenditure data
analysis.