The regression model has heteroskedasticity. The variance of error term depends on regressors. From boostrapping analysis, I got two different covariance matrices of $\beta$. The difference results from different resampling procedures. One is resampling only from residuals, and another one is resampling from the pair of residuals and regressors. The former covariance matrices is lower than the latter covariance matrices. I am trying to understand why difference resampling procedures make the former covariance lower, but I cannot figure it out. I appreciate if you give some help.