关于动物脑容量(x)与体型(y)相关性Bootstrap估计方法的疑问

Great question—this is a super common pitfall when first working with Bootstrap for correlation estimates, so let’s break this down clearly:

• You’re destroying the actual biological relationship
  When you resample x and y separately, you’re effectively mixing brain sizes from one animal with body sizes from another. For example, you might end up pairing a mouse’s tiny brain with an elephant’s massive body, which has no real-world meaning. The correlation you calculate from these random mismatches is completely artificial—it doesn’t reflect the true association between brain and body size in the animal population you’re studying.

• Bootstrap relies on resampling independent observation units
  The Bootstrap method works by creating new samples that mimic how your original data was collected. In this scenario, each independent observation is a single animal’s full set of measurements (x and y together). Resampling these paired units preserves the natural correlation structure present in your data, which lets you accurately estimate the variability of the correlation statistic itself.

• Separate resampling leads to meaningless results
  Imagine your original data has a strong positive correlation between brain and body size. If you shuffle x and y by resampling separately, you’ll break that positive association—you might even end up with spurious negative correlations in some Bootstrap samples. The resulting distribution of correlation coefficients won’t tell you anything about how the true population correlation varies; it’ll just show you the noise from random mismatches.

To put it simply: if you want to estimate the sampling distribution of a statistic that depends on paired relationships (like correlation), you can’t split up those pairs. You have to resample the complete, linked observations each time.

内容的提问来源于stack exchange，提问作者frightenedeyes