If I have a linear model
$ y_i = x_i'\boldsymbol\beta + \epsilon_i $
and I assume that OLS estimator of $\boldsymbol\beta$ is unbiased and consistent and Least absolute deviation (LAD) estimator of $\boldsymbol\beta$ is also consistent, what are the conditions for LAD to be more efficient than OLS?
Does it depend on sample size? Under which assumptions (if any) about the error terms $\epsilon_i$ is OLS more efficient than LAD?
