If the error terms $\epsilon_i$ are heteroskedastic, with $Var(\epsilon_i) = \sigma^2 w_i$, which estimation method is preferred over OLS to obtain efficient estimates?