3
$\begingroup$

I have found myself Googling this question more than once: ¿What is the relationship between the standardized multiple regression coefficient (the standardized partial slope) and the corresponding semi-partial correlation (also called the part correlation by some)? The answers that I find are not helpful, because they always provide an answer restricted to a multiple regression model with at most 2 predictors.

Note: The following similar question (Multiple regression or partial correlation coefficient? And relations between the two) addresses partial correlations and this question (Is there a difference between semipartial correlation and regression coefficient in multiple regression?) focuses only on the trivariate model (as indicated above).

So, I will start here with a summary of the two straightforward models.

For bivariate regression $$y=\alpha + \beta· x + \epsilon$$ the standardized slope, $\beta^*$, equals the correlation between $x$ and $y$: $$\beta^* = r_{x,y}$$

For trivariate regression $$y = \alpha + \beta·x + \gamma·z + \epsilon$$ the standardized partial slope for $x$ can be written as $$\beta^* = \frac{r_{y,x} - r_{y,z}·r_{x,z}}{1-r_{x,z}^2}.$$ When taken with the fact that the semi-partial correlation $r_{y(x|z)}$ is given by the equation $$r_{y(x|z)} = \frac{r_{y,x} - r_{y,z}·r_{x,z}}{\sqrt{1-r_{x,z}^2}}$$ we can obtain the following relationship between the standardized slope and the semi-partial correlation $$\beta^* = \frac{r_{y(x|z)}}{\sqrt{1-r_{x,z}^2}}$$

For multiple regression models with 3 or more predictors $$y = \beta_0 + \beta_1 · x_1 + \beta_2 · x_2 + \cdots + \beta_k · x_k + \epsilon$$ where $k>2$. My question is whether the following formulation is correct for multiple regression models with 3 or more predictors: $$\beta_i^* = \frac{r_{y(i|\langle i\rangle)}}{\sqrt{1-R_{i,\langle i\rangle}^2}}$$ where $r_{y(i|\langle i\rangle)}$ is the semi-partial correlation for y with i partialling out all other variables but the $i$-th, and $R_{i,\langle i\rangle}^2$ is the coefficient of determination obtained from predicting the $i$-th predictor from all of the other predictors in the model. (The notation $\langle i\rangle$ indicates the set of all predictor variables but the $i$-th.)

I am fairly confident my answer is correct, but I have not located a website to confirm this. So, my questions are: (1) ¿Is this correct? (2) ¿Can someone provide a proof? or (3) ¿Can someone provide easily accessible references?

$\endgroup$
3
  • 1
    $\begingroup$ It's been a while since you've asked this question: have you been able to locate an answer in the meantime? $\endgroup$
    – Anthony
    May 20 '21 at 11:28
  • 1
    $\begingroup$ I have been playing with a couple of numerical examples, and your formula appears to be correct. $\endgroup$
    – Anthony
    May 20 '21 at 12:31
  • 1
    $\begingroup$ I have not located any more information, and I haven't had the time to work out the math either...but I have been operating under the assumption that my rationale is correct. $\endgroup$
    – Gregg H
    May 20 '21 at 15:49
2
$\begingroup$

This paper lists an indirect way to convert between standardised $\beta$'s and partial and semi-partial coefficients via the $t$-statistic.

To wit, in a regression with $n$ samples and $p$ predictors, the partial-correlation coefficient of a predictor $X_f$ is given by $$ r_{\text{par}} = \frac{t_f}{\sqrt{t_f^2 + \text{df}}}\ , $$ where $t_f$ is the $t$-statistic of the regression coefficient $\beta_f$, and $\text{df}=n-p-1$. On the other hand, the semi-partial correlation coefficient is given by $$ r_{\text{s-par}} = \frac{t_f\sqrt{1-R^2}}{\sqrt{\text{df}}}\ , $$ where $R^2$ is for the full model, and the rest of the notation is as above.

$\endgroup$

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.