In a study with $N=1000$ units, a two-phase sample is taken. In phase 1, $n'=200$ units are measured for auxiliary variable $X$. In phase 2, $n=50$ units are measured for variable $Y$. The regression estimator is $\bar{y}_{lr} = \bar{y} + b(\bar{X}' - \bar{x})$. What does this method primarily improve compared to a simple mean estimator?