In a two-phase sampling design, we use a large first-phase sample $n'$ to estimate the mean of an auxiliary variable $\bar{X}'$. If the correlation between $X$ and $Y$ is $\rho_{xy}=0.8$, how much does the regression estimator reduce the variance compared to the simple mean?