Part of the difficulty in interpreting the results obtained with robust regression
estimators is that the authors are often vague about the properties of the error term
and, consequently, about what location function of the conditional distribution of y
is being estimated.
For example, in his seminal contributions, Huber (1973, 1981) just states that the
errors are independent with approximately identical distributions
However, Huber
(1973, p. 800) adds that the desired estimate of β “will in some sense generalize
a robust alternative to the sample mean,” suggesting that x0
iβ = E(yi), for fixed
regressors, or x0
iβ = E(yi|xi), for random regressors