Consider the model $ y = \beta_0 + \beta_1 x_1 + \beta_2 x_2 + \epsilon$.

  1. What is the relationship between the correlation coefficients $r_{y,x_1}$, $r_{y,x_2}$ and the regression coefficients $\beta_1$ and $\beta_2$?

  2. In particular, how to interpret a situation where a particular correlation coefficient is statistically significant but the corresponding regression coefficient is not statistically significant?

  3. I recall reading a concept called "partial correlation coefficients". Is that in some way relevant in the above context?

Let's assume that the variable $x_1$ and $x_2$ are centered, it will make things easier (nothing prevents you to do that before doing your regression). Then, it is straightforward to see that:

$r_{y, x_1}\sigma_y=\beta_1\sigma_{x_1} + \beta_2r_{x_1, x_2} \sigma_{x_2}$

and

$r_{y, x_2}\sigma_y=\beta_2\sigma_{x_2} + \beta_1r_{x_1, x_2} \sigma_{x_1}$

Hence, the relation also involves standard deviations terms and the correlation between $x_1$ and $x_2$.

This should answer your second question. For example, if $r_{x_1, x_2}=1$ and $x_1 = x_2$, any solution $\beta_1\sigma_{x_1} + \beta_2r_{x_1, x_2} \sigma_{x_2} = \sigma_y$ leads to the same linear model for $y$. As a consequence, $\beta_1$ (or $\beta_2$) can take arbitrary values (which are going to depend on the numerical implementation of the linear regression you are using).

Looking at the value of $r_{x_1, x_2}$ is then critical before making any relation between $r_{y, x_i}$ and $\beta_1, \beta_2$.

Your Answer

 
discard

By clicking "Post Your Answer", you acknowledge that you have read our updated terms of service, privacy policy and cookie policy, and that your continued use of the website is subject to these policies.

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