Take an example of a linear regression,

$Y= \beta X$

where X and Y values are z-score transformed ($\mu$ = 0, $\sigma$ = 1).

In this situation the correlation coefficient $r$ equals $\beta$ because the following equality

$\beta = r \sigma_y/\sigma_x $.

While deriving bootstrap confidence intervals, I was surprised by the observation that beta and r have different bootstrap distributions:

enter image description here

The mean of these two distributions coincides with each other. However the confidence intervals are different, while confidence interval of $r$ includes zero, this was not true for the confidence interval of $\beta$, suggesting a significant contribution. Based on the observation on $\beta$, one might conclude that the regression is significant, whereas $r$ is not different than zero.

It was interesting (well at least for me) to realize that the equality of $\beta$ and $r$ in the equation above, does not imply equality of their distributions. I would be happy to learn more what is really happening here.


This is because the linear regression estimator $\bar \beta$ makes use of the information that $X$ and $Y$ are centered whereas the estimator $\bar \rho$ for the correlation coefficient has to subtract the empirical mean of $X$ and $Y$. In this case, the variance of $\bar \beta$ is lower than the variance of $\bar \rho$ (in a similar fashion as to estimate the variance when the mean is know).

  • $\begingroup$ so but then the distribution of $\beta$ should have higher variance, not the $r$. $\endgroup$ – bonobo Jul 24 '15 at 11:28

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Not the answer you're looking for? Browse other questions tagged or ask your own question.