Notably, although the estimators do not have an explicit form, this
work employs liner programming techniques to solve the above minimization
problem.
Restated, one feature of the QR is the ability to trace the entire distribution
of dependent variables that are conditional on the independent
one. Comparing equation (5) with equations (2) and (3) reveals a
key feature of the QR technique: the estimator vector of b,h varies with
h. Moreover, by comparing the behaviors with different h, one can
characterize the non-uniform estimator vector, namely b,h, in various
performance quantile regimes. In addition, a comparison of equation
(3) with equation (5) reveals that the LAD estimator is a special
case of the quantile-varying estimator with a quantile of 0.5.
Notably, this study uses the matrix bootstrap method to estimate
standard errors for the coefficients in the QR model. In a Monte Carlo
study, Buchinsky (1994) first recommends bootstrap methods for relatively
small samples, because they are robust to changes in bootstrap
sample size relative to the data sample size. Further, Koenker and Hallock
(2001) propose using the percentile method to construct confi-
dence intervals for each parameter in bh. In comparison with standard
asymptotic confidence intervals, the bootstrap percentile intervals are
generally not symmetric around the underlying parameter estimate.
Further, the bootstrap procedures can also deal with the joint distribution
of various QR estimators, which allows the equality of slope
parameters to be tested across various quantiles.