In a two-phase sampling design where $n'$ is the first-phase sample size and $n$ is the second-phase sample size, if the regression estimator $\bar{y}_{lr} = \bar{y} + b(\bar{X}' - \bar{x})$ is used to estimate the population mean $\bar{Y}$, what is the primary condition for this estimator to be more efficient than the simple mean estimator $\bar{y}$?