# In bivariate linear regression is there a direct relationship between $n$, $r^2$ and coefficient error?

In bivariate linear regression is there a direct relationship between sample size $n$, coefficient of determination $r^2$ and $\sigma_\beta$ (the standard error of coefficient $\beta$)?

Assume data have been normalized so both target and predictor variable have $\sigma=1$.

Putting the question another way, does $\sigma_\beta$ tell me something different to $r^2$ or are they measures of the same thing? Or, is it possible to have a strong, certain but unreliable link between two variables (large $\beta$, small $\sigma_\beta$, but small $r^2$)?

(In multiple regression this doesn't apply as even with high $r^2$, $\sigma_\beta$ can indicate uncertainty as to which of the multiple predictors is causing the response).

EDIT

Just got this out of my software (without standardized data):

regression coeff 0.023
stderr of coeff 0.0046
p=0.000002

n=131
multiple r2=0.17

predictor std=22.5
target std=2.24


Standardized coefficient is presumably

$0.023*22.5/2.24 = 0.23$.

If standardized coefficient is the same as correlation, then

$r^2 = 0.23^2 = 0.053$

...not the same as the software gave. What am I doing wrong?

• Do you mean "coefficient of determination" for "correlation", & "multiple regression" for "multivariate regression"? – Scortchi Sep 6 '13 at 10:35
• Yes. (My bad on $r^2$ but isn't multiple regression the same as multivariate regression?) – Sideshow Bob Sep 6 '13 at 10:45
• No; multivariate regression means a multiple response (target or outcome or dependent variable). Having multiple predictors (independent variables) does not itself make a regression multivariate. – Nick Cox Sep 6 '13 at 10:47
• Your question is puzzling. If the predictor in bivariate regression has been standardised, then its coefficient equals the correlation: this is in essence an inevitable consequence. If not, then not in general. The way to think of this is in terms of units of measurement or dimensional analysis. A correlation, and hence its square, has no units, but a regression coefficient has units (units of y)/(units of x). Standardising washes out both units and leaves you with dimensionless numbers. – Nick Cox Sep 6 '13 at 10:53
• Ok so you are saying it's not possible, with standardized data, to have large $\beta$ and small $r^2$, because they are both the same thing - a measure of effect size. $\sigma_\beta$ meanwhile tells me the significance. Thanks btw, learning a few things here :) – Sideshow Bob Sep 6 '13 at 11:08

Here is a silly example from R (which I do not know well, but you can download it and it amounts to a lingua franca):

> y = c(23,32,45,54,67,75)
> x = c(1,2,3,4,5,6)
> lm(y ~ x)

Call:
lm(formula = y ~ x)

Coefficients:
(Intercept)            x
11.93        10.69

> cor(y, x)
[1] 0.998227
> sd(y)
[1] 20.02665
> sd(x)
[1] 1.870829
> 10.69 * sd(x) / sd(y)
[1] 0.9986273


There is some rounding error because I just took the printed result for the coefficient, but the principle is sound. Other software gives identical results.

• That's great, and works for me too with my data. Must be a bug somewhere in my preprocessing script. – Sideshow Bob Sep 6 '13 at 12:40