Author
Abstract
We consider the second-order asymptotic properties of the bootstrap of L_1 regression estimators by looking at the difference between the L_1
estimator and its first-order approximation, where the latter is the
minimizer of a quadratic approximation to the L_1 objective function. It is
shown that the bootstrap distribution of the normed difference does not
converge (either in probability or with probability 1) to the ``correct''
limiting distribution but rather converges in distribution to a random
distribution. A characterization of this random distribution is given.
Some applications and extensions are given.
Keywords