Consider a two-phase sampling design. We estimate the mean of $Y$ using a regression estimator $\bar{y}_{lr} = \bar{y} + b(\bar{X}' - \bar{x})$. What is the main condition under which this estimator is more efficient than a simple mean estimator $\bar{y}$?